Research On Automatic Image Refocusing Based On Depth Estimation

The goal of image refocusing is to refocus on a certain region in an unfocused image,making it as foreground which stand out in sharpness against the background region,the refocused image will be more aesthetic.With the prevalence of smart phones and pocket cameras,image refocusing has become a basic editing and pro-cessing method for its power in interesting object emphasis and image beautification,which has been widely used in practice.For example,image refocusing can be applied to taking enhanced photos function in smart phones,post-processing of photos func-tion in some softwares and so on.As an important topic of computer vision,image refocusing has been explored and studied by many researchers.However,existing image refocusing methods mostly require several images or motion videos to obtain more information for refocusing work.Moreover,as a more challenging work,single-image refocusing commonly suffers from high computational cost or depends on expensive hardware like light field camera.Therefore those current work cannot apply to mobile devices.According to the limitations,we present a novel technique which exploits depth map to perform single-image refocusing automatically and rapidly.The technique only needs user to specify an input image at the beginning,then estimates the depth of it.On the basis of depth information,the technique refo-cuses on the salient objects in the image.The whole process does not need additional devices and any other user intervention.Because of the algorithm is automatic and has low computational complexity,the technique can be applied to mobile devices.The primary contributions of this paper are:1.Improving a deep learning method based on generative adversarial nets,which helps us to achieve accurate depth information from a single image.Previous methods which do not utilize deep learning can easily produce inaccurate results owning to the lack of clues obtained from a single image.We study and explore the traditional generative adversarial nets,modify the objective function of con-ditional generative adversarial nets,improve the generator and discriminator to avoid limitations of previous methods.Experimental results demonstrate that the method can estimate the depth from a single image effectively and rapidly,which enhances refocusing work significantly.2.Presenting a novel method for single-image refocusing.The method leverages current saliency detection methods to locate the region of interest for a given im-age rapidly.Then the original image is softly segmented into layers and blurred with different confusion sizes according to the depth map calculated before.At last,the blurred layers are softly combined into a refocused image in which salient objects stand out from the background.Experimental results demonstrate that our method performs outstanding both in automatic photo refocusing and computational complexity.On the basis of the above research,we implement a prototype system to esti-mate the depth from a single image and then refocus automatically based on the depth.The system can meet our study requirements,demonstrating the effectiveness of the method.The system also indicates that our method can be applied to refocus scenes in real life.