No abstract

Image matching is the key step for image registration. Due to the existing nonlinear intensity differences between multisource images, their matching is still a challenging task. A fast matching approach based on dominant orientation of gradient (DOG) is proposed in this article, which is robust to nonlinear intensity variations. The DOG feature maps are constructed by extracting DOG feature of each pixel in the images in the first place. A template matching method is used to determine correspondences between images based on the feature representations. We define a similarity measurement, referred to as sum of cosine differences, which can be accelerated by fast Fourier transform. Subsequently, the subpixel accuracy can be achieved by fitting the similarity measurement using a quadratic polynomial modal. A new variable template matching (VTM) method has been developed to improve the matching performance. Experimental results confirm that the proposed matching approach is robust to nonlinear intensity differences and has time efficiency. The VTM method additionally improves the matching precision effectively. Index Terms-Dominant orientation of gradient (DOG), image matching, variable template matching (VTM).

Nonlocal means synthetic aperture radar (SAR) image despeckling approaches have attracted much research attention. However, high computational burden always limits its application in practice, especially using complex similarity measures. We present a fast patchwise nonlocal method using joint intensity and structure measures for SAR image despeckling. Nonlocal methods often define the similarity criterion only based on amplitude or intensity image. In order to preserve structure details, the structure information is also introduced into similarity measure by constructing gradient orientation feature map. The gradient orientation statistical test is performed to determine whether the patches contain the same structural components, and the similar patches are selected through the constant false alarm ratio strategy. Furthermore, we reorganize the patchwise nonlocal despeckling method and accelerate it using fast Fourier transform. Meanwhile, we utilize a Gaussian kernel to aggregate patchwise weights for each pixel in its patch area, so as to reduce the blur effect of classical patchwise nonlocal methods on details. The experiments have demonstrated that the proposed method is an efficient restoration method and has great structure and texture retention ability.