Defect detection in automotive glass based on modified YOLOv5 with multi-scale feature fusion and dual lightweight strategy

Automotive glass is one of the key components in manufacturing engineering, and the inspection of defects is essentially an important item of quality evaluation. Deep learning has become a promising technology and is very suitable for defect detection. However, a unified method to detect all kinds of defects is very difficult due to the few publicly available image datasets of automotive glass, and it is a challenge to deploy high-precision models in resource-constrained edge devices. Focusing on these problems, defect detection in automotive glass based on modified YOLOv5 with pseudo-tagging and dual lightweight strategy is developed in this paper. First, aiming at the problem that the defect samples are extremely lacking, combined pseudo-labeling and traditional data expansion methods have been explored to effectively increase the number of samples for meeting the requirement of deep learning model training, thus improving the performance of the detection model. Second, double lightweight modules, MobileNetV3, and Ghost module are introduced into the backbone and the neck network of the YOLOv5 model, respectively, for reducing the complexity of the model. In addition, a multi-scale feature fusion (M-SFF) module, which riches the semantic and spatial information of feature maps, is added to the output of the backbone to further improve the detection accuracy of the model. The effectiveness of each innovation module was verified through ablation experiment and horizontal comparative experimentally. Experimental results show that the complexity and deployment difficulty of the improved YOLOv5 model are significantly reduced, the floating-point operations per second (FLOPs) values and weight sizes are significantly smaller than those of other models, and the detection performance is also improved based on the original YOLOv5, an average accuracy increase of 3.9 % on average, precision slightly decreased, and recall rates increased by 7.6 % . The method is appropriate for applications in the defect inspection of automotive glass.