Infrared and visible image fusion based on discrete nonseparable shearlet transform and convolutional sparse representation

In order to overcome the shortcomings of traditional infrared and visible image fusion methods, a fusion method based on Discrete Nonseparable Shearlet Transform (DNST) and Convolution Sparse Representation (CSR) is proposed. Firstly, the source images are decomposed into approximate images and directional detail images using DNST. Compared with other multi-scale decomposition tools, DNST can better separate the overlapped important feature information in different scales. Secondly, the salient feature maps of the source images are employed to weight average approximate images, which can prevent the loss of the brightness and energy. CSR can deeply extract the salient features of the image, The l1 norm of multi-dimensional coefficients is used as the activity level measure to construct the Salient Feature Map (SFM), which can generate the weight distribution decision map of the approximate image. The rule of Coefficient absolute max-Gaussian filtering is used as fusion rule of the directional detail images. The decision map of initial weight distribution is obtained by the Coefficient absolute max rule, then the decision map is filtered by Gaussian filter to reduce the sensitivity of noise and increase the proportions of visible image information. Finally, the fused coefficients are reconstructed by the inverse DNST, and the fusion image is obtained. Experimental results demonstrate that the proposed fusion method can achieve superior performance compared with other typical fusion methods in the existing literature in both subjective vision and objective criteria evaluation. © 2021, Jilin University Press. All right reserved.