Objective Automatic registration of optical and synthetic aperture radar (SAR) images is challenging owing to the significant geometric and radiometric differences between optical and SAR images. In this study, the phase congruency algorithm was used to calculate the phase because of its radiation invariance, construct feature direction information and feature intensity information, and establish a local feature descriptor, i.e., maximum phase index map (MPIM). Consequently, the corresponding points are obtained from the input image using the correlation measure of MPIM. Moreover, the projection transformation is used to achieve registration. Experimental results show that the proposed method shows strong adaptability to the radiation difference between optical and SAR images and exhibits a high registration accuracy.
Methods In this study, a multimodal image matching algorithm based on phase congruency was proposed. First, the definition of phase congruency was presented based on the frequency domain transformation. Then, the energy concept was introduced into the phase congruency calculation and the energy function was calculated using the odd and even filters of the Log-Gabor wavelet filter. The phase congruency transformation was obtained and found to be consistent with human vision. Further, the maximum moment Mψ was calculated based on the phase congruency transformation and the corner points of the image were extracted using the Mψ graph. Thereafter, a multimodal image matching algorithm based on MPIM of the Log-Gabor filter response was obtained.
Results and Discussions As shown in Fig.2, the image texture is preserved and noise is suppressed after the phase congruency transformation. By comparing the Harris corner points extracted from the original image and Mψ based on the phase congruency transformation, the corner distribution extracted from Mψ is more reasonable and conducive for subsequent image matching. Then, to test the performance of the MPIM operator, the information extraction ability of the operator is assessed by changing the template size and observing the number of matching points. The matching results of MPIM, histogram of oriented gradients, and multi-innovation algorithms are observed by changing the template size from 20 to 100. The results show that the proposed MPIM operator-based method exhibits a strong ability to extract the image texture. To verify the matching ability of the proposed algorithm for multimodal images, four sets of experiments were designed to perform optical, SAR, LiDAR, and electronic map registration to evaluate the algorithm in terms of neutrosophic c-means and root mean square error. The resolution and time span of the experimental images are large, and the radiation difference is obvious; hence, image matching is difficult. Table 2 shows the registration results. The proposed MPIM algorithm can adapt to radiation changes and stably extract the image texture for registration, thereby achieving a good matching effect.
Conclusions The registration of optical and SAR images is difficult owing to the large nonlinear radiation difference between optical and SAR images. To solve this problem, a registration method based on the MPIM descriptor was constructed based on the phase congruency to extract the image texture and resist noise. A descriptive operator with radiation invariance and texture characterization was obtained. Using the intermediate results of phase congruency calculation and combining them with the Harris operator, MPIM, normalized cross correlation measure, and random sample consensus algorithm, the problem of large local similarity difference between optical and SAR images owing to radiation differences is solved. In the experiment, the Harris operator was used to compare the feature points extracted from the Mψ graph and the original image. Moreover, it is verified that the combination of Mψ graph and Harris operator can extract more convincing feature points. Using four sets of multisource image experiments with large differences, it is verified that the proposed method can register optical and SAR images with robustness and obtain high matching accuracy. This method is not only applicable to optical and SAR image matching but also to the other images with large radiation differences. However, the proposed algorithm shows weak resistance to image scale and rotation and cannot complete matching in the case of large rotation and scale differences between images. Furthermore, the phase congruency requires the image texture to be very rich and mismatching can easily occur when the texture information is not rich. Future researches should focus on this problem to improve the ability of operators to adapt to more complex environments.
