Rust Amazon Captcha Solver

An attempt to resolve Amazon.com captchas without using Tesseract OCR. Highly inspired by gopkg-dev/amazoncaptcha and a-maliarov/amazoncaptcha. We reuse the dataset from the Go library but in an uncompressed bincode format.

We have simplified the resolving process as much as possible, resulting in a library with less than 200 LoC. In terms of speed, in a release build on an M1 Mac, loading the library takes approximately 30ms (dataset loading), and resolving a captcha takes about 40ms.

Functional Schema

Amazon captchas are quite repetitive, with letters always being less than 32 pixels, and there are not millions of possible combinations. The dataset.bin contains most of the possibilities. When a letter does not match exactly, we use a similarity comparison function.

Example

use amazon_captcha_rs::new_solver; 

let image = image::open("testdata/caggpa.jpg").unwrap();

let solver = new_solver();
let response = solver.resolve_image(&image).unwrap();

assert_eq!(response, "caggpa");

Precision

We downloaded and resolved 100 captchas in the examples directory to test the precision. As of version 0.2.0:

Solved: 93/99
Precision: 93.94%

Changelog

Version 0.2.0

• Improve letter extraction method
• Add letter merging (used when first letter is cropped with last letter)
• Precision bumped to 93%.

Version 0.1.0

• Initial release (64% precision)