Core Concepts
The authors propose an unsupervised neural network approach, the TVflowNET, to efficiently approximate the solution of the total variation (TV) flow given an initial image and a time instance, without requiring ground truth data.
Abstract
The key highlights and insights of the content are:
The authors introduce a novel energy functional that, when minimized, yields the solution to the entire TV flow of an image up to a specific time. This allows learning a mapping from the product of the space of images and time to the space of images, rather than learning the solution for fixed initial and boundary conditions.
To avoid numerical instabilities related to the explicit form of the TV subgradient, the authors propose a loss functional that uses the pointwise characterization of the TV subgradient, allowing them to also learn the subgradient at any time.
The authors investigate the performance of three different TVflowNET architecture designs (Semi-ResNet, U-Net, Learned Gradient Descent) and four training regimes based on different image sizes. They demonstrate that the TVflowNET can successfully approximate the TV flow solution, retain TV flow properties like one-homogeneity, and enable efficient spectral TV decomposition.
The TVflowNET achieves a remarkable two orders of magnitude improvement in computation time compared to the classical model-driven approach, significantly enhancing the practicality and efficiency of TV flow solutions in real-world applications.
Stats
The TV flow solution can be approximated with high PSNR (up to 45.14 dB) and SSIM (up to 0.990) compared to the model-driven approach.
The learned diffusivity term φ can be approximated with PSNR up to 29.86 dB and SSIM up to 0.676.
Quotes
"The main difference between our approach and standard applications of PINNs is that our network learns the solution of the TV flow from an arbitrary initial image, as opposed to standard approaches that learn the solution of a PDE with fixed initial and boundary conditions."
"Notably, we achieve a remarkable two orders of magnitude improvement in computation time compared to the model-driven approach. This significant reduction in processing time enhances the practicality and efficiency of TV flow solutions via TVflowNET in real-world applications."