Study On Image Quality Assessment Methods For Certain Situations Via Perception Model

With the development of image acquisition methods and the performance improvement of terminal display equipment,different types of images,such as retargeted images,dehazing images and stereo images,have attracted more and more attention.However,in the process of data acquisition,transmission and display,images are inevitably affected by complex external environment,resulting in quality distortion,directly reducing human experience or indirectly interfering image processing algorithms.Therefore,it is an important subject in the field of computer vision to evaluate image quality of certain situations.In addition,building different models based on visual perception plays a key role in image quality evaluation.This thesis studies image quality assessment methods for certain situations via perception model.Firstly,retargeted image quality assessment method based on mapping perception model is established to evaluate the retargeting quality caused by the change of the ratio of horizontal or vertical resolution.Specifically,this method first extracts the sensitive geometric shape and content differences in the process of image retargeting,and uses the feature vector to represent the changes before and after image retargeting.Then,this thesis introduces Auto En Coder as regression method,and reduce the over fitting caused by insufficient data.Secondly,this thesis proposes an objective evaluation algorithm of dehazing image quality considering the balance of positive and negative utility,which can be defined as balance perception model.In the process of image processing in foggy environment,this method can provide threshold guarantee for image acquisition.Specifically,this thesis uses three different feature extraction methods for dehazing image quality assessment: positive effect feature,medium effect feature and negative effect feature.Furthermore,the improved deep belief network is used as the regression method of the evaluation algorithm,which improves the performance of objective evaluation.Thirdly,this thesis proposes stereo image quality evaluation algorithm based on binocular perception model from two different perspectives.On the one hand,according to the binocular stereo vision model,it can fully mine the differences between the processing patterns of binocular information and monocular information,which help distribute the weight.Then,IW-SSIM model and SW-SSIM model are introduced,and the objective evaluation model of stereo image quality is constructed.On the other hand,the feature of stereo scene is extracted from DCT domain,and the model is trained with deep belief network.Finally,the stereo image quality evaluation algorithm with good performance is constructed.The two models focus on feature extraction and quality regression,then explore and verify the guiding role of stereo perception from different perspectives.Fourthly,three application examples are used to verify the important role of different image quality assessment methods for certain situations in practical software and hardware system.Among them,the application of retargeted image quality evaluation in image retrieval field counteracts the influence of image resolution stretching on image retrieval accuracy;the application of dehazing image quality evaluation method improves the effectiveness of image dehazing algorithm,which plays a key role in improving performance of image contrast enhancement algorithm;the quality evaluation method for stereo image is applied to wearable auxiliary devices for visual impaired groups,which improves accuracy of object recognition.The above four aspects of research focus on image quality assessment methods for certain situations via perception model,which are carried out in detail from the theory to application levels.Based on the research content of this thesis,the contributions can be summarized as follows.Firstly,this thesis deeply explores the relationship between human visual processing and image quality evaluation methods,and explores the driving role of perception model in image quality evaluation of three different certain situations.Secondly,this thesis completes the retargeted image evaluation based on mapping perception,the dehazing image quality evaluation based on balanced perception and the stereo image quality evaluation based on binocular perception,which breaks through the data regression bottleneck of traditional algorithms.Thirdly,this thesis proposes to use the image quality evaluation model of specific situations in other areas of image processing to verify the application value of image quality evaluation.