Multi-Task Learning with Additive U-Net for Image Denoising and Classification

Researchers have developed a new neural network architecture called Additive U-Net (AddUNet) for image denoising and multi-task learning, as detailed in their paper published on arXiv on February 26, 2026. The innovation involves replacing traditional concatenative skip connections with gated additive fusion in U-Net architectures, which constrains shortcut capacity while maintaining consistent feature dimensionality throughout the network depth. This structural modification aims to improve information flow between encoder and decoder sections and stabilize joint optimization processes for multiple tasks. The research demonstrates how this architectural approach addresses fundamental challenges in multi-task learning scenarios where networks must balance competing objectives. The Additive U-Net represents a significant advancement in computer vision, particularly for applications requiring simultaneous image denoising and classification capabilities. By implementing additive skip fusion instead of concatenative skips, the researchers have created a more efficient architecture that preserves fixed feature dimensionality across network depth. This approach provides structural regularization that helps control the information flow between the encoder and decoder sections of the network. The paper documents extensive testing across single-task and multi-task scenarios, showing promising results for image denoising applications while maintaining performance on classification tasks. This innovation could have significant implications for medical imaging, autonomous vehicles, and other fields where high-quality image processing is essential.