Experimental Comparison of Registration Methods for Multisensor Sar-Optical Data

Abstract

Synthetic aperture radar (SAR) and optical satellite image registration is a field that developed in the last decades and gave rise to a great number of approaches. The registration process is composed of several steps: feature definition, feature comparison and optimization of a geometric transformation between the images. Feature definition can be done using simple traditional filtering or more complex deep learning (DL) methods. In this paper, two traditional approaches and a DL approach are compared. One can then wonder if the complexity of DL is worth to address the registration task. The aim of this paper is to quantitatively compare approaches rooted in distinct methodological areas on two common datasets with different resolutions. The comparison suggests that, although more complex, the DL approach is more precise than traditional methods.

Registration of Multisensor Images through a Conditional Generative Adversarial Network and a Correlation-Type Similarity Measure

“…A preliminary version of this work was published by the authors in the conference paper [11], while a subset of the experimental results appeared in the conference paper [12]. The present article extends these works by providing a more in-depth methodological analysis and by expanding the experimental evaluation of the performances with the use of two additional datasets associated with different areas and/or sensors.…”

“…Both such methods combine deep architectures for image-to-image translation only with feature-based registration schemes. In [12], an experimental comparison of deep cGAN-based registration and traditional approaches is presented. Furthermore, the method proposed in [63] uses Siamese networks and contrast learning [64] to learn a representation space and then applies a cross-correlation measure to detect horizontal alignments in the images.…”

Registration of Large Optical and SAR Images with Non-Flat Terrain by Investigating Reliable Sparse Correspondences

“…However, we think that the high response locations obtained by these Harris-like operators can only be considered 'salient' in a very limited local neighborhood, about 10 × 10 pixels for the normally applied parameter setting. It does not guarantee the saliency of the local image template, which is usually 100 × 100 to 200 × 200 pixels [11][12][13][14][15][16][17][18][19][20][21][22][23][24][25][26][27][28][29][30]. To this end, a deep learning-based feature point detector is proposed in [29], which uses a convolutional network to assess the 'goodness' of the local image patches for template matching.…”

“…Current studies on optical-SAR registration, no matter the handcrafted methods [7][8][9][10][11][12][13][14][15][16][17][18][19] or the deep learning-based ones [20][21][22][23][24][25][26][27][28][29][30], mainly focus on dealing with the vast radiometric and geometric disparity problem, which makes it quite difficult to obtain sufficient reliable CPs that are sparsely distributed across the input image pairs. After the putative CPs are obtained, outlier removal and image warping are mostly conducted under the assumption that the geometric relationship between the input optical-SAR image pairs can be depicted by a linear equation, such as the affine or projective transformation.…”

“…The HOPES method extracts the primary edge structure using the Gabor filters and conducts an edge fusion algorithm to obtain shared features from optical and SAR images. In recent years, many deep learning-based methods have come out [20][21][22][23][24][25][26][27][28][29][30], where diverse kinds of Siamese or pseudo-Siamese convolutional neural network architectures are designed to learn shared features from optical and SAR images. In [27], the CNN feature extractor produces pixelwise deep features, which mimic the handcrafted method.…”