Crate ruvector_cnn 

CNN Feature Extraction for Image Embeddings

This crate provides pure Rust CNN-based feature extraction with SIMD acceleration. It is designed for CPU-only deployment including WASM environments.

Features

• MobileNet-V3 Small/Large backbones
• SIMD acceleration (AVX2, NEON, WASM SIMD128)
• INT8 quantization support
• Pure Rust (no BLAS/OpenCV dependencies)
• Parallel batch processing with rayon (optional)

Example

use ruvector_cnn::{CnnEmbedder, EmbeddingConfig};

let embedder = CnnEmbedder::new(EmbeddingConfig::default())?;
let embedding = embedder.extract(&image_data, width, height)?;
println!("Embedding dim: {}", embedding.len());

Using MobileNet Backbone

use ruvector_cnn::embedding::MobileNetEmbedder;

// Create a MobileNetV3-Small embedder
let embedder = MobileNetEmbedder::v3_small()?;

// Extract features from normalized float tensor (NCHW format)
let features = embedder.extract(&image_tensor, 224, 224)?;
println!("Feature dim: {}", features.len()); // 576 for V3-Small

Modules

contrastive

Contrastive Learning Module

int8

INT8 quantization module for ADR-091

kernels

INT8 Quantized Kernels Module

layers

Neural Network Layers

quantize

INT8 Quantization Module (ADR-091)

simd

SIMD Backend Dispatch Module

Structs

CnnEmbedder

CNN Embedder for feature extraction

EmbeddingConfig

Configuration for CNN embedding extraction

Tensor

A multi-dimensional tensor with NHWC layout

Enums

CnnError

Errors that can occur during CNN operations.

Traits

EmbeddingExtractor

Embedding extractor trait

Type Aliases

CnnResult

Result type for CNN operations.