Aixin Zhang

A self-adaptive super-resolution method based on dictionary library

In recent years, with the theory of compressed sensing being proposed and applied widely, the sparse representation method has become one of the hotspots to handle the superresolution problem. Usually, this kind of algorithms use only one dictionary pair for all low-resolution patches, which makes the recovered results less satisfied due to its bad adaptability. To overcome such problem, in this paper, we propose a self-adaptive image super-resolution reconstruction algorithm in which different dictionary pairs are trained and used for recovery according to different types of low-resolution patches. All the dictionary pairs are stored in the dictionary library for reuse. We apply the proposed algorithm on human face images and generic images respectively. The results show that the algorithm is superior to the existing algorithms in reconstruction performance. Besides compared with that of generic images, the reconstruction performance of human face images is improved more greatly.

Perceptual Compressive Sensing Scheme Based on Human Vision System

Compressive Sensing (CS) theory has gained widespread attention due to its advantage of breaking through the limits of Nyquist sampling theorem. To make the CS more adaptive, some works based on the human vision system (HVS) have been conducted, but incurring some other problems at the same time, such as additional sensors, higher computation power and added experiments. To solve these problems a perceptual CS scheme based on the masking effect of human eyes towards image textures is proposed in this paper. It is consisted of three steps. First the image signal is represented sparsely through DWT transformation, and the masking matrix for different brightness changing areas is computed based on the DWT decomposition. Second the CS measurement with the masking matrix is performed on the image signal to get a lower-dimension sampling. And finally the image is reconstructed from its sparse sampling data. Through several experiments we can see that the proposed scheme performs better in both visual quality and PSNR assessment than common CS without any increasing of the computational complexity.

Design and Anti-collusion Performance Analysis of a Quantization Based Orthogonal Fingerprinting Scheme

Multimedia fingerprinting is a technique to trace illegal redistribution of digital representations of multimedia content and to combat multi-user collusion attack. Among existing collusion resistant fingerprinting schemes, spread spectrum (SS) orthogonal fingerprinting has the best collusion resistance in non-blind detection, but in blind detection it shows inadequate performance. In this paper, we propose a novel quantization based orthogonal fingerprinting (QOFP) scheme that possesses good performance in non-blind and blind scenarios. To generate one fingerprinted copy, we implement subtractive dither quantization to the host signal with a dither sequence. By letting the elements in all dither sequences follow independent and identically distributed uniform distribution over a quantization step size, we make fingerprints embedded in different copies mutually orthogonal. Compared with SS orthogonal fingerprinting, our proposed QOFP has almost identical collusion resistance in non-blind detection and much better collusion resistance in blind detection.

An efficient angle-based shape matching approach towards object recognition

The pixel-based contour map is one of the most common used shape representation methods for shape matching in object recognition field. However it is difficult to remain accurate and efficient at the same time when recognizing the objects with diversity of postures or different presence from different perspectives. To solve this problem, in this paper we propose an angle-based shape matching approach by introducing a new concept of angle-based features. Furthermore, the object recognition process adopting such angle-based shape matching approach is described in detail. With numerous experiments conducted on the Weizmann Horse dataset, we demonstrate that the proposed method is accurate, efficient and robust towards different poses and resolutions at the same time.