Comparison of Keypoint Detectors and Descriptors for Relative Radiometric Normalization of Bitemporal Remote Sensing Images

Abstract

This paper compares the performances of the most commonly used keypoint detectors and descriptors (SIFT, SURF, KAZE, AKAZE, ORB, and BRISK) in keypoint-based Relative Radiometric Normalization (RRN) of unregistered bitemporal multi-spectral images. The keypoints matched between subject and reference images represent possible unchanged regions and form a Radiometric Control Set (RCS). The initial RCS is further refined by removing the matched keypoints with a low crosscorrelation. The final RCS is used to approximate a linear mapping between the corresponding bands of the subject and reference images. This procedure is validated on five datasets of unregistered multi-spectral image pairs acquired by inter/intra sensors in terms of RRN accuracy, visual quality, quality and quantity of the samples in the RCS, and computational time. The experimental results show that keypoint-based RRN is robust against variations in spatial-resolution, illumination, and sensors. The blob detectors (SURF, SIFT, KAZE, and AKAZE) are more accurate on average than the corner detectors (ORB and BRISK) in RRN, with an expense of higher computational cost. The source code and samples of datasets used in this study are made available at https://github.com/ArminMoghimi/keypoint-based-RRN to support reproducible research in remote sensing.

Two-Fold and Symmetric Repeatability Rates for Comparing Keypoint燚etectors

“…6. Several keypoint evaluations have been proposed in the literature [5,28,29]. However, there is no consensus on a universally optimal detector for all possible image geometrical and photometric variations [23].…”

“…Keypoints can be defined as the significant image features used in various applications, including image matching, registration, remote sensing, computer vision, and robot navigation [1][2][3][4][5][6]. Over the last few decades, numerous keypoint detectors have been proposed, each with its own set of characteristics, computation methods, intended applications, geometrical transformation invariance, and immunity to image artifacts.…”

Enblending Mosaicked Remote Sensing Images With Spatiotemporal Fusion of Convolutional Neural Networks

“…In the experimental procedure, the newly proposed MOST method is compared with other mosaicking algorithms commonly used or recently proposed, i.e. mosaicking with the ENVI software (abbreviated as ENVI) [49], enblending (abbreviated as ENBLEND) [44], enfusing (abbreviated as ENFUSE) [50], quadratic programming (abbreviated as QP) [12], and keypoint based relative radiometric normalization (abbreviated as RRN) [21]. Three spatiotemporal fusion algorithms, Fit-FC [27], FSDAF [30], and EDCSTFN [42] which are among the best spatiotemporal fusion algorithms of different groups, are also compared using the proposed framework.…”

“…In addition to PIF, some studies have discussed the regression methods, such as orthogonal linear regression [17], iteratively weighted least square regression [18], random forest [19], Spectral-consistent regression [20], and so on. Recently, Moghimi et al [21] compared the performances of keypoint detectors and descriptors for PIF-based radiometric normalization.…”

Self-Similarity Features for Multimodal Remote Sensing Image Matching

“…Because of the different imaging mechanisms of various sensors, the characteristics of the same ground scene typically vary in different images. Therefore, multimodal image matching is a challenging task, as the radiometric differences are extremely significant [6], [7].…”