RJB-Net: Residual Deep Learning with Joint Bilateral Denoising Network for Remote Sensing Image Fusion

Remote sensing image fusion amalgamate information from panchromatic and multispectral remote sensing images to generate an optimum representative image. In this paper, we have proposed a novel method for image fusion known as Residual Deep Learning with Joint Bilateral Denoising Network (RJB-Net). The RJB-Net procedure is initiated with panchromatic image denoising using a joint bilateral filter to suppress noise and retain edge information. RJB-Net model is trained at image patches covering all major spectral classes. The denoised panchromatic image along with multispectral data is input to the residual deep learning network for inference and generates a fused image. The remote sensing datasets used for RJB-Net training and evaluation are Indian Cartosat-1 panchromatic and Resourcesat-2 multispectral imagery. The experiment section compares the RJB-Net denoising output with prominent spatial domain image filtering methods both qualitatively and quantitatively. It is found that RJB-Net fused images inject the spatial information from panchromatic images in a controlled manner while retaining the spectra characteristics of multispectral data. The merging performance of RJB-Net is compared with state-of-the-art remote sensing image fusion techniques that range from classical to recent deep learning model. The proposed method RJB-Net generates superior results among all fusion methods at different feature targets and compared the performance both visually as well as with image quality metrics.