Article Preview
TopIntroduction
In SAR image registration, feature-based methods are often recommended because of their advantages of robustness and adaptability (Jianwei et al., 2015, Deyun et al. 2017). The Scale-Invariant Feature Transform (SIFT) (Clémence et al 2017) and SURF (Yanshan et al., 2018) registration algorithms based on keypoint feature are receiving close review in virtue of its invariance to scale, rotation, and partial invariance to illumination changes and affine projection.
As a local region description algorithm, SURF is superior to SIFT algorithm in the detection rate of keypoints, the accuracy of feature description and the calculation efficiency (Herbert et al. 2006, Jianping et al. 2017). However, SAR imaging characteristics and target attributes cause low information consistency and serious interference, which reduces the accuracy of SURF descriptors in SAR images (Yan et al. 2017). In order to improve the registration performance of SURF algorithm, the variants of SURF are proposed. In (Bo et al. 2015), taking into account the structural features of SAR images, the coarse matching algorithm based on SURF extracted the keypoints of SAR images using Harris corner method, which improved the matching accuracy. In (Xiangsheng et al. 2011), Gradual Speeded-Up Robust Features (G-SURF) algorithm was proposed to solve the problem of image gradient information loss caused by the calculation of the Gaussian second order derivative in SURF. In G-SURF, the corresponding gradient information was added to the computation of feature vectors in order to obtain better robustness.
In addition, the SURF descriptor also has a simplified framework, which sets the up-right of the keypoints to the dominant orientation and calculates the feature vectors on the basis of it (Wenming et al. 2018). This simplified feature extraction method omits the step of dominant orientation assignment, and achieves better performance in SAR image registration. The precondition of the simplified method is that the registration image is already aligned in the azimuth. However, in SAR image registration, it is difficult to ensure that all registration images have accurate directional information, such as the following: the data sharing is not complete, or SAR image registration is used to correct the error of inertial guidance. Thus, the application scope of the algorithm of the simplified method is greatly reduced.
Therefore, in order to ensure the invariance to rotation of the algorithm, the calculation of the dominant orientation of keypoints in SURF and its variants is essential. However, the dominant orientation assignment based on gradient difference in SAR images is sensitive to speckle noise and geometric deformation. The multiplicative noise in SAR images leads to stronger gradient magnitude in homogeneous areas with high reflectivity than that with low reflectivity, and the geometrical distortion can result in the inaccurate gradient description (Kazutoshi et al. 2018, Chengliang et al. 2018). Thus, the stability and consistency of the dominant orientation assignment is unreliable. Furthermore, the SURF feature description based on the dominant orientation of keypoint is also unsafe. Thus, the performance of the SURF framework is degraded in SAR image registration.
In (Lina et al. 2017), the PSIFT descriptor, abandons the calculation of dominant orientations, and calculates the gradient magnitudes and orientations and further sampled in the radial and angular directions with different scales. The PSIFT descriptor with the invariance to rotation achieved excellent performance for SAR image registration. Based on the advantages of SURF in stability and efficiency, the PSURF feature descriptor for SAR image registration is introduced in this work. The PSURF feature descriptor inherits the advantages of the SURF descriptor, and can further reduce the computational complexity and increase the robustness of the SAR image registration. The rest of the paper is organized as follows. The original SURF algorithm is introduced in section 2. The PSURF descriptor, including descriptor analysis and descriptor construction is described in section 3. In section 4, experiments on SAR image registration are given to validate the effectiveness of the PSURF descriptor. Section 5 comes the conclusion of the paper.