oxidize-pdf 4.6.0

//! PDF image extraction functionality
//!
//! This module provides functionality to extract images from PDF documents with
//! advanced preprocessing for scanned documents.

use super::{OperationError, OperationResult};
use crate::graphics::ImageFormat;
use crate::parser::objects::{PdfArray, PdfName, PdfObject, PdfStream};
use crate::parser::{PdfDocument, PdfReader};
use std::collections::HashMap;
use std::fs::{self, File};
use std::io::{Read, Seek, Write};
use std::path::{Path, PathBuf};

#[cfg(feature = "external-images")]
use image::{DynamicImage, GenericImageView, ImageBuffer, ImageFormat as ImageLibFormat, Luma};

/// PDF transformation matrix (a, b, c, d, e, f)
///
/// Represents a 3x3 matrix: `[a c e; b d f; 0 0 1]` that transforms point `(x,y)` to `(a*x + c*y + e, b*x + d*y + f)`
#[derive(Debug, Clone)]
pub struct TransformMatrix {
    pub a: f64, // x scaling
    pub b: f64, // y skewing
    pub c: f64, // x skewing
    pub d: f64, // y scaling
    pub e: f64, // x translation
    pub f: f64, // y translation
}

impl TransformMatrix {
    #[allow(dead_code)]
    fn new(a: f64, b: f64, c: f64, d: f64, e: f64, f: f64) -> Self {
        Self { a, b, c, d, e, f }
    }

    /// Check if this matrix represents a 90-degree rotation
    #[allow(dead_code)]
    fn is_90_degree_rotation(&self) -> bool {
        // For 90-degree rotation: a ≈ 0, d ≈ 0, b and c are non-zero
        self.a.abs() < 0.001 && self.d.abs() < 0.001 && self.b.abs() > 0.001 && self.c.abs() > 0.001
    }

    /// Check if this matrix represents a simple scaling
    #[allow(dead_code)]
    fn is_simple_scale(&self) -> bool {
        // For scaling: b ≈ 0, c ≈ 0, a and d are scaling factors
        self.b.abs() < 0.001 && self.c.abs() < 0.001 && self.a.abs() > 0.001 && self.d.abs() > 0.001
    }

    /// Check if this is a matrix that needs rotation for proper OCR
    #[allow(dead_code)]
    fn is_fis2_like_matrix(&self) -> bool {
        // Some PDFs use 841.68 x 595.08 which are A4 dimensions (landscape fitting in portrait)
        // This indicates the image is landscape but being fit into portrait page
        (self.a - 841.68).abs() < 1.0
            && (self.d - 595.08).abs() < 1.0
            && self.b.abs() < 0.001
            && self.c.abs() < 0.001
    }
}

/// Preprocessing options for extracted images
#[derive(Debug, Clone)]
pub struct ImagePreprocessingOptions {
    /// Auto-detect and correct rotation
    pub auto_correct_rotation: bool,
    /// Enhance contrast for better OCR
    pub enhance_contrast: bool,
    /// Apply noise reduction
    pub denoise: bool,
    /// Upscale small images using bicubic interpolation
    pub upscale_small_images: bool,
    /// Minimum size to trigger upscaling
    pub upscale_threshold: u32,
    /// Upscale factor (2x, 3x, etc.)
    pub upscale_factor: u32,
    /// Convert to grayscale for better OCR on text documents
    pub force_grayscale: bool,
}

impl Default for ImagePreprocessingOptions {
    fn default() -> Self {
        Self {
            auto_correct_rotation: true,
            enhance_contrast: true,
            denoise: true,
            upscale_small_images: true,
            upscale_threshold: 300,
            upscale_factor: 2,
            force_grayscale: false,
        }
    }
}

/// Options for image extraction
#[derive(Debug, Clone)]
pub struct ExtractImagesOptions {
    /// Output directory for extracted images
    pub output_dir: PathBuf,
    /// File name pattern for extracted images
    /// Supports placeholders: {page}, {index}, {format}
    pub name_pattern: String,
    /// Whether to extract inline images
    pub extract_inline: bool,
    /// Minimum size (width or height) to extract
    pub min_size: Option<u32>,
    /// Whether to create output directory if it doesn't exist
    pub create_dir: bool,
    /// Preprocessing options for extracted images
    pub preprocessing: ImagePreprocessingOptions,
}

impl Default for ExtractImagesOptions {
    fn default() -> Self {
        Self {
            output_dir: PathBuf::from("."),
            name_pattern: "page_{page}_image_{index}.{format}".to_string(),
            extract_inline: true,
            min_size: Some(10),
            create_dir: true,
            preprocessing: ImagePreprocessingOptions::default(),
        }
    }
}

/// Result of image extraction
#[derive(Debug)]
pub struct ExtractedImage {
    /// Page number (0-indexed)
    pub page_number: usize,
    /// Image index on the page
    pub image_index: usize,
    /// Output file path
    pub file_path: PathBuf,
    /// Image dimensions
    pub width: u32,
    pub height: u32,
    /// Image format
    pub format: ImageFormat,
}

/// An extracted image whose encoded bytes remain in memory.
#[derive(Debug, Clone)]
pub struct ExtractedImageData {
    /// Zero-based page containing the image.
    pub page_number: usize,
    /// Zero-based image index within the page.
    pub image_index: usize,
    /// Image width in pixels.
    pub width: u32,
    /// Image height in pixels.
    pub height: u32,
    /// Encoding of `data`.
    pub format: ImageFormat,
    /// Encoded image bytes after optional preprocessing.
    pub data: Vec<u8>,
}

/// Resource bounds applied while extracting images in memory.
#[derive(Debug, Clone, Copy)]
pub struct ImageExtractionLimits {
    /// Maximum number of images delivered to the consumer.
    pub max_images: usize,
    /// Maximum encoded size of one delivered image.
    pub max_encoded_bytes_per_image: usize,
    /// Maximum total encoded size delivered during the operation.
    pub max_total_encoded_bytes: usize,
    /// Maximum width multiplied by height for one image.
    pub max_decoded_pixels_per_image: u64,
}

impl Default for ImageExtractionLimits {
    fn default() -> Self {
        Self {
            max_images: usize::MAX,
            max_encoded_bytes_per_image: usize::MAX,
            max_total_encoded_bytes: usize::MAX,
            max_decoded_pixels_per_image: u64::MAX,
        }
    }
}

/// Errors produced by bounded in-memory image extraction.
#[derive(Debug, thiserror::Error)]
pub enum ImageExtractionError {
    /// An existing PDF operation or image-processing step failed.
    #[error(transparent)]
    Operation(#[from] OperationError),
    /// A configured extraction bound would be exceeded.
    #[error(
        "Image extraction limit exceeded for {limit}: maximum {maximum}, attempted {attempted}"
    )]
    LimitExceeded {
        limit: &'static str,
        maximum: u64,
        attempted: u64,
    },
}

/// Result type for bounded in-memory image extraction.
pub type ImageExtractionResult<T> = Result<T, ImageExtractionError>;

impl ImageExtractionError {
    fn into_operation_error(self) -> OperationError {
        match self {
            Self::Operation(error) => error,
            error @ Self::LimitExceeded { .. } => {
                OperationError::ProcessingError(error.to_string())
            }
        }
    }
}

struct ImageExtractionBudget {
    limits: ImageExtractionLimits,
    images: usize,
    encoded_bytes: usize,
}

impl ImageExtractionBudget {
    fn new(limits: ImageExtractionLimits) -> Self {
        Self {
            limits,
            images: 0,
            encoded_bytes: 0,
        }
    }

    fn check_pixels(&self, width: u32, height: u32) -> ImageExtractionResult<()> {
        let pixels = u64::from(width) * u64::from(height);
        if pixels > self.limits.max_decoded_pixels_per_image {
            return Err(ImageExtractionError::LimitExceeded {
                limit: "decoded pixels per image",
                maximum: self.limits.max_decoded_pixels_per_image,
                attempted: pixels,
            });
        }
        Ok(())
    }

    fn check_image_slot(&self) -> ImageExtractionResult<()> {
        let attempted = self.images.checked_add(1).unwrap_or(usize::MAX);
        if attempted > self.limits.max_images {
            return Err(ImageExtractionError::LimitExceeded {
                limit: "image count",
                maximum: self.limits.max_images as u64,
                attempted: attempted as u64,
            });
        }
        Ok(())
    }

    fn consume(&mut self, bytes: usize) -> ImageExtractionResult<()> {
        if bytes > self.limits.max_encoded_bytes_per_image {
            return Err(ImageExtractionError::LimitExceeded {
                limit: "encoded bytes per image",
                maximum: self.limits.max_encoded_bytes_per_image as u64,
                attempted: bytes as u64,
            });
        }
        self.check_image_slot()?;
        let next_images = self.images + 1;
        let next_bytes =
            self.encoded_bytes
                .checked_add(bytes)
                .ok_or(ImageExtractionError::LimitExceeded {
                    limit: "total encoded bytes",
                    maximum: self.limits.max_total_encoded_bytes as u64,
                    attempted: u64::MAX,
                })?;
        if next_bytes > self.limits.max_total_encoded_bytes {
            return Err(ImageExtractionError::LimitExceeded {
                limit: "total encoded bytes",
                maximum: self.limits.max_total_encoded_bytes as u64,
                attempted: next_bytes as u64,
            });
        }
        self.images = next_images;
        self.encoded_bytes = next_bytes;
        Ok(())
    }
}

/// Image extractor
pub struct ImageExtractor<R: Read + Seek> {
    document: PdfDocument<R>,
    options: ExtractImagesOptions,
    /// Cache for already processed images
    processed_images: HashMap<String, PathBuf>,
}

impl<R: Read + Seek> ImageExtractor<R> {
    /// Create a new image extractor
    pub fn new(document: PdfDocument<R>, options: ExtractImagesOptions) -> Self {
        Self {
            document,
            options,
            processed_images: HashMap::new(),
        }
    }

    /// Visit every extracted image without writing to the filesystem.
    ///
    /// # Errors
    ///
    /// Returns an error when the document cannot be parsed, a configured limit
    /// would be exceeded, image processing fails, or the visitor returns an error.
    pub fn visit_images<F>(
        &mut self,
        limits: ImageExtractionLimits,
        mut visitor: F,
    ) -> ImageExtractionResult<()>
    where
        F: FnMut(ExtractedImageData) -> ImageExtractionResult<()>,
    {
        let page_count = self
            .document
            .page_count()
            .map_err(|error| OperationError::ParseError(error.to_string()))?;
        let mut budget = ImageExtractionBudget::new(limits);
        for page_number in 0..page_count as usize {
            self.visit_page_images(page_number, &mut budget, &mut visitor)?;
        }
        Ok(())
    }

    /// Extract every image as encoded bytes without filesystem writes.
    ///
    /// # Errors
    ///
    /// Returns an error when parsing or image processing fails, or when a
    /// configured extraction limit would be exceeded.
    pub fn extract_all_in_memory(
        &mut self,
        limits: ImageExtractionLimits,
    ) -> ImageExtractionResult<Vec<ExtractedImageData>> {
        let mut images = Vec::new();
        self.visit_images(limits, |image| {
            images.push(image);
            Ok(())
        })?;
        Ok(images)
    }

    /// Extract one page's images as encoded bytes without filesystem writes.
    ///
    /// # Errors
    ///
    /// Returns an error when the page cannot be read, image processing fails,
    /// or a configured extraction limit would be exceeded.
    pub fn extract_from_page_in_memory(
        &mut self,
        page_number: usize,
        limits: ImageExtractionLimits,
    ) -> ImageExtractionResult<Vec<ExtractedImageData>> {
        let mut images = Vec::new();
        let mut budget = ImageExtractionBudget::new(limits);
        self.visit_page_images(page_number, &mut budget, &mut |image| {
            images.push(image);
            Ok(())
        })?;
        Ok(images)
    }

    fn visit_page_images<F>(
        &mut self,
        page_number: usize,
        budget: &mut ImageExtractionBudget,
        visitor: &mut F,
    ) -> ImageExtractionResult<()>
    where
        F: FnMut(ExtractedImageData) -> ImageExtractionResult<()>,
    {
        let page = self
            .document
            .get_page(page_number as u32)
            .map_err(|error| OperationError::ParseError(error.to_string()))?;
        let resources = self
            .document
            .get_page_resources(&page)
            .map_err(|error| OperationError::ParseError(error.to_string()))?;
        let mut references = Vec::new();
        if let Some(resources) = resources {
            if let Some(PdfObject::Dictionary(xobjects)) = resources.get("XObject") {
                for object in xobjects.0.values() {
                    if let PdfObject::Reference(number, generation) = object {
                        references.push((*number, *generation));
                    }
                }
            }
        }

        let mut image_index = 0;
        for (number, generation) in references {
            let object = self
                .document
                .get_object(number, generation)
                .map_err(|error| OperationError::ParseError(error.to_string()))?;
            let PdfObject::Stream(stream) = object else {
                continue;
            };
            if !matches!(stream.dict.get("Subtype"), Some(PdfObject::Name(name)) if name.0 == "Image")
            {
                continue;
            }
            budget.check_image_slot()?;
            if let Some(image) =
                self.prepare_image_data(&stream, page_number, image_index, budget)?
            {
                budget.consume(image.data.len())?;
                visitor(image)?;
                image_index += 1;
            }
        }

        if self.options.extract_inline {
            for content in self
                .document
                .get_page_content_streams(&page)
                .map_err(|error| OperationError::ParseError(error.to_string()))?
            {
                self.visit_inline_images(&content, page_number, &mut image_index, budget, visitor)?;
            }
        }
        Ok(())
    }

    fn prepare_image_data(
        &self,
        stream: &PdfStream,
        page_number: usize,
        image_index: usize,
        budget: &ImageExtractionBudget,
    ) -> ImageExtractionResult<Option<ExtractedImageData>> {
        let Some(PdfObject::Integer(width @ 1..)) = stream.dict.get("Width") else {
            return Ok(None);
        };
        let Some(PdfObject::Integer(height @ 1..)) = stream.dict.get("Height") else {
            return Ok(None);
        };
        let (Ok(width), Ok(height)) = (u32::try_from(*width), u32::try_from(*height)) else {
            return Ok(None);
        };
        if self
            .options
            .min_size
            .is_some_and(|minimum| width < minimum || height < minimum)
        {
            return Ok(None);
        }
        budget.check_pixels(width, height)?;
        let color_space = stream.dict.get("ColorSpace");
        let bits = match stream.dict.get("BitsPerComponent") {
            Some(PdfObject::Integer(bits)) => *bits as u8,
            _ => 8,
        };
        let smask = self.extract_smask_alpha(&stream.dict, width, height);
        let first_filter = match stream.dict.get("Filter") {
            Some(PdfObject::Name(name)) => Some(name.0.as_str()),
            Some(PdfObject::Array(filters)) => filters.0.first().and_then(|filter| match filter {
                PdfObject::Name(name) => Some(name.0.as_str()),
                _ => None,
            }),
            _ => None,
        };
        let (data, format) = match first_filter {
            Some("DCTDecode") => (stream.data.clone(), ImageFormat::Jpeg),
            Some("FlateDecode" | "LZWDecode") | None => {
                let decoded = self.decode_image_stream(stream)?;
                (
                    self.convert_raw_image_data_to_png(
                        &decoded,
                        width,
                        height,
                        color_space,
                        bits,
                        smask.as_deref(),
                    )?,
                    ImageFormat::Png,
                )
            }
            Some("CCITTFaxDecode") => {
                let decoded = self.decode_image_stream(stream)?;
                (
                    self.convert_ccitt_to_png(&decoded, width, height)?,
                    ImageFormat::Png,
                )
            }
            Some(_) => return Ok(None),
        };
        #[cfg(feature = "external-images")]
        let data = if self.should_preprocess() {
            self.preprocess_image_data(&data, width, height, format)?
        } else {
            data
        };
        Ok(Some(ExtractedImageData {
            page_number,
            image_index,
            width,
            height,
            format,
            data,
        }))
    }

    fn visit_inline_images<F>(
        &self,
        stream_data: &[u8],
        page_number: usize,
        image_index: &mut usize,
        budget: &mut ImageExtractionBudget,
        visitor: &mut F,
    ) -> ImageExtractionResult<()>
    where
        F: FnMut(ExtractedImageData) -> ImageExtractionResult<()>,
    {
        let mut position = 0;
        while let Some(begin) = Self::find_bytes(stream_data, b"BI", position) {
            let Some(id) = Self::find_bytes(stream_data, b"ID", begin + 2) else {
                break;
            };
            let Some(end) = Self::find_bytes(stream_data, b"EI", id + 2) else {
                break;
            };
            budget.check_image_slot()?;
            let dictionary = String::from_utf8_lossy(&stream_data[begin + 2..id]);
            let (width, height) = self.parse_inline_image_dict(dictionary.trim());
            budget.check_pixels(width, height)?;
            let data = stream_data[id + 2..end].to_vec();
            budget.consume(data.len())?;
            let format = self
                .detect_image_format_from_data(&data)
                .unwrap_or(ImageFormat::Raw);
            visitor(ExtractedImageData {
                page_number,
                image_index: *image_index,
                width,
                height,
                format,
                data,
            })?;
            *image_index += 1;
            position = end + 2;
        }
        Ok(())
    }

    fn find_bytes(haystack: &[u8], needle: &[u8], start: usize) -> Option<usize> {
        haystack
            .get(start..)?
            .windows(needle.len())
            .position(|window| window == needle)
            .map(|position| start + position)
    }

    /// Extract all images from the document
    pub fn extract_all(&mut self) -> OperationResult<Vec<ExtractedImage>> {
        if self.options.create_dir && !self.options.output_dir.exists() {
            fs::create_dir_all(&self.options.output_dir)?;
        }
        let options = self.options.clone();
        let mut cache = std::mem::take(&mut self.processed_images);
        let mut images = Vec::new();
        let result = self.visit_images(ImageExtractionLimits::default(), |image| {
            images.push(Self::persist_image_data(&options, &mut cache, image)?);
            Ok(())
        });
        self.processed_images = cache;
        result.map_err(ImageExtractionError::into_operation_error)?;
        Ok(images)
    }

    /// Extract images from a specific page
    pub fn extract_from_page(
        &mut self,
        page_number: usize,
    ) -> OperationResult<Vec<ExtractedImage>> {
        if self.options.create_dir && !self.options.output_dir.exists() {
            fs::create_dir_all(&self.options.output_dir)?;
        }
        let options = self.options.clone();
        let mut cache = std::mem::take(&mut self.processed_images);
        let mut images = Vec::new();
        let mut budget = ImageExtractionBudget::new(ImageExtractionLimits::default());
        let result = self.visit_page_images(page_number, &mut budget, &mut |image| {
            images.push(Self::persist_image_data(&options, &mut cache, image)?);
            Ok(())
        });
        self.processed_images = cache;
        result.map_err(ImageExtractionError::into_operation_error)?;
        Ok(images)
    }

    fn persist_image_data(
        options: &ExtractImagesOptions,
        cache: &mut HashMap<String, PathBuf>,
        image: ExtractedImageData,
    ) -> OperationResult<ExtractedImage> {
        let key = format!("{:x}", md5::compute(&image.data));
        let allow_deduplication = !options.name_pattern.contains("{page}");
        let extension = match image.format {
            ImageFormat::Jpeg => "jpg",
            ImageFormat::Png => "png",
            ImageFormat::Tiff => "tiff",
            ImageFormat::Raw => "rgb",
        };
        let filename = options
            .name_pattern
            .replace("{page}", &(image.page_number + 1).to_string())
            .replace("{index}", &(image.image_index + 1).to_string())
            .replace("{format}", extension);
        let existing = allow_deduplication
            .then(|| cache.get(&key).cloned())
            .flatten();
        let path = existing.unwrap_or_else(|| options.output_dir.join(filename));
        if !allow_deduplication || !cache.contains_key(&key) {
            let mut file = File::create(&path)?;
            file.write_all(&image.data)?;
            cache.insert(key, path.clone());
        }
        Ok(ExtractedImage {
            page_number: image.page_number,
            image_index: image.image_index,
            file_path: path,
            width: image.width,
            height: image.height,
            format: image.format,
        })
    }

    /// Detect image format from raw data by examining magic bytes
    fn detect_image_format_from_data(&self, data: &[u8]) -> OperationResult<ImageFormat> {
        if data.is_empty() {
            return Err(OperationError::ParseError(
                "Image data too short to detect format".to_string(),
            ));
        }

        // Check for PNG signature (needs 8 bytes)
        if data.len() >= 8 && &data[0..8] == b"\x89PNG\r\n\x1a\n" {
            return Ok(ImageFormat::Png);
        }

        // Check for TIFF signatures (needs 4 bytes)
        if data.len() >= 4 {
            if &data[0..2] == b"II" && &data[2..4] == b"\x2A\x00" {
                return Ok(ImageFormat::Tiff); // Little endian TIFF
            }
            if &data[0..2] == b"MM" && &data[2..4] == b"\x00\x2A" {
                return Ok(ImageFormat::Tiff); // Big endian TIFF
            }
        }

        // Check for JPEG signature (needs 2 bytes)
        if data.len() >= 2 && data[0] == 0xFF && data[1] == 0xD8 {
            return Ok(ImageFormat::Jpeg);
        }

        // If data is too short for any meaningful detection
        if data.len() < 2 {
            return Err(OperationError::ParseError(
                "Image data too short to detect format".to_string(),
            ));
        }

        // Default to PNG for FlateDecode if no other format detected
        // This is a fallback since FlateDecode is commonly used for PNG in PDFs
        Ok(ImageFormat::Png)
    }

    /// Apply rotation transformation
    #[cfg(feature = "external-images")]
    #[allow(dead_code)]
    fn apply_rotation_transformation(
        &self,
        img: DynamicImage,
        matrix: &TransformMatrix,
    ) -> OperationResult<DynamicImage> {
        // Determine rotation direction based on matrix values
        // For 90-degree clockwise: a=0, b=1, c=-1, d=0
        // For 90-degree counter-clockwise: a=0, b=-1, c=1, d=0

        if matrix.b > 0.0 && matrix.c < 0.0 {
            Ok(img.rotate90()) // 90 degrees clockwise
        } else if matrix.b < 0.0 && matrix.c > 0.0 {
            Ok(img.rotate270()) // 90 degrees counter-clockwise (270 clockwise)
        } else {
            // Default to 90-degree rotation for landscape-in-portrait cases
            Ok(img.rotate90())
        }
    }

    /// Apply scaling transformation
    #[cfg(feature = "external-images")]
    #[allow(dead_code)]
    fn apply_scale_transformation(
        &self,
        img: DynamicImage,
        matrix: &TransformMatrix,
    ) -> OperationResult<DynamicImage> {
        let (current_width, current_height) = img.dimensions();

        // Calculate new dimensions based on scaling factors
        let new_width = (current_width as f64 * matrix.a.abs()) as u32;
        let new_height = (current_height as f64 * matrix.d.abs()) as u32;

        if new_width > 0 && new_height > 0 {
            Ok(img.resize(new_width, new_height, image::imageops::FilterType::Lanczos3))
        } else {
            // If scaling results in invalid dimensions, return original
            Ok(img)
        }
    }

    /// Parse inline image dictionary to extract width and height
    fn parse_inline_image_dict(&self, dict_str: &str) -> (u32, u32) {
        let mut width = 100; // Default width
        let mut height = 100; // Default height

        // Simple parsing - look for /W and /H parameters
        for line in dict_str.lines() {
            let line = line.trim();

            // Parse width: /W 123 or /Width 123
            if line.starts_with("/W ") || line.starts_with("/Width ") {
                if let Some(value_str) = line.split_whitespace().nth(1) {
                    if let Ok(w) = value_str.parse::<u32>() {
                        width = w;
                    }
                }
            }

            // Parse height: /H 123 or /Height 123
            if line.starts_with("/H ") || line.starts_with("/Height ") {
                if let Some(value_str) = line.split_whitespace().nth(1) {
                    if let Ok(h) = value_str.parse::<u32>() {
                        height = h;
                    }
                }
            }
        }

        (width, height)
    }

    /// Decode an image stream, resolving an indirect `/DecodeParms` (or `/DP`)
    /// first so that filter predictors are actually applied (issue #286).
    ///
    /// `PdfStream::decode` only sees the stream's own dictionary; when the
    /// decode parameters are stored as an indirect reference the predictor is
    /// silently skipped, leaving the per-row predictor bytes in the output.
    fn decode_image_stream(&self, stream: &PdfStream) -> OperationResult<Vec<u8>> {
        let parse_options = self.document.options();

        let needs_resolution = ["DecodeParms", "DP"].into_iter().any(|key| {
            stream
                .dict
                .0
                .get(&PdfName(key.to_string()))
                .map(Self::contains_reference)
                .unwrap_or(false)
        });

        let decode_result = if needs_resolution {
            let mut dict = stream.dict.clone();
            for key in ["DecodeParms", "DP"] {
                if let Some(obj) = dict.0.get(&PdfName(key.to_string())).cloned() {
                    let resolved = self.resolve_decode_params(&obj);
                    dict.0.insert(PdfName(key.to_string()), resolved);
                }
            }
            PdfStream {
                dict,
                data: stream.data.clone(),
            }
            .decode(&parse_options)
        } else {
            stream.decode(&parse_options)
        };

        decode_result
            .map_err(|e| OperationError::ParseError(format!("Failed to decode image stream: {e}")))
    }

    /// Whether an object is, or directly contains, an indirect reference.
    fn contains_reference(obj: &PdfObject) -> bool {
        match obj {
            PdfObject::Reference(_, _) => true,
            PdfObject::Array(arr) => arr
                .0
                .iter()
                .any(|e| matches!(e, PdfObject::Reference(_, _))),
            _ => false,
        }
    }

    /// Resolve indirect references inside a `/DecodeParms` value (the value
    /// itself, or each element of a per-filter array).
    fn resolve_decode_params(&self, obj: &PdfObject) -> PdfObject {
        let resolved = self.document.resolve(obj).unwrap_or_else(|e| {
            // Falling back to the unresolved reference means the predictor is
            // skipped and the image decodes to garbage — the original #286
            // symptom. Surface it instead of failing silently.
            tracing::warn!("Failed to resolve /DecodeParms reference: {e}");
            obj.clone()
        });
        match resolved {
            PdfObject::Array(arr) => PdfObject::Array(PdfArray(
                arr.0
                    .iter()
                    .map(|e| self.document.resolve(e).unwrap_or_else(|_| e.clone()))
                    .collect(),
            )),
            other => other,
        }
    }

    /// If `color_space` is an `[/Indexed base hival lookup]` array, resolve it
    /// into `(resolved_base, hival, palette_bytes)`.
    fn try_resolve_indexed(
        &self,
        color_space: Option<&PdfObject>,
    ) -> Option<(PdfObject, usize, Vec<u8>)> {
        let array = color_space?.as_array()?;
        let first = array.0.first()?.as_name()?;
        if first.0 != "Indexed" && first.0 != "I" {
            return None;
        }
        let base = self.document.resolve(array.0.get(1)?).ok()?;
        let hival = array.0.get(2)?.as_integer()?.max(0) as usize;
        let lookup = self.resolve_lookup_bytes(array.0.get(3)?)?;
        Some((base, hival, lookup))
    }

    /// Resolve the Indexed lookup table into palette bytes (it may be a string
    /// literal or an indirect stream).
    fn resolve_lookup_bytes(&self, lookup: &PdfObject) -> Option<Vec<u8>> {
        match self.document.resolve(lookup).ok()? {
            PdfObject::String(s) => Some(s.0),
            PdfObject::Stream(s) => s.decode(&self.document.options()).ok(),
            _ => None,
        }
    }

    /// Resolve the `/N` (component count) of an `[/ICCBased stream]` colour space.
    fn icc_components(&self, color_space: Option<&PdfObject>) -> Option<u8> {
        let array = color_space?.as_array()?;
        if array.0.first()?.as_name()?.0 != "ICCBased" {
            return None;
        }
        let stream = self.document.resolve(array.0.get(1)?).ok()?;
        let n = stream
            .as_stream()?
            .dict
            .0
            .get(&PdfName("N".to_string()))?
            .as_integer()?;
        // /N is 1, 3 or 4 for valid ICC profiles. Clamp so a malformed value
        // can't truncate (e.g. -1 → 255) and blow up a downstream allocation.
        Some(n.clamp(1, 4) as u8)
    }

    /// Convert raw image sample data to PNG format.
    ///
    /// Handles Indexed colour spaces (one palette index per pixel, expanded to
    /// the base colour) and computes the component count from the colour space
    /// (issue #286 — Indexed was previously treated as 3-component RGB).
    fn convert_raw_image_data_to_png(
        &self,
        data: &[u8],
        width: u32,
        height: u32,
        color_space: Option<&PdfObject>,
        bits_per_component: u8,
        smask_alpha: Option<&[u8]>,
    ) -> OperationResult<Vec<u8>> {
        // Resolve an indirect ColorSpace reference up front.
        let resolved_cs = color_space.and_then(|cs| self.document.resolve(cs).ok());
        let cs = resolved_cs.as_ref().or(color_space);

        // Indexed colour space: the data carries a single palette index per
        // pixel. Expand to the base colour space so the PNG is a real picture
        // rather than indices misread as grayscale.
        if let Some((base, hival, palette)) = self.try_resolve_indexed(cs) {
            let base_components = self.color_space_component_count(Some(&base)) as usize;
            // 8-bit indices are already one byte per pixel — borrow directly
            // instead of cloning; only sub-byte depths need unpacking.
            let indices: std::borrow::Cow<[u8]> = if bits_per_component == 8 {
                std::borrow::Cow::Borrowed(data)
            } else {
                std::borrow::Cow::Owned(unpack_indices(data, width, height, bits_per_component))
            };
            let pixel_count = (width as usize) * (height as usize);
            if indices.len() < pixel_count {
                return Err(OperationError::ParseError(format!(
                    "Indexed image data too small: expected {} indices, got {}",
                    pixel_count,
                    indices.len()
                )));
            }
            let rgb = expand_indexed(&indices[..pixel_count], &palette, base_components, hival);
            return self.encode_png_maybe_alpha(
                &rgb,
                width,
                height,
                base_components as u8,
                8,
                smask_alpha,
            );
        }

        // Non-indexed: component count from the colour space.
        let icc_n = self.icc_components(cs);
        let components = image_sample_components(cs, icc_n);

        // Calculate expected data size. Use usize arithmetic so large images
        // do not overflow the intermediate product (a u32 multiply wraps near
        // 4 GB and would let truncated data pass the check below).
        let bytes_per_sample = if bits_per_component <= 8 { 1 } else { 2 };
        let expected_size = (width as usize)
            * (height as usize)
            * (components as usize)
            * (bytes_per_sample as usize);

        // Validate data size
        if data.len() < expected_size {
            return Err(OperationError::ParseError(format!(
                "Image data too small: expected {}, got {}",
                expected_size,
                data.len()
            )));
        }

        // Convert to PNG format using simple PNG encoding
        self.encode_png_maybe_alpha(
            data,
            width,
            height,
            components,
            bits_per_component,
            smask_alpha,
        )
    }

    /// Number of colour components for a (resolved) colour space, resolving an
    /// `/ICCBased` `/N` when needed.
    fn color_space_component_count(&self, color_space: Option<&PdfObject>) -> u8 {
        let icc_n = self.icc_components(color_space);
        image_sample_components(color_space, icc_n)
    }

    /// Decode an image's `/SMask` into a per-pixel 8-bit alpha buffer sized to
    /// `width`×`height` (nearest-neighbour resized if the mask resolution
    /// differs). Returns `None` when there is no soft mask, or when the mask
    /// is not a plain 8-bit grayscale raster we can interpret (e.g. a DCT or
    /// 16-bit mask), in which case the image is emitted without alpha.
    fn extract_smask_alpha(
        &self,
        image_dict: &crate::parser::objects::PdfDictionary,
        width: u32,
        height: u32,
    ) -> Option<Vec<u8>> {
        let smask = image_dict.0.get(&PdfName("SMask".to_string()))?;
        let resolved = self.document.resolve(smask).ok()?;
        let stream = match &resolved {
            PdfObject::Stream(s) => s,
            _ => return None,
        };
        let dict = &stream.dict.0;
        // Validate sign before casting: a negative /Width or /Height would cast
        // to a huge u32 and (on 32-bit targets) wrap the `sw * sh` product,
        // producing a corrupt mask. Reject non-positive dimensions outright.
        let sw_i = dict.get(&PdfName("Width".to_string()))?.as_integer()?;
        let sh_i = dict.get(&PdfName("Height".to_string()))?.as_integer()?;
        if sw_i <= 0 || sh_i <= 0 {
            return None;
        }
        let sw = sw_i as u32;
        let sh = sh_i as u32;
        let sbpc = dict
            .get(&PdfName("BitsPerComponent".to_string()))
            .and_then(|b| b.as_integer())
            .unwrap_or(8);
        if sbpc != 8 {
            return None; // only 8-bit masks are supported
        }

        let gray = self.decode_image_stream(stream).ok()?;
        let expected = (sw as usize) * (sh as usize);
        // A shorter buffer means the mask is not a plain gray raster (e.g. DCT);
        // bail rather than misread it.
        if gray.len() < expected {
            return None;
        }
        let gray = &gray[..expected];

        if sw == width && sh == height {
            return Some(gray.to_vec());
        }
        // Nearest-neighbour resize to the base image's dimensions.
        let mut out = Vec::with_capacity((width as usize) * (height as usize));
        for y in 0..height {
            let sy = ((y as u64 * sh as u64) / height as u64) as usize;
            let row = sy * sw as usize;
            for x in 0..width {
                let sx = ((x as u64 * sw as u64) / width as u64) as usize;
                out.push(gray[row + sx]);
            }
        }
        Some(out)
    }

    /// Encode `samples` as PNG. When `alpha` is present and the samples are
    /// 8-bit grayscale or RGB, composite it as the alpha channel and emit an
    /// RGBA PNG (grayscale is expanded to RGB first); otherwise emit the image
    /// as-is. Images that are 16-bit, DCT-encoded, or have 4 components (CMYK or
    /// an already-RGBA base) are emitted without alpha (the soft mask is dropped).
    fn encode_png_maybe_alpha(
        &self,
        samples: &[u8],
        width: u32,
        height: u32,
        components: u8,
        bits_per_component: u8,
        alpha: Option<&[u8]>,
    ) -> OperationResult<Vec<u8>> {
        match alpha {
            Some(a) if bits_per_component == 8 && (components == 1 || components == 3) => {
                let pixel_count = (width as usize) * (height as usize);
                // Callers guarantee these: the non-indexed path validates
                // `data.len() >= width*height*components`, the indexed path feeds
                // exactly `pixel_count*components` expanded bytes, and `alpha` is
                // sized to the base image. The `unwrap_or` below stay as a release
                // safety net; the asserts surface a broken contract in tests.
                debug_assert!(
                    samples.len() >= pixel_count * components as usize,
                    "sample buffer too short: {} < {}",
                    samples.len(),
                    pixel_count * components as usize
                );
                debug_assert_eq!(a.len(), pixel_count, "alpha length must match pixel count");
                let mut rgba = Vec::with_capacity(pixel_count * 4);
                for i in 0..pixel_count {
                    let (r, g, b) = if components == 3 {
                        let p = i * 3;
                        (
                            *samples.get(p).unwrap_or(&0),
                            *samples.get(p + 1).unwrap_or(&0),
                            *samples.get(p + 2).unwrap_or(&0),
                        )
                    } else {
                        let v = *samples.get(i).unwrap_or(&0);
                        (v, v, v)
                    };
                    // Missing mask samples default to opaque.
                    let al = *a.get(i).unwrap_or(&255);
                    rgba.extend_from_slice(&[r, g, b, al]);
                }
                self.create_png_from_raw_data(&rgba, width, height, 4, 8)
            }
            _ => self.create_png_from_raw_data(
                samples,
                width,
                height,
                components,
                bits_per_component,
            ),
        }
    }

    /// Create PNG from raw pixel data
    fn create_png_from_raw_data(
        &self,
        data: &[u8],
        width: u32,
        height: u32,
        components: u8,
        bits_per_component: u8,
    ) -> OperationResult<Vec<u8>> {
        // Simple PNG creation - create a basic PNG structure
        let mut png_data = Vec::new();

        // PNG signature
        png_data.extend_from_slice(&[0x89, 0x50, 0x4E, 0x47, 0x0D, 0x0A, 0x1A, 0x0A]);

        // IHDR chunk
        let mut ihdr = Vec::new();
        ihdr.extend_from_slice(&width.to_be_bytes());
        ihdr.extend_from_slice(&height.to_be_bytes());
        ihdr.push(bits_per_component);

        // Color type: 0 = grayscale, 2 = RGB, 6 = RGBA
        let color_type = match components {
            1 => 0, // Grayscale
            3 => 2, // RGB
            4 => 6, // RGBA
            _ => 2, // Default to RGB
        };
        ihdr.push(color_type);
        ihdr.push(0); // Compression method
        ihdr.push(0); // Filter method
        ihdr.push(0); // Interlace method

        self.write_png_chunk(&mut png_data, b"IHDR", &ihdr);

        // IDAT chunk - compress the image data
        let compressed_data = self.compress_image_data(data, width, height, components)?;
        self.write_png_chunk(&mut png_data, b"IDAT", &compressed_data);

        // IEND chunk
        self.write_png_chunk(&mut png_data, b"IEND", &[]);

        Ok(png_data)
    }

    /// Write a PNG chunk with proper CRC
    fn write_png_chunk(&self, output: &mut Vec<u8>, chunk_type: &[u8; 4], data: &[u8]) {
        // Length (4 bytes, big endian)
        output.extend_from_slice(&(data.len() as u32).to_be_bytes());

        // Chunk type (4 bytes)
        output.extend_from_slice(chunk_type);

        // Data
        output.extend_from_slice(data);

        // CRC (4 bytes, big endian)
        let crc = self.calculate_crc32(chunk_type, data);
        output.extend_from_slice(&crc.to_be_bytes());
    }

    /// Simple CRC32 calculation for PNG
    fn calculate_crc32(&self, chunk_type: &[u8; 4], data: &[u8]) -> u32 {
        // Simple CRC32 - in a real implementation we'd use a proper CRC library
        let mut crc: u32 = 0xFFFFFFFF;

        // Process chunk type
        for &byte in chunk_type {
            crc ^= byte as u32;
            for _ in 0..8 {
                if crc & 1 != 0 {
                    crc = (crc >> 1) ^ 0xEDB88320;
                } else {
                    crc >>= 1;
                }
            }
        }

        // Process data
        for &byte in data {
            crc ^= byte as u32;
            for _ in 0..8 {
                if crc & 1 != 0 {
                    crc = (crc >> 1) ^ 0xEDB88320;
                } else {
                    crc >>= 1;
                }
            }
        }

        crc ^ 0xFFFFFFFF
    }

    /// Compress image data for PNG IDAT chunk
    fn compress_image_data(
        &self,
        data: &[u8],
        width: u32,
        height: u32,
        components: u8,
    ) -> OperationResult<Vec<u8>> {
        use flate2::write::ZlibEncoder;
        use flate2::Compression;
        use std::io::Write;

        let mut encoder = ZlibEncoder::new(Vec::new(), Compression::default());

        // PNG requires scanline filtering - add filter byte (0 = None) to each row
        let bytes_per_pixel = components as usize;
        let bytes_per_row = width as usize * bytes_per_pixel;

        for row in 0..height {
            // Filter byte (0 = no filter)
            encoder.write_all(&[0])?;

            // Row data
            let start = row as usize * bytes_per_row;
            let end = start + bytes_per_row;
            if end <= data.len() {
                encoder.write_all(&data[start..end])?;
            }
        }

        encoder
            .finish()
            .map_err(|e| OperationError::ParseError(format!("Failed to compress PNG data: {e}")))
    }

    /// Convert CCITT Fax decoded data to PNG (for scanned documents)
    fn convert_ccitt_to_png(
        &self,
        data: &[u8],
        width: u32,
        height: u32,
    ) -> OperationResult<Vec<u8>> {
        // CCITT is typically 1-bit monochrome
        // Convert 1-bit to 8-bit grayscale
        let mut rgb_data = Vec::new();

        // Calculate potential row strides - try multiple alignments
        let bits_per_row = width as usize;
        let min_bytes_per_row = bits_per_row.div_ceil(8);

        // Try different row stride alignments (1, 2, 4, 8, 16 byte alignment)
        let possible_strides = [
            min_bytes_per_row,              // No padding
            (min_bytes_per_row + 1) & !1,   // 2-byte aligned
            (min_bytes_per_row + 3) & !3,   // 4-byte aligned
            (min_bytes_per_row + 7) & !7,   // 8-byte aligned
            (min_bytes_per_row + 15) & !15, // 16-byte aligned
        ];

        // Try to detect the correct stride by checking data patterns
        let correct_stride =
            self.detect_correct_row_stride(data, width, height, &possible_strides)?;

        for row in 0..height {
            let row_start = row as usize * correct_stride;

            for col in 0..width {
                let byte_idx = row_start + (col as usize / 8);
                let bit_idx = 7 - (col as usize % 8);

                if byte_idx < data.len() {
                    let bit = (data[byte_idx] >> bit_idx) & 1;
                    // CCITT: 0 = black, 1 = white
                    let gray_value = if bit == 0 { 0 } else { 255 };
                    rgb_data.push(gray_value);
                } else {
                    rgb_data.push(255); // White for missing data
                }
            }
        }

        // Create PNG from grayscale data
        self.create_png_from_raw_data(&rgb_data, width, height, 1, 8)
    }

    /// Detect the correct row stride by analyzing data patterns
    fn detect_correct_row_stride(
        &self,
        data: &[u8],
        width: u32,
        height: u32,
        possible_strides: &[usize],
    ) -> OperationResult<usize> {
        let bits_per_row = width as usize;
        let min_bytes_per_row = bits_per_row.div_ceil(8);

        // If we don't have enough data for analysis, use minimum stride
        if data.len() < min_bytes_per_row * 3 {
            return Ok(min_bytes_per_row);
        }

        // Calculate expected total size for each stride
        for &stride in possible_strides {
            let expected_size = stride * height as usize;

            // If this stride gives us a size close to actual data length, use it
            if expected_size <= data.len() && (data.len() - expected_size) < stride * 2 {
                // Allow some tolerance

                return Ok(stride);
            }
        }

        // If no stride fits perfectly, calculate from data length
        let calculated_stride = data.len() / height as usize;
        if calculated_stride >= min_bytes_per_row {
            return Ok(calculated_stride);
        }

        // Fallback to minimum
        Ok(min_bytes_per_row)
    }

    /// Check if preprocessing should be applied
    #[allow(dead_code)]
    fn should_preprocess(&self) -> bool {
        self.options.preprocessing.auto_correct_rotation
            || self.options.preprocessing.enhance_contrast
            || self.options.preprocessing.denoise
            || self.options.preprocessing.upscale_small_images
            || self.options.preprocessing.force_grayscale
    }

    /// Apply image preprocessing
    #[cfg(feature = "external-images")]
    fn preprocess_image_data(
        &self,
        data: &[u8],
        width: u32,
        height: u32,
        format: ImageFormat,
    ) -> OperationResult<Vec<u8>> {
        // Load image using the image crate
        let img_format = match format {
            ImageFormat::Jpeg => ImageLibFormat::Jpeg,
            ImageFormat::Png => ImageLibFormat::Png,
            ImageFormat::Tiff => ImageLibFormat::Tiff,
            ImageFormat::Raw => {
                // For raw data, create a simple RGB image
                return self.preprocess_raw_image_data(data, width, height);
            }
        };

        let img = image::load_from_memory_with_format(data, img_format)
            .map_err(|e| OperationError::ParseError(format!("Failed to load image: {e}")))?;

        let mut processed_img = img;

        // Apply preprocessing steps
        processed_img = self.apply_rotation_correction(processed_img)?;
        processed_img = self.apply_contrast_enhancement(processed_img)?;
        processed_img = self.apply_noise_reduction(processed_img)?;
        processed_img = self.apply_upscaling(processed_img, width, height)?;

        if self.options.preprocessing.force_grayscale {
            processed_img = DynamicImage::ImageLuma8(processed_img.to_luma8());
        }

        // Encode back to bytes
        let mut output = Vec::new();
        processed_img
            .write_to(&mut std::io::Cursor::new(&mut output), img_format)
            .map_err(|e| OperationError::ParseError(format!("Failed to encode image: {e}")))?;

        Ok(output)
    }

    /// Preprocess raw image data
    #[cfg(feature = "external-images")]
    fn preprocess_raw_image_data(
        &self,
        data: &[u8],
        width: u32,
        height: u32,
    ) -> OperationResult<Vec<u8>> {
        // Create a simple grayscale image from raw data
        if data.len() < (width * height) as usize {
            return Err(OperationError::ParseError(
                "Raw image data too small".to_string(),
            ));
        }

        let img_buffer = ImageBuffer::<Luma<u8>, Vec<u8>>::from_raw(
            width,
            height,
            data[..(width * height) as usize].to_vec(),
        )
        .ok_or_else(|| OperationError::ParseError("Failed to create image buffer".to_string()))?;

        let img = DynamicImage::ImageLuma8(img_buffer);
        let mut processed_img = img;

        // Apply preprocessing
        processed_img = self.apply_rotation_correction(processed_img)?;
        processed_img = self.apply_contrast_enhancement(processed_img)?;
        processed_img = self.apply_noise_reduction(processed_img)?;
        processed_img = self.apply_upscaling(processed_img, width, height)?;

        // Encode to PNG
        let mut output = Vec::new();
        processed_img
            .write_to(&mut std::io::Cursor::new(&mut output), ImageLibFormat::Png)
            .map_err(|e| OperationError::ParseError(format!("Failed to encode image: {e}")))?;

        Ok(output)
    }

    /// Auto-detect and correct rotation
    #[cfg(feature = "external-images")]
    fn apply_rotation_correction(&self, img: DynamicImage) -> OperationResult<DynamicImage> {
        if !self.options.preprocessing.auto_correct_rotation {
            return Ok(img);
        }

        // Simple rotation detection based on aspect ratio and content analysis
        let (width, height) = img.dimensions();

        // If image is wider than it is tall but contains mostly vertical text,
        // it might need rotation. This is a simplified heuristic.
        if width > height * 2 {
            // Likely rotated 90 degrees - try rotating
            return Ok(img.rotate90());
        }

        // For now, return as-is. In a more sophisticated implementation,
        // we could use OCR or edge detection to determine optimal rotation.
        Ok(img)
    }

    /// Enhance contrast for better OCR
    #[cfg(feature = "external-images")]
    fn apply_contrast_enhancement(&self, img: DynamicImage) -> OperationResult<DynamicImage> {
        if !self.options.preprocessing.enhance_contrast {
            return Ok(img);
        }

        // Apply histogram equalization by adjusting brightness and contrast
        let enhanced = img.adjust_contrast(20.0); // Increase contrast by 20%
        Ok(enhanced.brighten(10)) // Slightly brighten
    }

    /// Apply noise reduction
    #[cfg(feature = "external-images")]
    fn apply_noise_reduction(&self, img: DynamicImage) -> OperationResult<DynamicImage> {
        if !self.options.preprocessing.denoise {
            return Ok(img);
        }

        // Simple blur to reduce noise
        Ok(img.blur(0.5))
    }

    /// Upscale small images for better OCR
    #[cfg(feature = "external-images")]
    fn apply_upscaling(
        &self,
        img: DynamicImage,
        original_width: u32,
        original_height: u32,
    ) -> OperationResult<DynamicImage> {
        if !self.options.preprocessing.upscale_small_images {
            return Ok(img);
        }

        let min_dimension = original_width.min(original_height);
        if min_dimension < self.options.preprocessing.upscale_threshold {
            let new_width = original_width * self.options.preprocessing.upscale_factor;
            let new_height = original_height * self.options.preprocessing.upscale_factor;

            return Ok(img.resize(
                new_width,
                new_height,
                image::imageops::FilterType::CatmullRom,
            ));
        }

        Ok(img)
    }
}

/// Extract all images from a PDF file
pub fn extract_images_from_pdf<P: AsRef<Path>>(
    input_path: P,
    options: ExtractImagesOptions,
) -> OperationResult<Vec<ExtractedImage>> {
    let document = PdfReader::open_document(input_path)
        .map_err(|e| OperationError::ParseError(e.to_string()))?;

    let mut extractor = ImageExtractor::new(document, options);
    extractor.extract_all()
}

/// Extract images from specific pages
pub fn extract_images_from_pages<P: AsRef<Path>>(
    input_path: P,
    pages: &[usize],
    options: ExtractImagesOptions,
) -> OperationResult<Vec<ExtractedImage>> {
    let document = PdfReader::open_document(input_path)
        .map_err(|e| OperationError::ParseError(e.to_string()))?;

    let mut extractor = ImageExtractor::new(document, options);
    let mut all_images = Vec::new();

    for &page_num in pages {
        let page_images = extractor.extract_from_page(page_num)?;
        all_images.extend(page_images);
    }

    Ok(all_images)
}

/// Number of colour samples per pixel carried by the *image data* for a colour
/// space.
///
/// For `Indexed` the data carries a single palette index per pixel (1). For
/// `ICCBased`, pass the profile's resolved `/N` via `icc_n` (defaults to 3 when
/// unknown). `DeviceN` reports the number of named colorants.
fn image_sample_components(color_space: Option<&PdfObject>, icc_n: Option<u8>) -> u8 {
    match color_space {
        Some(PdfObject::Name(cs)) => match cs.0.as_str() {
            "DeviceGray" | "G" | "CalGray" => 1,
            "DeviceRGB" | "RGB" | "CalRGB" | "Lab" => 3,
            "DeviceCMYK" | "CMYK" => 4,
            _ => 3,
        },
        Some(PdfObject::Array(array)) => {
            match array
                .0
                .first()
                .and_then(|o| o.as_name())
                .map(|n| n.0.as_str())
            {
                Some("Indexed") | Some("I") => 1,
                Some("Separation") => 1,
                Some("DeviceN") => array
                    .0
                    .get(1)
                    .and_then(|o| o.as_array())
                    .map(|names| names.0.len().max(1) as u8)
                    .unwrap_or(1),
                Some("ICCBased") => icc_n.unwrap_or(3),
                Some("CalGray") | Some("DeviceGray") => 1,
                Some("DeviceCMYK") => 4,
                Some("CalRGB") | Some("Lab") | Some("DeviceRGB") => 3,
                _ => 3,
            }
        }
        _ => 3,
    }
}

/// Expand one-index-per-pixel data into `base_components`-byte pixels using the
/// `lookup` palette (`(hival + 1) * base_components` bytes).
///
/// Indices greater than `hival` are clamped; a short palette is zero-padded so
/// the output length is always `indices.len() * base_components`.
fn expand_indexed(indices: &[u8], lookup: &[u8], base_components: usize, hival: usize) -> Vec<u8> {
    let mut out = Vec::with_capacity(indices.len() * base_components);
    for &idx in indices {
        let entry = (idx as usize).min(hival);
        let start = entry * base_components;
        for c in 0..base_components {
            out.push(lookup.get(start + c).copied().unwrap_or(0));
        }
    }
    out
}

/// Unpack packed samples (1/2/4/8 bits per component) into one byte per sample,
/// honouring PDF row alignment (each scanline starts on a byte boundary).
///
/// For `bits_per_component >= 8` the data is returned unchanged.
fn unpack_indices(data: &[u8], width: u32, height: u32, bits_per_component: u8) -> Vec<u8> {
    // Only the spec-valid packed depths {1, 2, 4} are unpacked. 8 (and the
    // defensive 0) pass through unchanged; any other value (e.g. a malformed
    // 3/5/6/7) would make the scanline shift underflow, so it also passes
    // through and the caller's size check rejects it cleanly.
    if !matches!(bits_per_component, 1 | 2 | 4) {
        return data.to_vec();
    }
    let bpc = bits_per_component as usize;
    let width = width as usize;
    let height = height as usize;
    let row_bytes = (width * bpc).div_ceil(8);
    let mask = (1u16 << bpc) - 1;
    let mut out = Vec::with_capacity(width * height);
    for row in 0..height {
        let row_start = row * row_bytes;
        for col in 0..width {
            let bit_index = col * bpc;
            let byte = row_start + bit_index / 8;
            let shift = 8 - bpc - (bit_index % 8);
            let value = data
                .get(byte)
                .map(|b| ((*b as u16) >> shift) & mask)
                .unwrap_or(0);
            out.push(value as u8);
        }
    }
    out
}

#[cfg(test)]
mod tests {
    use super::*;
    use tempfile::TempDir;

    fn name(s: &str) -> PdfObject {
        PdfObject::Name(PdfName(s.to_string()))
    }

    #[test]
    fn test_image_sample_components_device_color_spaces() {
        assert_eq!(image_sample_components(Some(&name("DeviceGray")), None), 1);
        assert_eq!(image_sample_components(Some(&name("DeviceRGB")), None), 3);
        assert_eq!(image_sample_components(Some(&name("DeviceCMYK")), None), 4);
        // Unknown name / missing colour space default to RGB (legacy behaviour).
        assert_eq!(image_sample_components(Some(&name("Weird")), None), 3);
        assert_eq!(image_sample_components(None, None), 3);
    }

    #[test]
    fn test_image_sample_components_indexed_is_one() {
        let indexed = PdfObject::Array(PdfArray(vec![
            name("Indexed"),
            name("DeviceRGB"),
            PdfObject::Integer(23),
            PdfObject::String(crate::parser::objects::PdfString(vec![0u8; 72])),
        ]));
        assert_eq!(image_sample_components(Some(&indexed), None), 1);
    }

    #[test]
    fn test_image_sample_components_iccbased_uses_n() {
        let icc = PdfObject::Array(PdfArray(vec![name("ICCBased"), PdfObject::Reference(5, 0)]));
        assert_eq!(image_sample_components(Some(&icc), Some(1)), 1);
        assert_eq!(image_sample_components(Some(&icc), Some(4)), 4);
        // Falls back to RGB when /N is unknown.
        assert_eq!(image_sample_components(Some(&icc), None), 3);
    }

    #[test]
    fn test_image_sample_components_devicen_counts_colorants() {
        let devicen = PdfObject::Array(PdfArray(vec![
            name("DeviceN"),
            PdfObject::Array(PdfArray(vec![name("Cyan"), name("Magenta")])),
            name("DeviceCMYK"),
            PdfObject::Reference(9, 0),
        ]));
        assert_eq!(image_sample_components(Some(&devicen), None), 2);
    }

    #[test]
    fn test_expand_indexed_maps_indices_to_palette_rgb() {
        // 3-entry RGB palette: red, green, blue.
        let palette = vec![255, 0, 0, 0, 255, 0, 0, 0, 255];
        let indices = [0u8, 2, 1];
        let rgb = expand_indexed(&indices, &palette, 3, 2);
        assert_eq!(rgb, vec![255, 0, 0, 0, 0, 255, 0, 255, 0]);
    }

    #[test]
    fn test_expand_indexed_clamps_out_of_range_index() {
        let palette = vec![10, 20, 30, 40, 50, 60]; // 2 entries, hival = 1
                                                    // Index 5 is past hival -> clamped to the last entry.
        let rgb = expand_indexed(&[5u8], &palette, 3, 1);
        assert_eq!(rgb, vec![40, 50, 60]);
    }

    #[test]
    fn test_unpack_indices_passthrough_for_8bit() {
        let data = vec![1, 2, 3, 4];
        assert_eq!(unpack_indices(&data, 2, 2, 8), data);
    }

    #[test]
    fn test_unpack_indices_4bit_two_pixels_per_byte() {
        // One row of 2 pixels at 4 bpc packed into a single byte 0xA3 -> [0xA, 0x3].
        let data = vec![0xA3];
        assert_eq!(unpack_indices(&data, 2, 1, 4), vec![0x0A, 0x03]);
    }

    #[test]
    fn test_unpack_indices_2bit_four_pixels_per_byte() {
        // One row of 4 pixels at 2 bpc packed into a byte 0b11_10_01_00 -> [3,2,1,0].
        let data = vec![0b1110_0100];
        assert_eq!(unpack_indices(&data, 4, 1, 2), vec![3, 2, 1, 0]);
    }

    #[test]
    fn test_unpack_indices_passthrough_for_unsupported_bpc() {
        // A malformed 3 bpc must not panic (shift underflow); it passes through.
        let data = vec![0xAB, 0xCD];
        assert_eq!(unpack_indices(&data, 4, 1, 3), data);
    }

    #[test]
    fn test_unpack_indices_1bit_respects_row_byte_alignment() {
        // 3 pixels per row at 1 bpc => each row occupies 1 byte (padded).
        // Row 0: 0b101_00000 -> 1,0,1 ; Row 1: 0b011_00000 -> 0,1,1
        let data = vec![0b1010_0000, 0b0110_0000];
        assert_eq!(unpack_indices(&data, 3, 2, 1), vec![1, 0, 1, 0, 1, 1]);
    }

    #[test]
    fn test_extract_options_default() {
        let options = ExtractImagesOptions::default();
        assert_eq!(options.output_dir, PathBuf::from("."));
        assert!(options.extract_inline);
        assert_eq!(options.min_size, Some(10));
        assert!(options.create_dir);
    }

    #[test]
    fn test_filename_pattern() {
        let options = ExtractImagesOptions {
            name_pattern: "img_{page}_{index}.{format}".to_string(),
            ..Default::default()
        };

        let pattern = options
            .name_pattern
            .replace("{page}", "1")
            .replace("{index}", "2")
            .replace("{format}", "jpg");

        assert_eq!(pattern, "img_1_2.jpg");
    }

    #[test]
    fn test_extract_options_custom() {
        let temp_dir = TempDir::new().unwrap();
        let options = ExtractImagesOptions {
            output_dir: temp_dir.path().to_path_buf(),
            name_pattern: "custom_{page}_{index}.{format}".to_string(),
            extract_inline: false,
            min_size: Some(50),
            create_dir: false,
            preprocessing: ImagePreprocessingOptions::default(),
        };

        assert_eq!(options.output_dir, temp_dir.path());
        assert_eq!(options.name_pattern, "custom_{page}_{index}.{format}");
        assert!(!options.extract_inline);
        assert_eq!(options.min_size, Some(50));
        assert!(!options.create_dir);
    }

    #[test]
    fn test_extract_options_debug_clone() {
        let options = ExtractImagesOptions {
            output_dir: PathBuf::from("/test/path"),
            name_pattern: "test.{format}".to_string(),
            extract_inline: true,
            min_size: None,
            create_dir: true,
            preprocessing: ImagePreprocessingOptions::default(),
        };

        let debug_str = format!("{options:?}");
        assert!(debug_str.contains("ExtractImagesOptions"));
        assert!(debug_str.contains("/test/path"));

        let cloned = options.clone();
        assert_eq!(cloned.output_dir, options.output_dir);
        assert_eq!(cloned.name_pattern, options.name_pattern);
        assert_eq!(cloned.extract_inline, options.extract_inline);
        assert_eq!(cloned.min_size, options.min_size);
        assert_eq!(cloned.create_dir, options.create_dir);
    }

    #[test]
    fn test_extracted_image_struct() {
        let image = ExtractedImage {
            page_number: 0,
            image_index: 1,
            file_path: PathBuf::from("/test/image.jpg"),
            width: 100,
            height: 200,
            format: ImageFormat::Jpeg,
        };

        assert_eq!(image.page_number, 0);
        assert_eq!(image.image_index, 1);
        assert_eq!(image.file_path, PathBuf::from("/test/image.jpg"));
        assert_eq!(image.width, 100);
        assert_eq!(image.height, 200);
        assert_eq!(image.format, ImageFormat::Jpeg);
    }

    #[test]
    fn test_extracted_image_debug() {
        let image = ExtractedImage {
            page_number: 5,
            image_index: 3,
            file_path: PathBuf::from("output.png"),
            width: 512,
            height: 768,
            format: ImageFormat::Png,
        };

        let debug_str = format!("{image:?}");
        assert!(debug_str.contains("ExtractedImage"));
        assert!(debug_str.contains("5"));
        assert!(debug_str.contains("3"));
        assert!(debug_str.contains("output.png"));
        assert!(debug_str.contains("512"));
        assert!(debug_str.contains("768"));
    }

    // Helper function to create minimal valid PDF for testing
    fn create_minimal_pdf(temp_file: &std::path::Path) {
        let minimal_pdf = b"%PDF-1.7\n\
1 0 obj\n\
<< /Type /Catalog /Pages 2 0 R >>\n\
endobj\n\
2 0 obj\n\
<< /Type /Pages /Kids [] /Count 0 >>\n\
endobj\n\
xref\n\
0 3\n\
0000000000 65535 f \n\
0000000009 00000 n \n\
0000000055 00000 n \n\
trailer\n\
<< /Size 3 /Root 1 0 R >>\n\
startxref\n\
105\n\
%%EOF";
        std::fs::write(temp_file, minimal_pdf).unwrap();
    }

    #[test]
    fn test_detect_image_format_png() {
        // Create a minimal valid PDF document for testing
        let temp_dir = TempDir::new().unwrap();
        let temp_file = temp_dir.path().join("test.pdf");
        create_minimal_pdf(&temp_file);

        let document = PdfReader::open_document(&temp_file).unwrap();
        let extractor = ImageExtractor::new(document, ExtractImagesOptions::default());

        // PNG magic bytes
        let png_data = b"\x89PNG\r\n\x1a\n\x00\x00\x00\x0DIHDR";
        let format = extractor.detect_image_format_from_data(png_data).unwrap();
        assert_eq!(format, ImageFormat::Png);
    }

    #[test]
    fn test_detect_image_format_jpeg() {
        let temp_dir = TempDir::new().unwrap();
        let temp_file = temp_dir.path().join("test.pdf");
        create_minimal_pdf(&temp_file);

        let document = PdfReader::open_document(&temp_file).unwrap();
        let extractor = ImageExtractor::new(document, ExtractImagesOptions::default());

        // JPEG magic bytes
        let jpeg_data = b"\xFF\xD8\xFF\xE0\x00\x10JFIF";
        let format = extractor.detect_image_format_from_data(jpeg_data).unwrap();
        assert_eq!(format, ImageFormat::Jpeg);
    }

    #[test]
    fn test_detect_image_format_tiff_little_endian() {
        let temp_dir = TempDir::new().unwrap();
        let temp_file = temp_dir.path().join("test.pdf");
        create_minimal_pdf(&temp_file);

        let document = PdfReader::open_document(&temp_file).unwrap();
        let extractor = ImageExtractor::new(document, ExtractImagesOptions::default());

        // TIFF little endian magic bytes
        let tiff_data = b"II\x2A\x00\x08\x00\x00\x00";
        let format = extractor.detect_image_format_from_data(tiff_data).unwrap();
        assert_eq!(format, ImageFormat::Tiff);
    }

    #[test]
    fn test_detect_image_format_tiff_big_endian() {
        let temp_dir = TempDir::new().unwrap();
        let temp_file = temp_dir.path().join("test.pdf");
        create_minimal_pdf(&temp_file);

        let document = PdfReader::open_document(&temp_file).unwrap();
        let extractor = ImageExtractor::new(document, ExtractImagesOptions::default());

        // TIFF big endian magic bytes
        let tiff_data = b"MM\x00\x2A\x00\x00\x00\x08";
        let format = extractor.detect_image_format_from_data(tiff_data).unwrap();
        assert_eq!(format, ImageFormat::Tiff);
    }

    #[test]
    fn test_detect_image_format_unknown() {
        let temp_dir = TempDir::new().unwrap();
        let temp_file = temp_dir.path().join("test.pdf");
        create_minimal_pdf(&temp_file);

        let document = PdfReader::open_document(&temp_file).unwrap();
        let extractor = ImageExtractor::new(document, ExtractImagesOptions::default());

        // Unknown format - should default to PNG
        let unknown_data = b"\x00\x01\x02\x03\x04\x05\x06\x07\x08";
        let format = extractor
            .detect_image_format_from_data(unknown_data)
            .unwrap();
        assert_eq!(format, ImageFormat::Png); // Default fallback
    }

    #[test]
    fn test_detect_image_format_short_data() {
        let temp_dir = TempDir::new().unwrap();
        let temp_file = temp_dir.path().join("test.pdf");
        create_minimal_pdf(&temp_file);

        let document = PdfReader::open_document(&temp_file).unwrap();
        let extractor = ImageExtractor::new(document, ExtractImagesOptions::default());

        // Too short data (less than 2 bytes)
        let short_data = b"\xFF";
        let result = extractor.detect_image_format_from_data(short_data);
        assert!(result.is_err());
        match result {
            Err(OperationError::ParseError(msg)) => {
                assert!(msg.contains("too short"));
            }
            _ => panic!("Expected ParseError"),
        }
    }

    #[test]
    fn test_filename_pattern_replacements() {
        let options = ExtractImagesOptions {
            name_pattern: "page_{page}_img_{index}_{format}.{format}".to_string(),
            ..Default::default()
        };

        let pattern = options
            .name_pattern
            .replace("{page}", "10")
            .replace("{index}", "5")
            .replace("{format}", "png");

        assert_eq!(pattern, "page_10_img_5_png.png");
    }

    #[test]
    fn test_extract_options_no_min_size() {
        let options = ExtractImagesOptions {
            min_size: None,
            ..Default::default()
        };

        assert_eq!(options.min_size, None);
    }

    #[test]
    fn test_create_output_directory() {
        let temp_dir = TempDir::new().unwrap();
        let output_dir = temp_dir.path().join("new_dir");

        let options = ExtractImagesOptions {
            output_dir: output_dir.clone(),
            create_dir: true,
            ..Default::default()
        };

        // In real usage, ImageExtractor would create this directory
        assert!(!output_dir.exists());
        assert_eq!(options.output_dir, output_dir);
        assert!(options.create_dir);
    }

    #[test]
    fn test_pattern_with_special_chars() {
        let options = ExtractImagesOptions {
            name_pattern: "img-{page}_{index}.{format}".to_string(),
            ..Default::default()
        };

        let pattern = options
            .name_pattern
            .replace("{page}", "1")
            .replace("{index}", "1")
            .replace("{format}", "jpg");

        assert_eq!(pattern, "img-1_1.jpg");
    }

    #[test]
    fn test_multiple_format_extensions() {
        let formats = vec![
            (ImageFormat::Jpeg, "jpg"),
            (ImageFormat::Png, "png"),
            (ImageFormat::Tiff, "tiff"),
        ];

        for (format, expected_ext) in formats {
            let extension = match format {
                ImageFormat::Jpeg => "jpg",
                ImageFormat::Png => "png",
                ImageFormat::Tiff => "tiff",
                ImageFormat::Raw => "raw",
            };
            assert_eq!(extension, expected_ext);
        }
    }

    #[test]
    fn test_extract_inline_option() {
        let mut options = ExtractImagesOptions::default();
        assert!(options.extract_inline);

        options.extract_inline = false;
        assert!(!options.extract_inline);
    }

    #[test]
    fn test_min_size_filtering() {
        let options_with_min = ExtractImagesOptions {
            min_size: Some(100),
            ..Default::default()
        };

        let options_no_min = ExtractImagesOptions {
            min_size: None,
            ..Default::default()
        };

        assert_eq!(options_with_min.min_size, Some(100));
        assert_eq!(options_no_min.min_size, None);
    }

    #[test]
    fn test_output_path_combinations() {
        let base_dir = PathBuf::from("/output");
        let options = ExtractImagesOptions {
            output_dir: base_dir,
            name_pattern: "img_{page}_{index}.{format}".to_string(),
            ..Default::default()
        };

        let filename = options
            .name_pattern
            .replace("{page}", "1")
            .replace("{index}", "2")
            .replace("{format}", "png");

        let full_path = options.output_dir.join(filename);
        assert_eq!(full_path, PathBuf::from("/output/img_1_2.png"));
    }

    #[test]
    fn test_pattern_without_placeholders() {
        let options = ExtractImagesOptions {
            name_pattern: "static_name.jpg".to_string(),
            ..Default::default()
        };

        let pattern = options
            .name_pattern
            .replace("{page}", "1")
            .replace("{index}", "2")
            .replace("{format}", "png");

        assert_eq!(pattern, "static_name.jpg"); // No placeholders replaced
    }

    #[test]
    fn test_detect_format_edge_cases() {
        let temp_dir = TempDir::new().unwrap();
        let temp_file = temp_dir.path().join("test.pdf");
        create_minimal_pdf(&temp_file);

        let document = PdfReader::open_document(&temp_file).unwrap();
        let extractor = ImageExtractor::new(document, ExtractImagesOptions::default());

        // Empty data
        let empty_data = b"";
        assert!(extractor.detect_image_format_from_data(empty_data).is_err());

        // Data exactly 8 bytes (minimum for PNG check)
        let exact_8 = b"\x89PNG\r\n\x1a\n";
        let format = extractor.detect_image_format_from_data(exact_8).unwrap();
        assert_eq!(format, ImageFormat::Png);

        // Data exactly 4 bytes (minimum for TIFF check)
        let exact_4 = b"II\x2A\x00";
        let format = extractor.detect_image_format_from_data(exact_4).unwrap();
        assert_eq!(format, ImageFormat::Tiff);

        // Data exactly 2 bytes (minimum for JPEG check)
        let exact_2 = b"\xFF\xD8";
        let format = extractor.detect_image_format_from_data(exact_2).unwrap();
        assert_eq!(format, ImageFormat::Jpeg); // JPEG only needs 2 bytes
    }

    #[test]
    fn test_complex_filename_pattern() {
        let options = ExtractImagesOptions {
            name_pattern: "{format}/page{page}/image_{index}_{page}.{format}".to_string(),
            ..Default::default()
        };

        let pattern = options
            .name_pattern
            .replace("{page}", "5")
            .replace("{index}", "3")
            .replace("{format}", "jpeg");

        assert_eq!(pattern, "jpeg/page5/image_3_5.jpeg");
    }

    #[test]
    fn test_image_dimensions() {
        let small_image = ExtractedImage {
            page_number: 0,
            image_index: 0,
            file_path: PathBuf::from("small.jpg"),
            width: 5,
            height: 5,
            format: ImageFormat::Jpeg,
        };

        let large_image = ExtractedImage {
            page_number: 0,
            image_index: 1,
            file_path: PathBuf::from("large.jpg"),
            width: 2000,
            height: 3000,
            format: ImageFormat::Jpeg,
        };

        assert_eq!(small_image.width, 5);
        assert_eq!(small_image.height, 5);
        assert_eq!(large_image.width, 2000);
        assert_eq!(large_image.height, 3000);
    }

    #[test]
    fn test_page_and_index_numbering() {
        // Test that page numbers and indices work correctly
        let image1 = ExtractedImage {
            page_number: 0, // 0-indexed
            image_index: 0,
            file_path: PathBuf::from("first.jpg"),
            width: 100,
            height: 100,
            format: ImageFormat::Jpeg,
        };

        let image2 = ExtractedImage {
            page_number: 99,  // Large page number
            image_index: 255, // Large index
            file_path: PathBuf::from("last.jpg"),
            width: 100,
            height: 100,
            format: ImageFormat::Jpeg,
        };

        assert_eq!(image1.page_number, 0);
        assert_eq!(image1.image_index, 0);
        assert_eq!(image2.page_number, 99);
        assert_eq!(image2.image_index, 255);
    }
}

#[cfg(test)]
#[path = "extract_images_tests.rs"]
mod extract_images_tests;