YOLOv5-LED 水母识别分类算法

In response to issues such as low accuracy，poor real-time performance，and limited recognition types in existing jellyfish recognition algorithms，a novel algorithm named YOLOv5-LED for jellyfish recognition and classification is proposed. Firstly，G-Conv module，G-BottleNeck module，and G-C3 module are designed，forming the basis for a Ghost-based feature extraction module. Subsequently，a four-scale feature detection head structure is introduced，along with a feature fusion structure based on bidirectional cross-scale PANet and the incorporation of CBAM attention mechanism，resulting in the creation of a new jellyfish detection and recognition algorithm model YOLOv5-LED. Finally，improvements are made to the IOU loss function by introducing a distribution loss function based on KL divergence to replace the cross-entropy loss function，and the candidate box generation algorithm is enhanced. Moreover，a method based on Cluster NMS is introduced to replace the weighted NMS algorithm in YOLOv5. Experimental results show that with a threshold of 0.5∶0.95，the average detection precision of YOLOv5-LED improved by 2.7% compared to the base YOLOv5. The parameter count decreased by 13.6%，and the computational load reduced by 7.2%. These improvements not only enhance precision but also reduce parameters and computational complexity，achieving network lightweighting.