DM-YOLO for MLCCs’ automatic defect detection

Abstract

Defect detection for multi-layer ceramic capacitors (MLCC) is crucial. MLCCs’ defects are characterized by multi-scale and long-tailed distribution. In order to more accurately locate and identify defects in MLCC, this work developed an automatic sampling and sorting device for MLCC, which is characterized by a high level of automation and high-definition sampling of microelectronic components. Then, proposing a Deep Learning based model DM-YOLO (Decouple-trained multiscale-boosted YOLO) for defect detection. In this method, Decoupled Training is exerted on the model’s Detection head for improving its performance from long-tailed effect. For multi-scale targets detection, this model uses the Efficient Implementation of Universal Inverted Bottleneck in Layer-aggregation Network (EULN) module designed in this work, which provides flexible receptive field as needed. Besides, Bi-RepGFPN is used to enhance the feature fusion effect among feature maps from different scales and repair the loss of image characteristic information caused by the network’s increasing depth. The experiments demonstrate that our model achieves an mAP0.5 of 0.942 and an mAP0.5:0.95 of 0.615 on the MLCCs’ dataset. And the recall rate of DM-YOLO on this dataset achieves 0.925, meeting the requirements for the task of MLCCs’ defect detection.