Detecting unseen targets during inference in infrared imaging

Performing reliable target recognition in infrared imagery is a challenging problem due to the variation of the signatures of the targets caused by changes in the environment, the viewpoint or the state of the targets. Due to their state-of-the-art performance on several computer vision problems, Convolutional Neural Networks (CNNs) are particularly appealing in this context. However, CNNs may provide wrong classification results with high confidence. Robustness to disturbed inputs can be mitigated through the implementation of specific training strategies to improve classification performances. But they would generally require retraining or fine-tuning the CNN to face new forms of disturbed inputs. Besides, such strategies do not necessarily tackle novelty detection without training an auxiliary classifier. In this paper we propose two solutions to give the ability of a trained CNN to deal with both adversarial examples and novelty detection during inference. The first approach is based on one-class support vector machines (SVM) and the second one relies on the Local Outlier Factor (LOF) algorithm for example detection. We benchmark our contributions on SENSIAC database for a pre-trained network and evaluate how they may help mitigate false classifications on outliers and adversarial inputs.