Efficient Image Interpolation via Anchored Neighborhood Weighted Nonlinear Regression

We present an efficient and effective learning-based nonlinear image interpolation method in this paper. Our proposed method utilizes a single-hidden layer feed-forward neural network to mapping the local nonlinear relationship between high resolution patches and low resolution patches directly. In training phase, the training samples are firstly divide into 4 groups based on different known pixels pattern. For each group, a low resolution dictionary is trained and neighborhood are searched in the whole group data set for each atom which is regarded as an anchor. The searched neighborhoods for each anchor especially are then used for learning nonlinear regression model. In the interpolation phase, in order to further improve the interpolation performance, we weight the regression models of related anchors according to the similarity between input patch and anchors. With the off-line learned regression models, our method can achieve high computational efficiency. Extensive experimental results demonstrate the superior performance of our interpolation method compared with the state-of-the-art interpolation methods.