Study On Image Fusion For Remote Sensing Based On GPU

Recently, there exist a number of satellites and many modern aerial digital cameras, which provide multispectral image together with panchromatic image. In this context, image fusion plays an important role since it effectively combines the spatial texture of panchromatic image and the spectral information of multispectral image. Even though the study of image fusion technique has been done for decades and numerous achievements have been developed, researchers always focused on color distortion and texture loss. With the development of earth observation technique, acquisition methods of remote sensing image data are numerous and data volume is also growing quickly. Nevertheless, image fusion algorithms are data-massive and computation-intensive, the processing performance cannot meet heavy tasks of current production if common serial processing techniques are adopted.In the field of photogrammetry and remote sensing, it has been a trend to achieve high performance based on the CPU-GPU co-processing platform. On the basis of the study of image fusion algorithms, this thesis makes a research on the GPU parallel processing scheme of image fusion algorithms to meet the rapid image fusion requirements of large image. It mainly includes the following:1. The research status of image processing in the field of remote sensing and the image fusion parallel algorithm are briefly summarized. The key technology of CUDA programming, methods of performance analysis and performance optimization are introduced. Aimed at image processing, several optimizing strategies are provided, the procedure of image processing is built in the CUDA environment.2. Methods of three commercial software are analyzed. After merging the panchromatic image and multispectral image of GaoFen-1, fusion results are compared visually. The methods of assessment and quantitative metrics are introduced in details. The domestic research status of fused image quality evaluation are thoroughly analyzed, then many problems are pointed out.3. Comparing the performance of three resampling methods in pansharpening process. Then the preprocessing of image fusion based on CUDA is realized by improving the resampling procedure. An edge enhancement algorithm with GPU acceleration which can be used in image fusion is proposed. This algorithm doesn't affect the performance, meanwhile, the edge information with certain weights is added to the original panchromatic image. The fused image with slightly sharper spatial detail can be produced.4. The parallel processing framework for large image fusion is designed. It realizes the CPU-GPU co-processing of image fusion by the use of GPU libraries. On the basis of analyzing HPF and HCS algorithms, both of them are processed by the decomposition of algorithm. Finally, it reaches the desired effect.5. Aimed at the problem of color distortion, the Block-Regression algorithm is improved by the method of histogram adjustment. The parallel processing procedure of improved algorithm based on GPU is put forward. The bottleneck of performance is optimized by using shared memory, which makes full use of the advantage of GPU. According to the test results of GaoFen-1 and Gao Fen-2 images, the performance of improved algorithm is analyzed. From the view of the fusion effect, the improved algorithm can effectively reduce the loss of color and texture. Comparing the CPU and GPU, the improved algorithm based on GPU can achieve a high speed-up ratio. Comparing with two commercial software, the result proves the superiority of image fusion based on GPU, which can satisfy the mass production of fused image.