A Method For Image Quality Assessment Based On Visual Attention And Structure Distortion

Image Quality Assessment plays a basic and key role in computer vision and machine learning. It can be widely used in many areas such as image quality control,image processing criterion,system parameter optimization and so on. Meanwhile,Image Quality Assessment is also a difficult issue in computer vision because of the kinds of distortions generated from image acquisition and image processing,which makes it difficult to have a unified evaluation method to solve such problem. Apart from this,Image Quality Assessment is even more a visual perception issue,it is too subjective to give a concrete description. So it is a very challenging task to design an objective and accurate Image Quality Assessment.The object of this thesis is to incorporate visual attention explicitly with traditionally Image Quality Assessment to enhance their performance,which mostly based on error sensitivity statistics or HVS. The former only focus on the difference between the distorted image and reference image,without considering the impact of visual perception,so there are great discrepancies between the objective score and Mean Optional Score. Through integrated simulation of the characteristics of HVS, the later metrics could better ensure the consistency of the objective and subjective evaluation,however,subject to the algorithm complexity,large amount of computation,and the current level of understanding of the HVS mechanisms,we still need further study. Therefore,in order to combing these two ideas, this paper proposed an algorithm based on error statistics by adding visual attention mechanisms to play their advantages and improve performance of the algorithm.First of all,the basic concept and theoretical framework of Image Quality Assessment are listed, the new framework which incorporates with visual attention was proposed, and the rationality that visual attention could improve the performance of algorithm is also discussed. Then,two basic theories this papers based on are described in detail,they are context-aware Saliency detection and structural similarity index,and the theoretical analysis and experimental proof are also carried out. At last,we design three programs which can combine the Saliency Map and distortion map in allusion to the characteristics of the two maps,and finally discovered an algorithm which is more efficient and accurate.