Research On Image Dehazing Algorithm Based On Deep Learning

In hazy weather,there are a large number of dust particles in the atmosphere.Light will be absorbed and scattered by these particles in the transmission process,which makes the image received by imaging equipment appear low contrast,gray color,poor imaging quality and other problems.Therefore,it is of great research significance and application value to remove the interference brought by haze from degraded images,recover the details and features of original images as much as possible,and improve the contrast of images.The current mainstream image defogging technologies mainly include image enhancement,physical model and deep learning.The fog removal method based on image enhancement has no formation principle for fog,which will cause local excessive enhancement,color distortion and other problems.The defogging mode based on physical model relies on prior information.Due to the complex causes of haze in real scenes,this defogging mode is difficult to accurately estimate the parameters of the physical model of atmospheric scattering,which will further affect the defogging effect of images.Compared with the above two methods,deep learning has stronger generalization ability.Therefore,this thesis studies image defogging algorithm based on deep learning,and the specific research work is as follows:An encoder-decoder based image defogging algorithm is proposed for aiming at the problems of parameter estimation in atmospheric scattering model.The algorithm is independent of the atmospheric scattering model and can directly restore the fogged image to the non-fogged image from end to end.The network architecture based on U-Net reduces the information loss caused by the network during training and enables the neural network to focus on the more important information by combining the intensive connection and attention mechanism.By introducing the enhancement module,the image is reconstructed on four different scales to realize the enhancement of the image details.Finally,a comparative experiment is conducted between the synthetic image with fog and the real image with fog.The results show that the algorithm proposed in this paper achieves better fog removal on the public image data set,and the restored image is more detailed.Aiming at the difficulty of collecting paired image data sets with and without fog in the same scene,a Cycle GAN algorithm based on attention mechanism is proposed.By adding residual attention module into the generator,the network can better deal with the situation of uneven fog distribution.The residual attention module is mainly composed of channel attention module and spatial attention module,which can adaptively assign weights to different channels and spatial pixels,so that the neural network can pay attention to more important pixels and feature channels,thus improving the learning ability of the network.The discriminator is a combination of global discriminator and local discriminator,so that the generator can obtain features of different scales in the process of training,so as to enhance the generation ability of the model.The color loss function is introduced to reduce the color distortion.Finally,a comparative experiment is conducted between the synthetic image and the real image.The results show that the algorithm proposed in this paper not only has better defogging effect on the public image data set,but also enhances the image details.