Lightweight Leather Surface Defect Inspection Model Design for Fast Classification and Segmentation

Automated inspection of leather surface defects is critical in evaluating product quality, yet manual inspection is still time-consuming and error-prone. Conventional automated methods, on the other hand, exhibit high computational complexities, are rigid in dealing with varied defects, and often require extensive manual parameter tuning. To counter these challenges, we propose a lightweight model integrated with symmetry for efficient defect classification and segmentation. The model consists of a streamlined semantic segmentation network that uses depthwise separable convolution and symmetric padding to preserve edge features while eliminating deconvolution layers, thus considerably reducing computational overhead. Moreover, a discrimination network automates defect detection without requiring manual thresholds, and a segmentation suggestion stage refines defect masks for practical cutting recommendations. Experimental results demonstrate a 96.75% detection accuracy and 89.41% mean pixel accuracy, achieving performance comparable to state-of-the-art models (e.g., KMDNet, U-Net) while reducing training time by 40% and model size by 60%. The symmetry-driven architecture enhances computational efficiency (0.05 s/img) and robustness across multiple defect types. Furthermore, the modular design enables independent updates for new defect types without requiring full retraining, addressing a major limitation of prior methods. These findings highlight the potential of symmetry-based architectures in industrial quality control, offering a scalable and efficient solution for automated defect detection.