A fast algorithm for learning-based super-resolution reconstruction of face image

Considering that the interpolation method for image zooming yields blocky or blurred results and the results of learning-based super-resolution algorithm are satisfactory but time-consuming, in this paper, we introduce a novel fast algorithm for learning-based super-resolution reconstruction of face image and implement it. First, we generate an optimized training database; second, for each test low-resolution color face image, we locate the skin region of YIQ space´s Y channel and apply super-resolution reconstruction algorithm to this region of interest; third, we break it into patches, search for N closest candidates for each patch, and under maximum a posteriori criterion, get best-matching candidate for each patch; finally, we get residual details by merging all patches, and then the final reconstructed high-resolution face image after integrating it with interpolated results.