L0范数约束的全变分图像融合

Image fusion via L0-regularized total variation

针对图像融合中传统方法易引起图像细节和边缘丢失的问题,提出了一种L0范数约束的全变分(TV-L0)图像融合方法。对源图像进行均值滤波,根据滤波前后的图像获取源图像的边缘特征,通过特征图的梯度获取源图像对应的权重,基于此权重,由源图像建立目标图像并由源图像的梯度建立目标梯度。进一步基于目标图像建立变分方程的保真项,并基于目标梯度建立方程的正则项。通过引入变量,基于快速傅里叶变换,实现了融合模型的快速求解。实验结果表明:与基于离散余弦谐波小波变换图像融合方法相比,提出的方法多焦点图像融合和红外与可见光图像融合中的平均梯度(AG)分别提高了0. 71%和15. 71%,标准差(SD)分别提高了4. 97%和4. 77%,从源图像转移到融合图像的边缘信息量(QUV/F)分别提高了0. 94%和6. 13%,从源图像到融合图像丢失的边缘信息量(LUV/F)分别降低了5. 47%和17. 37%。提出的方法所获取的融合图像在主观视觉效果和客观评价函数两方面均优于传统方法。

Aiming at problem of loss of image detail and edge information in image fusion caused by conventional fusion method,a novel image fusion method via L0-regularized total variation is proposed. An average filtering is performed on source images to extract edge feature,and the weights of source images are calculated by gradient of feature map. Based on target image,and acts as the fidelity term which penalizes the departure from the initial image. A target gradient map is also made to regularize the gradient of fused image. By introducing auxiliary variables,an alternating minimization algorithm is used to get the fused image. Experiments are conducted by multi-focus images as well as infrared and visible images,results of fused images are computed by some quality metrics. Comparing with discrete cosine harmonic wavelet transform fusion method,average gradient is increased by0. 71 % and 15. 71 %,and standard deviation is increased by 4. 97 % and 4. 77 %,as well as QUV/Fis increased by 0. 94 % and 6. 13 %,respectively,the LUV/Fis decreased by 5. 47 % and 17. 37 %. It concludes that the proposed algorithm obtains remarkable results both in visual perception and objective metrics.