Xnet: Single Image Deblurring Based on Multi-Scale Multi-Branch Feature Fusion

Single image deblurring is a critical task in the field of computer vision. The presence of blur in images can significantly degrade the accuracy and effectiveness of downstream tasks, including 3D modeling, target identification, instance segmentation, and related applications. The U-Net architecture, built on convolutional neural networks, has demonstrated exceptional performance in single-image deblurring tasks, establishing itself as a widely adopted framework for addressing image deblurring challenges. The traditional U-Net employs a single encoder branch and a single decoder branch to handle image deblurring tasks. This approach often results in the loss of some valuable image features, which play a crucial role in the deblurring process. To extract more image features from the input, we propose a network architecture with two encoder branches and two decoder branches. Additionally, to obtain more effective feature information from multi-scale inputs, we propose the multiscale feature extraction (MFE) module to handle these multi-scale inputs. Furthermore, we propose the EMA Feature Fusion (EMAFF) module to connect the outputs of the two encoder branches, and the Cross-stage symmetric feature fusion (CSFF) module to manage skip connections. Finally, we incorporate EMA attention module into both the encoder and decoder modules, which helps enhance the network's deblurring capability. Comprehensive evaluations conducted on the GoPro and Turbulence-Blur datasets reveal that the proposed network surpasses existing state-of-the-art approaches in terms of precision and performance.