Search

Improving denoising algorithms based on nonlocal self-similarity (NSS) to cope with increasing noise levels has become difficult. This is primarily because of difficulty in accurately grouping similar image patches solely on original spatial-domain of noisy images. To solve this problem, we propose to group similar patches on transform-domain learned from clean natural images. In this paper, we introduce a denoising algorithm comprising principal component dictionary (PCD)-based patch grouping and a low-rank approximation process. In the proposed algorithm, PCD learns from clean natural images and uses the knowledge gained to guide similar patches grouping results in noisy images. Patch grouping is directly implemented on PCD-based transform-domain. And, external knowledge and internal NSS prior are used jointly for image denoising. The results of experiments conducted indicate that the proposed denoising algorithm outperforms several state-of-the-art denoising algorithms, especially in heavy noise conditions. [ABSTRACT FROM AUTHOR]

• A robust deep low-rank feature learning and encoding model for cross-age face recognition is proposed. • A manifold-constrained low-rank approximation algorithm is proposed to learn clean low-rank feature matrix for representation. • A locality-constrained feature encoding framework is proposed to represent the cross-age face images of the same identity by similar codewords. • The proposed method outperforms other state-of-the-art methods in terms of recognition rate. Cross-age face recognition (CAFR) is a challenging task, due to significant intra-personal variations. Furthermore, the training and testing data may contain random noise components. To address these issues, this paper proposes a deep low-rank feature learning and encoding method. Firstly, our method employs manifold learning in the low-rank optimization, which preserves the global and local structure of the data samples, while learning the clean low-rank features. Secondly, we encode the low-rank features using our locality-constrained feature encoding method, which learns an age-insensitive codebook from training data, and enables the intra-class samples to share the same local bases in a codebook. In the testing stage, the gallery and probe features are encoded by the learned codebook, which represents the images of the same identity by similar codewords for recognition. Furthermore, the periocular region of human faces is investigated for CAFR. Extensive experiments on five datasets demonstrate the effectiveness of our method. [ABSTRACT FROM AUTHOR]

Abstract: We present a dictionary learning algorithm which is tailored to the block-based image prediction problem. More precisely, we learn two related sub-dictionaries and , the first one () for approximating known samples in a causal neighborhood of the block to be predicted and the other one () to approximate the block to be predicted. These two dictionaries are learned so that representation vectors computed by approximating the known samples using will lead to a good approximation of the block to be predicted when used together with . Because of its simplicity, this method can be used for on-the-fly learning of dictionaries. The proposed method has first been evaluated for intra prediction. It has then been applied in a complete image compression algorithm. Experimental results show gains up to 3dB in terms of prediction compared to the H.264/AVC intra modes and up to 2dB in terms of rate-distortion performance. [Copyright &y& Elsevier]

Abstract: A computational approach for obtaining a similarity measure between qualitative descriptions of shape, colour and size of objects within digital images is presented. According to the definition of the qualitative features, the similarity values determined are based on conceptual neighbourhood diagrams or interval distances. An approximate matching algorithm between object descriptions is defined and applied to tile mosaic assembling and results of previous approaches are improved. [Copyright &y& Elsevier]

Enhancement of remotely sensed images is a challenging problem, since the enhanced image has to have an improved contrast and edge information while preserving the original radiance values as much as possible. In this paper, a scale aware enhancement method based on rolling guidance is proposed for remotely sensed images. For each scale, a guidance image is defined and the approximation image is provided by an iterative joint filtering of the approximation and guidance images. Then the extracted details are amplified through an adaptive scheme and added to the final level approximation layer to provide the resulting enhanced image. A comparative study between the proposed methods with classical edge preserving filters and traditional methods have been carried out by using several criteria. The proposed methods have an average of 12% improvement for contrast gain (CG) metric and 81% improvement for enhancement measurement (EME) metric compared to the closest comparison method. • A remote sensing scale-aware image enhancement method is proposed. • A multiscale decomposition scheme based on rolling guidance is developed. • The images are enhanced by weighting the detail layers in the reconstruction step. • An adaptive weighting scheme is proposed to provide a better local enhancement. • The proposed method has better enhancement, while preserving spectral content. [ABSTRACT FROM AUTHOR]

Abstract: Let S be a set of n points in the plane. We derive algorithms for approximating S by a step function of size k < n, i.e., by an x-monotone rectilinear polyline with k < n horizontal segments. We use the vertical distance to measure the quality of the approximation, i.e., the maximum distance from a point in S to the horizontal segment directly above or below it. We consider two types of problems: min-ε, where the goal is to minimize the error for a given number of horizontal segments k and min-#, where the goal is to minimize the number of segments for a given allowed error ε. After O (n) preprocessing time, we solve instances of the latter in O (min{k log n, n}) time per instance. We can then solve the former problem in O (min{n 2, nk log n}) time. Both algorithms require O (n) space. Our second contribution is an approximation algorithm for the min-ε problem that computes a solution within a factor of 3 of the optimal error for k segments, or with at most the same error as the k-optimal but using 2k −1 segments. Furthermore, experiments on real data show even better results than what is guaranteed by the theoretical bounds. Both approximations run in O (n log n) time and O (n) space. [Copyright &y& Elsevier]

This paper proposes to apply coarse-grained parallel genetic algorithm (CGPGA) to solve polygonal approximation problem. Chromosomes are used to represent digital curves and genes correspond to points of curves. This method divides the whole population into several subpopulations, each of which performs evolutionary process independently. After every migration interval number of generations, these subpopulations exchange their information with each other. Inspired by the designing theory of ensemble learning in machine learning, this paper further improves the basic CGPGA through adopting different but effective genetic algorithms, respectively, in different subpopulations. Both the diversity among different subpopulations and the accuracy in each individual subpopulation are ensured. Experimental results, based on four benchmark curves and four real image curves extracted from the lake maps, show that the basic CGPGA outperforms the used genetic algorithm, and further the improved CGPGA (ICGPGA) is more effective than the basic CGPGA, in terms of the quality of best solutions, the average solutions, and the variance of best solutions. Especially for those larger approximation problems, the ICGPGA is more remarkably superior to some representative genetic algorithms. [ABSTRACT FROM AUTHOR]