Stereoscopic 3D Image Resizing Method Based On Seam Carving And Cropping

With the rapid development of stereoscopic 3D media technologies and contents,image resizing has become an important topic in computer vision.Image resizing studies how to adapt an input image to various display screens with different sizes and aspect ratios while avoid introducing distortion as much as possible.The seam carving method is very suitable for processing natural scene images and the cropping method can avoid introducing distortion to the image content in the cropped region.However,there are still some problems with the seam carving and cropping methods.For example,for seam carving method,severe visual distortion may be introduced by repeated removal or insertion of seams within a same region of the image.lImage cropping method tends to suffer "monocular object" and "window violation"problems,which usually confuse our brain,leading to visual fatigue.To solve these problems,this paper studies the stereoscopic 3D image resizing method based on seam carving and cropping.The contributions mainly include:An accumulative energy-based seam carving(AESC)method is proposed for 2D image resizing.Since the geometric consistency and 3D structure of 3D image are of great complicated and should be considered when using stereo seam carving,we first solve the problem of visual distortion caused by repeated removal or insertion of scams within a same region in 2D images.When removing or inserting a certain seam,we distribute the energy of each pixel on the optimal seam to its adjacent pixels so as to avoid the extreme concentration of seams,especially within a texture region.in addition,in order to reduce the geometric distortion of salient objects and avoid serious distortion occurring in the background region,we modify the energy function by adding the image saliency and the edge information detected by Canny operator in the background region.This method lays the foundation for the following proposed stereo seam carving method.An accumulative energy-based stereo seam carving(AESSC)method is proposed for3D image resizing.Firstly,we distribute the energy of each pixel on the seam to its adjacent 8-connected pixels when removing or inserting a seam.Secondly,since the information of each pixel is anisotropic,we use the Sobel operator to detect the direction that has the maximum edge information and continue the energy accumulation along this direction.Finally,we incorporate 3D structure consistency constraint in the energy function and adopt a pixel visibility maintenance method.Experimental results show that the proposed method can effectively reduce visuai distortion for 3D images while maintaining the geometric consistency.An example learning-based stereo image cropping(ELSIC)method which can enhance visual quality is proposed for 3D image resizing.First of all,some high-quality professional stereoscopic images which are similar to the input image with a given scene are retrieved and served as the learning examples.The similarity between two 3D images is calculated by measuring the distance using the GIST feature and the color histogram feature.Then,we proposed an example learning-based method to evaluate the quality of each cropping window and obtain an optimal cropping window.The quality of each cropping window considers the composition,the depth distribution,the information loss and the visual comfort according to the rules borrowed from stereoscopic photography.Finally,we use horizontal image translation(HIT)method for the left and right view to adjust the depth distribution,this is imitated from the way to get high-quality cropping results by professionals.Experimental results show that the proposed method can effectively obtain the visual comfortable results after cropping.