Dual-Branch Network with Adaptive Rational Nonlinear Function for Image Deblurring

In the field of deep learning-based image deblurring, while the single transformer architecture possesses strong global modeling capabilities, it still exhibits certain limitations in local feature extraction. To further improve the effectiveness of blind image deblurring, we present ARFF-Transformer-CNN Network (ATCNet) which integrates the local feature extraction capabilities of CNN into existing transformers, proposing a dual-branch network architecture that combines CNN and transformer. Specifically, the CNN branch effectively extracts local image features such as edges and textures through multi-layer convolutional operations, while the Transformer branch captures global image information like long-range dependencies and global contextual information through self-attention mechanisms. We also propose the adaptive rational activation feed forward(ARFF) module as the feedforward layer in the transformer. This module is a feedforward network component that combines the learnable activation functions of KAN networks with depthwise separable convolutions, aiming to enhance the nonlinear representation capability of features. Experimental results demonstrate that this dual-branch network architecture achieves excellent deblurring performance across multiple datasets, significantly improving image clarity and detail representation.