AMD-Net: lightweight component defect detection network with attention-guided adaptive compressor and multi-kernel grouped convolution

Abstract In modern industrial fields, the quality of the components is crucial for system performance. The current defect detection methods are computationally demanding and require a large number of parameters. A novel lightweight defect detection method is proposed, named AMD-Net, which consists of three innovative modules: attention-guided adaptive compressor (AGAC), multi-kernel grouped convolution (MGC) and dynamic shared convolution head (DSCH). AGAC utilizes adaptive weight downsampling and spatial attention mechanisms to effectively preserve key features while reducing parameter counts, MGC merges multiscale kernels with grouped convolutions to enhance the detection of multiscale defects, DSCH incorporates a dynamic scale adjustment mechanism and shared convolution layers, reducing parameters and boosting computational efficiency. The model size was reduced to 65% of its original, only 4.1 MB. To improve accuracy, knowledge distillation was employed for the lightweight structure, and the mean average precision (mAP) improved by 1.9% over the original model, reaching 91 %.