基于结构-纹理分解和VGG深层网络的红外与可见光图像融合

Infrared and Visible Image Fusion Based on Structure-Texture Decomposition and VGG Deep Networks

针对当前红外与可见光图像融合方法中存在低频信息未充分利用,以及高频细节易混入噪声等问题,提出一种基于结构-纹理(ST)分解和VGG深层网络的红外与可见光图像融合方法。首先用均值滤波将输入图像分解为高-低频子带,并引入ST对低频子带二次分解,采用平均法则与邻域空间频率分别对结构和纹理进行预融合;同时,将输入图像送入VGG网络得到多层特征映射,并用Sigmiod函数实现高频子带的归一化预融合;最后利用预融合的高频、低频结构和低频纹理3个频带进行图像融合重建。实验结果表明,提出的方法能够融合图像的深层细节特征,有效保留纹理细节并抑制噪声,且在噪声评估、结构相似性、均方误差、峰值信噪比等客观指标方面具有明显优势。

To address the problems of underutilization of low-frequency information and easy mixing of high-frequency details with noise in the current infrared and visible image fusion methods,an infrared and visible image fusion method based on structure-texture(ST) decomposition and VGG deep network is proposed.First,the input image is decomposed into high-low frequency subbands using mean filtering,and ST is introduced to re-decompose the low-frequency subbands.The structure and texture are pre-fused by absolute maximum and neighborhood spatial frequency,respectively.Subsequently,the input image is input into the VGG network to get the multi-layer feature maping,and the Sigmiod function is used to realize the normalized prefusion of the high-frequency subband.Finally,the pre-fused high-frequency,low-frequency structure,and low-frequency texture are used for image fusion and reconstruction.Experimental results show that the proposed algorithm can fuse the deep detail features of images,retain texture details,and suppress noise effectively,and has significant advantages in noise assessment,structural similarity index measure,mean square error,peak signal to noise ratio,and other objective indexes.