Supervised Hashing with Kernel Based on Feature Fusion for Remote Sensing Image Retrieval

Abstract

Hashing has emerged as an influential solution to improve high-resolution remote sensing images retrieval (HRRSIR) performance. Most of the existing hashing methods adopt single feature to retrieve images. However, it is difficult for single feature to express the highly complex geometrical structures and spatial patterns of high-resolution remote sensing images. To address these issues, we propose a supervised hashing with kernel based on feature fusion method, which is called FKSH. FKSH mainly includes feature extraction, feature fusion and hash learning. Firstly, GoogLeNet and VGG16 are selected to learn features. In order to keep more spatial information, the features are extracted with the original input size and keep the output form of the three-dimensional tensor. Then max-pooling is performed on the tensor to retain the salient features. Secondly, the features from GoogLeNet are copied to uniform the dimension with that from VGG16. Thus, the features from GoogLeNet and VGG16 can be fused by element-wise addition. Finally, the high-dimensional fused features are mapped to compact binary codes by hashing, and the compact binary codes are adopted to retrieve remote sensing images. Experiments on the two benchmarked datasets clearly shows our superiority.