Knowledge Distillation-Based Multi-Scale Feature Matching for Industrial Anomaly Detection in Intelligent Manufacturing

The demand for industrial intelligent manufacturing and e-commerce is growing, and high-quality industrial intelli-gent manufacturing products are the key factors for the success of e-commerce, and intelligent manufacturing is closely related to deep learning. Surface defect detection of industrial products is an important research direction of industrial intelligence, which is crucial to improve the level of industrial intelligent manufacturing and e-commerce. However, despite the superior performance of advanced methods on datasets, the inference speed (FPS) and anomaly segmentation maps of most models are still not ideal when observed after experiments. Therefore, in this paper, under the premise of not affecting model performance, we use L2 norm model pruning technique to accelerate the inference speed of the model. Specifically, we pruned 5% of the third convolutional layer of the state-of-the-art teacher-student feature pyramid matching model used for industrial anomaly detection. The model is trained and tested on datasets, achieving good experimental performance and inference speed (FPS). Instead of using AUROC, which is commonly used by most methods, this paper adopts the PRO metric to ensure complete and accurate evaluation. Due to the optimization operation performed on the model, our method achieves accurate and fast pixel-level anomaly detection. On the MVTAD anomaly detection dataset, our method provides highly competitive results in terms of accuracy and inference speed (FPS).