CAPTCHA Recognition 🔐🤖

A Deep Learning Approach using CRNN + CTC Loss

---

📌 Overview

This project implements a deep learning pipeline to recognize text-based CAPTCHAs using a Convolutional Recurrent Neural Network (CRNN) combined with Connectionist Temporal Classification (CTC) loss.

• Character set: Uppercase A–Z, Lowercase a–z, Digits 0–9 (62 symbols).
• Input size: Images are resized to 200×50, converted to grayscale, enhanced with CLAHE, and normalized.
• Output: Sequence of 5 predicted characters decoded with CTC.

---

📂 Repository Structure

.
├── model.py            # CRNN model (inference + training with CTC)
├── preprocess.py       # Dataset loader & preprocessing
├── train.py            # Training & evaluation script
├── test.py             # Inference (single image / folder)
├── export_safe.py      # Convert trained model to safe format
├── requirements.txt    # Dependencies
├── README.md           # Documentation
└── saved_model/        # Saved checkpoints (.keras)

---

📊 Dataset

• Source: Kaggle — Captcha Dataset (123k images)
• Filenames encode ground truth labels (e.g., aB3xQ.png → "aB3xQ").
• Fixed length: 5 characters (MAX_CHARS=5 in preprocess.py).

Update dataset path in preprocess.py:

DATA_DIR = "/path/to/dataset"

---

🏋️ Training

Run locally:

python train.py

Training details

• Preprocess dataset → Train/Test split
• Train CRNN with CTC loss (30 epochs)
• Save checkpoints → saved_model/crnn_ctc_best.keras
• Export final model → saved_model/crnn_ctc_final.keras

Colab / GPU training

• The model was trained using a GPU (Google Colab environment).
• A Jupyter Notebook (training_colab.ipynb) is included in this repository with the full pipeline:
  • Download the dataset from Kaggle
  • Configure GPU usage
  • Train the CRNN + CTC model
  • Export a safe version of the trained model (export_safe.py)
  • Run inference on sample images

---

📈 Evaluation

Metrics:

• Exact Match Accuracy (all 5 chars correct).
• Character Error Rate (CER).

Results (Kaggle dataset, 62-class setup):

• 🎯 Exact Match Accuracy: 0.8066
• ✂️ CER: 0.0769

---

🔄 Export Safe Model

During training, the CRNN model includes a Lambda layer (collapse_height).
This can sometimes cause issues when reloading the model in different TensorFlow/Keras versions.

To make the model portable, use export_safe.py to rebuild it with a registered custom layer (CollapseHeight) and save it again:

After training, convert the final model into a safe format (replaces Lambda with a registered custom layer):

python export_safe.py   --in_keras ./saved_model/crnn_ctc_final.keras   --out_keras ./saved_model/crnn_ctc_final_safe.keras

The exported *_safe.keras model is easier to load across different TensorFlow/Keras versions.

---

🔍 Inference

Single image:

python test.py --model saved_model/crnn_ctc_final_safe.keras --image path/to/captcha.jpg

Folder of images:

python test.py --model saved_model/crnn_ctc_final_safe.keras --folder path/to/images --limit 50

Extra flags:

• --no_auto_crop → disable smart cropping
• --force_invert → force color inversion
• --strong → stronger preprocessing for thin strokes
• --save_debug out/ → save debug images

---

🖼️ Examples

---

⚙️ Notes

• To use lowercase + digits only, update CHARS in preprocess.py and adjust NUM_CLASSES in model.py.
• Ensure CNN time steps ≥ label length (automatically checked in train.py).

---

👨‍💻 Author

Mohamed Saad
💼 [https://github.com/msaad-dot]

---