Bullet Hole Detection in a Military Domain Using Mask R-CNN and ResNet-50

Small arms shooting practices and competitions are routine activities in the military domain. The shooting group or bullet group analysis serves as a metric for the precision of a weapon, the shooter's accuracy, and consistency, and as a method for improving or refining one's shooting abilities. This analysis mechanism, however, is either manual or semi-automatic, employing image processing-based algorithms such as template matching, histogram equalization, white balancing, median and gaussian altering, peak detection, and image subtraction in an indoor setting, which is incapable of adapting to environmental conditions such as humidity, temperature, ambient light, wind speed, and rain, among others. Recent advancements in artificial intelligence or deep learning techniques explored ways to facilitate automation in various sectors. In this paper, we have used such deep learning approaches to automize the shooting system in real-time within a military domain and achieved success in resolving the traditional image processing drawbacks. Our proposed methodology has two phases. The first phase uses Mask R-CNN a conceptually simple, flexible, and general framework for object instance segmentation to extract the target region from the environment, and in the second phase, we fed the output segmented target of the first phase to ResNet-50 a convolutional neural network architecture to detect the bullet holes. Several experiments have been conducted on real-time datasets and the results show 0.87 of average precision using mask R-CNN to segment the target and ResNet-50 give 0.80 to detect bullet holes.