Learning-based super-resolution for remote sensing images via adaptive reconstruction regularization

• Remote sensing image super-resolution algorithm is proposed based on online dictionary learning. • Reconstruction regularized parameter is adaptively solved by the U-curve algorithm. • High-frequency patches are separately reconstructed to obtain sharper edges and crispier textures. • Our super-resolution results are averagely 0.07 higher in PSNR and 0.0043 higher in SSIM. Dictionary learning has made a great breakthrough in single image super-resolution (SR) for remote sensing images. Most of these learning-based methods adopt the empirical training regularized parameter and reconstruction regularized parameter for all input images. This ignores the feature differences among the input low-resolution (LR) images and finally limits the SR performance. To address this problem, we propose an adaptive reconstruction regularization (ARR) algorithm. The reconstruction regularized parameter is adaptively solved by the U-curve algorithm corresponding to each LR remote sensing image, leading to the optimal reconstruction sparse coefficients. To enhance the edges and corners, the high-frequency patches are separately reconstructed after feature extraction. The final SR image is obtained by the superposition of the high-frequency part and the Bicubic up-scaling part. According to the experimental results, our proposed ARR can obtain sharper edges and crispier textures in the recovered SR images than several state-of-the-art methods, demonstrating superior SR performances.