Attention-Like YOLO:嵌入类注意力机制的YOLO算法  

Attention-Like YOLO:YOLO Detection Algorithm that Embeds Attention-Like Mechanism

作　　者：胡朝海 李自胜[1,2] 王露明 HU Chaohai;LI Zisheng;WANG Luming(School of Manufacturing Science and Engineering,Southwest University of Science and Technology,Mianyang 621000;Key Laboratory of Engineering Testing Technology,Ministry of Education,Mianyang 621010)

机构地区：[1]西南科技大学制造科学与工程学院,绵阳621000 [2]制造过程测试技术教育部重点实验室,绵阳621010

出　　处：《计算机与数字工程》2023年第9期1973-1978,共6页Computer & Digital Engineering

摘　　要：YOLOv3算法满足了大多数任务的实时性和检测精度要求,但对于精度要求更高(大于80%)的任务,未能实现较好的检测效果。针对上述问题,论文提出了一种类注意力机制(Attention-Like)。该机制输入两个分辨率大小不同的特征图,首先利用Padding对小特征图进行上采样,采样后的特征图通过Sigmoid函数运算得到上采样权值,其次将上采样权值作用于大特征图以获得过渡特征图,利用卷积对过渡特征图进行下采样,然后通过Sigmoid函数运算得到下采样权值,最后将下采样权值作用于小特征图,通过该方法增强小特征图的几何信息。将Attention-Like嵌入YOLOv3的骨干网络DarkNet-53,实现了Attention-Like YOLO检测算法。实验表明,该算法的平均精确度均值最高达到了82.8%,有效提升了检测精度。The YOLOv3 algorithm meets the real-time and detection accuracy requirements of most tasks,but it is difficult to achieve better detection results for tasks with higher accuracy requirements(greater than 80%).In response to the above problems,this paper proposes a kind of Attention-Like.This mechanism inputs two feature maps with different resolutions.First,Padding is used to upsample the small feature map.The sampled feature map is calculated by the Sigmoid function to obtain the upsampling weight,and then the upsampling weight is applied to the large feature map.To obtain the transition feature map,convolution is used to down-sample the transition feature map,and then the down-sampling weight is obtained through the Sigmoid function operation,and finally the down-sampling weight is applied to the small feature map,and this method is used to enhance the geometric information of the small feature map.Attention-Like is embeded into DarkNet-53,the backbone network of YOLOv3,and the Attention-Like YOLO detection algorithm is realized.Experiments show that the mean average precision of the algorithm is up to 82.8%,which effectively improves the detection accuracy.

关 键 词：目标检测 YOLOv3算法 类注意力机制 AL-YOLO DarkNet-53 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]