FASCNet: An Edge-Computational Defect Detection Model for Industrial Parts

Online inspection of industrial parts becomes increasingly important for factories to improve production quality, where small sizes and high computations increase difficulties in the defect detection process. In order to solve these issues, we propose a defect detection model to identify detailed defects with edge computations, named fast attention segmentation classification network (FASCNet). In the model, we design skip connection attention (SCA) with edge average attention (eAA), edge sum attention (eSA), and attention for segmentation (AS) to catch complex features of extremely tiny defects. Additionally, global mixed pooling (GMP) operation is explored to adaptively obtain severe mapping into low dimensional feature domains. Furthermore, a tensor freeze decomposition (TFD) is discovered to reduce model computation and complexity for edge devices. Finally, we achieve an average precision (AP) of 97.86% and giga floating point operations (GFLOPs) of 64.4495 on the real-world sprocket surface data set, which has 16.06% of GFLOPs of the current state-of-the-art method while only lowering the AP by 0.58%. On the public data set, we achieve an AP of 98.83% and GFLOPs of 92.0379 on the Severstal Steel data set. The experimental results indicate that our model performs more effectively than other state-of-the-art approaches in terms of both accuracy and computational cost simultaneously.