A Denoising and Structured Segmentation Method for IC Images via Integration of YOLOv5 and Lightweight UNet

The chip industry is a critical foundation for national security and economic development, with integrated circuit (IC) reverse engineering serving as a key component in the semiconductor value chain. IC reverse engineering involves multiple steps, including layer-by-layer imaging using a Scanning Electron Microscope (SEM), device identification, gate-level netlist extraction, and functional inference. Among these, the segmentation of electrical components and metal lines from the background of IC images is a prerequisite for device recognition and related procedures. During the imaging process, significant noise often arises due to circuit damage, which may cause genuine circuit regions to be misidentified as background, thereby interfering with subsequent structural recognition and design processes. To effectively address this issue, this study proposes the use of the YOLOv5 detection model to identify and remove noise regions in circuit images. By treating noise as a detection target and performing network training and performance evaluation, experimental results show that the method achieves 87.26 % detection precision on the test set, demonstrating its effectiveness in noise identification. After denoising, a lightweight segmentation network, MobileNetV2-UNet, is further applied for structured segmentation of the entire IC image. Considering the possibility of residual noise, the proposed model integrates noise-robust mechanisms and specific optimizations for blurred edges. Experiments were conducted on a proprietary IC image dataset provided by an industry partner. The final results show effective segmentation of metal wiring regions, with an mean Intersection over Union (mIoU) of 88.62 % and a mean Pixel Accuracy (MPA) of 94.03 %, outperforming many mainstream methods. These results validate the proposed approach's potential and practical value in complex IC image segmentation tasks.