ADVANCED CRACK DETECTION IN BUILDING STRUCTURES USING PIX2PIX AND U-NET ARCHITECTURES

Traditional crack detection methods, including manual inspection and classical computer vision techniques, often suffer from inconsistencies due to variations in lighting, surface texture, and environmental noise, resulting in unreliable and inefficient evaluations for large-scale inspections. This study presents a deep learning-based crack detection system designed to enhance segmentation accuracy and structural coherence in building inspections. The approach integrates a Pix2Pix conditional GAN framework with a U-Net generator, trained on the CrackSeg9k dataset comprising 9,255 crack images collected under diverse conditions. While U-Net provides strong baseline segmentation through its encoder-decoder structure and skip connections, its outputs can suffer from discontinuities and misclassification. To address these limitations, Pix2Pix adversarial learning refines the segmentation by enforcing continuity and edge sharpness through a discriminator-guided loss. The training pipeline includes a tailored augmentation strategy and was optimized using TensorFlow with GPU acceleration. Through staged experimentation, the best performance was observed at 100 epochs with a batch size of 8, achieving a mean IoU of 74.9% and F1-score of 71.1%. The proposed method demonstrates the viability of GAN-based refinement for crack segmentation and provides a robust foundation for integration with robotic inspection systems.