Understanding Convolutional Neural Networks for OCR and CAPTCHA Recognition

The author, an algorithm engineer at Qunar.com with a background in Java development, shares his experience developing OCR solutions and recent work on end‑to‑end image recognition.

Machine learning problems are categorized into regression (predicting continuous values) and classification (predicting discrete categories), which sets the stage for discussing neural networks.

Convolution is presented as a mathematical operation that extracts features from input images using a kernel that slides over the pixel matrix, producing feature maps.

Pooling (down‑sampling) methods such as max pooling reduce computational load and provide translation invariance by summarizing regions of feature maps.

Fully connected layers connect every neuron from the previous layer to the next, aggregating extracted features; techniques like sparse connections and weight sharing reduce the massive number of parameters.

The classic MNIST handwritten digit dataset and the LeNet‑5 architecture are introduced, detailing each layer (C1, S2, C3, S4, C5/C6) and how they process 32×32 images through convolutions, pooling, and fully connected stages.

Applying these concepts to CAPTCHA recognition, the article describes building a TensorFlow model (code shown as images) that trains on thousands of labeled CAPTCHA images, achieving about 87% accuracy on a test set, which can be boosted to over 99% with repeated request strategies.

The author concludes by noting more advanced techniques (e.g., bidirectional RNNs with CTC) for harder CAPTCHAs and encourages further study of loss functions, activation functions, and mathematical foundations.