Research On Image Scaling And Image Quality Assessment

With popularization of digital acquisition technology and signal processing theory, more and more images stored in digital form. Digital image processing technology originated in the 1920s. Image scaling and image quality assessment are two of the key technologies of digital image processing.The main purpose of image scaling was to change the image resolution of the source video signals to meet the demands of the terminal display devices. In the field of consumer electronics, image scaling was implemented on hardware. The hardware design considered more about real time, low cost and low power techniques. Therefore, the research of image scaling paid more attentions on the algorithms and its hardware implementations.Traditionally, image quality was evaluated by human. So the subjective image quality assessment was natural and consistent with the subjective feelings. But the subjective method was expensive and it was also too slow for the real-world applications. It also cannot be processed automatically by computer. The objective image quality assessment was more convenient, faster and easier to implement. So it was benifical to many other digital image processing technologies which needed to control or monitor the quality of the processed images. However, the result of objective methods was sometimes not so precise and it would differe from the subjective feelings. Therefore, how to evaluate the image quality more consistent with subjective sense was the main research of objective image quality assessment.This paper mainly focused on the key technology of image scaing and image quality assessment. It can be summarized as follow:(1) An edge adaptive four-point piecewise parabolic scaling algorithm and its architecture were presented, in which the pixels along or nearby edges were interpolated by edge direction oriented method. It can preserve the details in images and prevent the edge to be blurry. In hardware implementation, an efficient VLSI architecture based on Farrow structure was developed. Experiments show that the proposed algorithm can be scaled by arbitrary scaling ratios as well as preserving edges in image. Its hareware cost is lower than cubic interpolation algorithm, and it can process the image no larger than 2048×1536.(2) A scaling engine using reconfigurable interpolation filter based on farrow structure was proposed. The engine is designed to perform the arbitrary scaling ratios with the image resolution smaller than 2560×1920 pixels. It could scale up or down, respectively, in horizontal or vertical directions. It was composed of four functional units and five line buffers, which was more competitive than conventional ones. A strict fixed-point strategy is applied to minimize the quantization errors of hardware realization. Experimental results show that the engine provides a better image quality and a comparatively lower hardware cost than reference implementations. In addition, the reconfigurable featues make the engine better meet the various demands of users either in the performance or the cost of design.(3) A full reference image quality assessment based on Gabor filter was discussed. Gabor filter can simultaneously reach the theoretically limited lower bound in time and frequency on the product of its bandwidth and duration. Images were firstly decomposed with different wavelengths and different orientations by Gabor filters. After extracting the features from the decomposed images and pooling them together, the structural similarity index of global image is achieved. Finally, following the evaluation procedure of the Video Quality Experts Group, the objective scores from the structural similarity index was computed. Experimental results show that our proposed assessment can correlate well with human subjectivity in all kinds of distortions. The performance of our proposed assessment was comparable with state-of-the-art methods.(4) A novel reduced-reference image quality assessment algorithm using divisive normalization-based non-tensor product wavelet transforms was proposed. The algorithm applies the non-tensor product wavelet transforms to decompose image. Subsequently, Gaussian scale mixtures model was applied to furtherly process the coefficients of each subband. The normalized coefficients were highly fitted with the zero-mean Gaussian distribution. Finally, city-block distance was measured for image quality degrades. The data rate of our metric was much lower than the existing reduced-reference image quality assessments, and its performance was also better which was close to the full reference image quality assessment. So it became a new choice for objective quality assessment.