Research On Key Technologies Of Intelligent Texture Image Completion

With the rapid development of computer vision technology, image completion has gradually become a research focus in the field. The purpose of image completion is to repair the image with missing information and make sure that it is non-detectable for an observer who does not know the original image. Given the fact that images with rich texture details are of complex characteristics, the related research faces a new challenge. By analyzing the characteristics of texture image, the dissertation intends to study the key technologies of texture image completion from the aspects of target-region extraction, image completion and image quality assessment. This research will build the theoretical base for object removal, error concealment and special effects in images and videos.In the stage of target-region extraction, a coarse-to-fine method is proposed to automatically find the defective regions in texture image. Gaussian filtering and threshold operation are first employed to acquire the rough contour of defects, and then a fuzzy Chan-Vese active contour model is established to drive the initial rough contour move towards the target and stop at the ideal location. Furthermore, the dissertation studies the issue of high-contrast texture extraction from fine-grained background. A multiple-channel active contour model which contains fuzzy information and adjustment factor is presented at this stage. Edge and intensity features, filtered by nonlinear diffusion, are taken as two feature channels for region segmentation. During implementation, the feature with larger differences between texture regions is taken as the leading term that drives the unsupervised evolution of segmentation curve. Experimental results demonstrate that the proposed methods can effectively solve the problem of target-region extraction in images which contain a variety of defects and texture details.In the research of image texture completion, an image completion algorithm based on patch associated matching and low-rank matrix super resolution is proposed to handle block artifacts caused by pixel value mutation between adjacent patches. With the relevance of adjacent patches taken into consideration, the implementation of noisy low-rank matrix filling ensures the continuity of changes in the detail information of texture and color. On the other hand, another image completion method based on dynamic-scale patch matching and layer-wise optimization is also proposed to reduce structure discontinuities. The proposed method implements a dynamic-scale strategy to gain the best candidate patches through two patch searching processes. Furthermore, image completion problem is abstracted as a chain optimization problem based on the relevance of adjacent patches and solved through dynamic programming layer by layer from the outside to the inside until the whole image is completely repaired. Experimental results demonstrate both the effectiveness and efficiency of the proposed algorithm for various natural images.Finally, an image quality assessment method using visual attention mechanism and image statistical characteristics is proposed to handle the issue of objective image completion quality measure. Based on the analysis of human visual perception, image quality assessment is translated into human attention to different regions in the image. According to visual saliency map and intrinsic characteristics of the repaired image, structure measure index and texture measure index are presented to form the final quality assessment system on the basis of hierarchical human perception. Experimental results demonstrate that the proposed method has high correlation with human subjective assessment result.