Sparse Training Data-Based Hyperspectral Image Super Resolution Via ANFIS Interpolation

Hyperspectral image super resolution aims to improve the spatial resolution of given hyperspectral images, which has become a highly attractive topic in the field of image processing. Existing techniques typically focus on super resolution with sufficient training data. However, restricted by data acquisition conditions, certain hyperspectral images or band images are very different to obtain, resulted in insufficient training data. In order to solve this problem, a new hyperspectral image super resolution method is proposed in this paper in an effort to conduct the super resolution task over insufficient (sparse) training data, by applying the recently introduced ANFIS (Adaptive Network-based Fuzzy Inference System) interpolation method. Particularly, the training dataset is divided into several subsets. For the subsets with sufficient training data, the relevant ANFIS models are trained using standard ANFIS learning algorithm, while for the subsets with sparse training data, the corresponding ANFIS models are interpolated through the use of ANFIS interpolation. Experimental results indicate that compared with the methods using sufficient training data, the proposed method can achieve very similar result, showing its effectiveness for situations where only sparse training data is available.

I. Introduction

Hyper-Spectral Image (HSI) represents a breakthrough in the field of imaging. Different from traditional two dimensional images, a hyperspectral image is a three dimensional image cube containing both spatial information and spectral information. Owing to the high spectral resolution, HSIs are widely used in various real world applications [1]. However, due to the hardware and budget constraints, hyperspectral images are usually obtained with low spatial resolution. In order to improve the spatial resolution of a given low resolution HSI, and to provide high quality images for the subsequent high level tasks, hyperspectral image super resolution (HSI SR) techniques have been developed recently [2]. HSI SR is an extension of the traditional 2D natural image super resolution. Thus, many techniques designed for natural images can be exploited in HSI SR, by considering an HSI as a set of two dimensional band images [3], with the band images enlarged either independently or collaboratively. Amongst the various HSI super resolution algorithms, the learning based ones are the most commonly adopted. These methods work by learning appropriate representations to describe the relationship between a low resolution (LR) image and the corresponding high resolution (HR) image, via mechanisms such as learning dictionaries [4], linear/non-linear mappings [5] and convolutional neural networks [6]. Fuzzy rule-based systems have also been designed to make interpretable descriptions of such LR-HR relationships [7], [8], which can produce not only mappings with explicit physical meanings, but also a transparent structure so that it would be much easier to analyse and adjust the resulting model and SR outcomes.