Study On Image Inpainting Algorithm With Global Optimization Based On Exemplar Structure Features

In the field of digital image processing,image inpainting technology has emerged as an important research topic since the beginning of the new century.After nearly two decades of development,it has become a research hotspot in computer graphics and computer vision.It has penetrated into all aspects of modern people's daily life.The exemplar-based image inpainting algorithm has a good effect in restoring large damaged areas,and it is divided into the matching-based method and the globaloptimization-based method according to different inpainting methods.Among them,the image restoration problem is regarded as an optimization problem in the global-optimization-based method,and all missing pixels are filled by optimizing the energy function at the same time which has took good care of the global characteristics of the image,and it solves the error accumulation problem easily caused by the greedy matching strategy in the matching-based method.It starts with the structural features of the image,and explores how to better extract and measure the structural feature information of the exemplar,and use it as a priori information to guide the image restoration process,so two improved algorithms are designed.Starting from the two key points of the construction of global energy equation and the selection of candidate labels,and better playing the role of data term and smooth term in the global energy equation,a image inpainting algorithm combining structure feature with information entropy is designed for the inaccurate and incomplete extraction of image structure information in the energy equation of existing algorithms.The method first introduces the structural tensor in the data term to represent the local features of the image,and then measures the anisotropy of the structural features between the exemplar patches.Then,based on the color information,the information entropy of the average information amount in the pixel patch is introduced to construct the smooth term to maintain the consistency of the neighborhood information;finally constructs the weighted global energy equation and uses the graph cut algorithm to obtain the energy optimal solution,at the same time,multi-resolution layer inpainting and adaptive determination of exemplar block size are used to reduce the algorithm complexity.The experimental results show that this algorithm can maintain the consistency of the reconstructed image structure and the clarity of the texture information,and obtains satisfactory results.Starting from the grayscale gradient which is the easiest to extract the edge information of the image structure,comprehensively considers various structural feature extraction methods,and constructs a gradient feature to extract and measure the structural information of the image,so an image inpainting algorithm using HSI gradient statistics feature is designed.The method maps the image to the HSI color space to extract the gradient features of the image,and uses the statistic such as gradient mean and variance to redefine the structural similarity between the sample patches,and constructs the global energy of the weighted data term and smooth term so that the global equations make the construction of data term and smooth term more favorable for the extraction and measurement of image structure information.At the same time,the algorithm limits the label matching range to improve the label matching accuracy and reduce the algorithm complexity.According to the experimental results,this algorithm can reconstruct different types of damaged images well,and ensure that the structural parts extend correctly and completely from the known area to the damaged area,and the joint effect is weakened,which is more in line with human visual requirements.In addition,in order to improve the practicability and operability of the proposed algorithms,an image inpainting demonstration system was designed by using MATLAB GUI to output and display the inpainting results directly.