This project is an image-based hCaptcha solver using a convolutional neural network (CNN) in TensorFlow. It classifies images to identify specific poses, potentially helping solve CAPTCHA challenges by distinguishing between target and non-target poses.
-
convert.py: Prepares images for training by applying grayscale conversion and Gaussian blur. It processes images in both training and validation folders to enhance model accuracy. -
train.py: Sets up and trains the CNN model with data augmentation (rotation, shifts, shear, zoom, and flipping) to improve robustness. The model includes convolutional, max pooling, and dense layers, with a binary classification output layer. -
validate.py: Validates the model on a separate dataset. Loads images, applies Gaussian blurring, resizing, and normalization, and uses the model to predict each image's class.
- Data Augmentation: Enhances model generalization by randomly altering images during training.
- Custom Preprocessing: Applies Gaussian blur to reduce noise in images before validation.
- Binary Classification: Distinguishes between two classes (e.g., "sitting" and "jumping").
TODO: 1, TEST GAN + VAE SYNTHETIC IMAGE GENERATOR
150 epochs 100% accuracy
