Anomaly Detection for Surface of Industrial Products Based on Patch Kernel Density Estimation

Visual anomaly detection is a crucial problem in computer vision, commonly employed in industrial product quality inspections. A new model called Patch Kernel Density Estimation (PaKDE) is proposed in this paper. The proposed model integrates a pre-trained deep convolutional neural network (CNN), an embedding vector generation module, a subsampling module, a kernel density estimation (KDE) classifier module, and an anomaly map generator module to realize anomaly detection for surface of industrial products. Firstly, a pre-trained deep CNN is used to extract image features. Then, embedding vectors of features are generated and subsampled. Finally, a KDE classifier is trained using representative features. Experiments on the industrial inspection benchmark dataset show that our method achieves state-of-the-art performance: 89.6% on image-level area under the receiver operating characteristic curve (AUROC) and 85.0% on pixel-level AUROC.