Study Of Imaga Retargeting Quanlity Assessement Based On Saliency-driven Classification And Reconstructed Structure Distortion

With the different diverse applications and user’s demand,screens of multimedia terminals have various resolutions or aspect ratios.As the original images with a fixed-size are displayed on different devices,they may need to make appropriate adjustments,to reduce quality degradation or the screens are wasted caused by the mismatch of resolution or aspect ratio.Content-aware image retargeting methods are effective to resize important content with satisfactory reconstruction quality,which aim at guaranteeing the quality of visual important contents during rescaling.However,the qualities of the resized images produced by various retargeting methods are different given an original image.Therefore,to select appropriate image retargeting methods for different applications and improve the performance of an image retargeting approach,evaluating the quality of retargeted images effectively is a primary task.Subjective approaches are adopted in the early quality assessments for retargeted images.Though the subjective quality evaluation is effective,such a measurement is difficult to be applied to real-time scenes,because so time consuming and high cost are needed.To replace subjective evaluation,and simulate human perception of important content distortion,the objective evaluation algorithms based on salient region distortion have been proposed to assess the quality of retargeted images.However,Most of the existing image retargeting quality assessment methods rarely consider the influence on the quality with different area and amount of salient region in retargeted images,it may cause the weak mismatch between the subjective and objective assessment results.Also,there are some quality assessment methods ignore the difference structure between original image and retargeted image,which may cause salient regions of resized images detected are often difficult to match the corresponding contents of the original images,the performance of the evaluation algorithm is reduced easily.In addition,most of the existing evaluation algorithms mainly evaluate the structural distortion by the difference of structure and texture information between original image and retargeted image,which may not only need to ensure the matching of the structural information between original image and retargeted image,but also consume a lot of time to extract and compare the structural information.To solve these issues,in this paper,the image retargeting quality evaluation algorithms based on salient region distortion and structure distortion are deeply studied and innovated.The research results are as follows.A novel image retargeting quality assessment method based on saliencydriven classification has been proposed.Firstly,the saliency information of original image are concerned to develop an adaptive image classification method,because of various distortion characteristics,different metrics are designed to evaluate the quality of different types of retargeted images.Furthermore,a novel salient region extraction method have been proposed for retargeted images,which makes full use of the backward registration information from the retargeted image to the original image to extract the salient region of the retargeted image.This method could benefit to accurate distortion measurement in important areas.The Retarget Me and CUHK databases are used to verify the performance of the proposed method.The experimental results show that the proposed method is obviously better than the existing methods.Specifically,in the Retarget Me database,it is seen that the average Kendall Rank Correlation Coefficient(KRCC)value of the proposed method is 0.589,the stability value is 0.119;the proposed method is also evaluate the performance on the CUHK database,in which the Pearson Linear Correlation Coefficient(PLCC)value of the proposed method is0.7601,and the evaluation square error is 8.284,which is lowest than the compared works.Besides,the novel salient region extraction approach for resized images designed in this paper can facilitate the visually prominent areas of the resized image to match the original contents.In this thesis,an image retargeting quality evaluation algorithm based on the distortion of the reconstructed structure is proposed.Firstly,to effectively measure the structural distortion in retargeted image,this thesis utilizes the backward registration between retargeted image and original image to complete information reconstruct.Secondly,according to distribution of the reconstructed information,two methods based on the global reconstruction information distribution and the salient reconstruction information are proposed for measuring structural distortion.Moreover,these two distortion metrics are incorporated into the saliency-driven classification evaluation method.The Retarget Me and CUHK databases are used to proven the performance of the proposed method.The experimental results show that the evaluation algorithm proposed in this paper can effectively evaluate the global and salient structural distortions for retargeted images.The KRCC of the proposed method and the subjective results in the Retarget Me database is as high as 0.654,which is better than all current image retargeting quality evaluation algorithms.In addition,the structural texture information is no required to be separately detected and compared of original image and retargeted image in the proposed,which saves the time of evaluation processing.