An efficient defect detection method for MLCC based on EHL-DETR deep learning model

Multilayer ceramic capacitor (MLCC) may present various shapes, sizes, and subtle defects during manufacturing, which can affect its performance and functionality. However, existing detection methods often fail to meet the precision and speed required for MLCC defect detection. To address these issues, this paper builds an automated optical detection system (AODS) and proposes a deep learning model called EHL-DETR for defect identification and localization in MLCC images. First, we propose an efficient gate block (EGB) which enhances spatial feature extraction to improve the model’s multi-scale recognition capabilities. Second, the HiLo attention mechanism is introduced to achieve weighted fusion of local and global feature information, enhancing the model’s ability to detect subtle defects. Furthermore, the lightweight multi-scale feature pyramid network (LMSFPN) is proposed to enhance the fusion of multi-scale features and detect the defect more accurately. Experimental results demonstrate that EHL-DETR performs excellently on the MLCC dataset, achieving mAP₅₀, mAP_50-95, precision, and recall of 95.0%, 59.4%, 93.5%, and 90.1%, respectively, and a frame rate of 98.8 FPS, meeting the real-time defect detection requirements for MLCC.