Single Image Dehazing Based On Deep Learning

In the weather with thick haze,the captured images will be degraded due to the light scattering caused by the atmosphere particles such as haze,fog and mist.It not only affects the image quality,but also seriously affects the processing of outdoor visual processing systems such as monitoring field,military surveying,and automatic driving.Therefore,it is necessary to dehazing to get high quality image to ensure that outdoor visual processing systems can work normally in the haze weather.This thesis carry out in-depth research on image dehazing using deep learning network.Our main study contents are summarized as follows.(1)A dehazing algorithm based on multi-scale convolutional network is proposed,for the sake of solving the problem that the features related haze density extracted by traditional dehazing algorithm such as dark channel prior are only applicable under certain circumstances and can describe the haze density of complex scenes.Firstly,uses the front part of network to automatically extract the features relevant to the haze density.Secondly,uses convolution kernels of different scales to fuse these features to achieve the effect of feature fusion in the traditional dehazing algorithm.Then uses the mean pooling and maximum pooling to retain the high frequency information in the process of transmission prediction.Finally,uses two fully connected layers to learn the nonlinear mapping relationship between the extracted features and the transmission.The experimental results show that the extracted features are closely related to their haze density,and the dehazing results are better than other existing algorithms on both synthetic images and natural images.(2)An Inception Dehazing Network with Self-Attention is proposed.There are three unknowns of transmission,atmospheric,and haze-free image in the process of image dehazing,but most algorithms only optimize the transmission or the haze-free image ignoring the internal relationship between these three unknown,resulting in bad dehazing result.In terms of this issue,this method simultaneously optimizes the transmission,atmospheric,and hazefree image.Firstly,uses two inception networks with self-attention to estimate the transmission and the atmospheric respectively where Inception module can automatically select the most suitable convolution kernel size so that the estimated transmission rate can maintain local consistency better and the introduced self-attention mechanism enhances the strong correlation region with the target transmission or atmospheric to make the two networks to learn their target more easily.Secondly,bring the predicted transmittance and atmospheric into the physical model to obtain the dehazing image.Finally,use the joint discriminator to judge the true or false of the sample pair consisting of the dehazed image and the estimated transmission to achieve further optimization for both of them.The experimental results show that this algorithm achieves good dehazing result on both synthetic images and natural images.(3)A dehazing GAN based on the constraints of the physical model is proposed.Most dehazing algorithms perform dehazing through the image degradation model,but it is theoretical only,and may deviate from the real physical model.In terms of this issue,this method aims to generate a more realistic model than the image degradation model.Firstly,through the generating network,the dehazed image is obtained by mapping the input of haze image.Secondly,bring this dehazed image into the physical model to obtain the generated hazy image.Then,use the discriminant network to judge the true and false of this generated hazy image to achieve the guiding role of the generation network training,so that the generation network becomes a more realistic model than the image degradation model through the guiding role of this discriminant network.Moreover,during the training process,the introduction of the weighted mean square error to alleviate the over-enhancement of the dehazing image.The experimental results show that this algorithm achieves good dehazing result on both synthetic images and natural images,and the haze is removed in both the regions with dense haze and the regions with few haze and the dehazed image looks natural.