Pixel prediction by context based regression

We propose a pixel prediction algorithm, which learns a regression function corresponding to each context. A context refers to a group of pixels, that have similar correlations with its neighboring pixels. We propose to form a pixel´s feature vector by its neighboring pixels´ ratios, so that they better capture the pixel properties described by the regression weights. Then we use K-means clustering to classify the feature vectors of all pixels into several contexts. Clustering reduces pixel randomness within each context, thus reducing prediction error. We apply three regression algorithms, the least square, quantile and lasso regression, which assume different loss function and regularization. Experimental results demonstrate that all context based regression methods have outperformed conventional pixel predictors. Among them, quantile regression, which assumes l1-norm loss function has the best result. It has 3.1% less bits per pixel (bpp) than least square prediction with 12 neighboring pixels.