An industrial defect detection method based on mixed noise synthesis

Abstract

Deep learning-based methods have significantly reduced the cost of traditional manual quality inspection while enhancing accuracy and efficiency in industrial defect detection. As a result, these methods have become a prominent research focus in computer vision for intelligent manufacturing. They are increasingly applied in various production and operational contexts, including automated inspection, intelligent monitoring, and quality control. This paper presents a novel method called mixed noise synthesized defect detection, designed to identify multiple types of defects in industrial products. The proposed method employs a generative adversarial network architecture composed of a defect synthesizer, a defect discriminator, a feature extractor, and a multi-scale patch adaptor. By leveraging the feature extractor and multi-scale adaptor, the method effectively captures normal feature distributions and synthesizes realistic defect features through mixed noise synthesis, thereby significantly reducing reliance on labeled data. In addition, the defect discriminator uses a dual evaluation strategy that combines adversarial loss with Kullback–Leibler divergence to assess input features and quantify defect severity. Comprehensive experiments on benchmark anomaly detection datasets demonstrate that the method achieves high performance, with image-level and pixel-level area under the receiver operating characteristic curve scores of 99.8% and 99.4% for texture categories, and 96.7% and 98.3% for object categories, substantially outperforming state-of-the-art methods. The source code is publicly available at https://github.com/ah-ke/MNS-Defect.git.