A Review of Deep Learning-Based 15Visual Inspection Techniques for Chip Surface Defects

Abstract

With the continuous improvement of semiconductor chip integration, chip surface defects have a serious impact on chip performance and reliability. Traditional detection methods such as manual inspection and traditional machine vision have problems such as strong subjectivity, reliance on manual feature design, and poor adaptability, which are difficult to meet the high-precision detection needs of advanced chips. Deep learning has shown obvious advantages in automatic feature extraction, small target detection and complex background adaptation, and has gradually become the core technology of chip surface defect visual inspection. This paper systematically combs the types and imaging characteristics of chip surface defects, classifies and summarizes deep learning-based detection methods from four aspects: classification, object detection, segmentation and Transformer architecture, and analyzes the key technical difficulties such as multi-scale defects, complex texture interference and data imbalance in the detection process. On this basis, the current challenges in detection accuracy, cross-process generalization, data annotation and industrial deployment are discussed, and the future development trends such as multi-modal fusion, weakly supervised learning, model lightweight and intelligent quality control closed-loop are prospected. The review aims to provide theoretical reference and technical support for the research and engineering application of chip surface defect detection based on deep learning.