基于交替非负最小平方投影梯度NMF的TINST域图像融合

Fusion method based on alternating non-negative least squares using projected gradient for NMF in TINST domain

为了克服现有基于多尺度分解(MSD)的图像融合方法存在的某些缺陷,增强多模态图像的融合精度,提出了一种平移不变不可分离剪切波交换(TINST)结合交替非负最小平方投影梯度非负矩阵分解(NMF)的图像融合方法。首先利用TINST对源图像进行多尺度、多方向分解,得到一个低频子带和多个高频方向子带系数;然后将低频子带系数看作原始观测数据,采用交替非负最小平方投影梯度NMF算法进行融合,得到包含特征基的融合低频子带系数,将高频方向子带系数作为脉冲耦合神经网络(PCNN)的外部输入激励,经点火处理和判决选择算子选择,得到融合高频方向子带系数;最后应用逆TINST重构融合后的子带系数,得到融合图像。采用多组多模态图像进行融合实验,并对融合结果进行了客观评价。试验结果表明,本文提出的融合方法在主观和客观评价上均优于其他MSD融合方法。

For overcoming some defects of existing image fusion methods based on multi-scale decomposi- tion（MSD） and enhancing the accuracy of multi-modality images, an adaptive fusion method based on al- ternating non-negative least squares using projected gradient for non-negative matrix factorization （NMF） are proposed in translation invariance nonseparable shearlet transform （TINST） domain. First- ly,the multi-scale and multi-direction decomposition for source images is performed by TINST, and a low-frequency and some high-frequency subbands are obtained. Secondly,the low-frequency subband co- efficients are regarded as original observed data, which are imposed to iterative operation by alternating non-negative least squares using projected gradient for NMF algorithm to obtain fused low-frequency subband coefficients. High-frequency directional suband coefficients are used as external input excitation in pulse coupled neural networks （PCNN）,and after the fire processing and compare-selection operator computing, fused high-frequency directional suband coefficients are obtained. Finally,all the fused subba- nds are reconstructed to an image by inverse TINST. Some fusion experiments on several sets of different modality images are done, and objective performance assessments are implemented for fused images. The experimental results indicate that the proposed method outperforms a few existing typical fusion techniques based on multi-scale decomposition （MSD） in subjective and objective assessments.