Performance Evaluation of Boosted 2-Stream TCRNet

Abstract

Target detection in infrared imagery is a particularly challenging problem due to the presence of terrain clutter. The TCRNet-2 CNN architecture was introduced to combat this issue and has been shown to perform better than conventional networks such as faster RCNN and YOLOv3. In this paper, we evaluate the performance of the Boosted 2-Stream TCRNet in detail (including robustness to range variations, performance under day and night conditions) and compare it with that of YOLOv5. A MWIR dataset released by DSIAC is used for training and testing the network. We also propose the MWIR target classifier that recognizes the 10 classes in the NVESD dataset and achieves an accuracy of 65.72% which is state-of-the-art to date.