A Review of Deep Learning Methods for Wafer Defect Detection with a Focus on YOLO-series Detectors

Abstract

The semiconductor industry faces increasing demands for precise and efficient wafer defect detection as device miniaturisation continues. Traditional manual inspection methods have proven inadequate for modern high-density wafer manufacturing, necessitating automated detection solutions that can operate reliably at scale. This review examines the evolution of wafer defect detection methodologies, tracing the progression from conventional image processing techniques through classical machine learning approaches to state-of-the-art deep learning frameworks, with particular emphasis on the YOLO (You Only Look Once) series of object detectors. The paper analyses the architectural trajectory from YOLOv1 through YOLOv10, highlighting key innovations including anchor-free detection, feature pyramid networks, and lightweight convolutional modules. Special attention is given to recent advances that integrate clustering–template matching with improved YOLO architectures, exemplified by the CTM-IYOLOv10 framework, which achieves 98.1% detection accuracy while reducing inference time by 23.2% and model size by 52.3% relative to the baseline YOLOv10 model. The review synthesises findings from the recent literature spanning 2015 to 2024, providing comprehensive performance comparisons across detection accuracy, computational efficiency, and real-time capability. Current challenges and promising future research directions are discussed, including self-supervised learning, few-shot detection, explainable artificial intelligence, and edge deployment optimisation.