Stereoscopic Image Quality Assessment By Analyzing Binocular Perceptual Properties

The increasing demands of high quality 3D images in recent years have let stereoscopic image quality assessment(SIQA)become a very important research field among modern image processing techniques.However,the study upon SIQA still remains immature due to the technical difficulties brought about by the complexity of human visual system(HVS).The generation of human binocular vision does not equal to a simple summation of visual signals from both eyes,but introduces additional binocular effects.For instances,binocular rivalry causes unequal attention distribution for both eyes resulting from differences in image contents of left and right views,artificial stereoscopic scenes lead to discomfort in visual sensation,etc.These effects are potentially influencing human perception upon stereoscopic image quality.Subjective SIQA is time and labor consuming and non-repeatable,thus,designing objective methods to correlate with subjective sensation as highly as possible,is the most important topic in the field.Based on the perceptual properties of human binocular vision,this paper focuses on the design of SIQA methods,and proposes:1.A full-reference SIQA method based on the hierarchical structure of HVSBy analyzing the structures and properties of different hierarchies in HVS,an SIQA method is proposed,which operates with the help of reference images,and utilizes the strategy that extracts and quantifies multiple image features and maps them to image quality.Meanwhile,influences of binocular rivalry are taken into consideration on pixel level to ensure its performance.Through experimental results on public SIQA databases,the proposed method shows better consistency with human subjective sensation and better stability among various distortion types,in comparison to existing representative full-reference methods.2.A no-reference SIQA method considering binocular rivalry and visual comfortAnother SIQA method is proposed in the paper that can be operated completely without reference images.Based on the adaptive image segmenting results,the method extracts and quantifies image quality related features on multiple spatial scales.Also,binocular effects including binocular rivalry and visual discomfort are estimated on these scales,and visual saliency computation models are adopted,so that the performance of the method is further improved.Experimental results on public databases show that,in terms of accuracy and stability,the method is not only superior comparing to representative NR methods,but is also quite competitive against FR ones.