Intelligent Image Processing And Inpainting For Ancient Fresco Preservation

The preservation of endangered ancient frescos is an urgent and complex task, however, the traditional methods are inefficient with a lack of ancient frescoes preservation personnel, which hinders the development of ancient frescoes preservation cause. As a result, the introduction of computer technology, especially image processing technology becomes very important, for the ancient fresco preservation assisted by image processing technology can not only reduces damages caused by human factor but also greatly improve efficiency.In this paper, the research on intelligent image processing and inpainting technologies for the preservation of ancient fresco was used to solve the ancient fresco line drawings generation, ancient fresco disease recognition, ancient fresco damaged area segmentation and restoration issues. The existing image processing technology can hardly solve these problems for its low connection with the characteristics of ancient fresco and information of related art and culture. Its shortcomings mainly include:Firstly, the existing line drawings generation techniques don't take the specific style information existed in ancient frescos into account. Secondly, the existing feature extraction and recognition technology don't consider the inconsistencies of brightness and angle and noise characteristics of ancient fresco images. Thirdly, the existing image segmentation technology cannot effectively integrate high-level semantic knowledge with underlying clues. At last, the existing inpainting technology used for the virtual restoration of ancient frescos has problems of error propagation and accumulation or lack of considering structural information.In order to solve the problems mentioned above, this dissertation focuses on key technologies of intelligent image processing and inpainting for ancient fresco preservation. The main research topics are as follows:1. An ancient fresco line drawing generation method was proposed. As most of Chinese frescoes are traditional meticulous painting, which contains line drawing strokes, an ancient fresco line drawing extraction algorithm using color and shape information was presented, color and shape features were firstly used to extract the line drawing strokes, then a SVM classifier was trained and used to remove the cracks which might be misidentified as line drawing strokes. Then, an interactive line drawing generation algorithm was presented. The algorithm first got the contour of a fresco image which was composed of many connected curves, and then rendered the strokes by learning styles from examples. As both the expertise of the artist and user interaction information were used, the method could get line-drawings of any ancient frescoes.2. An ancient fresco diseases recognition method via texture feature and support vector machine was proposed. Firstly, a multi-resolution gray-scale and rotation invariant texture feature extraction with wavlet local binary patterns was presented, which had advantages of computational simplicity and strong antinoise ability. Then, a new tree-structured support vector machine for multi-class classification was presented. The distances measured at the kernel space and fuzzy clustering were used to construct the tree-structured multi-class SVM, which had more classification accuracy comparable to some popular multi-class SVM approaches. At last, the improved feature exactor and classifier were used to implement the ancient fresco diseases recognition algorithm.3. An efficient ancient fresco damaged area concurrent detection and segmentation algorithm based on graph cut was proposed. We cast the problem of damaged areas segmentation within a fresco image as that of estimating a probabilistic model which consisted of an object category model in addition to the grid CRF. While low-level cues provided bottom-up information, the damaged areas category model based on relevance vector machine incorporated top-down information about the color and texture of the damaged areas. In contrast to most existing algorithms which trained top-down and bottom-up modules separately, our method took into account both bottom-up and top-down cues simultaneously. 4. Two ancient fresco image inpainting algorithms via discrete optimization were proposed. As greedy synthesis based image inpainting algorithms might cause visual inconsistency, and some other global optimization based image inpainting algorithms might not consider structure information, two fast ancient fresco image inpainting algorithms via discrete optimization were proposed. The first algorithm formulated the image inpainting problem as minimization of a weighted energy function, which was optimized using an Expectation Maximization(EM)-like algorithm. The second algorithm formulated the image inpainting problem as a graph labeling problem, which was optimized using belief propagation algorithm, besides, a non-local mean based label candidates reduction was presented to improve the performance of belief propagation.