No-reference Image Quality Assessment Based On Feature Extraction In Contourlet Domain

With the development of nowadays, it is more often for images to acquisition, transmission and processing. Each of these processes needs to evaluate image quality, so image quality assessment(IQA) has been a hot issue in the field of image processing. IQA can be divided into subjective and objective quality assessment. The subjective quality assessment is most reliable as human beings are the ultimate recipient, but it is time consuming and cannot be embedded in an automation system. As a result, it is essential to develop objective quality assessment which can generally be divided into full-reference(FR), reduced-reference(RR) and no-reference(NR) IQA. FR and RR methods need all or part of the information of original images. However, it is difficult to obtain this information. So the study of no-reference image quality assessment(NR-IQA) becomes particularly important. This paper is mainly about the NR-IQA. Following are the main work:1. By studying the basic principles of dual-tree complex wavelet transform(DTCWT), we put forward a NR-IQA method using DTCWT. Firstly, we calculate the energy value of real and imaginary sub-bands after DTCWT, then using support vector machine(SVM) to establish the NR-IQA model. Experimental results show that the method has linearity against the subjective perception, and is superior to NR-IQA model based on discrete wavelet transform(DWT).2. By learning the contourlet transform, we propose a new NR-IQA model in contourlet domain. Number of features is extracted: energy of sub-band coefficients within scales and energy differences between scales, as well as measurements of the statistical relationships of pixels within and across scales. Then features are fed to SVM which learns to predict image quality. The features are also performed with a two-stage framework. Experimental results of both metrics show that the method has high linearity against the human subjective perception and low time complexity.3. We propose a NR-IQA method based on contourlet and spatial domain according to the different characteristics of the two domains. We extract the scale feature mutual information, direction feature energy and pixel features shape, variance and mean after fitting of Gaussian function. The prediction model is accomplished by SVM. Experimental results show that the method is consistent with subjective score better and outperforms the other reported related methods in the literature.