siddhartamukherjee/NEU-DET-Steel-Surface-Defect-Detection

Implements multiple segmentation architectures (Unet, FPN) with interchangeable encoders (ResNet, Inception, Xception) via the segmentation_models.pytorch library, enabling comparative analysis across encoder-decoder combinations. Evaluates models using domain-specific metrics (IoU, Dice Coefficient, Dice Positive/Negative) and separates validation data across all six defect classes to prevent data leakage. Built entirely in PyTorch with CUDA support, featuring modular utility scripts for mask extraction from XML annotations, data loading, training orchestration, and real-time inference workflows.

131 stars. No commits in the last 6 months.