Feature Extraction and Selection Supported ANFIS Interpolation and its Applications

Abstract

Adaptive network-based fuzzy inference system (ANFIS) is a general estimator and is widely employed in various applications to solve non-linear problems, such as in single image super resolution (SISR) techniques. It exploits features of the fuzzy rule-based inference system together with the neural network’s strength of optimised weights in the intermediate layers. However, most of the existing research on ANFIS is focused on applications where sufficient training data is available. In contrast, in certain real-world situations, it is difficult to obtain sufficient data to undertake this necessary training process. Due to various budget or hardware constraints, it is challenging to always have sufficient training image data. Such a shortage of training data results in the degradation in performance of the poorly trained ANFIS models. Subsequently, such an SISR technique also requires a lot of computational time and resource. In an effort to solve such problems, this thesis proposes two types of ANFIS interpolation approaches. In the first instance, a set of given image training data is split into various subsets of sufficient and insufficient (sparse) training data subsets. A typical ANFIS training process is applied for those subsets that involve sufficient data, whereas ANFIS interpolation is employed for the rest that contains only sparse data. The current literature does not contain approaches which can accomplish this. Consequently, the proposed sparse data-based approach is compared against the state-of-the-art sufficient data-based SISR methods. In the second instance, a pre-processing phase containing highly efficient feature extraction and selection algorithms (exploiting evolutionary and fuzzy-rough feature selection algorithms) is included to the existing algorithm, whilst aiming to reduce the computational time and resources for existing SISR ANFIS applications. Experimental results demonstrate the significantly improved performance of the proposed approach compared to standard ANFIS model, with the average PSNR increased from 33.22 to 35.12 dB with the Set5 SISR benchmark datasets. Moreover, Martian SISR method with sparse training data is then developed with the proposed approach. The overall results indicates its average PSNR increased from 36.69 to 39.81 dB, hence its efficacy is proved under the determined constraints.

Date of Award	2023
Original language	English
Awarding Institution	Aberystwyth University
Supervisor	Qiang Shen (Supervisor) & Changjing Shang (Supervisor)