CAPTCHA Recognition System

CS4243 Mini Project - CAPTCHA Recognition using CRNN + CTC Loss

Upload a CAPTCHA image to see the model's prediction with confidence score.

Model Architecture:

• ResNet-based CNN feature extraction (4 layers, 2 blocks each)
• Bidirectional LSTM (hidden_size=384, 2 layers)
• CTC Loss for alignment-free training

Performance:

• Sequence Accuracy: 55.6%
• Character Accuracy: 85.82%
• Trained on 7,777 samples with heavy augmentation

Features:

• Confidence scoring: Shows prediction reliability
• Multiple predictions: Shows top 3 alternatives when confidence < 60%
• Smart warnings: Alerts when visual ambiguity exists (0/o, i/1/l confusion)
• Real-time inference: Results in <1 second

Training Details:

• 14 iterations of systematic experimentation
• Data augmentation: rotation, shear, black lines, noise
• Regularization: dropout, weight decay, early stopping