Weak texture remote sensing image matching based on hybrid domain features and adaptive description method

Weak texture remote sensing image (WTRSI) has characteristics such as low reflectivity, high similarity of neighbouring pixels and insignificant differences between regions. These factors cause difficulties in feature extraction and description, which lead to unsuccessful matching. Therefore, this paper proposes a novel hybrid-domain features and adaptive description (HFAD) approach to perform WTRSI matching. This approach mainly provides two contributions: (1) a new feature extractor that combines both the spatial domain scale space and the frequency domain scale space is established, where a weighted least square filter combined with a phase consistency filter is used to establish the frequency domain scale space; and (2) a new log-polar descriptor of adaptive neighbourhood (LDAN) is established, where the neighbourhood window size of each descriptor is calculated according to the log-normalised intensity value of feature points. This article prepares some remote sensing images under weak texture scenes which include deserts, dense forests, waters, ice and snow, and shadows. The data set contains 50 typical image pairs, on which the proposed HFAD was demonstrated and compared with state-of-the-art matching algorithms (RIFT, HOWP, KAZE, POS-SIFT and SIFT). The statistical results of the comparative experiment show that the HFAD can achieve the accuracy of matching within two pixels and confirm that the proposed algorithm is robust and effective. In order to solve the matching problem of WTRSIs, this paper proposes a new matching algorithm based on the hybrid domain as well as the adaptive neighbourhood descriptor. Experimental results confirm the effectiveness of the proposed HFAD in processing WTRSIs, and its matching results are better than those of state-of-the-art matching algorithms.image

Automated summary

This paper proposes a new matching algorithm for weak texture remote sensing images (WTRSI) based on hybrid domain and adaptive neighbourhood descriptors. Experimental results confirm the effectiveness and superiority of the algorithm in processing such images.