Weidong Yan

Feature extraction of hyperspectral images based on preserving neighborhood discriminant embedding

A novel manifold learning feature extraction approach—preserving neighborhood discriminant embedding (PNDE) of hyperspectral image is proposed in this paper. The local geometrical and discriminant structure of the data manifold can be accurately characterized by within-class neighboring graph and between-class neighboring graph. Unlike manifold learning, such as LLE, Isomap and LE, which cannot deal with new test samples and images larger than 70×70, the method here can process full scene hyperspectral images. Experiments results on hyperspectral datasets and real-word datasets show that the proposed method can efficiently reduce the dimensionality while maintaining high classification accuracy. In addition, only a small amount of training samples are needed.

Robust Registration of Remote Sensing Image Based on SURF and KCCA

The Speeded Up Robust Features (SURF) description’s success for computer vision applications makes it an attractive solution for image registration problem. For remote sensing images, SURF feature matching is often impacted by similar structures. To overcome the mentioned problem, we propose to use spatial relationship along with the SURF descriptor for remote sensing image registration. Firstly, a putative set of correspondences is obtained based on distances between SURF feature descriptors. Secondly, the spatial relationship of matched features is accomplished based on Kernel Canonical Correlation Analysis (KCCA), and then an influence function is established by the spatial relationship to figure out false matches. Introduction of spatial relationship to the SURF descriptors not only reduces the number of false matches but also help to keep the number of correct matches. Experimental results show an overall significant reduction of the mismatches while maintaining a high rate of correct matches.

Super-resolution from a single image based on local self-similarity

Super-resolution from a single image plays an important role in many areas. However, it is still a challenging work, especially in the high-resolution image’s quality and the algorithm’s efficiency. To obtain high-resolution images, a new single image super-resolution technique that extends existing learning-based super-resolution frameworks is presented in this paper. We don’t use any external example database or image pyramid to learn the missing details, and propose a single image SR method by learning local self-similarities from the original image itself. To synthesize the missing details, we design new filters which based on principles that model the super-resolution process, and use the new filters to establish the HR-LR patch pairs using the original image and its downsampled version. To obtain the SR image, we adopt a gradual magnification scheme to upscale the original image to the desired size step by step. In addition, to control the iterative error, we use the original image to guide the details added. Experimental results demonstrate that the proposed method is very flexible and give good empirical results.

Robust image registration using adaptive coherent point drift method

Abstract. Coherent point drift (CPD) method is a powerful registration tool under the framework of the Gaussian mixture model (GMM). However, the global spatial structure of point sets is considered only without other forms of additional attribute information. The equivalent simplification of mixing parameters and the manual setting of the weight parameter in GMM make the CPD method less robust to outlier and have less flexibility. An adaptive CPD method is proposed to automatically determine the mixing parameters by embedding the local attribute information of features into the construction of GMM. In addition, the weight parameter is treated as an unknown parameter and automatically determined in the expectation-maximization algorithm. In image registration applications, the block-divided salient image disk extraction method is designed to detect sparse salient image features and local self-similarity is used as attribute information to describe the local neighborhood structure of each feature. The experimental results on optical images and remote sensing images show that the proposed method can significantly improve the matching performance.

Feature matching based on unsupervised manifold alignment

Feature-based methods for image registration frequently encounter the correspondence problem. In this paper, we formulate feature-based image registration as a manifold alignment problem, and present a novel matching method for finding the correspondences among different images containing the same object. Different from the semi-supervised manifold alignment, our methods map the data sets to the underlying common manifold without using correspondence information. An iterative multiplicative updating algorithm is proposed to optimize the objective, and its convergence is guaranteed theoretically. The proposed approach has been tested for matching accuracy, and robustness to outliers. Its performance on synthetic and real images is compared with the state-of-the-art reference algorithms.

Spectral feature matching based on partial least squares

We investigate the spectral approaches to the problem of point pattern matching, and present a spectral feature descriptors based on partial least square (PLS). Given keypoints of two images, we define the position similarity matrices respectively, and extract the spectral features from the matrices by PLS, which indicate geometric distribution and inner relationships of the keypoints. Then the keypoints matching is done by bipartite graph matching. The experiments on both synthetic and real-world data corroborate the robustness and invariance of the algorithm.

Local discriminant non-negative matrix factorization feature extraction for hyperspectral image classification

Non-negative matrix factorization (NMF) ignores both the local geometric structure of and the discriminative information contained in a data set. A manifold geometry-based NMF dimension reduction method called local discriminant NMF (LDNMF) is proposed in this paper. LDNMF preserves not only the non-negativity but also the local geometric structure and discriminative information of the data. The local geometric and discriminant structure of the data manifold can be characterized by a within-class graph and a between-class graph. An efficient multiplicative updating procedure is produced, and its global convergence is guaranteed theoretically. Experimental results on two hyperspectral image data sets show that the proposed LDNMF is a powerful and promising tool for extracting hyperspectral image features.

Feature matching using modified projective nonnegative matrix factorization

We present a novel matching method to find the correspondences among different images containing the same object. In the proposed method, by considering each feature point-set as a matrix, two point-sets are projected onto a common subspace using modified projective nonnegative matrix factorization. The core idea of the proposed approach is to jointly factorize of the two feature matrices and the matching operate on embeddings of the two point-sets in the common subspace. Furthermore, it is robust to noise due to the merit of the subspace method. The proposed approach was tested for matching accuracy, and robustness to noise. Its performance on synthetic and real images was compared with state-of-the-art reference algorithms.

Description of Salient Features Combined with Local Self-Similarity for SAR Image Registration

Local feature descriptor plays an important role in image representation and is helpful to further image processing. This paper proposes a local feature descriptor based registration method for synthetic aperture radar (SAR) images. The proposed method starts with identifying evenly distributed features by applying the divided salient image disk (SID) extraction method. To describe the shape content of local neighborhood, local self-similarity (LSS) descriptor is built in the local normalized region with a suitable size for every detected feature. Finally, the correspondence is found by measuring the similarity between LSS descriptors. The registration experiments on SAR images demonstrate that the proposed method can be applied to SAR image registration.

Single image super resolution based on multiscale local similarity and neighbor embedding

Image quality and algorithm efficiency are the two core problems of super resolution (SR) from a single image. In this paper, we propose a novel single image SR method by using multiscale local similarity and neighbor embedding method. The proposed algorithm utilizes the self similarity redundancy in the original input image, and does not depend on external example images or the whole input image to search and match patches. Instead, we search and match patches in a localized region of the image in each level, which can improve the algorithm efficiency. The neighbor embedding method is used to generate more accurate patches for reconstruction. Finally, we use the original image and filters we design to control the iterate errors which caused by layered reconstruction, which can further improve the quality of SR results. Experimental results demonstrate that our method can ensure the quality of SR images and improve the algorithm efficiency.