Beyond Deep Residual Learning for Image Restoration: Persistent Homology-Guided Manifold Simplification

By Woong Bae, Jaejun Yoo, Jong Chul Ye. 2016

198 citations recorded by the source corpus.

The latest deep learning approaches perform better than the state-of-the-art signal processing approaches in various image restoration tasks. However, if an image contains many patterns and structures, the performance of these CNNs is still inferior. To address this issue, here we propose a novel feature space deep residual learning algorithm that outperforms the existing residual learning. The main idea is originated from the observation that the performance of a learning algorithm can be improved if the input and/or label manifolds can be made topologically simpler by an analytic mapping to a feature space. Our extensive numerical studies using denoising experiments and NTIRE single-image super-resolution (SISR) competition demonstrate that the proposed feature space residual learning outperforms the existing state-of-the-art approaches. Moreover, our algorithm was ranked high in the NTIRE competition with 5-10 times faster computational time compared to the top ranked teams.

Subfields: Computer Vision and Pattern Recognition, Media Technology

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Media Technology

Citations in the corpus

The Perception-Distortion Tradeoff (2017, background): “[2] (“primary” variant), SRResNet variants which optimize MSE and VGG2,2 [20], SRGAN variants which optimize MSE, VGG2,2, and VGG5,4, in addition to an adversarial loss [20], and ENet [34] (“PAT” variant).”
Deep Convolutional Framelets: A General Deep Learning Framework for Inverse Problems (2017, background): “Note that (91) is in fact the existing deep learning denoising algorithm [70, 9, 4], whereas the first iteration of (92) corresponds to the existing deep learning-based inpainting algorithm [70, 73].”
Deep Learning for Image Super-Resolution: A Survey (2019, methodology): “Bae et al. [116] ﬁrstly combine WT with deep learning based SR model, take sub-bands of interpolated LR wavelet as input and predict residuals of corresponding HR sub-bands.”
Neural Nearest Neighbors Networks (2018, methodology): “Note that we refrain from building our N 3 Net for SISR from more recent networks, e. g. MemNet [38], MDSR [25], or WDnCNN [3], since they are too costly to train.”
Real Image Denoising With Feature Attention (2019, methodology): “[10] employed persistent homology analysis [24] via wavelet transformed domain to learn the features in CNN denoising.”
Photorealistic Style Transfer via Wavelet Transforms (2019, background): “There have been several attempts to incorporate both classical signal processing and deep learning approaches, including feature reduction [16], network compression [11, 16], superresolution [2], classifications [3, 7, 23, 27, 30] and image denoising [14].”
NTIRE 2017 Challenge on Single Image Super-Resolution: Dataset and Study (2017, background): “[2] is the third winner of NTIRE 2017 challenge.”
Deep Learning on Image Denoising: An overview (2019, unknown): “(2017) [12] CNN Gaussian image denoisng, super-resolution CNN with wavelet domain, and residual learning (RL) for image restoration Jin et al.”
NTIRE 2017 Challenge on Single Image Super-Resolution: Methods and Results (2017, background): “The basic design and ideas are introduced in detail in [2].”