Anomaly Segmentation of Leather Defects using Modified U-Net Architectures

An automated real-time leather inspection system based on machine vision needs defect detection algorithms that are accurate and fast to aid the requirements of the leather industry. Precise indication of defective areas on the leather needs to be highly accurate for two reasons: to avoid defective regions in final graded leather and to eliminate wastage of defectless regions. In this work, we have proposed two modified U-Net architectures to reduce the number of learning parameters without affecting the segmentation accuracy. These models were compared with state-of-the-art segmentation architectures, namely U-Net (with VGG16/ResNet34 encoders) and YOLO11. The models were evaluated based on performance metrics, which found that the proposed models provide results comparable to the classical U-Net and YOLO11 segmentation architectures.