基于改进YOLOv3的快速车标检测方法

A fast vehicle logo detection method based on improved YOLOv3

文章提出一种基于改进YOLOv3算法的快速车标检测方法(network of vehicle logo detection,NVLD)。该方法使用以深度可分离卷积为基础的骨干网络提取车标特征;改进了YOLOv3的上采样方式和特征金字塔结构,使用反卷积的方式调整特征图大小以提取出更多的特征信息,并利用更多的浅层特征信息提高小目标的检测精度;在损失函数部分增加IOU损失以更准确地进行边界框回归。与最新的目标检测方法进行比较,实验结果表明该文方法的检测速度更快,是YOLOv3的2倍;准确率更高,比其他算法的平均精度均值(mean average precision,mAP)高出0.5%~2.8%。

A fast vehicle logo detection method based on improved YOLOv3 algorithm termed NVLD(network of vehicle logo detection)is proposed.This vehicle logo detection method uses a network based on depthwise separable convolution to extract features.The up-sampling and feature pyramid of YOLOv3 are improved by using deconvolution to adjust the size of the feature layer to extract more feature information,and more shallow feature information is used to improve the detection accuracy of small targets.The intersection over union(IOU)loss is added into the loss function to more accurately perform bounding box regression.Compared with the state-of-the-art detection methods,the experimental results show that the detection speed of the proposed object detection network is faster,which is 2 times that of YOLOv3.Its accuracy is higher,and the mean average precision(mAP)value is 0.5%2.8%higher than those of other algorithms.