Research On Image Resizing Based On Depth And Saliency Feature

With the emergence of various digital terminals,the same image will be displayed more often on screens with different size and resolution.This leads to the requirements of image retargeting technique.Image retargeting is the process to adjust the aspect ratio of the source electronic image according to the screen size,so that the important information cannot be lost and the subject area keeps without seriously distortion.This paper proposes an algorithm to retarget image combing the depth clue,edge sensitive saliency and image energy.In order to solve the problems of loss of important information,distortion of image subject area,and incomplete information of non-subject area structure,we propose a novel algorithm to scale the target image based on the multi-view image features.Specifically,the content of this paper is listed as follows:(1)Improve the traditional region-based saliency calculation methods and propose contour-sensitive saliency features.Keeping important information is essential criteria for evaluating image retargeting algorithms.The traditional saliency computation only focus on the importance of image area without concerning the edge information,in that way,the retargeting result will lead to contour distortion.In this paper,we propose the contour-sensitive significance calculation for this phenomenon and solve the problem of deformation around edges.(2)The depth information constraint is introduced to the retargeting process.The structural information of the target image is represented by the relationship between various objects in the image.Among them,image depth information(position relation)is one of the measurement criteria to describe the structure information.In this paper,the depth information of the image is calculated,and the constraint of depth information is added during the scaling process,so that the scaled image still has similarity in structure with the original image.(3)For large-scale image scaling,a novel algorithm for maintaining the structure of non-theme regions is proposed to overcome the problems of over-protection of theme area.In large scale image scaling,the over-protection issue will cause distortion around significant objects.In order to solve it,the derivative of depth or saliency is used to balance the ratio of theme and non-theme area,which can effectively avoid the distortion of the non-significant area at the edge of the significant area.The main contributions of this paper are listed as follows:(1)Combining depth-of-field features to propose contour-sensitive significance calculation,which can calculate the significant features of the object edge in the image according to thepositional discrimination,the color discrimination,the saliency fusion algorithm,and the sparse learning method to obtain the blurred depth descriptor.(2)A non-theme region structure preservation algorithm is proposed.The conventional algorithm is mainly aimed at the preservation of salient region information.However,this paper proposes the maintenance of non-theme area to ensure the searched energy line is globally optimal.