A Pixel-Level Segmentation Convolutional Neural Network Based on Global and Local Feature Fusion for Surface Defect Detection

Surface defect detection (SDD) is a fundamental task in the smart industry to ensure product quality. Due to the complexity and diversity of the industrial scenes and the low contrast and tiny sizes of the defect, it is still difficult to accurately segment the defect. To overcome these issues, this research studied the pixel-level segmentation convolutional neural network based on multi-scale features, global mapping, feature pyramid and attention mechanisms (MMPA-Net). First, the low- and high-level features are extracted as the multiscale network to enrich the defect features information. Second, the global and local feature fusion (GLF) with the global mapping branch (GMB) module is developed to gradually refine the defect details to promote the detection of defects with different sizes and shapes. Third, deep supervision is applied to the global feature map and multiscale prediction maps to train MMPA-Net. MMPA-Net has been conducted on three public SDD datasets, and the results show that MMPA-Net has achieved state-of-the-art results on the intersection over union (IoU) by comparing with other deep learning (DL) methods (NEU-Seg: 86.62%, DAGM 2007: 87.94%, MT: 84.23%).

Automated summary

Surface defect detection is a crucial task in the smart industry. This research proposes a pixel-level segmentation convolutional neural network, MMPA-Net, based on multi-scale features, global mapping, feature pyramid, and attention mechanisms. MMPA-Net achieves state-of-the-art results on three public SDD datasets, outperforming other deep learning methods in terms of intersection over union (IoU).