摘要
为了解决航拍小汽车检测精度和速度问题,论文在YOLO v3算法的基础上,设计了一种新的网络结构,并提出了一种改进YOLO v3算法。首先用混合深度卷积核代替单一卷积核,设计了一种新的特征提取网络。其次,对YOLO v3的FPN网络进行了改进,将深度特征映射和浅层特征映射融合,减小了卷积核的感受野。最后,在设计损失函数时,用GIOU代替IOU计算损失。结果表明,改进后的算法更准确、速度更快。
In order to solve the problem of detecting accuracy and speed of cars from aerial photographing,a new network structure is designed based on YOLO v3 algorithm in this paper,and an improved YOLO v3 algorithm is proposed for detecting cars from aerial photographing.Firstly,the single convolution kernel is replaced by the mixed depth-wise convolution kernel,and a new feature extraction network is designed.Secondly,the FPN network of YOLO v3 is improved,and the deep feature map and the shal⁃lower feature map are fused to reduce the receptive field of convolution kernel.Finally,when designing the loss function,GIOU is used for calculating the loss instead of IOU.The results show that the improved algorithm is more accurate and faster.
作者
王茂琦
李军
马佶辰
徐康民
WANG Maoqi;LI Jun;MA Jichen;XU Kangmin(School of Automation,Nanjing University of Science and Technology,Nanjing 210094)
出处
《计算机与数字工程》
2022年第4期775-779,795,共6页
Computer & Digital Engineering