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

ABSTRACT

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.

Additional information

Funding

This work was supported by the National Natural Science Foundation of China [Grant No. 41801288 and 61866028], and by the Natural Science Foundation of Jiangxi Province [Grant No. 20202BAB212011 and 20202BABL202030], and by the Key Research and Development Project of Jiangxi Province [Grant No. 20192BBE50073 and 20203BBGL73222].