Deep Learning for Weak Feature Defect Detection in MLCC Inspection System

Multilayer ceramic capacitors (MLCCs) are susceptible to production defects with weak feature strengths, varying sizes and uncertain edges, which challenges existing inspection models in meeting industrial requirements. To address these challenges, this article proposed WAC-YOLO, a novel detect method designed for real-time online detection of MLCC multifaceted appearance defects. The proposed network incorporates the multiple attention wavelet-based downsampling (MAWD) module in the backbone, which separates feature map across different frequencies and directions, and selectively enhances it using tailored attention mechanisms. Additionally, the channel convolution cross stage partial (CCCSP) module fuses global and local information through dual branches and different kernel size convolution, merging and weigh interchannel information to improve the detection of weak features and multiscale defects. Furthermore, the adoption of generalized focal loss v2 (GFLv2) leverages defect bounding box distribution statistics to enhance classification accuracy. Experimental results on the MLCC six-surface defect dataset demonstrate the effectiveness of WAC-YOLO, achieving mean average precision (mAP)50 of 91.0%, mAP95 of 54.0%, ${F}1$ score of 89.3%, and detection speed of 98.2 frame/s, meeting real-time application requirements.