改进金字塔和跳跃连接的YOLOv5目标检测网络

针对YOLOv5网络模型的金字塔结构不能有效捕捉目标的跨尺度特征信息和存在梯度消失现象以及易导致目标检测精确率不高的问题,利用FPT (feature pyramid transformer)的注意力机制有效提取目标的跨尺度特征信息,把原YOLOv5网络模型中的FPN (feature pyramid network)和PAN (path aggregation network)结构替换为FPT,在FPT结构的两端加入跳跃连接(skip connection)并引入新的Mish激活函数,从而提出一种改进金字塔和跳跃连接的YOLOv5目标检测网络模型YOLO FS.在PASCAL VOC和MS COCO数据集上的对比实验结果表明,基于YOLO FS网络的目标检测在平均检测准确率、召回率和F1值上均有明显提升.

YOLOv5 object detection network with improved pyramid and skip connection

The pyramid structure of the YOLOv5 network model can not effectively capture the cross-scale feature information of the object and the gradient disappears which causes the problem that the accuracy of object detection is not high. The attention mechanism of FPT(feature pyramid transformer) can be used to extract the cross-scale feature information of the object, and then the FPN(feature pyramid network) and the PAN(path aggregation network) in the original YOLOv5 network model are replaced by the FPT. Skip connection is added at both ends of the FPT structure and an improved Mish activation function is introduced. Thus, an improved object detection network model, namely YOLO FS, is proposed. The experimental results on PASCAL VOC and MS COCO datasets show that the average detection accuracy, recall and F1-score of object detection based on the proposed network are significantly improved.