摘要
目标检测在计算机视觉中具有广泛的应用,而YOLOv5是目标检测领域中的经典模型。然而,YOLOv5的参数量较大,不适合应用于自动驾驶等领域,因此,基于YOLOv5改进了一个轻量级的目标检测模型。首先,使用ShuffleNetv2替换了原有的CSPDarknet53主干网络,从而降低了网络计算量。其次,使用轻量级增加精度的架构Stem Block。再次,在特征提取网络的末端加入了Coordinate Attention,使其更好地聚焦图像中重要信息。最后,使用二元交叉熵损失函数,突出类别互斥的差异。实验结果表明,相比于YOLOv5方法,改进的模型mAP只降低0.08,fps达到了91。因此改进的模型在大幅度降低参数和计算量的同时,取得了理想的检测精度。
Object detection is widely used in computer vision,and YOLOv5 is a classical model in the field of object detec-tion.However,YOLOv5 has a large number of parameters and is not suitable for applications such as autonomous driving and edge wear.Therefore,this paper improves a lightweight target detection model based on YOLOv5.First,ShuffleNetv2 is used to replace the original CSPDarknet53 backbone network,thereby reducing the amount of network computation.Sec-ond,use the lightweight,precision-enhancing architecture Stem Block.Again,Coordinate Attention is added to the end of the feature extraction network to better focus on important information in the image.Finally,the binary cross entropy loss function is used to highlight the difference of category mutual exclusion.The experimental results show that compared with the YOLOv5 method,the improved model mAP is only reduced by 0.08,and fps reaches 91.Therefore,the improved model in this paper has achieved the ideal detection accuracy while greatly reducing the parameters and computation.
作者
厉振坤
付芸
LI Zhenkun;FU Yun(School of Opto-Electronic Engineering,Changchun University of Science and Technology,Changchun 130022)
出处
《长春理工大学学报（自然科学版）》
2023年第6期70-76,共7页
Journal of Changchun University of Science and Technology(Natural Science Edition)
基金
吉林省重点科技计划项目(20170204015GX,20180201049YY)。
