Deep Learning-Based Surface Defect Detection Technology for Microdevices

Surface defect detection on microdevices is challenged by extremely small defect sizes, complex background interference, and strict requirements on both detection accuracy and computational efficiency. The objective of this study is to develop a lightweight yet high-precision detection framework suitable for resource-constrained industrial deployment. To this end, this paper proposes LiteKANformer, a multi-module lightweight Transformer-based architecture. Based on LiteKANformer, a novel detection framework named LKF-YOLO is constructed by embedding global contextual modeling into the backbone and optimizing multi-scale feature fusion in the neck network. Experimental analysis is conducted on a PCB surface defect dataset and a semiconductor chip defect dataset. Compared with the C3TR module, LiteKANformer achieves comparable detection accuracy while reducing the parameter count by approximately 3.1% and improving the inference frame rate by 2.3%. Furthermore, the proposed LKF-YOLO framework outperforms other mainstream detection models on the PCB dataset in terms of accuracy, recall, and real-time performance. The main novelty of this work lies in the co-design of activation representation, normalization strategy, and computation primitives within a unified lightweight Transformer block, providing an effective solution that balances detection precision and deployment efficiency for microdevice surface defect detection.