Fast super-resolution algorithms using one-dimensional patch-based training and directional interpolation

• Published in: IET image processing Vol. 6; no. 5; pp. 548 - 557
• Main Authors: KANG, Y. U, JEONG, S.-C, SONG, B. C
• Format: Journal Article
• Language: English
• Published: Stevenage Institution of Engineering and Technology 01.07.2012; The Institution of Engineering & Technology
• Subjects: Algorithms; Applied sciences; Computation; Dictionaries; Exact sciences and technology; Information, signal and communications theory; Interpolation; Matching; Signal and communications theory; Signal representation. Spectral analysis; Signal, noise; Synthesis; Telecommunications and information theory; Training; Automatic classification; Dictionaries; High resolution; Similarity; Superresolution; K means algorithm; Artefact; Two dimensional model; Low resolution; Signal classification; Adaptive method; Learning; Interpolation; Simulation; One dimensional model; Miniaturization; Fast algorithm; Content analysis; Signal to noise ratio
• ISSN: 1751-9659; 1751-9667
• DOI: 10.1049/iet-ipr.2010.0489

This study proposes, fast super-resolution algorithms to up-scale an input low-resolution image into a high-resolution image. Conventional learning-based, super-resolution algorithms require large memory space to store a huge amount of synthesis information, and they require significant computation because of the large number of two-dimensional matching operations. To mitigate this problem, the authors train a dictionary using one-dimensional, patch-based training and K-means clustering at the learning phase, and they use one-dimensional matching and interpolation, based on the trained dictionary at the synthesis phase. Such one-dimensional, content-adaptive interpolation is applied separately in horizontal and vertical directions. In addition, the authors propose a hybrid algorithm, in which directional interpolation is utilised for vertical interpolation to further reduce the dictionary size and the so called staircase artefact. Simulation results show, that the proposed algorithm has higher peak-to-peak, signal-to-noise ratio and structure similarity values, while providing significantly smaller dictionary size and faster computation than the latest learning-based, super-resolution algorithm.