Cascade Spatial Feature Transformation and Blur Kernel Estimation for Super-Resolution

The research presents an innovative deep learning approach that employs a multi-stage convolutional network combined with deblurring mechanisms to achieve enhanced image super-resolution results. Compared to other CNN-based approaches, our method has two key advantages: (1) The cascade neural network retains its original structure even when the training set changes, reducing the occurrence of false textures during super-resolution reconstruction. Additionally, we introduce cascaded connections within residual networks to address feature mismatches. (2) The Spatial Feature Transform (SFT) layers are employed to handle multiple blur kernels, mitigating artifacts in image super-resolution. Consequently, our approach leverages blur kernel information to improve the robustness of convolutional networks while ensuring multi-level spatial feature trans-formations that preserve the original image characteristics. Finally, extensive experiments on both real and synthetic datasets validate the effectiveness of the proposed model, demonstrating superior performance in handling complex blur degradations.