Image Defogging Algorithm Based On Convolutional Neural Network Applicable To Different Fog Concentration

The presence of large amounts of suspended particulate matter in the air can scatter light to form haze,resulting in poor visibility,low contrast and blurred image quality in images acquired from outdoor scenes,which seriously affects the visual effect.The enhancement of visible light images to improve the effect of hazy weather on image quality is of great research importance.With the development of deep learning theory,the defogging algorithms based on convolutional neural networks have obtained good defogging effect with their powerful feature extraction ability,but most of them are not universally applicable,and there exists the phenomenon of excessive defogging for thin fog images or incomplete defogging for dense fog images.Based on the study of defogging algorithms,two image defogging algorithms based on convolutional neural networks are proposed.Select some clear images from Middlebury Stereo Datasets and combine with self-built datasets to artificially add fog through an atmospheric scattering model.During the fogging process,set different atmospheric light values and scattering coefficients to form a thin fog image dataset and a thick fog image dataset respectively.A direct end-to-end multi-scale single image dehazing network based on the fusion of global and local features is proposed for thin fog images.The network is divided into global feature extraction module,multi-scale feature extraction module and deep fusion module.The global feature extraction module extracts global features describing the image edge contours;the multi-scale feature extraction module extracts image features at different scales to obtain richer structural properties;the deep fusion module extracts local features describing the image content through convolutional layers,and finally fuses local features with global features through skip connection to form an overall image defogging scheme.The proposed algorithm is simulated and the results show that the method has good effect on the processing of thin fog images,can effectively extract image features,and the reconstructed images have realistic and natural colours.Relative to the thin fog image,a neural network with more layers can better obtain the mapping relationship between the dense fog image and the original image,but the deeper network may lead to the problems of gradient disappearance,shallow layer feature loss and high computational complexity.The proposed algorithm is based on the U-Net structure to achieve compression of feature size with a larger number of network layers,reduce the number of model parameters.The convolutional layers are densely connected to form dense blocks,and the output features of the first dense block are fused with the output features of the last dense block,effectively solving the problem of feature loss and gradient disappearance.Multi-scale convolution is introduced into the network,which can effectively improve the model learning accuracy.Simulations of the proposed algorithm show that the algorithm can effectively solve the problem of incomplete defogging of dense fog images,and the reconstructed images retain more information of the original images.