AnomalySeg: Deep Learning-Based Fast Anomaly Segmentation Approach for Surface Defect Detection

• DOI: 10.3390/electronics13020284
• Publication Date: 2024-01-08T10:21:38Z
• Authors: Yongxian Song; Wenhao Xia; Yuanyuan Li; Hao Li; Minfeng Yuan; Qi Zhang
• Keywords: Upsampling; Sørensen–Dice coefficient; Feature (linguistics)

Product quality inspection is a crucial element of industrial manufacturing, yet flaws such as blemishes and stains frequently emerge after the product completed. Most research has utilized detection models avoided segmenting networks due to unequal distribution faulty information. To overcome this challenge, work presents rapid segmentation-based technique for surface defect detection. The proposed model based on modified U-Net, which introduces hybrid residual module (SAFM), combining an improved spatial attention mechanism feedforward neural network in place remaining downsampling layers, except first layer encoder, applies decoder structure. Dilated convolutions are also incorporated obtain more information about feature defects reduce gradient vanishing problem model. An loss function with Dice focal introduced alleviate small segmentation problem. Comparative experiments were conducted different methods, revealing that coefficient (DSC) evaluated by approach better than previous generic benchmarks KolektorSDD, KolektorSDD2, RSDD datasets, fewer parameters FLOPs. Additionally, displays higher precision recognizing characteristics minor flaws. This paper proposes practical effective anomaly identification, delivering considerable improvements over methods.