Research On Single Image Dehazing Algorithm

In the haze weather conditions,there will be a great mount of suspended particles in the air,such as dust,water vapor and so on,what's more,the scattering of these suspended particles will make the images collected by camera becoming blurry.However,the information of the degraded image is greatly reduced,and it is hard to extract the effective feature from haze images,it also affects the effectiveness of the outdoor computer vision system.Therefore,it is significant to develop the dehazing algorithm to improve the quality of haze images.The main work of this paper is as follows:(1)In recent years,many novel dehazing algorithms based on atmospheric scattering models had been proposed by many scholars at home and abroad,but these algorithms all depend on effective assumptions or prior information.Among these algorithms,Dr.He proposed the DCP theory and obtained a amazing dehazing result,however,some research shows that the DCP prior information still has limitations.Through the detailed analysis and experiment of the DCP,this paper finds the algorithm cannot effectively remove the fog in the sudden depth of field region.Aiming at the deficiencies of the DCP algorithm,this paper proposed a new dark channel confidence calculation method,and used the fusion dark channel instead of the original dark channel,the proposed algorithm can adaptively compensates and corrects the estimated transmission rate of the region.The experimental results show that the improved algorithm in this paper obtains accurate atmospheric light values in a variety of scenarios,and the effective combination of fusion dark channel and dark channel confidence can obtain a more refined medium transmittance.Finally,in comparison with other algorithms,the algorithm in this paper has also achieved more excellent results.(2)Dehazing algorithms based on physical models has always been an illconditioned problem,and solving this problem usually requires suitable and effective conditions.However,deep learning can implicitly learn the correlation of parameters in the physical model,which can replace these prior information.By constructing a suitable dehazing network and training the network to learn the relationship between the parameters in the model and the foggy image,an effective defogging network can be realized.Octave convolution is a new type of convolution,which can separate high and low frequency features from feature images and convolve these features,and then fuse these features of different frequencies.This convolution method can effectively reduce the amount of network calculations,based on this feature,this paper used Octave convolution instead of the traditional regular convolution,and design an end-to-end defogging network based on the self-encoder network architecture.The proposed network realizes the extraction of multi-scale image features by constructing the residual Octave convolution group,and carries out the weighted fusion of the extracted multi-scale features with different proportions.This paper uses the RESIDE data set for network training,selects the appropriate frequency doubling convolution low-frequency feature selection ratio through ablation experiments,and uses it for subsequent dehazing experiments.The experimental results show that the defogging network designed in this paper has achieved a good dehazing effect and reduced the amount of model calculations to a certain extent.