Akademik Veri Yönetim Sistemi
Communications Faculty of Sciences University of Ankara Series A2-A3: Physical Sciences and Engineering, cilt.65, sa.1, ss.69-78, 2023 (Hakemli Dergi)
Camouflage, which is used as an art of hiding by living things in nature, started to be used in the military field in the 19th century with the widespread use of long-range firearms. When factors such as different nations, environment and climate are considered, we come across camouflages in various colors and patterns. Over time, the camouflage patterns adopted and used by countries or unions have become national identity. This study is on the classification and segmentation of camouflaged soldiers of 5 countries with deep learning. While the similarity of the camouflaged area with the background makes segmentation difficult, it becomes difficult to classify each camouflage pattern due to the cut of the fabric and the different locations of the pattern pieces on each soldier. There are different studies in the literature that are referred to as camouflage or pattern classification. The mentioned studies are in the form of segmentation of camouflaged object or classification of camouflaged objects of different types. Since the segmented and classified objects in this study are camouflaged soldiers, what is expected from the deep learning algorithm is to classify the objects mainly according to the camouflage pattern, not their outlines. In the study, 861 camouflaged soldier images were collected for 5 countries (Türkiye-Azerbaijan, USA, Russia, China, France) and polygonal labeling was made. Türkiye and Azerbaijan are considered a class as they have similar camouflages. For the solution of the problem, military camouflage classification was discussed with the Mask R-CNN algorithm, which is widely used today for object detection, segmentation and classification, and the importance of deep learning algorithms has been proven with such a difficult problem. The training resulted in 0.005219 classification loss and 0.03985 masking loss. The classification and segmentation success rate of the study is 95%.