Multifocus Image Fusion Method Based On Multi-scale Transformation Studies

Multi-focus image fusion technique can solve the problem that not all the targets in the same image are clear in case of imaging in the same scene with the same optical sensor due to the differences of field depth. With the development and continuous improvement of multi-scale analysis tools, image fusion has been made rapid progress. Currently, the selection of high-frequency coefficients in the frequency domain is one of the most difficult and hottest issues in multi-focus image fusion. This dissertation focuses on new methods on image fusion in frequency domain via some advanced multi-scale analysis tools.The main contributions are summarized as follows:(1)Making full use of the features of NonsubSampled Contourlet Transform(NSCT) such as anisotropism and translational invariance, an Artificial Bee Colony(ABC) based method on image fusion is proposed which may provides with an optimal over-all performance. Firstly, several objective evaluation standards on image fusion are selected and an index to compute an over-all property of fused image is constructed. Secondly, after the original image is decomposed by NSCT, the coefficients of low-frequency and high-frequency in NSCT domain are obtained. Then, according to the features of NSCT and regional energies, swarm intelligence of ABC is introduced to determine the proportion of high frequency coefficients, i.e. the optimal weights. Finally, high-frequency coefficients and the average of low-frequency coefficients are used to reconstruct the best fused image. Experimental results show that the method is superior to several fusion methods like principal component analysis, pyramid decomposition method, wavelet transform based method etc.,(2) An adaptive multi-focus image fusion method is suggested based on structural similarities and local energies in Curvelet domain. In this method, the original image is decomposed with Curvelet transform and the coefficients of low-frequency and high-frequency in Curvelet domain are obtained. Then, the low-frequency coefficients are gained by averaging the two groups of low-frequency coefficients, while the high-frequency coefficients are decided by the structural similarity depending on their local energies adaptively. Finally the fused image is reconstructed by inverse Curvelet transform. Experimental results indicate the proposed method not only can efficiently integrate the information, but also is superior to some widely-used methods such as the principal component analysis method, Laplacian pyramid based methods and wavelet transform based methods, when fused images are compared in terms of entropy, standard deviation and clarity.(3)A Curvelet-domain method for multi-focus image fusion is presented based on gray relational degrees of structural similarity. In this method, source images are respectively decomposed by discrete Curvelet transform. Then, after high frequency coefficients are divided into small blocks, gray relational degrees of grey theory are used to compute the structural similarities among these blocks, and the fused high-frequency coefficients are selected based on the degrees of structural similarities, while the arithmetic mean method is used to fuse the low-frequency coefficients. Finally, a fused image is reconstructed with the fused coefficients by performing the inverse Curvelet transform. Experiment results show the proposed method is superior to about10widely-used methods including principal component analysis method, maximum or minimum grayscale method, pyramid based methods and wavelet transform based methods in terms of entropy, standard deviation and clarity.