关键词: 目标检测, 入侵物种, YOLOv8n, 膨胀卷积, 轻量化, 损失函数

Abstract: The spread of invasive species poses a severe threat to ecosystems, weakening biodiversity and exerting widespread negative impacts on natural resources, agricultural production, and economic activities. To address the challenges of invasive species identification and insufficient detection accuracy, this paper proposes an improved YOLOv8n-based detection algorithm, IS-YOLO (Invasive Species-YOLO). First, the Bottleneck in the C2f module of the YOLOv8n backbone network is replaced with an improved FADC module, significantly enhancing feature extraction capability and the network's representation ability. Second, a shared convolution strategy is adopted in the detection head, replacing the decoupled head approach, using shared convolution and group normalization to achieve lightweight and precise improvements. Third, a new loss function, PIoUv2, is introduced to optimize bounding box regression, improving detection performance and model convergence speed. Finally, experimental results on the Species196-L public dataset demonstrate that the improved model achieves a mean average precision (mAP@0.5) of 59.7%, with a model size of 5.1 MB and a frame rate of 289 FPS. Compared to the YOLOv8n algorithm, the detection accuracy improved by 0.6 percentage points, the model size reduced by 1.9 MB, and FPS nearly doubled. The proposed algorithm combines high accuracy and efficiency, providing technical support for real-time monitoring and management of invasive species.

Key words: object detection, invasive species, YOLOv8n, dilated convolution, lightweight, loss function