Target Detection Algorithm Based on Improved YOLOv8 for Hynobius Amjiensis

Hynobius amjiensis is a critically endangered species endemic to eastern China with important scientific and ecological values, but its population reproduction is not optimistic. Currently, the mainstream research method is to assess population reproduction through human survey of oocyst number, which is difficult to observe individuals, and human intervention can easily damage the local ecology. Therefore, we used deep learning target detection to detect Hynobius amjiensis underwater. There are many difficulties in target detection in the actual underwater environment, such as poor image quality, many disturbances, small target of Hynobius amjiensis, easy to be occluded and abundant posture changes, which bring great difficulties to target detection. To make the target detection model of Hynobius amjiensis more robust, this paper proposes a target detection model of Hynobius amjiensis based on improved YOLOv8x, which introduces three innovative methods. First, the CBAM attention mechanism is fused with the C2f module to form the CBC2f fusion enhancement module, which enhances the ability to extract feature information and capture key features. Secondly, an enhanced BiFPN network is used to replace the PANet network for feature fusion to improve the information fusion capability between different layers. Finally, a shallow feature layer is added to improve the scale range of feature extraction. The results indicates that our proposed YOLO-XN model achieves greater performance compared to the original YOLOv8x model on the Hynobius amjiensis dataset. Specifically, the precision, recall, mAP, and mAP scores reach 94.6%, 90.1%, 94.6%, 70.1%, which are increased by 1.6%, 2.3%, 1.3%, and 2.0%, respectively. It demonstrates the superior performance of our improved YOLO-XN model in detecting Hynobius amjiensis and provide an effective auxiliary method for intelligent wildlife monitoring.