| Image anomaly detection aims to identify samples that do not conform to expected pattern.Due to the rareness of anomalous samples,image anomaly detection has often been formulated as an unsupervised or semi-supervised problem,that is to say,model must learn how to distinguish anomaly through only normal samples,which brings big challenge for model design and training.Autoencoder is a classic model for image anomaly detection with a pair of encoder and decoder.The encoder learns the representations of images and the decoder reconstruct them to original images,and the training in normal images help autoencoder reconstruct only normal images well,so we can use reconstruction error to detect anomaly.During the experiment,we find the error is related to the entropy of image.A normal image with higher entropy may have higher error than an anomalous image with lower entropy.Meanwhile,because the input is nearly identical to the target,the model may just reconstruct image through compression and decompression,which helps model reconstruct some anomalous images well.Both of them result in a lower performance for anomaly detection.Based on above analysis,we propose two novel anomaly detection methods to solve existing problems,and the main work is listed below:(1)Propose an autoencoder model which estimates reconstruction error with multi view.The core contribution is designing a better and comprehensive method for error estimation.Specifically,on the basis of pixel-wise mean square error,the proposed method adds structure similarity and feature similarity to estimate the error from local view and global view.What’s more,several algorithms are proposed to measure the difficulty of image reconstruction,which is used to normalize the error.With the help of proposed method,the performance gains a lot without altering the structure of autoencoder.(2)Propose a two-stage autoencoder with pretrained model.The core contribution is to force the encoder to learn unique semantic representations from normal images with the help of the two-stage training strategy.Specifically,we use RotNet to train the encoder independently,and then we train the decoder for image reconstruction with the weight of encoder keeping frozen.Meanwhile,an extra image refinement module is proposed to mend the flaw of reconstruction,which consists of an encoder,a decoder and an information reduction module.The reduction module forces encoder to send minimal information for image refinement.Several experiments show the advantage of proposed method. |
