The Theory And Method Of Remote Sensing Image Fusion Based On Sparse Representation

In order to obtain a high reliability, a clearer fused image, take the fusiontechnology to integrate the complementary information and remove the redundantinformation. The fused imagery promotes the ability of human to describe the image. Itis also more conducive to the follow-up computer processing.This paper takes sparse representation theory as basis to exploratory study ofremote sensing image fusion. The main contents of this paper include:1. This paper applies sparse representation theory on remote sensing image fusion,and uses training dictionary to express images. The training dictionary has betteradaptability and flexibility, and also can well represented the remote sensing imageswhich have the complex features. The SFIM algorithm can reasonable inject the spatialdetails of high-resolution image into low-resolution image. This paper takes the SFIMalgorithm on the intensity component and panchromatic image. This paper chooses thespatial frequency maximum as the fusion rule to fusion the coefficients of sparserepresentation. This fusion rule takes full advantage of the local features, effectivelyenhanced the spatial details of the fused image. The experimental results prove that theproposed method improve the spatial resolution and better maintain the spectralcharacteristics.2. Due to the algorithm of the remote sensing image fusion based on trainingdictionary can not well maintain the spectral information. We analyze the advantages ofwavelet transform. The wavelet transform can preserve the spectral characteristics. Wecombine wavelet transform and sparse representation theory to fuse panchromatic imageand multi-spectral images. We analyze the characteristics of high-frequency andlow-frequency data, and then take the different fusion methods on the high and lowfrequency data which with different data characteristics. Due to the low frequency datais the approximation of the original image, the coefficients are not equal to0, can’t beseen as “sparse”. Hence, the low-frequency images obtained their coefficients by sparserepresentation which using the training dictionary. The high-frequency data reflects thesignificant information of the original images, most of its coefficients close or equal to zero, can be seen as “sparse”, could be directly fusion. This paper takes the imageinformation maximum as the fusion rule. The fusion rule can consider the correlation ofthe coefficient and retains the edges and details. Through multiple sets of experimentaland analysis, the algorithm can enhanced the spatial resolution of the fusion image andmaximize the retention of the spectral characteristics.