Shulou(Shulou.com)06/02 Report--

This article shows you what the application paper of GAN in low-level vision is like, the content is concise and easy to understand, it can definitely brighten your eyes. I hope you can get something through the detailed introduction of this article.

1 [Image separation, rain / reflection / shadow, etc.] Deep Adversarial Decomposition: A Unified Framework for Separating Superimposed Images

Separating image layers from a single mixed image is an important but challenging task. In this paper, a unified framework "deep adversarial decomposition depth confrontation decomposition" is proposed for image separation. Methods dealing with the mixture of linear and nonlinear under confrontation training can achieve the best results in a variety of computer vision tasks (de-rain, de-reflection / shadow).

2 [Image Restoration] Learning Invariant Representation for Unsupervised Image Restoration

An image restoration method based on unsupervised learning is proposed in this paper. High-quality images with better details can be reconstructed by decoupling representation, adapting to the resistance domain and obtaining invariant representation from noise data, assisted by effective self-supervised constraints.

3 [defog] Domain Adaptation for Image Dehazing

In recent years, image de-fog based on learning method has achieved the most advanced performance. However, most methods train defog models on synthetic blurred images. Because of domain offset domain shift, these models are difficult to be extended to real blurred images. For this reason, a domain adaptation paradigm is proposed, which is composed of an image conversion module and two image defog modules. Specifically, firstly, the bidirectional conversion network is applied to bridge the gap between the synthetic domain and the real domain by converting the image from one domain to another. Then, the images before and after conversion are used to train the two image defog networks with consistency constraints.

4 [Image reconstruction, restoration, super-score] EventSR: From Asynchronous Events to Image Reconstruction, Restoration, and Super-Resolution via End-to-End Adversarial Learning

The event camera Event cameras has many advantages over the traditional camera, but when the intensity images is reconstructed from the event stream, the output is low resolution (LR), noisy and not realistic. For this reason, a new end-to-end pipeline is proposed, which can reconstruct the LR image from the event stream, enhance the image quality and upsample the enhanced image, which is called EventSR. The method is unsupervised and applies adversarial learning. About the lab video https://youtu.be/OShS_MwHecs

5 [attention mechanism, hyperscore] Learning Texture Transformer Network for Image Super-Resolution

Image Super Resolution (SR) restores realistic textures from low-resolution (LR) images, and migrates related textures to LR images by using high-resolution images as a reference (Ref). However, the existing SR methods ignore the use of attention mechanism to migrate high-resolution (HR) textures from Ref images. This paper proposes TT (Texture Transformer Network) SR, in which LR and Ref images are represented as queries and keywords in the converter, respectively. TTSR consists of four closely related modules, which are optimized for image generation tasks, including DNN learnable texture extractor, correlation embedding module, hard attention module for texture migration and soft attention module for texture synthesis. This design encourages joint feature learning across LR and Ref images, in which deep feature correspondence can be found through attention, thus accurate texture feature information can be transmitted / migrated.

6 [self-supervision, super-score] PULSE: Self-Supervised Photo Upsampling via Latent Space Exploration of Generative Models

The main purpose of single-image super-resolution is to construct high-resolution (LR) images from corresponding low-resolution (HR) inputs. In the previous method, which is usually supervised, the training target usually measures the average distance in the pixel direction between the Super Resolution (SR) and the HR image. Optimizing such indicators usually leads to blurring. In this paper, a new super-resolution algorithm PULSE (sampling on photos through potential space exploration) is proposed, which accomplishes this task in a completely self-supervised way.

7 [decoupling representation, multimode image conversion, superdivision, repair] Nested Scale-Editing for Conditional Image Synthesis

An image synthesis method is proposed, which can provide hierarchical navigation in the potential code space. For small local or very low resolution images, the method always outperforms the latest techniques in generating sampled images closest to GT.

The above content is what the application paper of GAN in low-level vision is like. Have you learned the knowledge or skills? If you want to learn more skills or enrich your knowledge reserve, you are welcome to follow the industry information channel.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.