cvCNN

cvCNN is a C++ training project for handwritten digit recognition based on a small convolutional neural network built with LibTorch and OpenCV.

It trains on the handwritten digit images generated by the companion project cvData. This repository does not create the dataset itself: it expects the preprocessed images and labels exported by cvData. As shown in the image above, in this setup the project can achieve a 100% recognition rate on a 100-digit test sample after training on 6900 handwritten digits.

Workspace Layout

cvCNN and cvData must be located in the same parent workspace directory, because cvCNN uses relative paths to load the dataset.

Expected layout:

workspace/
├── cvCNN
└── cvData

At runtime, cvCNN reads:

• ../../cvData/img/out/binning/
• ../../cvData/img/out/labels.txt

What The Program Does

The executable:

• loads handwritten digit images produced by cvData
• reads the associated labels
• builds a train/test split
• trains a small CNN for digit classification
• evaluates the model on the test set
• exports a visualization of the first convolution kernels to img/kernels_conv1.png

Requirements

• CMake 3.18+
• A C++17 compiler
• OpenCV
• LibTorch

The current CMakeLists.txt expects LibTorch to be available at:

/opt/libtorch

If your LibTorch installation is elsewhere, update Torch_DIR in CMakeLists.txt.

Build

From the cvCNN repository root:

mkdir -p build
cd build
cmake ..
cmake --build .

Run

Make sure cvData has already generated the dataset in img/out/.

Then run:

cd build
./cvCNN

Running from the build/ directory is important because the dataset paths are resolved relative to the executable location.