基于改进YOLOv3的合成孔径雷达图像中建筑物检测算法

传统合成孔径雷达（SAR）图像中建筑物检测算法主要是在特定场景下通过人工提取特征进行特定类别的建筑物检测，存在平均检测精度不高、检测效率低的问题，为此提出一种基于改进YOLOv3的SAR图像中建筑物检测算法，通过深度学习实现建筑物的自动检测。制作SAR图像中建筑物数据集，针对建筑物的尺寸特点，通过改进的K均值聚类算法重新设置先验框大小；在结构上借鉴深度神经网络的聚合残差转换思想，将YOLOv3骨架网络中用于构建特征层的单路卷积残差模块改进为多路卷积残差模块，提高通道信息利用率的同时降低计算量；加入浅层特征融合模块，增加特征图中建筑物的形状特征所占比重，在特征融合层之前，使用转置卷积进行上采样，增加细节特征；使用改进YOLOv3算法进行建筑物检测模型的训练，并在测试集上进行测试。实验结果表明，相比原始YOLOv3算法，改进YOLOv3算法在SAR图像中建筑物数据集上平均检测精度提高了9.2%，召回率提高了6.3%，同时保持了较快的检测速度。

A SAR Image Building Detection Algorithm Based on Improved YOLOv3

Since the traditional synthetic aperture radar （SAR）image building detection algorithm is mainly to detect the specific buildings by manually extracting the features in specific scenarios, it always has low average detection accuracy and low detection efficiency. A SAR image building detection algorithm based on improved YOLOv3 is proposed to realize the automatic detection of buildings through deep learning. The SAR image building dataset is produced, and the sizes of priori anchors are re-set by the improved K-means clustering algorithm according to the size characteristics of the buildings. Then the structure of the aggregated residual transformations for deep neural networks is used to construct the feature layer of YOLOv3 skeleton network. The single-channel convolution residual module is improved into a multi-channel convolution residual module to increase the channel information utilization and reduce the computation load. And a shallow feature fusion module is added to increase the contour shape characteristics of buildings in the feature map. The feature layer is upsampled to add detail features before the feature fusion layer through the transposed convolution. The improved YOLOv3 algorithm is used to train the building detection model which is tested on the test dataset. Experimental results show that the improved YOLOv3 algorithm improves the average detection accuracy on SAR image building dataset by 9.2% and the recall rate by 6.3% compared with the original YOLOv3 algorithm, while maintaining a fast detection speed. Key