Research On Defogging Algorithm Based On Neural Network

The existence of haze not only affects people’s daily life,but also increases the difficulty for outdoor vision machines to obtain clear images,which is not conducive to the continuous development of subsequent machine vision tasks.Therefore,the problem of image dehazing has become an urgent problem to be solved in the field of machine vision,and the research on the direction of image dehazing is of great practical significance.Based on the atmospheric scattering model,the traditional dehazing algorithm uses prior knowledge and certain constraints to perform mathematical statistics on image features.However,due to the limitation of human statistical calculation,it may cause inaccurate parameter estimation,thus affecting the restored image.the quality of.In recent years,deep learning networks have made significant progress in the field of image processing,especially neural networks have powerful feature information extraction capabilities,which overcome the time-consuming and labor-intensive problems of traditional algorithms.However,due to the limitation of the training data set,the neural network has a low quality of image processing with complex fog structure in the real environment,such as local fog residue,color distortion,loss of edge information and other problems.In order to solve the above problems,this dissertation analyzes and verifies two existing classical algorithms,the dark channel dehazing algorithm and the gamma correction algorithm.The study found that the "dark channel dehazing algorithm" has fewer parameters,less computation,and the overall image has a good restoration effect.However,when there are large flat areas of pigment in the processing image,its transmittance estimation ability is reduced,and the output image is extremely high.It is prone to halo,color distortion,etc.The "gamma correction algorithm" has the ability to improve the visual perception of the image and highlight important detail information,but the method essentially only improves the contrast effect of the image,and it is easy to lose edge information.Aiming at the advantages and disadvantages of dark channel defogging algorithm and gamma correction algorithm in fog map processing,this dissertation proposes that based on the codec network to dehaze an image,which combines the advantages of the two algorithms for image processing,and combines the features of the two images into fusion.Layers are used for feature fusion,and finally a clear image is restored by feature reconstruction through the decoding network.The network aims to effectively combine traditional algorithms and deep learning networks to improve the robustness of the network system.During the simulation experiment,it was found that due to the complex formation mechanism of fog in the real environment and the uneven distribution of fog concentration,the image has fog residue at the junction of the depth of field.Therefore,on this basis,a feature fusion dehazing network based on dual-attention module is proposed,which integrates the channel attention module and pixel attention module of the parallel structure into the dense block structure of the encoding network.Among them,the channel attention module can efficiently allocate limited network resources and avoid unnecessary background information occupying resources;the pixel attention mechanism pays more attention to the associated feature information of uneven fog distribution in the image,and is not suitable for images with uneven fog distribution in the real environment.It has stronger defogging ability and enhances the practicability of defogging network.Extensive experiments on synthetic datasets and real datasets show that the feature fusion dehazing network based on dual-attention module has better dehazing performance than existing dehazing algorithms.The three objective evaluation indicators of similarity and dehazing efficiency have been improved,which has certain engineering application value.