Research On Image Anomaly Detection Algorithms Based On Deep Generative Networks With Object-Aware

Image anomaly detection aims to identify and localize the anomalies in images,which plays an important role in many fields such as industrial defect detection and medical disease screening.With the development of deep learning,breakthroughs have been made in artificial intelligence based image anomaly detection technology,profoundly having affected people’s production and lifestyle.Methods based on deep learning heavily rely on a large amount of data with high-quality annotations.However,compared with normal data that are easily to be collected,abnormal data often have the characteristics of unknown,rare and heterogeneous.In addition,the process of labeling data is time-consuming and experience-dependent.Therefore,semi-supervised image anomaly detection algorithms,which leverage only normal data to train models for detecting anomalies,have been extensively studied and have become a hot topic in the field of computer vision.Reconstruction-based methods are one of the mainstream anomaly detection techniques.Such methods train reconstruction models by minimizing the difference between normal images and their reconstructed ones,and detect anomalies based on the large difference between untrained abnormal samples and their reconstruction images.However,due to the limited constraints on the latent space of the reconstruction models and the single means of detecting anomalies,there are still many problems that have not been resolved.To this end,this thesis takes retinal OCT images as the main research object,and conducts research on image anomaly detection algorithms based on deep generative networks with object-aware.The main research contents are summarized as follows:(1)In order to accurately detect abnormal images,this thesis proposes spatial-contextual variational autoencoder with object-aware correction for image anomaly detection.Firstly,to remove noisy image background,the retinal regions are extracted based on the gray distribution prior of retinal OCT images.Secondly,for the problem of limited constraints in latent space,a spatial variational autoencoder and a contextual variational autoencoder are constructed based on a multi-dimensional latent space and a one-dimensional latent space,modeling the local and global distribution of normal images,respectively.Finally,to reduce high reconstruction errors of normal regions caused by strong latent space constraints,an ablation-based methods is proposed to stably localize coarse abnormal regions in unsupervised way.(2)For the sake of accurately segmenting abnormal regions in images,this thesis proposes semantic augmentation variational autoencoder with spatial-channel object-aware for image anomaly segmentation.First of all,for the problem that the semantic distribution of normal images is difficult to describe,a semantic augmentation variational autoencoder is proposed.It obtains semantic augmentation latent variables by random sampling in the latent space,and then constructs distribution consistency constraints based on an image discriminator and a feature discriminator to achieve end-to-end self-supervised semantic data augmentation.Secondly,to segment abnormal regions from both local and global perspectives,this thesis proposes spatial-channel object-aware based image anomaly segmentation method.On the one hand,local abnormal changes are captured by comparing the difference between the multi-level fusion features extracted from the test images and their reconstruction images.On the other hand,based on the idea of ablation,the channel Gaussian smoothing weights of the features output by the last convolutional block of the encoder are calculated,and the abnormal regions that deviate from the global distribution of normal images are localized by capturing the feature maps that focus on the abnormal regions.Extensive experiments on multiple retinal OCT datasets are conducted to validate the effectiveness of the algorithms proposed in this thesis.The experimental results show that the proposed algorithms have significant improvement in detecting abnormal images and segmenting abnormal regions compared to existing advanced algorithms in the field,demonstrating the effectiveness of the proposed algorithms.