Research On Fusion Algorithm Of Remote Sensing Images

During the recent years, with the fast growth of computer processing capacity and the rapid development of spatial information science and sensor technology, remote sensing technology have been extend from the pilot phase to a multi-source, multi-channel and multi-level stage. Remote sensing satellites provide people with a variety of data which may from different areas, at different time and with different sensors. These data can have the coherence and complementarity and there is also redundancy. How to effectively use this data, to find useful information out of it, is now the focus of research work of remote sensing technology.Multi-source remote sensing image fusion combines the images of different scenes together to enhance the resolution of image, highlighting the objective characteristics, classifying accuracy and dynamically monitoring objects by using their coherence and complementarity. Multi-spectral and panchromatic image fusion is a now a focus of current research, and has a wide range of applications. The integration of these two can not only get a higher image resolution, but also able to better maintain the spectral information which provides a favorable precondition for the subsequent image classification, image segmentation, object recognition and tracking. This paper focuses on the multi-spectral and panchromatic images fusion, several new strategies are proposed based on previous fusion methods.The main contributions of this thesis are:(1) Based on the summary and analysis of existing research results, this thesis gives the three levels of multi-source remote sensing image fusion and respectively, for these three levels, group and induce their classic methods. For the Pixel-level fusion for remote sensing image, this thesis sums up the method for subjective and objective evaluation.(2) The thesis proposed new image fusion algorithm which based on the combination of wavelet transform and IHS transform. The spatial resolution of image can be well improved after the IHS transform fusion but with loss of spectrum; wavelet transform has good spatial and frequency domain locality, fused image can maintain more spectral information, but the detail of image is lost, together with the ringing effect. Therefore, the combination of IHS transform and wavelet transform results an improvement for both of the spatial and spectrum information.(3) A wavelet packet transform image fusion algorithm based on Soble operator is proposed. Mallat wavelet transform is the most commonly used wavelet transforming methods, but it only processes the output of the low-pass filter. However, the wavelet packet transform processes not only the output from the low-pass filer and also the higher one. The wavelet packet transform can gather more detailed information and have a stronger distinguish ability for time-frequency resolution than wavelet transform. Therefore, we use of wavelet packet transform combined with IHS transform. For the low-frequency part of wavelet packet decomposition, using energy-based approach, we focus on the extraction of contour information. For the high-frequency part, we introduce the Sobel operator to image feature extraction, focusing on improving the image details. Fused image has better spectral information, at the same time, the image details, edge characteristics are enhanced, and thereby image clarity is increased.(4) This thesis proposed a remote sensing image fusion algorithm based on non-sampling Contourlet (NSCT) transform and pulse coupled neural network (PCNN). Wavelet transform has good spatial and frequency domain localization features, but it can not accurately represent the edge direction information. Meanwhile, Contourlet transform has better multi-resolution, localized, directional, and the anisotropy . But during the transformation process, the need for image down-sampling operation, which does not have translational invariance, fused image will appear Gibbs phenomenon. NSCT transform not only have more flexible multi-resolution, multi-dimensional image representation ability, and solve the Contourlet Transform aliasing. With Pulse coupled neural network (PCNN) which has global coupling and pulse synchronization, you can retain more image details. Therefore, we combine of the PCNN and NSCT. Firstly, we use NSCT to decompose the source image. The resulting low-frequency coefficients will be fused by using the method based on local fusion energy. The high-frequency part will be fused by PCNN-based approach. Experimental results show that the method improves the spatial resolution to the max level and at the same time keep the spectrum information.