基于多尺度池化和范数注意力机制的遥感图像检索

遥感图像内容丰富，一般的深度模型提取遥感图像特征时容易受复杂背景干扰，对关键特征的提取效果不佳，并且难以表达图像的空间信息，该文提出一种基于多尺度池化和范数注意力机制的深度卷积神经网络，在通道层面与空间层面自适应地给显著特征加权。首先，在多尺度池化通道注意力模块中，结合空间金字塔池化的思想，对每个通道上的特征图进行不同尺度的最大池化。接着，采用自适应均值池化将尺寸不同的特征图转换为统一尺寸，以便通过逐像素相加的方式来关注不同尺度的显著特征。然后，在范数空间注意力模块中，将各通道对应同一空间位置的像素构成向量，通过计算向量组的L1范数和L2范数，获得具有空间信息的特征图。最后，采用级联池化的方法优化高层特征，并将该高层特征用于遥感图像检索。在UC Merced, AID与NWPU-RESISC45 3个数据集上进行实验，结果表明该文所提注意力模型，关注了不同尺度的显著特征，结合了空间信息，提高了检索性能。

Remote Sensing Image Retrieval Based on Multi-scale Pooling and Norm Attention Mechanism

Remote sensing images have rich content, and then the features extracted by the general depth model are easily interfered by the complex background. The key features can not be extracted well, and it is difficult to express the spatial information of the image. A deep convolutional neural network based on multi-scale pooling and norm attention mechanism is proposed, which weights adaptively salient features at the channel level and the spatial level. First, in the multi-scale pooling channel attention module, the max pooling of different scales is performed on the feature map of each channel based on spatial pyramid pooling. Next, the feature maps of different sizes are transformed to a uniform size by adaptive average pooling. Thus the salient features of different scales can be paid attention by element-wise addition. Then, in the norm spatial attention module, the pixels corresponding to the same spatial position of each channel are formed into vectors, and the feature map with spatial information is obtained by calculating the L1 norm and L2 norm of the vector group. Finally, the cascaded pooling method is adopted to optimize the high-level features, and the high-level features are used for remote sensing image retrieval. Experiment are conducted on UC Merced data set, AID data set and NWPU-RESISC45 data set. The results show that the proposed attention model improves the retrieval performance by concerning the salient features of different scales and combining the spatial information.