Deep learning for anomaly detection in computer vision

Abstract

Recently, deep learning (DL) inspired algorithms have performed remarkably well on many tasks such as machine translation, speech recognition and image classification etc. However, existing state-of-the-art algorithms struggle to learn the discriminative signals between normal and abnormal classes in an anomaly setting. This is due to the fact that these signals are subtle which often comes in form of little deviations in color, shape, structure etc. To tackle this problem, I propose a more efficient approach named AdeNet that requires lower computation and storage, making it more practical for use on edge devices. Anomaly detection use cases often involves the problem of class imbalance, a case where there are overwhelming samples (majority class) of one or more classes as compared to the others (minority class). This problem also contributes to the inability of DL algorithms to learn distinguishing signals. To address this, I propose an Encoder-based Generative Adversarial Network (eGAN) that leverages on pre-trained model to learn a separable distribution of these classes. Another associated problem to detecting anomalies is zero-shot learning (ZSL). This occurs as a result of the fact that it is practically infeasible to present to our model all possible instances of anomalies during training. Yet, we want models that are robust to new unseen out-of-distribution (OOD) samples during inference. Here, I employ the concept of contrastive learning (CL) to tackle this problem by using pretext task that learns to push embeddings of dissimilar classes far apart, and pull embeddings of similar classes. This seemingly simple concept forces the network to learn salient visual signals that are generalizable to identifying zero-shot instances.