ANFIS Interpolation and Its Application for Image Super Resolution

Abstract

Fuzzy inference systems work through manipulating fuzzy if-then rules. Amongst the various existing fuzzy inference systems, the adaptive network based fuzzy inference system (ANFIS) has become one of the most powerful and popular tools for finding solutions to highly non-linear problems. Whilst promising for practical applications, most existing research related to ANFIS focuses on how to learn such an inference system with sufficient training data. However, in some real situations it is very hard or even impossible to get sufficient data for the required learning process. The shortage of training data significantly degrades the performance of such learned ANFIS models. In light of this, the concept of ANFIS interpolation is proposed in the thesis, to deal with the problem of ANFIS construction with insufficient training data. It works by interpolating two well trained ANFISs in neighbouring domains where sufficient training data is available, in an effort to improve the performance of the constructed ANFIS in the current problem domain where there is only limited (or sparse) training data. Two types of ANFIS interpolation method are developed, including: 1) An initial method via group based rule interpolation, working through interpolating a group of fuzzy rules with the help of a rule dictionary; 2) An improved method via evolutionary process, in which the interpolated rules act as an initial population and are updated subsequently through the use of a genetic algorithm. Experimental results demonstrate the significantly improved performance of the proposed methods over the original non-interpolation ANFIS model, with the averaged root mean squared error (RMSE) reduced to 1.63 ± 0.39 from the original 3.71 ± 0.98. Having recognised the capability of ANFIS in producing effective non-linear mappings, it is applied to address the problem of image super resolution in this thesis, in an effort to learn the non-linear mapping between a low resolution image and a high resolution one. Firstly, a natural image super resolution method with full training data is developed by learning multiple ANFIS mappings, showing the effectiveness of ANFIS model when sufficient training data is provided. A hyperspectral image super resolution method with sparse training data through the use of the proposed ANFIS interpolation methods is then developed, validating the efficacy of the interpolated ANFIS model when only sparse training data is available.

Date of Award	2021
Original language	English
Awarding Institution	Aberystwyth University
Supervisor	Changjing Shang (Supervisor) & Qiang Shen (Supervisor)