基于CNN的非下采样剪切波域多聚焦图像融合-《郑州大学学报(工学版)》

文章信息/Info

Title:: Multi- focus Image Fusion Based on Convolution Neural Network in Non-sampled Shearlet Domain

作者:: 刘帅奇; 王洁; 安彦玲; 李子奇; 胡绍海; 王文峰; 1. 河北大学电子信息工程学院;2. 河北省机器视觉工程技术研究中心;3. 北京交通大学信息所;4. 中国科学院新疆生态与地理研究所数字图像处理实验室

Author(s):: Shuaiqi Liu1; 2; Wang Jie1; 2; An Yanling1; 2; Li Ziqi 1; 2; Hu Shaohai 3Wang Wenfeng 4; 1. School of Electronic Information Engineering, Hebei University; 2. Hebei Provincial Machine Vision Engineering Technology Research Center; 3. Information Institute of Beijing Jiaotong University; 4. Digital Image Processing Laboratory of Xinjiang Institute of Ecology and Geography, Chinese Academy of Sciences

Keywords:: image fusion; Multi-focus image fusion; non-subsampled shearlet transform; convolutional neural network; guide filter

摘要:: 结合非下采样剪切波变换的时频分离优良特性,提出了一种基于CNN的NSST变换域图像融合算法。首先，对源图像进行NSST分解，其次对分解的低频系数进行基于卷积神经网络（CNN）的融合策略，然后对分解的高频系数进行基于向导滤波的改进加权的拉普拉斯能量和（SMLD）模取大融合策略，然后将根据不同融合规则融合后的频率系数进行NSST反变化获取输出的清晰目标图像。本文所提出的图像融合算法不仅使融合后的图像充分保留了相应源图像的低频整体信息特征和高频细节信息特征，而且提高了最终融合图像空间信息特征的连续性。实验结果充分表明，该图像融合方法不仅可以获得更利于人眼接受的视觉效果图，且充分有效的提高了融合图像的客观性能评价指标。

Abstract:: In this paper, a new multi-focus image fusion algorithm is proposed based on convolution neural network in non-subsampled Shearlet (NSST) domain by using the advantages of time-frequency of NSST. Firstly, the source image is decomposed by NSST. Secondly, the fusion strategy based on the convolution neural network (CNN) is applied to the low frequency coefficients of the decomposition. Then, the improved weighted sum of Laplace energy based on the guided filtering are carried out to the high-frequency coefficients of the decomposition. Finally, the fused image can be gotten by inverse NSST transform. The algorithm fully preserves the information of the source image and improves the continuity of the image space. Experimental results show that the fusion algorithm can not only achieve better visual effects, but also improve its objective evaluation index.