Image Reconstruction Via Combining Patches Based On Point Sampling And EDGE-Directed Ideas

Image interpolation is the basic technology in the image processing and plays an important role in the applications. Traditional image interpolation methods are mainly based on the low-pass filter theory. But they can not deal with the sharply changed regions, such as edges and textures and it can cause the edges and textures blurred. To solve these problems, edge-directed interpolation theory is proposed. However, the interpolated images have low average quality in measure of PSNR and it can lead to high computational complexity.This paper proposes a new algorithm for image reconstruction via combining patches based on point sampling and edge-directed method (PSE). Usually, the traditional methods use the image data to construct fitting surfaces directly. As a result, the accuracy of the interpolation may not be assured. That's because the image data is obtained by area sampling method, and the pixel is regarded as an average value of certain area. Therefore, for the given image data, the precise point sampling values should be computed first. And then the point sampling data is used to construct the fitting surface patches. Through this way, the fitting patches have high interpolation precision and less errors. Last, we propose the weighted functions for different patches. The whole image surface is constructed by combining all the local patches with weighted functions. For the smooth areas (low-activity areas), we compute the interpolated pixels on the fitting surface by area pointing method.And for the non-smooth areas, which can be also called high-activity areas such as edge and textures, the method above can not produce better results. For this reason, a model is proposed to deal with the edges and textures. Traditional edge-directed approach can solve the problem with blurred edges very well, but the average quality of the image is not very good which can be seen from the PSNR measure. In addition, the uncontrolled radical changes may be produced during the computation the model parameters of the traditional methods. For example, the pixel may be negative. In this paper, a new edge-directed approach is presented to obtain the model parameters a=(a1,a2,a3,a4) and b=(b1,b2,b3,b4). Different from the existing edge-directed approaches, the relationships between the inner members of parameters are taken into account which can solve this problem.Through lots of experiments about the new approach on the standard images by different methods such as bi-cubic, PCC and SAI interpolation, it can be concluded that the PSE method proposed in this paper has best results in both PSNR measure and the objective visual quality. The interpolated images produced by PSE method have high PSNR values than the competed methods, and they do not have blurred edges and textures. Therefore, the experiments for testing the efficiency of the new approach show that the interpolated images reproduced have best interpolation precision and the best objective visual quality through the competed methods.