Adaptive band-based super-resolution reconstruction of video sequence

This paper presents an adaptive band-based super-resolution (SR) algorithm based on discrete wavelet transform (DWT). The original video sequence is divided into high-resolution key frames (KFs) and low-resolution non-key frames (NKFs). In general, DWT separates each frame (including key and upsampled non-key frames) into four bands (LL, LH, HL, and HH). With constructing conditional SR algorithm, quality of non-key frames is iteratively enhanced by reconstruction of lower-frequency bands and compensation of sampling error (or different data). After upsampling NKFs, corresponding information in higher-frequency bands can be compensated with adaptive prediction for motion search. Comparing with the previous work, PSNR of the proposed SR method with mixed upsampling has quality gain up to 0.2 dB.