Research On Image Inpainting Technique

With the development of various social technologies, digital image processing techniques involves almost all aspects of human life and work. The conception of digital image inpainting was proposed in an international academic conference. As a branch of digital image processing techniques, digital image inpainting thchnique has developed very fast these years. This thesis studies digital image inpainting thchnique in depth.Digital image inpainting is to restore the integrity of the damaged image. It fills in the designated area in the digital image and tries to make the repaired image look like that it has never been broken. So this technique needs to make the transition between the filled region and the original region natural.The conception and development of digital image inpainting are introduced firstly. Then some non-texture-based inpainting methods are studied in this paper, including BSCB model, CDD model and an inpainting algorithm based on RBF (radial basis functions). Next, inpainting algorithms based on texture synthesis are studied. By analysing Criminisi’s algorithm, key points for inpainting both texture information and structure information are found. Then, an improved method based on Criminisi’s algorithm is proposed. It improves the calculation of priority and micro-filling-order. For the calculation of priority, more gradient information of known pixels in the target patch is used. There are two micro-filling-orders. One is to fill in the block which has the point with the highest priority in center at each time. The other is to fill in the block which has the point with the lowest priority in center at each time. Experiments show that better results can be obtained by using the improved algorithm.Finally, digital image inpainting technology is combined with face images and two inpainting methods for face images are proposed. The first face image inpainting method is based on a single image. It inpaints face images by using Criminisi’s method flexibly. This process fills in the block witch has the point with the highest priority in center at each time. But it chooses the source region in a reasonable way for facial images. And the calculation of priority is also suitable to repair facial images. For the human face is symmetrical, this method does horizontal mirror transformation to damaged facial images and uses the known area of the result image as source region. And it also narrows the scope for searching the best matched patch. The calculation of priority takes the position into account to make the filling order from the middle to sides. The advantage of this method is that it is a simple algorithm and is easy to carry out. But it can not inpaint facial severely damaged images. In order to inpaint seriously damaged facial, the second method for face images is proposed. It is based on ICA (independent component analysis). Before inpainting, this method extracts base images from an image library using ICA algorithm. Base images contain basic features of the face. For a damaged face image a set of coefficients should be determined first. And then the results of inpainting can be obtained by base images and these coefficients. Compared with the method based on a single image, this method could inpaint seriously damaged face images.