Super-resolution using a local matching designed for suppressing errors of reconstructed edges

A novel registration algorithm for multi-frame super-resolution (SR) is proposed. SR requires known LR pixels shifted by non-integer amounts of pixels. Such LR pixels can be acquired from non-target frames using registration. A method suitable for implementing registration in consumer products is the block-matching algorithm (BMA). However, when the BMA is used, in the case of processing a straight edge, it is shown that the arrangement tends to be irregular. It is also shown that the quality of reconstructed images becomes poor when the arrangement of the registered pixels is irregular. The proposed method is based on the BMA but is designed for improving the arrangement of the registered pixels. The experimental results show that (1) the pixels are registered near each position to be reconstructed, and (2) the quality of reconstructed images is improved. To make the proposed method practical for consumer products, the search is performed using a fast block-matching algorithm.