为满足仿生机器鱼目标检测的需要，在YOLOv5基础上提出了一种轻量级检测算法，降低算法复杂度并提高精度。首先对YOLOv5s模型进行改进，通过GhostConv和C3Ghost模块降低参数量和计算量。其次，引入CA和CoordConv模块增强特征提取和目标位置感知能力，采用soft NMS减少使用传统非极大抑制(Non maximum suppression，NMS)带来的漏检、误检，同时使用MPDIoU简化相似性比较，提升检测精度和召回率。最后，所提出方法在目标检测数据集上的试验结果表明，改进的YOLOv5网络体积更小、精度更高，证明了该算法的有效性和优越性。
Research on Bionic Robotic Fish Object Detection Algorithm based on Improved YOLOv5
To meet the needs of object detection for biomimetic robotic fish,a lightweight detection algorithm based on YOLOv5 was proposed to reduce algorithm complexity and improve accuracy.First,improvements were made to the YOLOv5s model by using GhostConv and C3 Ghost modules to reduce the number of parameters and computational load.Second,CA and CoordConv modules were introduced to enhance feature extraction and target position perception capabilities,and soft NMS was used to reduce missed and false detections caused by traditional Non-Maximum Suppression(NMS).Additionally,MPDIoU was used to simplify similarity comparison,improving detection accuracy and recall rate.Finally,experimental results on the object detection dataset showed that the improved YOLOv5 network is smaller in size and higher in accuracy,demonstrating the effectiveness and superiority of the proposed algorithm.
万方数据
维普
