Image Super-Resolution via Dense Convolutional Neural Network in Curvelet Transform Domain

In recent years, Convolutional Neural Network (CNN)-based image super-resolution methods have made significant progress in reconstruction performance. However, they still suffer from issues such as blurry reconstructed images and loss of texture details. This paper proposes an image super-resolution method that combines the curvelet transform with a cascaded dense convolutional neural network, transforming the traditional spatial domain pixel prediction problem into a sub-band coefficient prediction problem in the curvelet transform domain. Through multi-scale geometric analysis, the curvelet transform can represent curved edges and texture details in images more sparsely. The proposed dense network contains multiple Residual Dense Blocks (RDBs), which fully utilize features at different levels through dense connections, thereby addressing the long-term dependency problem in deep networks. Experimental results on standard datasets such as Set5, Set14, and BSD100 demonstrate that the proposed method achieves optimal objective evaluation metrics (PSNR and SSIM) for 2 ×, 4 ×, and 8 × super-resolution tasks, while also producing more realistic texture details in subjective visual effects. Furthermore, evaluation on image classification tasks indicates that images reconstructed by this method are more beneficial for high-level vision tasks.