A Method to Create Training Dataset for Dehazing with Cyclegan

Haze usually blurs the characteristics of images shotted in adverse weather conditions. It brings many challenges for computer vision such as object detection. Due to the lack of effective training dataset of remote sensing image, the dehazing usually utilize the physical model rather than the deep learning approach based on a large number of training dataset. An effective method to create training dataset may provide some new ideas for remote sensing image dehazing. In this paper, a novel approach based on the visual features rather than the physical counterparts is developed to create a training dataset for remote sensing images dehazing. Firstly, 400 haze images and 400 clear images with size of 240×240 pixels were gathered from Landsat 8. Secondly, the dataset was utilized to train the cycle-consistent generative adversarial network (CycleGAN) to derive an image transform model which can convert a clear image into a haze one. Thirdly, 4000 clear images with size of 240×240 pixels were collected from Landsat 8 and the images were transformed into the haze images with our model. Finally, the ratio of training dataset and testing dataset is set as 4:1. The haze images created by us were selected as the input and the original clear images were chosen as the output to train the convolutional neural networks to derive the dehazing models. The created dataset and the dehazing models derived were tested, and the experimental results showed that our approach is better to keep the brightnessinformation of the original image, but it is not good at keeping the chroma information.