Research On Image Clearness Method In Hazy Scenario

In the hazy scenario,there are a large number of tiny water droplets suspended in the air,and light will scatter and refract when passing through these water droplets.Therefore,the image acquired by the imaging system in the hazy scenario will be degraded,which is usually manifested as the whole image is white,the contrast and color saturation of the image are reduced,resulting in the loss of some details of the image,which makes the image features difficult to recognize.It not only reduces the visual effect of the image,but also increases the difficulty of the subsequent processing of the image,resulting in a variety of outdoor monitoring systems that depend on the work of the imaging system,and the application of the detection system is greatly limited.Therefore,it is of great practical significance to study how to clear the blurred and hazy image obtained by the imaging system in hazy scenario,which is of great practical significance for the restoration of atmospheric degraded image and the enhancement of image visual effect,and has a great practical application prospect.Based on the analysis of the research status of image dehazing at home and abroad,this thesis makes an deeply study on the technology of image dehazing,which includes three aspects: the research of image clearness method based on sky segmentation and dark channel prior,the research of image clearness method based on multi-scale convolution neural network and the research of image clearness method based on conditional generative adversarial network.The main research and contributions are as follows:(1)In view of the problem that the original dark channel prior algorithm is easy to cause different degrees of sky color distortion and low overall brightness of the image when processing the hazy image with large area of sky,this thesis proposes a single image dehazing method based on sky segmentation and dark channel prior.When the original dark channel prior algorithm is used to calculate the transmission of the hazy image,it does not distinguish the sky region and the non-sky region of the image.However,because the dark channel value of the sky region is not close to zero,the estimated transmission of the sky region is not accurate.Therefore,this thesis adopts the idea of sky segmentation,uses the gaussian mixture model to model the hazy image,and uses the EM algorithm to optimize the model parameters.Then the hazy image is divided into sky region and non-sky region.According to the different concentration of haze in the sky region,the region is divided into light haze area,medium haze area and dense haze area,and the transmission is estimated again.Finally,the image dehazing is realized by combining with the atmospheric scattering model,and the local tone mapping is carried out to enhance the visual effect.The experimental results show that the method in this thesis can achieve good clearness for the hazy image with large area of sky,and there is no color distortion in the dehazed image,and the overall visual effect of the image is good.(2)In order to solve the problem that traditional image dehazing methods need to based on certain assumptions or prior conditions,and rely on the manual way to extract the relevant features of haze,and the effect of dehazing is not stable due to the limitation of dehazing scenario,this thesis proposes a single image dehazing method based on multi-scale convolutional neural network.In this method,firstly,a full convolutional neural network model is designed to learn the mapping relationship between the hazy image and its transmission by multi-scale feature extraction.Secondly,the atmospheric optical value is calculated according to the transmission obtained by the network model.Finally,the clearness of the hazy image is realized by combining with the atmospheric scattering model.The experimental results show that this method overcomes the limitation of the traditional dehaze method by prior information or assumptions and dehaze scenario,and achieves good image clearness.(3)In view of the fact that most of the dehazing methods depend on the atmospheric scattering model,it is necessary to estimate the transmission first,and then calculate the atmospheric optical value to realize the image dehazing.This process is easy to be affected by the intermediate parameters,which leads to the dehazing problem.In this thesis,a single image dehazing method based on conditional generative adversarial network is proposed.In this method,a conditional generative adversarial network for image dehazing is constructed.Through the training of mutual confrontation between the generator and the discriminator,the mapping relationship between the hazy image and the clear image is directly learned,so as to realize the image dehazing.The experimental results show that this method can effectively achieve image dehazing,and has good performance in image detail information preservation.