Lightweight Improvement of Multi-Scale Object Detection Model Based on Dual Attention Mechanism for Industrial Inspection

Industrial inspection scenarios present distinct challenges compared to civilian applications, characterized by significant scale variations in foreground objects and high inter-class similarity, unlike the more predictable scales and distinct target types typically encountered. Furthermore, these scenarios demand high accuracy and real-time performance. To address these challenges, this paper proposes a lightweight improvement method for multi-scale object detection models using local and global attention mechanisms. Based on the widely used YOLOv5 model, the method enhances feature extraction by adjusting the network architecture and incorporating attention modules. Meanwhile, detection layers for shallow features of the model are added, and the number of anchors in each layer is increased to improve the detection capability for multi-scale target defects. In addition, ordinary convolutions are replaced with Ghost convolutions to reduce the number of parameters in the network, achieving lightweight improvement. Finally, the training process is completed through the optimization of the objective function to obtain the model. This study uses two datasets of new energy battery cell surface images from different processing scenarios and evaluates the model based on four metrics: miss rate (false negative rate), false positive rate, inference speed, and model size (number of parameters). The results show that the miss rates for Datasets A and B are 0.26% & 0.39% and 0.41% & 0.67% , inference speeds are 90 FPS and 86 FPS, and the model size is 7.12 M. These results demonstrate that the proposed method significantly improves multi-scale object detection accuracy while maintaining a compact model (7.12 M) and high inference speed (86-90 FPS), meeting the stringent requirements for high-accuracy real-time detection in industrial environments.