PIndNet: A pixel-wise industrial defect inspection network using multiple pyramid feature aggregation

Automatic industrial defect inspection is essential in production processes which remains challenging primarily due to the low contrast, ambiguous defect boundaries, and anisotropic background. To address the problem, we propose a novel pixel-wise industrial defect inspection network using multiple pyramid feature aggregation. First, a new multiple pyramid feature aggregation module is proposed to extract multi-scale features using global attention block which aggregate spatial and semantic features. Second, a spatial information extraction module is proposed to strengthen the spatial information interaction and integrates the location information. Furthermore, a bilateral feature aggregation module is proposed to model the feature autocorrelation. Experiments were conducted on three industrial defect datasets. Experimental results show that global attention block can aggregate spatial and semantic features efficiently. Effective spatial and channel features are essential for industrial defect detection. Bilateral feature aggregation module can model feature autocorrelation and eliminate redundant without additional computation.