A Research Of The Hyperspectral Image Fusion Algorithms

Fusion technology is an effective way to reduce the difficulty caused by the huge data and strong relativity of hyperspectral image in image processing. Through research of the remote sensing image we can get more complete, objective, and essential cognition for the target. Because of the especial characters of the hyperspectral image, the way and aim in hyperspectral image fusion are both different to other image fusions. Based on the hyperspectral image processing, we discuss some traditional and new fusion algorithm. And based on the multi resolution wavelet composition, this paper proposed a new fusion rule which use pulse-coupled neural networks (PCNN) on high frequency coefficients selected.This thesis mainly talks about three parts:This paper introduced what is remote sensing image, the conception of remote sensing image fusion and the development situation in the national and inter-national. Besides the aim of the fusion and the image three fusion levels are also presented. In the paper, the character of the experiment image data and how to estimate the quality of the fusion image are both important and have discussion in this paper. In order to emphasize the intention of the fusion, compared to use classic statistics speciality to evaluate fusion condition, in this paper, also adopted classification accuracy way to evaluate the fusion effects.This paper showed some fusion algorithms which developed from image enhancement algorithms as follows: highpass filtering fusion method, HIS transform, and principle component analysis transform. Besides, take more consideration about wavelet transform in this paper, the first and second generation wavelet transform are both discussed in the paper. After divided the whole image data space into several subspaces by a way named adaptive subspace decomposition technology, to compose and decomposed each band image in the subspace level. In the low frequency coefficients levels, we use average value weights or variance weights. Each level of different high frequency coefficients chooses local deviation or local power criterions to fusion. After use such different fusion criterion, different fusion results can be achieved. So can evaluate these fusion images, use values of the evaluation can judge the effects of these fusion algorithms.On the base of the decomposition of hyperspectral image by the second generation wavelet algorithm which predicts and updates datasets on rectangle and quincunx grids, use a new fusion algorithm which introduce a new neural network named pulse-coupled neural networks (PCNN) to select the high frequency coefficients to achieve fusion aim. Through reasonable parameters enactment, the final results shows it's advantage compare the algorithms which use local deviation or local power criterions to select high frequency coefficients. So, the new algorithm presented in the paper can makes best use of the information in remote sensing images to be fused and prevents the loss of image information.