The Algorithm Research And Implementation Of Multi-band Image Fusion

Multi-band image fusion is to take a fusion process, according to a particularalgorithm, with different bands images which are generated by different sensors forthe same target. The processing results contain more comprehensive information thanany signal one, and can reflect the reality more accurately. Multi-band image fusion isan emerging discipline combines sensors, image and signal processing, parallelcomputing, artificial intelligence and so on. It is widely used in target detection,airport navigation, medical imaging and other fields, which has a great significancefor both the construction of national defense and the national economy.In this paper, I have made an in-depth study about the fusion and pre-processingalgorithms for multi-band images. This paper builds a fusion system for multi-bandremote sensing images, which includes the models of image enhancement, noisereduction and pixel and feature level image fusion. Also, the simulation verificationhas been made to detect the performance of the fusion system. In this paper, I take theenhancement process by using the improved Lee enhancement algorithm, while theresults of traditional histogram equalization algorithm and nonlinear transformationare compared. Combined the logarithmic transformation and the neighbor informationof pixels, the improved can eliminate the unbalanced illumination, adjust the overalldynamic range and significantly enhance the details such as edges and textures whichhelp in improving the recognizability of the image. For the salt and pepper noiseproblem exists in remote sensing images, this paper presents a fast and efficientde-noising algorithm. The algorithm increases the noise detection mechanism, anddistributes the filter template pixels with different weights, and final replaces thenoise point by the median filtering result. In this way, we can effectively remove saltand pepper noise in remote sensing image, and maintain the integrity of the texture atthe same time. For the Gaussian noise presented in remote sensing images, theK-SVD clustering training algorithm is used to construct the over-complete dictionaryof the source image. Through this algorithm, the effective information can beseparated from the Gaussian noise, and we can finally obtain the well de-nosingresults.This paper takes the feature-level fusion strategy in MATLAB, and uses theCanny and Sobel operators to extract the outline and the edge information after pretreatment and fuse the feature images based on the gradient pyramidtransformation algorithm. Meanwhile, we take the GeForce GTX650graphics as thedevelopment platform to utilize the high-speed parallel computing capability of theGUGPU. By writing CUDA programs and making a reasonable allocation ofcomputing unit in Visual Stdio10.0environment, we achieve the rapid real-timefusion processing.