Dual Flow: Joint Image-Measurement Interpolation for Inverse Problems
Dual Flow guarantees that its reconstructed image and denoised measurement agree under the underlying forward model.
The noisy observation initializes the measurement state; the model then updates image and measurement estimates together using their current values. Its network learns to solve an image-denoising problem with operator-dependent, correlated noise. The authors report state-of-the-art distortion and perceptual results across tested linear inverse problems, with shorter reconstruction times than competing methods. They say the only inference hyperparameter requiring adjustment is the number of steps.
HF Daily Papers' note
The noisy observation initializes the measurement state; the model then updates image and measurement estimates together using their current values. Its network learns to solve an image-denoising problem with operator-dependent, correlated noise. The authors report state-of-the-art distortion and perceptual results across tested linear inverse problems, with shorter reconstruction times than competing methods. They say the only inference hyperparameter requiring adjustment is the number of steps.
HF Daily Papers' note
score 4