Research On Image Inpainting Algorithm Based On Sparse Representation

Image is the most universal information carrier in human's daily life,and digital image inpainting technology always is the research hotspot in the field of digital image processing.The main idea of digital image inpainting is utilizing the known information to fill and repair the damaged area,making the repaired image satisfy the visual requirement of human.Image inpainting algorithm based on sparse representation utilizes the sparse property of image,and finds the sparse representation of image by the over-complete dictionary,then reconstructs the image through the obtained sparse coefficient to achieve the purpose of image restoration.This thesis studies image inpainting algorithm based on sparse representation,and the main work and achievements are as follows:1.A framework of two layers patch extracting and an improved framework for dictionary updating are proposed to improve the K-SVD image inpainting algorithm.Two layers patch extracting framework decompose the damaged image firstly to get single image patch,then we will decompose the single patch second time,and the single patch will be the object of inpainting operation each time.Then all the single patches will be parallelly inpainted through the K-SVD image inpainting algorithm,and integrated to get the final inpainted image.The improved framework for dictionary updating is applying the Closed-form solution to update the trained dictionary once more in the stage of dictionary updating.Experiments on several real images can indicate that the improved K-SVD image inpainting algorithm can enhance the inpainting performance compared with the original algorithm.2.In order to improve the shortcoming of the dictionary learning algorithm with high computation,which needs long time to train,the idea of ensemble learning is introduced to the procedure of dictionary training,and utilizes the Bootstrap sampling strategy and Bagging idea to design a multi-dictionaries fusing framework.The framework firstly utilizes the bootstrap sampling on the damaged image to generate multi-subsets,then parallelly train the subsets by K-SVD dictionary learning algorithm to obtain multi-dictionaries,then utilizes these dictionaries to inpaint the damaged image and can obtain multi-results,finally applies the Bagging strategy to average all the results to get the final inpainted image.Experiments on several real images can indicate the framework can keep the same inpainting performance compared with the original algorithm,and reduce the time-consuming sharply,enhancing the efficiency of dictionary learning.