Research Of Multi-sensor Image Fusion Algorithm

At present, with the rapid development of image sensor thechnology,multi-sensor image fusion also has a rapid development, and has become an importantbranch of information fusion. Image fusion can synthesize complementaryinformation of the multi-sensor image in the same scene, and remove redundantinformation. So it can get a more comprehensive and accurate scene image, therebyimprove the accuracy of the subsequent treating. This thesis has studied some existingclassical algorithms in the related fields and also has improved some relatedalgorithms. The main results are as follows:(1) A new feature-level visible image fusion algorithm is proposed. The sourceimage is divided directly into blocks, and the relevant characteristics of thecorresponding region that can reflect the capability of these images are extracted. Thedifference between these relevant regional characteristics for the correspondingregion from multi-sensor images is used as the input of support vector machine model,and then the regional images are classified based on SVM algorithm. Finlly the moreclearer image block in the corresponding region is adopted to reconstruct the fusedimage. Experiments results are given to show the merit of the proposed approach,compared to the traditional pixel-level fusion image effect.(2) In the process of classifying the above regional images, an improved PSOalgorithm using the adaptive inertia weight is provided to optimize the key parametersof SVM. Since the parameters of SVM have a larger impact on their learningperformance, and the system performance may be deteriorated if the unsuitable valuesof these parameters are selected, so this paper develops an adaptive inertia weightfactor in the PSO algorithm, in which the inertia factor is constructed by the numberof iterations, the local fitness and global fitness and then is benefical to search theglobal optimal parameters, and the fitness function is constructed by using theaccuracy of classification and the number of support vector machine. Experimentsresults are given to show the effectiveness of our method which may reduce thecomplexity of the model and improve the accuracy of classification, and thus improvethe integration effect of the above feature-level visible image fusion algorithm.(3) A fusion method for the visible and infrared images based on NonsubsampledContourlet Transform(NSCT) and Compressive Sensing(CS) is proposed. Firstly, the NSCT was preformed on the the visible and infrared source images, and thus the lowfrequency subband coefficient and varieties of directional bandpass subbandcoefficients can be obtained. Secondly, the low frequency subband coefficients isfused based on the ‘weighted averaging’ scheme, while the bandpass subbandcoefficients is fused based on standard deviation’s adaptive weighting scheme underthe compressive sensing theory framework. Finally, the fused image is obtained byperforming the inverse NSCT on the combined coefficients. Experiments results aregiven to show the effectiveness of image fusion because that the proposed method canpick up more information from source images.