Autoencoder-based Local Descriptor for Target Recognition in Infrared Images

Target recognition systems consist of two stages: feature extraction and classification. To improve the performance of these systems, we require high-quality features. The features extracted from RGB images are generally rich in information, which is not the case for infrared images. The processing of this later is challenging, mainly due to their properties of low contrast, competitive background and lack of edges. Traditional hand-crafted feature extractors are limited by these challenges. In this work, an autoencoder-based local descriptor is explored for target recognition in infrared images. Firstly, we generate the local descriptors using a convolutional autoencoder. Next, a global representation is built from these local descriptors, based on the bag of visual words model. Finally, we fed the global descriptor to a classifier. The evaluation of this approach is carried out on two thermal datasets: FLIR thermal starter dataset and VAIS dataset.