dizlexic/captcha-ocr-trainer

Captcha OCR Trainer is a Python toolkit for training an AI model that reads distorted text from images, specifically targeting the short scrambled-character sequences used in CAPTCHAs. It produces a trained model you can then use to automatically decode those challenge images. The model architecture is called CRNN, a Convolutional Recurrent Neural Network. The convolutional portion extracts visual features from the image, a BiLSTM layer handles the character sequence, and a technique called CTC Loss guides training without needing exact character positions marked in the data. The training code is written in PyTorch, and training is tuned for recognising sequences of four distorted characters. To train the model you place your CAPTCHA images in a folder and create a label file (labels.jsonl) that maps each filename to its correct text string, along with a verification flag. Running the training script loads that dataset, trains for a configurable number of epochs, and exports the finished model as an ONNX file. ONNX (Open Neural Network Exchange) is a portable format that lets you run the model using ONNX Runtime, a fast inference engine, without needing PyTorch installed in production. The training script accepts flags to set the number of epochs and the minimum and maximum character length. A separate test script lets you verify the trained model by feeding it a single image or pointing it at an entire dataset directory, and it compares predictions against your labels. Requires Python 3.10 or higher.