基于改进可变形卷积的FDC-YOLO v8水下生物目标检测方法研究  

Research on FDC-YOLO v8 Underwater Biological Object Detection Method Improved by Deformable Convolution

作　　者：袁红春[1] 李春桥 YUAN Hongchun;LI Chunqiao(School of Information,Shanghai Ocean University,Shanghai 201306,China)

机构地区：[1]上海海洋大学信息学院,上海201306

出　　处：《农业机械学报》2024年第11期140-146,共7页Transactions of the Chinese Society for Agricultural Machinery

基　　金：国家自然科学基金项目(41776142)。

摘　　要：水下生物目标检测是实现水下机器人自动化捕捞的关键性技术。针对水下生物目标检测任务中存在的目标重叠、遮挡以及目标尺度小而导致的误检、漏检等问题,提出了一种基于改进YOLO v8n的水下生物目标检测算法FDC-YOLO v8。首先,在主干网络中使用融合可变形卷积网络的FDC模块,以增强模型特征提取能力,提升其提取特征的丰富度。其次,引入融合分数阶傅里叶变换和空间注意力机制的FrSAConv模块,进一步分离多样目标特征,增强模型对多种特征的感知能力。最后,引入Wise-IoU损失函数作为模型边界框损失函数,以更好地解决目标不平衡以及尺度差异的问题。使用RUIE数据集进行实验,水下生物包括海胆、海星、海参、扇贝。实验结果表明,改进后的FDC-YOLO v8的平均精度均值达到85.3%,较基准模型提升2.6个百分点,推理速度达到769 f/s,在目标重叠、遮挡以及小尺度目标的水下生物目标检测中有更好的表现。Underwater biological target detection is a crucial technology for achieving automation in underwater robotic fishing.Aiming to address issues such as object overlap,occlusion,and false detections,missed detections caused by small object scales in underwater biological object detection tasks,an underwater biological object detection algorithm,FDC-YOLO v8 was proposed based on an improved YOLO v8n.Firstly,the FDC module was incorporated,which utilized deformable convolution networks in the backbone network to enhance the model’s feature extraction capability and enrich the diversity of extracted features.Secondly,the FrSAConv module,integrating fractional Fourier transform and spatial attention mechanism,was introduced to further separate diverse object features and enhance the model’s perceptual ability towards various features.Finally,the Wise-IoU loss function was introduced as the bounding box loss function to better address issues related to object imbalance and scale differences.The experiments were conducted by using the RUIE dataset,which included four types of underwater organisms:echinus,starfish,holothurian,and scallops.Experimental results demonstrated that the improved FDC-YOLO v8 achieved an mAP of 85.3%,a 2.6 percentage points improvement over the baseline model.The inference speed can reach 769 frames per second,showcasing better performance in underwater object detection of marine organisms with challenged such as object overlap,occlusion,and small-scale objects.

关 键 词：水下生物识别 目标检测 YOLO v8n Wise-IoU 可变形卷积网络 分数阶傅里叶变换 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]