RAM-SFE : A Novel Unsupervised Anomaly Detection Method Based on Random Anomaly Multi-Scale Feature Extraction

Anomaly detection is an important application in large-scale industrial manufacturing. However, most of the current surface defect detection methods based on image reconstruction use normal samples to train autoencoder (AE), which is easy to mistakenly identify normal pixels as abnormal pixels. Aiming at the shortcomings of surface defect detection methods based on image reconstruction, an anomaly detection method based on random anomaly multi-scale feature extraction (RAM-SFE) is proposed in this paper. Firstly, a data enhancement method (CNP) with cutting and noise paste is proposed, which enables the model to be trained with false abnormal data and enhances the robustness of the model. Secondly, a new image quality assessment scheme is proposed by combining gradient intensity Similarity deviation (GMSD), Structure Similarity Index Measure (SSIM) and Smooth L1, which solves the problem that the key information and texture details are not sufficiently concerned. At the same time, this method adopts multi-scale feature extraction technology, and introduces attention mechanism to make the model pay more attention to the sensitive region. Experiments show that compared with other anomaly detection methods, the proposed method achieves the best anomaly detection performance in many categories.