LAACNet: Lightweight adaptive activation convolution network-based defect detection on polished metal surfaces

Abstract

After the metal workpiece has been polished, there may still be small defects on the surface, which will adversely affect the quality of the product and the service life and availability of the metal in severe cases. To solve this problem, the industrial sector has been seeking more lightweight and efficient solutions. Therefore, this thesis proposes a Lightweight Adaptive Activation Convolution Network (LAACNet). Firstly, this thesis proposes a lightweight convolution module that adopts features concatenation to realize intra-channel and inter-channel information transfer and fusion. Secondly, this thesis proposes an adaptive activation convolution module with the enhanced nonlinear expression of the module, which makes the deep neural network expression more powerful. This thesis proposes a spatial channel coordinate attention module to capture the long-range dependencies between image pixels better. Finally, this thesis introduces a loss function that can optimize performance in target classification and localization tasks. Experiments were conducted on the self-made datasets Metal Surface Defect-Detection (MSD-DET) and GC10-Detection (GC10-DET), achieving Mean Average Precision 50 (mAP50) of 86.3% and 66.8%, respectively. The detection performance of this model is superior to other methods. In the ablation experiment, this thesis verified the effectiveness of each module. This thesis validated the Northeastern University-Detection (NEU-DET) dataset, achieving a mAP50 of 76.0%. The results show that LAACNet exhibits excellent robustness and generalization performance in surface defect detection. In addition, the method significantly reduces the number of model parameters, providing an effective choice for lightweight and efficient solutions.