Single Image Dehazing Method Based on Semi-Training Color Stripping

Abstract

In this paper, a semi training color stripping dehaze-net (STCSDN) is proposed which is a new method of dehazing with strong adaptability. It mainly depends on two important properties of convolution neural network in the process of dehazing. One is that the learning speed of convolutional neural network for contour and shadow information is faster than that for color information. The other is that network is not sensitive to haze concentration. Based on property one, STCSDN extracts the haze free gray image from the hazy color image by using the semi trained generator as the sketch extraction module through the CycleGAN. The gray image only contains the image contour and shadow information, and discards the original color information interfered by haze information. This method has strong adaptability, visibility and authenticity, and can be applied to any scene. Based on property two, STCSDN can process hazy images with different concentrations and get better results. Through the simulation experiments on different types of dehazing data sets, it is proved that the STCSDN proposed in this paper can remove the influence of haze, restore the image details and enhance the visual effect.