EA-YOLO: An Endangered Animal Detection Algorithm Incorporating Attention Mechanisms
Kai Yan, Junping Qin, Lei Wang, Hao Yan Sun, Shulan Sun, Xiaole Shi · 2024
Endangered animal resources are crucial for biodiversity conservation and research. The detection and identification of endangered animals is a significant means to achieve this. However, the recognition accuracy of endangered animals is often affected by complex environmental conditions in the field. This paper proposes an improved algorithm named Endangered Animals-You Only Look Once (EA-YOLO) to address the issues of low target recognition accuracy and lack of feature extraction caused by complex environmental conditions and small target proportions in the picture. The algorithm is based on YOLOv8. This paper proposes two improvements to YOLOv8. Firstly, a multi-scale object detection head is introduced to fuse more image feature information, enhancing the network's ability to perceive and localize small targets. Secondly, a Global Attention Mechanism (GAM) module is added to learn richer feature information. The network can focus on important features and ignore irrelevant ones, leading to improved performance. The EA-YOLO model was evaluated on the LoTE-Animal dataset through extensive experiments. Results indicate that the improved model achieved an average accuracy of 95%, which is a 9% improvement over the original algorithm while maintaining competitive detection accuracy.