三尺度分解和稀疏表示的红外和可见光图像融合

Three-Scale Deconstruction and Sparse Representation of Infrared and Visible Image Fusion

为提高对含噪声源图像的处理效果,以及提高融合图像的对比度、结构信息等,提出一种基于三尺度分解和稀疏表示的红外和可见光图像融合算法。首先,为增强对噪声的去除效果,并且维持源图像的结构和边缘特性,利用滚动引导滤波器对源图像进行分解,将源图像分解为基础层和细节层;其次,为充分利用基础分量中的细节和能量,同时降低模型的复杂性,构造结构-纹理分解模型,将基础层再次分解为基础结构层和基础纹理层。然后分析三个分量的不同特点,使用不同的融合规则对三个分量分别进行融合,针对细节分量,其含有主要的噪声成分,但含噪声程度又不一样,因此根据图像含噪声的程度自适应确定稀疏融合去噪参数,从而同时实现对细节分量的融合和去噪,并且能够有效地提高计算效率;针对基础结构分量,其包含的细节特征较少,直接采用基于视觉显著图的加权平均技术进行预融合;针对基础纹理分量,由于其包含了视觉上重要的信息或图像特征,如边缘、直线和轮廓等活动信息,能够反映出原始基础图像的主要细节,因此采用主成分分析方法进行预融合,最终通过重构细节、基础结构和基础纹理层来得到融合图像。为验证所提方法的有效性,选取了多组红外和可见光图像进行试验,并与近期的五种方法CNN、FPDE、ResNet、IFEVIP、TIF进行了对比,采用主观和客观的形式对结果进行分析。实验结果表明,同其他图像融合算法进行对比分析,该方法能够兼顾含噪声和无噪声图像的融合,在有无噪声的情况下均能够将源图像的细节、亮度和结构保留到融合图像中,而且能有效地消除噪声。

To improve the processing performance of source images with noise and improve the contrast and structural information of the fused images,an infrared and visible image fusion algorithm based on three-scale decomposition and sparse representation is proposed.First,to enhance the noise removal effect and maintain the structure and edge characteristics of the source image,the rolling guidance filter is used to decompose the source image,and the source image is decomposed into the base layer and the detail layer.Secondly,in order to make full use of the details and energy in the base component and reduce the complexity of the model,a structure-texture decomposition model is constructed.The base layer is decomposed into the base structure layer and base texture layer again,Then by analyzing the different characteristics of the three components,different fusion rules are utilized to fuse the three components respectively.The detail component contains the main noise components,but the noise level is different.Therefore,the sparse fusion denoising parameter is adaptively determined according to the noise level of the image to realize the fusion and denoising of the detail component at the same time and can effectively improve computational efficiency;for the base structure components,which contain fewer detail features,the weighted average technology based on the visual saliency map is directly used for pre-fusion;for the base texture components,because they contain visually important information or image features,such as edges,lines and contours and other activity information,which can reflect the main details of the original base image,the principal component analysis method is used for pre-fusion.Finally,the fused image is obtained by reconstructing the detail,base structure and texture layers.In order to verify the effectiveness of the proposed method,several groups of infrared and visible images were selected for experiments and compared with five recent methods,including CNN,FPDE,ResNet,IFEVIP and TIF,and the results were analyzed in a subjective and objective form.The experimental results show that,compared with other image fusion algorithms,this method can consider the fusion of noise-perturbed and noise-free images and can retain the details,brightness and structure of the source image in the fusion image with or without noise.Moreover,can effectively eliminate noise.