Perceptual Image Anomaly Detection

Abstract

We present a novel method for image anomaly detection, where algorithms that use samples drawn from some distribution of "normal" data, aim to detect out-of-distribution (abnormal) samples. Our approach includes a combination of encoder and generator for mapping an image distribution to a predefined latent distribution and vice versa. It leverages Generative Adversarial Networks to learn these data distributions and uses perceptual loss for the detection of image abnormality. To accomplish this goal, we introduce a new similarity metric, which expresses the perceived similarity between images and is robust to changes in image contrast. Secondly, we introduce a novel approach for the selection of weights of a multi-objective loss function (image reconstruction and distribution mapping) in the absence of a validation dataset for hyperparameter tuning. After training, our model measures the abnormality of the input image as the perceptual dissimilarity between it and the closest generated image of the modeled data distribution. The proposed approach is extensively evaluated on several publicly available image benchmarks and achieves state-of-the-art performance.

Anomaly Detection in Medical Imaging With Deep Perceptual Autoencoders

“…The reconstruction error, thus, indicates the abnormalities. The latest methods broadly extend this idea by employing different combinations of autoencoders and adversarial losses of GAN's (OCGAN [16], GANomaly [52], ALOCC [53], DAOL [22] PIAD [18]), variational or robust autoencoders [37], energy-based models (DSEBM [13]), probabilistic interpretation of the latent space [54], [55], bidirectional GANs [56], memory blocks [14], etc. The main difficulties of such approaches are: choosing an effective dissimilarity metric and searching for the right degree of compression (the size of the bottleneck).…”

“…In our paper, we evaluate and compare the strongest SOTA approaches ( [15], [18], and [19]) on the two aforementioned medical imaging tasks. We find these methods either to struggle detecting such types of abnormalities, or to require a lot of time and resources for training.…”

“…Recent related studies [29], [31], [32] showed the effectiveness of deep features as a perceptual metric between images (the perceptual loss), and as a score of anomaly [18]. Also, the use of the perceptual loss for training autoencoders has been very popular in a variety of tasks [18], [29], [32]- [36] except -inexplicably-in the task of image anomaly detection where it has been somewhat dismissed so far. Trained only on normal data, the autoencoders tend to produce a high reconstruction error between the input and the output when the input is an abnormal sample.…”

“…Lastly, we propose a new approach to the basic setup of anomaly detection model. Most approaches [13], [16]- [18], [20] prescribe not to use any anomaly examples during the model setup, dismissing the questions of optimization and of hyperparameter selection for such models. However, in reality, some types of abnormalities to detect are actually known (for example, the most frequent pathologies on the chest X-rays).…”

Anomaly Detection With Particle Filtering for Online Video Surveillance

“…The deep model effectively detected the abnormal events in surveillance video. In [14]- [16] denoising auto-encoders and GANs were used to adversely learn latent representations for one-class novelty detection. A deep convolution neural network while utilizing ImageNet for feature extraction was used with transfer-level learning for an unsupervised anomaly detection in medical images [17].…”