A Robust Pansharpening Algorithm Based on Convolutional Sparse Coding for Spatial Enhancement

Abstract

Pansharpening (PS) is a prominent remote sensing image fusion technique. It yields high-resolution multispectral (HRMS) images, which are imperative for the applications, such as recognition and detection. The PS methods based on conventional sparse representation induce blurring effects and are unable to preserve the essential spatial details in the fused outcome. In this article, to overcome these drawbacks, a robust fusion scheme is proposed based on convolutional sparse coding (CSC). The source images are decomposed into its constituent texture and cartoon components. The sparse coefficient maps are acquired from texture components by adapting CSC. Texture components are fused using activity level measurement, whereas averaging mechanism is used to fuse the cartoon components. The HRMS image is reconstructed by combining the fused components in proportion to the gradient information. Impact of number of filters on quality metrics estimation is analyzed. Comprehensive experiments are performed on the images acquired from distinct sensors. The proposed method is evaluated in terms of visual analysis and the quantitative metrics with reduced-scale and full-scale experiments. Extensive evaluations manifest the capability of the proposed method of maintaining the balanced tradeoff and retaining the desired spatial and spectral details.