A two-stage lightweight network for real-time IC surface defect detection

In the semiconductor industry, detecting minute surface defects on Integrated Circuits (ICs) is a significant challenge due to low accuracy and inefficiency. To address this issue, we present the design, implementation, and validation of a complete, end-to-end automated inspection system. At its core is a novel, two-stage lightweight detection workflow deployed on an NVIDIA Jetson NX embedded platform. The workflow first utilizes Tiny-UNet for precise Region of Interest (ROI) segmentation and then performs defect detection within these ROIs using our novel YOLO11n-C3GDW model. Experimental results demonstrate the outstanding performance of our integrated system: it achieves a state-of-the-art 78.2% mAP@0.5 with only 2.09M parameters and demonstrates practical real-time capability with a throughput of 10 pieces/min. This work delivers a validated, production-ready solution for real-time Automated Optical Inspection (AOI) systems, striking an exceptional balance between high accuracy and computational efficiency on edge devices.