A multi-level feature fusion method based on pooling and similarity for HRRS image retrieval

High-resolution remote sensing (HRRS) images contain complex visual contents and rich detailed information. This paper proposes a multi-level feature fusion method to improve the feature representation for HRRS image retrieval. Firstly, in order to obtain the multi-scale information of HRRS images, mid-level features and high-level features of VGG16 and GoogLeNet are extracted with different input sizes. Then a feature transformation method is proposed to adjust the size and the number of different feature maps, so that the distinct mid-level features and distinct high-level features can be fused separately using element-wise addition. There is a large amount of redundancy in the fusion features, thus small-region max-pooling method is adopted to aggregate the mid-level fusion feature, and global max-pooling method is used to aggregate the high-level fusion feature. Finally, an adaptive weight allocation method based on similarity is proposed to combine mid-level feature and high-level feature. Experimental results show that the multi-level feature fusion is an effective method to enhance the feature representation, thereby improving the retrieval performance.