Multi-focus Image Fusion Based On Self Similarity And Compressed Sensing

Image fusion is a technique involving sensors, signal processing, image processing, computer vision and artificial intelligence. High performance fusion algorithm can reveal the interest tendency of mufti-source information, improve the information capacity of images and highlight features, thus providing the observer with a more accurate, comprehensive and reliable analysis. Based on the features of mufti-source image, current image fusion methods can mainly be classified into mufti-focus image fusion, fusion of visible image and infrared image, and fusion of multicultural image and panchromatic image.Recently the image fusion based on sparse representation has attracting increasing attentions from more and more researchers at home and abroad. Its basic idea is as follows: Firstly, the sparse representation coefficients of source images are calculated under an over-complete dictionary. Secondly, the sparse representation coefficients are fused by using some fusion rule. Finally, the over-complete dictionary is multiplied by the fused sparse coefficients to obtain the fusion image. Compared to traditional fusion methods in spatial domain and transform domain, it has properties on sparsity, feature maintained, and separability.As the mufti-focus image fusion method based on the compressed sensing theory has a bad performance in edge preservation, a new image fusion method is proposed in this paper. The method combines the compressed sensing theory with a scheme for determining the fusion weight which is realized by selecting image areas and pixels based on the self-similarity of images. The realization process is: firstly, decompose wavelets of the source mufti-focus image; secondly, observe the high-frequency components within a Gaussian matrix, and calculate the self-similarity of the observed quantity and the decomposed low-frequency coefficients, and then make weighted fusions; thirdly, reconstruct the fused high-frequency coefficients; and finally, obtain the fused image through the inverse wavelet transform of reconstructed high-frequency and low-frequency coefficients. Experimental results show that this method improves visual effects of the mufti-focus image fusion, and at the same sampling frequency, the edge preservation and information entropy of this method are higher than those of the domain gradient fusion, respectively with a percent of 27% and 24%, and the maximum absolute values fusion, respectively with a percent of 27% and 24%, which are based on the compressed sensing theory.