Fast Multi-focus Image Fusion Based On Region Division

The principle of multi-focus image fusion is to fuse the focus pixels in multiple images recorded by multi-sensors with different focus depths in a fixed scene into one full-focus fusion image.The research of image fusion has always been an attractive scientific discussion.This thesis makes an in-depth research on this technology and proposes corresponding improvement methods.1.This thesis proposes a multi-focus image fusion algorithm based on super pixel-level convolutional neural network(sp-CNN).In this method,multi-scale super pixel segmentation is firstly applied to the source image to obtain the super pixels.Secondly,the sp-CNN is proposed to acquire the initial decision maps.Thirdly,according to the similarities and differences of the multiple initial decision maps,the uncertain region is reclassified by spatial frequency to obtain the phase decision map.At last,the final decision map is achieved to fuse the source images by post-processing the phase decision graph with morphology.Experimental results show that the proposed method achieves the goal of reducing time complexity and attain better fusion effect compared with the state-of-the-art fusion methods which utilize overlapping blocks.2.Under the Joint Photographic Experts Group(JPEG)framework,this thesis proposes a novel focus measure(FM)in Discrete Cosine Transform(DCT)domain for fast multi-focus image fusion.Firstly,the DCT coefficients of multi-focus source images in JPEG format are read directly.Then,a FM only using the low-order DCT coefficients is proposed.According to the proposed FM,the initial decision map is constructed.Finally,the final decision map is refined with morphological operation,and the fused image is achieved by a simple fusion rule.According to the analysis of experimental results,the method proposed in this thesis obviously decrease the storage capacity and computational complexity of the fusion algorithm,and at the same time still has competitive or even better fusion results in terms of visual perception and objective evaluations.Therefore,the method in this thesis is very suitable for real-time application in wireless vision sensor network(WVSN).