Impact of coherence and occurrence in the training of CNN for anomaly detection

Details

Supervisors

De Vleeschouwer, Christophe ; Jacques, Laurent

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en mathématiques appliquées

Abstract

The CNNs and other Deep Learning methods are more and more used to solve supervised machine learning tasks such as classification. Most of the researches about these methods concern the improvement of the model itself or the analyse of the behaviour of CNNs for a given problem. However, in this work, we consider the impact of the training data on the problem of anomaly detection using CNNs. We constructed for this purpose synthetic images that consist of a background similar for all images and a foreground, or anomaly, which appears only on certain images. We observed the impact of two main parameters on the classification accuracy of a CNN model. First, we looked at the influence of the coherence between background and foreground signals. The coherence is a kind of measure of the similarity between two sets of signals. Then, we looked at the importance of the frequency of occurrence of each class in the training set. We noticed that images with high coherence are faster to learn but have an unstable generalisation on unknown data, while images with low coherence, a priori easier to distinguish, have a slower learning but a good generalisation on unknown data. An in-depth analysis of the learning behaviour, through notably the performances of a trained network on images containing only foreground, highlighted that the different behaviours are due to a difference in the features learned by the CNN. On images with low coherence, the used CNN learns essentially the foreground, while on images with high coherence it learns the background as well. Surprisingly, it is not evident that the observed behaviour of the CNN is linked to the coherence between background and foreground. Other indicators such as the evolution of the coherence inside the network, or the spatial support of the foreground, or even the frequency of occurrence of discriminating elements in the training dataset, seem more appropriate. Nevertheless, this work illustrates that the content of the training dataset impacts significantly the behaviour of a CNN.