Segmentação de fissuras em dormentes de concreto protendido usando processamento de imagens e deep learning

Abstract

Crack segmentation in railway sleepers is crucial for ensuring railway safety and reliability. This dissertation proposes a complete methodology, covering the entire pipeline: from image acquisition, through pre-processing, segmentation, and post-processing, to presenting the results in a web application with interactive maps and dashboards. The study explores different deep learning-based segmentation architectures for segmenting thin cracks in sleeper images, comparing U-Net, ENet, SegNet, and a proposed modified ENet, S-ENet. The models were trained using Binary Cross-Entropy (BCE) loss over 150 epochs with an early stopping criterion based on the F1 score. Experimental results show that U-Net achieves the highest performance across evaluation metrics: 0.704 (F1), 0.7 (R), 0.743 (P) e 0.8 (B-F1), while ENet delivers approximately 15× faster inference speed: 0.134 seconds versus 1.96 seconds of U-Net. The S-ENet improves efficiency by being 1.5× faster (0.093 seconds) than the original ENet, while maintaining competitive segmentation quality. Despite slight numerical differences in evaluation metrics, visual results remain comparable. These findings highlight the trade-off between accuracy and computational efficiency, emphasizing the potential of lightweight architectures for real-time sleeper crack segmentation.