tvunnet

tvunnet is a term used in machine learning to refer to a class of neural network architectures that extend the U-Net family by incorporating total variation (TV) regularization to improve image reconstruction and denoising tasks. The term can denote several related designs rather than a single standardized model.

In these architectures, a U-Net backbone provides the encoder–decoder structure with skip connections, while TV regularization is integrated either as an explicit loss term during training or as a differentiable module within the network. Common approaches include adding a total variation loss to the objective function to promote piecewise-smooth outputs, or inserting TV-denoising blocks to suppress noise while preserving edges. Some variants also explore temporal TV for sequence data.

tvunnet variants are used in contexts such as medical imaging (MRI, CT reconstruction from undersampled data), rapid or low-dose imaging, and other image restoration tasks where edge preservation is important. They aim to combine the localization capability of U-Net with the edge-preserving effects of TV regularization.

The landscape is diverse and there is no single canonical tvunnet configuration. Practical considerations include tuning the weight of the TV term relative to data fidelity, computational overhead, and the availability of paired training data.

See also: U-Net, total variation, image denoising, medical image reconstruction, deep learning.