A Review of Deep Learning-Based Methods for Industrial Defect Detection

Abstract

Industrial visual inspection has evolved from methods based on handcrafted features to data-driven approaches based on convolutional detectors, detection Transformers, and vision-language models. This review classifies representative methods according to task output and training-data availability and compares two-stage detectors, YOLO-style one-stage detectors, DETR-style end-to-end detectors, normal-only anomaly detection, few-normal-shot learning, and zero-shot approaches. All claims and quantitative results are drawn from verifiable primary publications. The review identifies the applicability limits of each method family with respect to small defects, class imbalance, domain shift, and edge deployment, and provides deployment-oriented recommendations for model selection and evaluation.