Advancing semiconductor defect detection with integrated deep learning and color scale preprocessing

The increasing complexity and miniaturization of integrated circuits (ICs) have amplified the impact of minor manufacturing defects, which can significantly degrade chip performance and system reliability. Accurate classification of these subtle defects is crucial but challenging due to the intricate features in chip imagery. In this paper, we introduce VSRNet, an enhanced network architecture that leverages the strengths of VGGNet, integrated with Squeeze-and-Excitation (SE) attention mechanisms and residual connections (Res). To further boost classification accuracy, we develop a novel image preprocessing technique that amplifies subtle chromatic aberration, making minor defects more detectable. Experimental evaluations demonstrate that VSRNet consistently outperforms traditional approaches across multiple metrics, with notable improvements observed when incorporating the proposed preprocessing method. This work offers a robust solution for fine-grained defect classification in complex semiconductor manufacturing environments.