针对传统U-Net对于目标小、分辨率低和背景复杂的遥感图像的飞机检测率低问题,提出一种轻量级多尺度注意力U-Net模型(LWMSAU-Net)。该模型由相互对应的编码子网络和解码子网络组成,编码子网络采用多尺度模块,在编码和对应的解码模块之间使用残差跳跃连接模块,将图像的浅层特征与深层特征融合,通过增加浅层特征的权重,更多地保留飞机图像的边缘和细微结构特征,最后的编码模块采用残差注意力连接模块,连接编码子网络和解码子网络,加强对小尺度飞机目标的检测。解码路径在每个模块反褶积将特征图的大小乘以2,使特征图的数量减半,并与对称编码路径的特征图相结合。与U-Net相比,LWMSAU-Net的层数减少1,在遥感飞机图像数据集上进行实验,结果表明该方法能够有效检测遥感图像飞机,准确率可达94.72%。

As for the low aircraft detection rate by the traditional U-Net due to small aircraft targets, low resolution and complex background, a lightweight multi-scale attention U-Net model (LWMSAU-Net) is proposed. The model consists of encoding and decoding subnetworks corresponding to each other. The encoding subnetwork adopts multi-scale modules, and the residual jump connection module is used between the encoding and the corresponding decoding module to fuse the shallow features and deep features of the image, the more of the edge and fine structural features of the aircraft image is preserved by increasing the weight of shallow features and preserving. The last encoding module adopts residual attention connection module to connect encoding subnetwork and decoding subnetwork to strengthen the detection of small scale aircraft targets. The decoding path consists of 4 modules, where each deconvolution multiplies the size of the feature graph by 2, halving the number of feature graphs, and then combines with the feature graph of the symmetric encoding path. Compared with U-Net, the number of layers of LWMSAU-Net is decreased by 1. Experiments on remote sensing aircraft image dataset shows that the proposed method can effectively detect aircraft targets in remote sensing images with an accuracy of 94.72%.