Simple-FPN: An Image Anomaly Detection and Localization Network based on SimpleNet and Feature Pyramid

We propose a neural network that is simple to understand, easy to implement and deploy, called Simple-FPN. This network is mainly used to detect and locate anomaly in images. The neural network mainly includes the following four parts: (1) a feature extraction part, including a pre-trained feature extraction network and a corresponding feature pyramid structure, (2) a pre-adaptation part, which maps the features obtained by feature extraction to the feature distribution space formed by target image set, (3) an anomaly feature generator adding Gaussian noise to the extracted features, and (4) the anomaly discriminator part is used to distinguish anomaly and normal features. The anomaly feature generator needs to be included during training, but is not required during inference. The design of this neural network mainly relies on three assumptions: (1) mapping common features in the non-target environment to the target environment helps avoid deviations in data distribution, (2) generating anomaly features in the feature domain is more efficient than in the image space, (3) features at different scales in the image can be more effectively extracted through feature pyramid. Simple-FPN has better performance than some previous methods. In the data set MVTec AD, Simple-FPN’s AUROC can reach an average of 99.4%. In addition, Simple-FPN’s inference speed can reach 89 FPS when tested on a 4060 GPU (only calculating inference time). Simple-FPN has excellent performance among many other anomaly detection networks.