CDNet: Efficient Real-Time Chip Defect Detection for Consumer Electronics

Ensuring the quality and reliability of semiconductor chips in consumer electronics is crucial due to their widespread use in devices ranging from smartphones to home appliances. Traditional defect detection methods, which often rely on manual feature extraction and conventional image processing, struggle to handle complex defect scenarios and do not satisfy real-time operational demands. To address these challenges, this paper introduces an advanced real-time detection model, the Chip Defects Network (CDNet), which leverages a deep learning framework adapted from YOLOv10. CDNet enhances detection accuracy and computational efficiency through a novel architecture that includes a Bi-directional Improved Multi-Branch Auxiliary Feature Pyramid Network (BIMAFPN) for superior feature fusion, a Layered Multi-Scale Feature Module (LMSFM) for efficient multi-scale feature processing, and an Efficient Lightweight Detection Head (ELDH) designed to reduce model complexity while maintaining high precision. We also construct a custom dataset of 1674 chip images to evaluate our model, representing four prevalent types of defects. Extensive testing shows that CDNet outperforms the baseline YOLOv10s model by reducing parameters by 22.2%, increasing mean average precision (mAP50) by 1.6%, and improving inference speed by 10.7%. Comparisons with other state-of-the-art algorithms demonstrate CDNet’s superior performance, confirming its practicality for real-time defect detection in the fast-paced consumer electronics sector. Our code and data are available at: https://github.com/NGI-vision/CDNet.