Research On Key Issues Of Image Retargetting And Quality Evaluation

Digital images have become one of the mainstream contents in the current media communication for the rich meaning and visualization presentation.The diversity of terminal display devices has imposed higher requirements on multi-scale display of images and pleasant visual experiences.Therefore,it is a challenging work about how to satisfy the user's visual perception to the greatest degree.In this thesis,we study the issues of image retargeting and quality assessment of retargeted images.The details are as follows:1.A content aware fast hybrid image retargeting algorithm is proposed.Firstly,based on the important regions and the refined boundaries,the images are partitioned into unequal strip regions.For each strip,the object size could be estimated based on the importance.A fast framework is designed for non-continuous seam carving.It can obtain all of the seams by one time scanning.And the neighboring relationship matrix and touch bound matrix could be obtained simultaneously.And then the seam context could be utilized to update the energy of each seam in the stage of seam removing.Finally,the image quality of the strip removing a strip could be quickly estimated based on the relationship between the removed seams and the important sub-images.If the image quality is smaller than the assigned threshold,the non-continuous seam carving is stopped and the current strip is resized to the object size by using scaling.The experimental results demonstrate that the proposed approach achieves good performance in term of image quality and efficiency comprehensively.2.An objective quality assessment algorithm integrating the global and local features is proposed.First of all,the image is preprocessed,which includes feature points extraction and matching and saliency map generalization.Then image is partitioned into the triangulations based on the matching feature points.And next,the global quality is evaluated based on three factors including shape deformation factor,position offset factor and angle change factor.Furthermore,the local quality is evaluated by computing the SSIM of matching feature points.Finally,the global and local evaluation results are integrated into the final image quality.The proposed method makes up for the current image quality evaluation method features such as a single feature,and it can capture the image quality changes more comprehensively.3.An image quality assessment method based on Triplet network is proposed.Using the advantage of deep learning in feature extraction,the feature learning is optimized.First,according to the structural characteristics of the triplet network,the image data set is augmented and grouped.Then the network is used to train and learn the input triplet images,and the network parameters are continuously optimized to obtain a better quality evaluation model.4.An image quality evaluation system is designed,which mainly includes two parts,a subjective evaluation module and an objective evaluation module.The subjective evaluation module is used to subjectively evaluate the results of different retargeting algorithms and to obtain the user's subjective evaluation data.The objective evaluation module is used to evaluate the image quality by running the evaluation model.It usually uses the original image as a reference to calculate the similarity of the resized images of two retargeting algorithms and gives the evaluation results.