Study Of Objective Image Quality Assessment Algorithms

Image quality assessment plays an important role in relevant fields of image processing. The problem of image quality evaluation is involved in many applications, such as image compression, communication, storage, enhancement, watermarking etc. A good quality assessment metric can be used to guide the construction and adjustment of image processing systems, or to optimize the processing algorithms by adjusting the parameters. The most reliable way of assessing the quality of an image is subjective evaluation, because human beings are the ultimate receivers in most applications. The mean opinion score (MOS), which is a subjective quality measurement obtained from a number of human observers, has been regarded for many years as the most reliable form of quality measurement. But it is inconvenient to most applications. So, an appropriate objective image quality assessment metric is needed designing for approximating the subjective perception. According to the availableness of the original image, the objective image quality metrics can be classified into three types, full-reference model (FR), no-reference model (NR) and reduced-reference model (RR).For the image quality assessment with FR model, a structural-similarity-based (SSIM) metric is introduced with some modifications. Then, a content-based image quality metric (CBM) is presented through making full use of three essential factors of distortion, the"amount of error", the"location of error", and the"structure of error". The experimental results illustrate that the proposed metric approximates the MOS more closely than the metrics of the SSIM.Furthermore, to develop an image quality assessment with RR model, one of the spare signal expansion methods, named contourlet transform is introduced. Then, inspired by a RR image quality metric based on wavelet-domain natural image Statistic Model, a new contourlet-based metric is proposed by analyzing the amount of contourlet coefficients. The experimental results illustrate that the proposed metric which need a few data approximates the MOS more closely than the metrics of the wavelet-based metric.