Study On Image Quality Assessment Based On Sparse Representation And Local Features

Image Quality Assessment(IQA) plays an important role in visual signal processing and communication, such as image acquisition, image segmentation, image fusion, biomedical imaging, etc. The quality of images affects the accuracy of our information acquisition and user experience. With the fast-growing demand for image-based applications, how to get efficient and reliable evaluations of image quality becomes particularly important. The goal of IQA is to obtain evaluations that corresponding with human visual system(HVS), namely, designing effective IQA models to mimic the quality predictions of observers and providing reasonable quality scores.This paper studies on IQA from two perspectives, which are sparse representation and local features. Based on these observations, three effective IQA methods are presented.1. A new sparse representation-based full-reference IQA model(QASD) is proposed based on the construction of adaptive sub-dictionaries. Natural images are first utilized to train an overcomplete dictionary, then this dictionary is used to represent the reference image. For the distorted image, it is represented using new dictionaries(i.e., subdictionaries), which are obtained based on the already used basis vectors in the representation of reference images. Sparse representation can extract higher level features in images, but it is not sensitive to weak distortions, thus three additional features are employed for more effective IQA, namely, gradient, color, and luminance. The proposed method is not sensitive to training images, so a universal dictionary can be used to evaluate the quality of common images. Compared with other full-reference IQA models, the proposed method produces the state-of-the-art results, and it delivers consistently well performances across different image quality databases.2. A no-reference image blur assessment method based on feature points is proposed. First extract feature points from the blurred image and re-blurred image, then the two feature point maps are divided into blocks to generate block-wise quantity map respectively. Later, the block-wise quantity maps are combined to compute the feature point quantity similarity map. Finally, an overall blur score is achieved by pooling this similarity map with a visual saliency map. The proposed method outperforms several no-reference image blur metrics, as well as some representative general-purpose blind image quality metrics.3. A no-reference image blur assessment method based on saliency guided gradient similarity(SGGS) is proposed. A blurred image is first reblurred by iterative low-pass filtering, producing a heavily blurred image. With this reblurred images as reference, gradient similarity is computed. Finally, visual saliency is employed in the pooling and quality score is generated. The proposed metric produces blur scores with a fixed range of(0, 1), and it features fast computation. Compared with other no-reference IQA models, this method delivers better consistency with the HVS.