Research on Anomaly Detection Algorithm Based on Pixel-level Histogram Distribution Variations of Integrated Features

With the rapid development of industrial automation and intelligent manufacturing, the requirements for the quality and precision of industrial parts are increasingly stringent. As a crucial link in ensuring product quality, the importance of industrial part defect detection cannot be overstated. This paper proposes an innovative method for the detection of defects in industrial parts, which relies on deep learning technology and places special emphasis on the accuracy of feature extraction and the refinement of model training. The core of the study is the use of the deep neural network model WideResNet50 as the backbone network for image feature extraction, and the introduction of the L1 truncated loss function significantly enhances the model’s ability to recognize subtle feature differences in positive samples. Furthermore, this paper innovatively integrates the KAN (Kernelized Attention Network) layer into the model training process, replacing the traditional Multi-Layer Perceptron (MLP) network layer structure. Experimental results demonstrate that this method not only significantly accelerates the convergence speed of model training but also achieves a remarkable improvement of 1% in accuracy. This achievement not only brings new technological breakthroughs to the field of industrial part defect detection but also provides a new perspective and methodology for the application of deep learning in industrial visual inspection.