Applying neural network on image up-sampling to promote the efficiency of texture matching

The objective of image up-sampling is to produce the correlated high resolution image from a low resolution image. In this project, we propose to apply neural network on texture database mapping to achieve a more time efficient image up-sampling algorithm. A texture database of high-resolution images will be built in advance before the up-sampling procedure starts. The proposed method consists of two stages: 1. Train a set of neural networks (NNs) for classifying textures; 2. Match each pixel to its corresponding NN of high-resolution textures. In the first stage, we train a set of neural networks (one NN for each kind of texture) from the high-resolution texture database. After the set of NNs is well trained, a high resolution pixel value can be obtained by passing a low resolution pixel into a NN of its corresponding texture. In the second stage, for an input low resolution image, we firstly segment it according to its textures (texture segmentation), then match each pixel to its corresponding NN of high-resolution textures in the database. A super-resolution image can be obtained after all pixels have been matched. In order to avoid excessive human artificial results, the outputs of the proposed method have to be down-sampled and compared with the original low resolution image for further adjusting. The up-sampled image will not be output until the difference between the down-sampled image and original image is within a predefined constraint.