Feature Analysis of Injurious bird Images using Heat Map

Abstract

In this paper, we discuss injurious bird recognition system that we have developed. Currently, crows (scientific name: Corvus), which are wild injurious bird, cause various kinds of damage in japan, such as damage to crops, as well as producing feces, noise, and garbage. Crows cause the largest amount of damage to crops among pest birds, and the amount of damage is enormous at around 40,000 tons per year. Our research group has been developing a repellent system that uses drones to drive away injurious bird, and we are now investigating the possibility of applying this system to crows. However, to apply this system to crows, it must be able to recognize the target crow with high accuracy, without background or other wild birds affecting its results. In this study, we developed a CNN-based crow recognition system and investigated its accuracy. As a result, we were able to obtain a 100% precision ratio for all classes in terms of the conformity experiment. In the recognition experiment, recognition rate is more than 87% for all classes. For the results, we used Grad-CAM to examine which parts of the crow were characterized, and we were able to confirm certain rules regarding the parts of the crow that were captured by CNN learning and the causes of misrecognition.