Research on appearance defect detection system for SE-SSD electrolytic capacitors

Abstract In the defect detection task of electrolytic capacitors, there are challenges such as diverse target scales, complex background environments, and relatively small defect sizes. This article further improves the SSD object detection algorithm and proposes a new object detection algorithm SE-SSD. Firstly, replacing the original VGG-16 network with the lightweight MobileNetV3 network equipped with an SE (Squeeze and Excitation) module significantly enhance the model’s expressive power and improve detection speed while ensuring model lightweighting. Secondly, embedding a CBAM (Convolutional Block Attention Module) attention module in the network structure effectively integrates feature information of different scales and enhances the model’s detection ability for targets of various sizes. Then introduce the Focal Loss loss function, which can significantly reduce the weight of negative samples and enhance the ability to extract feature information. Finally, by building a defect electrolytic capacitor image acquisition device, the construction of the sample dataset was completed, achieving the detection and recognition of six types of defects in electrolytic capacitors: Deformation, Scratch, Broken_leg, Damaged_Side, Damaged_Top and Burns. The experimental results show that the improved SE-SSD object detection algorithm has increased the mAP value by 3.62% compared to the basic model, with accuracy and recall rates reaching 84.32% and 80.45%, respectively. It effectively suppresses the occurrence of false positives and false negatives, and demonstrates better object detection performance than other mainstream detection methods, providing reliable technical support for intelligent production on the production line.