YOLO-SAG: An improved wildlife object detection algorithm based on YOLOv8n

Abstract

Wildlife conservation is crucial for maintaining biodiversity, ensuring ecosystem balance and stability, and fostering sustainable development. Currently, the use of infrared camera traps to monitor and capture photos of wildlife is a vital methodology in protecting and researching wildlife, and automatic detection and identification of animals within captured photographs are paramount. However, factors such as the complexity of the field environment and the varying sizes of animal targets lead to low detection accuracy, while high-precision detection models are hindered by high computational complexity and sluggish training speeds. This paper proposes a wildlife target detection algorithm based on improved YOLOv8n - YOLO-SAG, which aims to balance accuracy and speed. Training stability is enhanced by introducing the Softplus activation function, which increases detection accuracy; incorporating the AIFI enhances intra-scale feature interaction, reducing missed and false detections. Integrating the GSConv and VoV-GSCSP module lightens neck convolutions, reducing computational redundancy and balancing the computational and parametric quantities brought by the AIFI. Experimental results on a self-made wildlife dataset indicate that the YOLO-SAG achieves 94.9%, 90.9%, 96.8%, and 79.9% in Precision, Recall, mAP@0.5, and mAP@0.5–0.95, respectively, which are 3.4%, 3.3%, 3.2%, and 4.9% higher than the original YOLOv8n. Inference and post-processing times reach 1.2 ms and 0.5 ms, a speedup of 25% and 54.5%, respectively, and the computation volume is only 7.2 GFLOPs, an 11.1% decrease.