Military camouflaged object detection with deep learning using dataset development and combination

Abstract

Camouflaged object detection (COD) is one of the emerging artificial intelligence technologies. COD identifies objects that require attention and time to detect with human eyes due to the similarity in texture or color to the surrounding environment. Despite the importance of camouflage and its detection in military, there is a lack of military camouflaged object detection research. Previous studies point out that the general COD has not been well studied due to the lack of camouflaged datasets, and the situation is worse in the military domain. This study aims at tackling the challenge in two directions. First, we carefully assemble the military camouflaged object (MCAM) dataset, including camouflaged soldiers and people as well as camouflaged military supplies for military COD. The experiment shows that MCAM can generate better performance results than the other benchmark datasets (CAMO, COD10K). Second, military (MCAM) and nonmilitary camouflage datasets (benchmark datasets) are combined and tested to overcome data scarcity. The experiment shows that the nonmilitary camouflage datasets are effective for military COD at a certain level, and a proper combination of military and nonmilitary camouflage datasets can improve the detection performance.