摘要
针对YOLOv3模型在进行目标检测时无法充分利用丰富的上下文信息导致目标漏检、误检,且模型参数量大的问题,提出一种轻量级目标检测算法。首先,该算法使用VGNetG网络作为骨干网络进行主干替换,这有利于减少模型计算参数量;其次,采用特征尺度感知模块FSSA进一步融合主干信息特征,使模型上下文信息对齐上采样的高级特征信息;最后,改进的交叉注意力模块ICC-Attention使用两个连接图来代替常见的单个密集连接图,并通过交叉路径法有效提取所有像素的上下文信息。改进后的算法比YOLOv3模型缩减一倍模型参数量,提高了算法参数效率与检测速率。算法在PASCAL VOC2007测试集上验证,实验结果表明,平均检测精度达到84.1%,参数量为5.37 M,检测速度为47帧/s,改进后的算法可有效改善YOLOv3模型中的漏检、误检与参数量大的问题,同时在检测速度方面可以满足实时性的要求。
Aiming at the problems of miss detection and false detection when the YOLOv3 algorithm performed detection,a lightweight object detection algorithm was proposed to deal with the problems which caused by the inadequate utilization of rich context information with a large number of model parameters.First,the VGNetG network was employed as the backbone network for replacement,which helped to reduce the number of model computational parameters;second,Feature Select Scale Aware(FSSA)module was adopted to further fuse the information features of the backbone to align the model contextual information to the high-level features of the upsampling;finally,the Improved Criss Cross Attention(ICC-Attention)utilized two connected graphs to efficiently extract the contextual information of all pixels by the cross path method instead of the common single dense connected graph.The improved algorithm exponentially reduced the number of model parameters and increased the efficiency of algorithm parameters and detection rate.The algorithm was validated on the PASCAL VOC2007 test set,and the experimental results showed that the average detection accuracy reached 84.1%,the number of parameters was 5.37 M,and the detection speed was 47 FPS.The improved algorithm effectively improved the problems of missed detection,false detection and a large number of parameters of the YOLOv3 algorithm.At the same time,the algorithm could meet the requirements of real-time in terms of detection speed.
作者
王燕妮
贾瑞英
WANG Yanni;JIA Ruiying(School of Information and Control Engineering,Xi’an University of Architecture and Technology,Xi’an 710055,China)
出处
《探测与控制学报》
CSCD
北大核心
2023年第5期98-105,共8页
Journal of Detection & Control
关键词
目标检测
轻量级
尺度感知
上下文信息
object detection
lightweight
scale awareness
contextual information
