Research And Implementation Of Overwater Iamge Dehazing Based On CycleGAN

In the water scene where the fog is frequent,the light is scattered due to the influence of water vapor,so that the contrast,color or other characteristics of the captured images are changed to form a foggy degraded image.Overwater hazy images have a limiting effect on application scenarios such as ship navigation,fishing boat fishing,and hull identification.Therefore,designing a suitable algorithm for overwater image dehazing has important scientific significance and practical value.However,the existing methods of image dehazing are designed for land scenes,and there are few research on the task of overwater image dehazing.At the same time,image quality assessment play an important role in promoting the development of image dehazing.At present,image quality assessment used in image removal can not evaluate the quality of dehazed images very well.Considering the above two points,the main work of this paper are summarized as follows:Firsly,in view of the shortcomings of existing image dehazing datasets,a new dataset called Hazy Water Dataset is proposed for overwater image dehazing task.This dataset includes large-scale real unpaired training data and two test sets for qualitative and quantitative evaluation,respectively.By comparison of 9 kinds of image removal algorithms such as DCP,it is found that the existing image dehazing methods which designed for land scene are not suitable for water surface scenes.Secondly,according to the unpaired data in Hazy Water Dataset,an overwater image dehazing method based on Cycle GAN is proposed,which named W-Dehaze GAN.WDehaze GAN uses two generators and two discriminators to convert from hazy images to haze-free images.The resize convolution used in the generator can effectively improve the quality of the generated images.The loss function includes adversarial Loss,cyclic consistency loss,and perceptual loss,where perceptual loss can constrain the quality of the generated images from semantic perspective.The experimental results show that the W-Dehaze GAN proposed in this paper is superior to other image dehazing algorithms in real images and synthetic images.Thirdly,aiming at the shortcomings of the dehazed image quality evaluation,a nonreference dehazed image quality assessment based on deep learning is proposed.The algorithm first uses the MSCN coefficient to normalize the brightness of the input image,then takes it into the convolutional layer of VGG-16 to extract features,and finally uses the full connection layer to regress to a quality evaluation value.The algorithm combines natural scene statistics with deep learning to implement image quality assessment.Experiments show that the proposed dehazed image quality assessment algorithm has high consistency with subjective scores and is superior to traditional quality assessment methods.In the end,an effective method is proposed for overwater image dehazing and the quality assessment of dehazed images,respectively,and a single image dehazing and evaluation system is designed and implemented.This system encapsulates the overwater image dehazing algorithm and image quality evaluation algorithm which proposed in this paper into a convenient and executable software.