Abstract:In recent years, convolutional neural networks have been widely used in image defogging and achieved remarkable performance. However, most of the learning-based defogging methods using pairs of synthetic foggy and corresponding ground-truth images for training. Due to the gap between the distribution of real-word foggy and synthetic foggy images, there are limits to apply these defogging methods to autonomous driving in practice. CycleGAN is a two-way GANs network that could using unpaired real-word images to train image defogging models. However, there are several problems when CycleGAN is directly used for image defogging: (1) Using two different distributed data to train the same generator will confuse the learning of generators, which reduce the convergence speed and defogging results of the network. (2) The generator ignores the global features exploration which is very important for learning the scene and lighting information and needs the post-processing to restore the original image scales, which will result in a decrease in the quality of the generated image. To address these issues, we propose a Scale-free Heterogeneous CycleGAN (SH-CycleGAN) to utilize unpaired real-word images for boosting image defogging. The SH-CycleGAN contains a Heterogeneous Learning CycleGAN (HLCG) framework, and a generator with a Global Features Fusion module and an Adaptive Pooling module(GFFAP). Specifically, in the proposed HLCF framework, each BatchNorm layer learning in the generator is independent, which can solve the problem of confusion caused by learning different distributed data. Furthermore, the development of the GFFAP model can deal with image of any scales and improve the generated images. The experiments compared with eight state-of-the-art image defogging methods on both synthetic and real-world images demonstrate that our proposed method outperforms state-of-the-art performance and obtains more pleasing visually defogging results.

DefogGAN: Predicting Hidden Information in the StarCraft Fog of War with Generative Adversarial Nets

Attack Deterministic Conditional Image Generative Models for Diverse and Controllable Generation

No One Can Escape: A General Approach to Detect Tampered and Generated Image

Scale-free Heterogeneous Cyclegan for Defogging from a Single Image for Autonomous Driving in Fog.

Principled Knowledge Extrapolation with GANs.

Image Defogging Algorithm Based on Semi-Supervised Learning Neural Network

Single-Image Defogging Algorithm Based on Improved Cycle-Consistent Adversarial Network

StarCraft Adversary-Agent Challenge for Pursuit–evasion Game

Real-Time Strategy Game Win-Lose Prediction Model Using 3D Convolutional Neural Networks

Efficient Reinforcement Learning for StarCraft by Abstract Forward Models and Transfer Learning

GRIP-GAN: an Attack-Free Defense Through General Robust Inverse Perturbation

A Remote Sensing Image Defogging Method Based on Improved CycleGAN Network

FISS GAN: A Generative Adversarial Network for Foggy Image Semantic Segmentation

Defending against GAN-based Deepfake Attacks via Transformation-aware Adversarial Faces

Jointly Defending DeepFake Manipulation and Adversarial Attack Using Decoy Mechanism

Image dehazing algorithm based on improved generative adversarial network

Generative Adversarial Trainer: Defense to Adversarial Perturbations with GAN

GanDef: A GAN based Adversarial Training Defense for Neural Network Classifier

Adversary agent reinforcement learning for pursuit-evasion

Defense-GAN: Protecting Classifiers Against Adversarial Attacks Using Generative Models

GAN-Based Content Generation of Maps for Strategy Games