基于YOLOv8的江豚性行为检测与识别模型

Yangtze Finless Porpoise Sexual Behavior Object Detection Model Based on YOLOv8

摘要: 对江豚性行为的研究有助于此濒危物种的保育工作。传统方式依靠人眼观察，存在效率低、成本高、易疲劳等问题。随着深度学习技术的发展，中国科学院水生生物研究所配备的视觉监控设备可借助目标检测与识别技术提高观测效率。本文首先创建YFPSB数据集，包含2901张图像，涵盖五种行为类别。针对水下画质模糊、江豚与背景混淆等问题，本文建立了YOLOv8n-DBTA模型，首先通过自动伽马校正和暗通道去雾对图像预处理，接着设计双主干特征融合框架，增加更多特征融合路线，保留更多底层细节和特征以提升检测能力，然后设计任务对齐检测头，将位置与类别信息交互，学习江豚行为与位置的关系，最后进行模型实验分析，结果表明：模型在mAP_0.50和mAP_0.50:0.95指标分别达到97.9%和79.1%，优于其他主流模型。

Abstract: Research on the sexual behavior of Yangtze finless porpoises contributes to the conservation of this endangered species. Traditional methods rely on human observation, which is inefficient, costly, and prone to observer fatigue. With advancements in deep learning technology, the visual monitoring system at The Finless Porpoise Pavilion at the Institute of Hydrobiology, Chinese Academy of Sciences, can leverage object detection and recognition techniques to improve observation efficiency. In this study, we first constructed the YFPSB dataset, consisting of 2901 images covering five behavioral categories. To address challenges such as underwater image blurriness and the confusion between finless porpoises and the background, we establish the YOLOv8n-DBTA model. First, we apply automatic gamma correction and dark channel dehazing for image preprocessing. Then, we design a dual-branch feature fusion framework that introduces additional feature fusion paths, preserving more low-level details and features to enhance detection capability. Additionally, we develop a task-aligned detection head to enable interaction between positional and categorical information, allowing the model to learn the relationship between porpoise behavior and location. Finally, we conduct model experiments and analysis. The results demonstrate that our model achieves 97.9% and 79.1% on mAP_0.50 and mAP_0.50:0.95, respectively, outperforming other mainstream models.