Study On Multi-focus Image Fusion Based On Multi-scale Transform

Multi-focus image fusion means that the clear regions of source images areextracted and the corresponding clear regions are integrated into a fusion image inwhich every region is clear. Therefore, how to accurately extract the clear regionsdirectly decides the fusion quality of the multi-focus image. In general, studyingmulti-focus image fusion may be from three levels: pixel-level, feature-level anddecision level. In the thesis the multi-focus image fusion are studied by usingmulti-scale transforms, such as Q-Shift Dual-Tree Complex Wavelet andNonsubsampled Contourlet, based on the pixel-level image fusion and some imagefusion algorithms are proposed.Firstly, the features of wavelet transform are studied and Q-Shift Dual-TreeComplex Wavelet Transform(Q-Shift DT-CWT) is used for multi-focus image fusions.According to the principle of optical imaging, which is "In multi-focus image, thehigh frequency coefficients of clear image are far greater than it of fuzzy image andthe low frequency coefficients of clear image are not less than it of fuzzy image", andthe correlation between high and low frequency coefficients, the method of selectivitymaximum is used for image fusion. So the rule of local area standard deviationmaximum selectivity is used for the low frequency and the rule of module absolutevalue sum maximum selectivity is used for the high frequency. By the subjectiveobservation and the objective evaluation, the experimental results show that theproposed algorithm can effectively realize the multi-focus image fusion, well preservethe image detail information, avoid the artifacts caused by incorrect pixel selectivityin wavelet transform.Secondly, the contourlet transform which is able to realize image sparserepresentation is studied because wavelet transform is not the most sqarse imagerepresentation method and Nonsubsampled Contourlet Transform(NSCT) is used formulti-focus image fusions. According to the principle of NSCT, the rule of regionenergy maximum selectivity is used for the low frequency and the rule of absolutevalue sum maximum selectivity is used for the high frequency. From the differentfusion rules and the different multi-resolution analysis, it is proved that the proposedalgorighm is effective and feasible. By the subjective observation and the objectiveevaluation, the experimental results show that the proposed algorithm obtained betterfusion quality than the algorithm based on Q-Shift DT-CWT proposed in before. Thirdly, the color multi-focus image fusion is studied by Q-Shift DT-CWT andNSCT in IHS. After IHS transformation, the I, H and S are separated from the colorimage. The I component is corresponded to the gray image and it is irrelevant to thecolor information, so only the I component is fused with rules which are proposed inbefore. By the subjective observation and the performance parameters, such asEntropy, Average Gradient, Mutual Information and Root Mean Square Error, toevaluate the experiment, the results show that the studied algorithms not only reducedthe computational complexity, but also reduced the uncertain factors of the color biasor distortion.