Evaluation of Active Learning for Anomaly Detection in Images

During the last decade, the scientific research community has made astonishing steps in the development and improvement of algorithms that exploit huge amount of data with the aim of making machines perform tasks such as classification, object detection and semantic segmentation. Usually, the strategy used to train this models is the supervised-learning technique that requires labeled datasets. In fact, the progresses made in the Machine Learning and Deep Learning field where enabled by a key factor: the presence of famous benchmark datasets already labeled. Those famous dataset usually contains a large amout of data. The issue is that those dataset does not allow to create models useful for real application. The manufactury industry trends is clearly in the direction of digitalization. In fact, many companies try to incorporate those powerful algorithms inside the production/manufacturing processes. The problem is that a strict bottleneck is present: the absence of taylored datasets. The activity of creating a dataset is resource consuming. Creating an annotated dataset present time and financial costs. A solution is presented by Active Learning Strategies in which the main objective is to reduce as much as possible the burdain of the human driven annotation activity by selecting the most representative and useful data to train a model. The literature, the research and the testing process of the Active Learning techniques was made exploiting the usual benchmark datasets, but in this work we present state of the art techniques of Active Learning applied to the Anomaly Detection field to see if those strategies are robust in more realistic and challenging context. Moreover, due to inconsistent results and conclusion from the researchers com- munity and due to the lack of presence of a stable methodology to quantify the improvement in the Active Learning field a stable framework is presented to enable results comparisons in a fair way.