改进yolov3的遥感图像建筑物检测方法

Building Detection in Remote Sensing Images Based on Improved yolov3

针对遥感图像中小型建筑物检测率低的问题,提出一种改进yolov3的小型建筑物检测算法。首先,利用k-means++聚类分析数据集上的先验框尺寸信息,筛选出最优的Anchor Box,使定位更加精准,降低网络损失。其次,在yolov3网络结构的基础上,将第11层浅层特征与网络深层特征融合,生成一个尺度为104×104的新特征图层,用于提取更多小型建筑目标特征。再次,加入Coordinate Attention机制,用于提高网络对图像中有用信息的敏感度。最后,加入CIOU边框回归损失,为边界框提供移动方向以及更准确的位置信息,加快模型收敛。将上述方法应用于文中数据集,结果表明,改进后的yolov3平均检测速度为23.39帧/s,mAP为93.9%,在牺牲部分检测速度的情况下,有效地提升了小型建筑物检测的精度。

In order to address the problem of low detection rate of small and medium buildings in remote sensing images,an improved yolov3 small building detection algorithm is proposed.First,the anchor box size on the k-means++cluster analysis data set was used to screen out the optimal Anchor Box,making the positioning more accurate and reducing the network loss.Second,the shallow features of the l1"layer and the deep features of the network were fused on the bases of yolov3 network structure to generate a new 104x104 feature layer used for extracting more fea-tures of small buildings.Third,the Coordinate Attention mechanism was added to improve the sensitivity of network to the useful information in images.Finally,CIOU frame regression loss was introduced to provide the information of movement direction and more accurate position for the bounding box,accelerating the convergence of the model.This method was applied to the data set herein.The result suggests that the average detection speed of improved yolov3 is 23.39 frames/s with a mAp of 93.9%and that the detection accuracy of small buildings is effectively improved by sacrificing partial detection speed.