基于改进YOLO v8模型的水下小型生物目标检测算法研究

Target Detection Algorithm for Underwater Small Organisms Based on Improved YOLO v8 Model

摘要:

针对传统检测算法在水下小型生物目标检测中存在的定位精度低以及难以识别微小目标和重叠目标的问题,提出了一种改进的YOLO v8小目标检测算法(WLDA – YOLO v8)。设计了轻量级多尺度卷积模块LEM,替换原网络中C2f模块的标准卷积,降低了计算复杂度,提高了模型在复杂环境下的特征提取能力;融合了ATSS样本分配策略和动态检测头DyHead,实现动态自适应的训练样本选择;集成改进的WIoU v3s边界框回归损失函数,以提高模型对水下小目标的定位精度。在自建数据集上的综合测试中,WLDA – YOLO v8的mAP50 – 95比原始YOLO v8s提高5.5个百分点,同时参数量降低至7.9×10?,浮点运算量由1.98×101?降至1.59×101?,大幅降低了模型复杂度。研究结果为提高水下小型目标检测的实时性和准确性提供了有效途径。

Abstract:

Under water environment is complex and dynamic. When light propagates in water, it undergoes severe attenuation, scattering and color cast, leading to prevalent defects in underwater images, including low contrast, strong noise and blurred details, which restricted the application of underwater target detection technology in marine resource exploration, ecological monitoring and underwater engineering. To tackle the challenges of low localization accuracy and difficulties in detecting small and overlapping targets in traditional detection algorithms for underwater biological targets, WLDA – YOLO v8, an improved version of the YOLO v8 algorithm optimized for small-object detection was proposed. A lightweight multi-scale convolution module (LEM) was designed to replace the original C2f module in the network, reduce computational complexity and enhance feature extraction in complex environments. To enable dynamic and adaptive training sample selection, the ATSS sample assignment strategy and the dynamic detection head (DyHead) were integrated. Additionally, the incorporation of an improved WIoU v3s bounding box regression loss function improved the model's localization accuracy for small underwater objects. Compared with the original YOLO v8s, some comprehensive tests on the custom data set showed that mAP50 – 95 of WLDA – YOLO v8 was improved by 5.5 percentage points, while the number of parameters was reduced to 7.9×10? and GFLOPs was decreased from 1.98×101? to 1.59×101?, significantly reducing model complexity. The experimental results verified that the proposed algorithm significantly elevated the accuracy and efficiency of the model in underwater target detection tasks, as well as its convergence speed and training stability.