Research On Digital Image Objective Quality Assessment

Vision is the most important way of human beings to access the world with the development of technology, we are easy to acquire, storage, compress, and display the multimedia information. Digital images, videos, are abruptly increasing in our daily life. In the steps of acquisition, processing, coding, storage, transmission and reconstruction of digital images, image quality is ofetn distorted. An image or picture with distorted quality cann't satisfy the applications. In the other words, we need pay attention to visual information fedelity. Thereby, how to evaluate the quality of a given image and how to design a simple and efficient algotithm to evaluate is an important problem.Image quality assessment (IQA) can be classified as three categories, full-reference(FR), reduced-reference(RR), and no-reference(NR), according to whether a perfect reference image be used. In this article, firstly, we study a local significance map of an image according to characteristics of human visual system(HVS). Then a FR IQA based on the map is proposed. Secondly, structural information is studied for RR IQA. Finally, a NR IQA model describes that how to use local kurtosis to pridict image quality. Our work can be concluded as follows:1) Firstly, a full reference model for image quality assessment is considered. According to human visual system, the different regions of an image have different significance when it is observed. In general image processing, we divide an image to two part:smooth and texture parts.. However, masking theory indicate that the structure is the most important element in an image. An area with much structural information is more noticable than others (in sub-section 2.3.6). Accordingly, distinguish the strucural area from texture part is necessary. If we can divide the three area in an image, a map of local significance can get easily. Then, different regions have different weights. Obviously, We can pridict image quality accurately with the map.2) The human eye can recognize objects in an image easily because of its characteristic of obtaining the structure from it. Iin transmission, compression and other types of image processing, the structure information distorted result in distortion of the whole image quality. In some case, the cost of acquiring reference image is considerable, but some key information can be obtain form ancillary channel. So we give research a model to evaluate image quality without full reference, only a few key information of the reference image. In our model, statistics information of image structure is the key information.3) Lastly, we are focus on no reference case. In general, the images are non-stationary signal, but we take them as a tationary process in many image processing. Here, the statistical theory can be applied to analyze various characteristics of the image. This paper analyzes the two-dimensional image signal. Under some supposes, the dataset of images can be regarded as a two-dimensional random field. Then local kurtosis is calculated in every blocks after some decorrelated steps and a separate process. The sum of the local kurtosis is the metric for an image.The studies show that image quality assessment algorithms should be combined with a specific distortion type or application. In IQA algorithms, characteristics of HVS are integrated into the models to improve their performances. A FR IQA model with image ROI is proposed and verified with good performance in this aritcle. In our RR IQA model, structural information of the image is useful to evaluate the distortion. When reference image cann't be access, pooling local kurtosis as the model in chapter 5 is a good choice.