Research On Content-aware Image Retargeting Based On Multi-operator

Due to diversity of display devices,how to adjust an image to satisfy the display demand of various resolution and proportions has gained extensive research efforts.Content-aware image retargeting techniques aims at preserving the quality of visual important objects during resizing.Many image retargeting approaches have been proposed to achieve this object.However,existing methods mostly focus on performing retargeting towards only a single direction one time.This might cause deformation of visually important area in resized images since the aspect ratio of important objects can't be maintained,especially for the case that the size of the important contents in an original image is larger than the target resolution of a resized image.On the other hand,the images transmitted and stored through internet are often encoded by using discrete cosine transform(DCT)technique,hence the images must be decoded fully before retargeting by using the image retargeting approaches in pixel domain.And then the resized images are encoded fully again.This may lead to a great consumption of terminal resource.If the image retargeting is performed in the DCT domain,computational complexity will be reduced.In this thesis,a multi-operator image retargeting algorithm is proposed which can preserve the visual important contents in pixel domain.Moreover,a multi-operator image retargeting algorithm in DCT domain is also presented.The detail are as follows.A novel multi-operator image retargeting scheme is proposed,which combines a seam carving(SC)method with a uniform scaling technique,and a gradient vector flows(GVF)based seam carving.Specifically,to preserve the aspect ratio of visual important contents as resizing an image,the image towards both vertical and horizontal directions are stretched and then a similarity transform by using an indirect SC and a uniform scaling method is performed.After that,a direct SC approach with gradient vector flow(GVF)is employed to shrink the image.But in most cases,the image both change the height and width of the image are resized.So,to deal with a general case of image resizing,develop a bidirectional image retargeting scheme is developed,which can perform retargeting in both horizontal and vertical directions simultaneously and change the image to any size and proportion.The experiments are conducted in the public image database.And the experiments of single-dimensional and Bi-dimensional image retargeting demonstrate that the proposed schemes are superior than the existing image retargeting algorithm in subjective quality.The average objective mass index of this algorithm is larger than the existing algorithm,which indicates that the objective quality of this algorithm is better.the computational complexity of single-dimensional image retargeting algorithm is similar to GVF based SC algorithm,and is much smaller than other multi-operator algorithms.the computational complexity of Bi-dimensional image retargeting algorithm is less than that of GVF based SC algorithm and existing multi-operator retargeting algorithm.Currently,the image transmission and storage in the internet is compressed used the joint photographic expert group(JPEG)standard,if use pixel domain scaling algorithm,the JPEG image must be completely decompression to pixel domain.This greatly increases the computational complexity of the terminal device.In order to reduce the resource consumption of fully encoding and decoding during the actual application,a content-aware image retargeting scheme based on the DCT domain is presented.Firstly,the compression bit stream is decoded to the DCT domain,and the DCT coefficient matrix of the image is obtained.Then a saliency map of the image is calculated by using the DCT coefficients.Then the image in non-retargeting direction is stretched,which may help to reduce the proportion of visual object in the image and to avoid the visual object deformation.Most importantly,since the scaling in DCT domain which may delete directly 8 rows or 8 columns will bring a noticeable sawtooth,a new image similarity transformation method in DCT domain is proposed.In this method,similarity transformation is performed on the basis of a DCT coefficient instead of a whole DCT block.Finally,GVF based SC scheme in DCT domain is used to adjust the image size to achieve the target size.The experiments are conducted in the public image database.And the experiments demonstrate that the proposed schemes the subjective quality of our algorithm is better than the existing image retargeting algorithm in DCT domain,and it is also better than the image retargeting algorithm in pixel domain.The objective quality index of this algorithm is larger than the existing algorithm,which indicates that the objective quality of this algorithm is better.In addition,the compute time in the process of retargeting is less than the multi-operator scheme,SC algorithms and GVF based SC algorithms in pixel domain,at the same time omit the time complexity of completely decoding and encoding.