Research On Single Image Dehazing Algorithm Based On Deep Neural Network

The appearance of fog and haze will not only affect the presentation of outdoor scenery,but also reduce the sharpness of outdoor shooting and cause image color distortion.Therefore,it is necessary to find efficient dehazing technology to improve the sharpness of images.Hazy images often suffer from color distortion,blur and other visible visual quality degradation,affecting the performance of some advanced visual tasks.Therefore,single image dehazing has always been a challenging and significant problem.Convolutional neural network has been widely used in image dehazing task,but the limitations of convolutional operation limit the development of dehazing.Nowadays,Transformer has shown superior globality in the development of computer vision,and the location association does not grow with the deepening of the network.On the other hand,existing deep learning-based dehazing methods place more emphasis on dehazing and less on image color restoration.Therefore,many dehazing methods are often troubled by color oversaturation,insufficient color or color cast.Aiming at these problems,this thesis uses deep learning technology to carry out an in-depth study of single image dehazing methods.The main contents of this thesis are as follows:(1)We propose a progressive method incorporating color layers.It gradually recovers the image by repeatedly invoking an auxiliary network.The RGBA image information captured by the soft color segmentation is used as the input for the auxiliary learning.Specically,we first introduce the gated recurrent unit in the feature extraction module,which can effectively extract image features while preventing model overfitting.Next,local features are extracted in the residual learning module by combining the recurrent layer and residual blocks.Finally,the composite module integrates the features to produce a clean image with rich details.In addition,recursive computation is used at each stage to reduce network parameters while improving performance.Extensive experimental results demonstrate that this method outperforms existing methods in both quantitative and qualitative evaluations.(2)A two-stage dehazing method based on Transformer is proposed.Specfically,in the first stage,the encoder and decoder is improved to combine Transformer and convolutional neural network to achieve basic feature extraction.Each layer in the encoder only models the local relationship,continuously reduces the resolution of the feature map,and expands the receptive field.In addition,an inter-block supervision mechanism is added between encoder unit and decoder unit for refining the features and improving the efficiency of feature delivery by supervising their selection.In the second stage,the original resolution block is used to extract the local features,and performs feature fusion and interaction.In addition,to ensure the delivery of the real feature signals in the first stage and improve the transmission efficiency of the network,we add fusion attention mechanism between stages.It adds the residual image of the early input features to the images acquired in the first stageo be delivered to the next stage.Ablation experiments show that our two-stage network has significant benefits for image quality and visual effects.Experimental results on RESIDE,O-HAZE,and I-HAZE datasets show that our method is superior to advanced methods in dehazing effectiveness.