C-Neural-Engine OCR

A lightweight neural network for optical character recognition (OCR), written in pure C.

Features

• Single hidden layer neural network with sigmoid activation
• Mini-batch gradient descent training
• Binary model serialization for fast loading

Optimizations

• Flattened weight matrices: 2D weights stored as contiguous 1D arrays in row-major order for better CPU cache utilization.
• Mini-batch gradient descent: Gradients accumulated across samples before weight updates, improving convergence.
• Inlined activation functions: static inline eliminates function call overhead.
• BLAS-ready kernels: Matrix operations abstracted for easy swap to optimized BLAS routines.
• Zero-copy serialization: Weights written/read as raw binary, no parsing overhead.

Building

make

For debug builds with sanitizers:

make debug

Usage

Training

Place your training images in the letters/ directory:

letters/
  A/
    sample1.png
    sample2.png
  B/
    sample1.png
  ...

Each image must be 32x32 pixels in grayscale. Run:

./ocr-engine

The trained model is saved to model.bin.

Inference

./ocr-engine path/to/image.png

Outputs the predicted letter (A-Z) with confidence score.