Research And Application Of Single Image Dehazing Algorithm Based On Deep Neural Network

Single image dehazing in computer vision is always a challenging task because of its ill-posed property.With the application of computer vision tasks such as face recognition,object detection and semantic segmentation in people’s production and life more and more widely,as the upstream visual task of these tasks,single image haze removal has gradually become a hot research field.First of all,due to the problem of detail loss in the process of dehazing in the existing dehazing methods and the problem that the heterogeneous haze in the image cannot be removed,we proposed a dehazing model based on deep neural network.The network uses encoder-decoder structure and two core modules to achieve high quality image restoration.Image detail loss is a common problem in image restoration,a feature enhancement module based on the Strengthen-Operate-Subtract(SOS)enhancement strategy is designed to improve image quality.The module innovatively incorporates multi-scale potential features to supplement the missing signal.In addition,in order to eliminate the uneven haze in the image,a new hybrid attention unit is proposed,which intensifies the important information in the image in multiple dimensions and highlights the main objects in the image from the background.A large number of evaluation and simulation results show that the proposed dehazing model achieves excellent dehazing performance.Secondly,most of the current dehazing methods do not learn much about object structure and ignore the importance of image prior.To solve this problem,we explore an image prior that can protect the image structure and introduce it into the model to guide highquality image reconstruction.According to the research,the residual-channel prior can protect the image structure.Therefore,we apply the residual-channel prior to the fog image restoration task,and propose a structure maintenance dehazing module and a residualchannel enhancement module based on the guidance of the residual-channel prior.In addition,in order to unify the image features into the same feature space,we design an adaptive normalization module based on the hollow convolution technology.The adaptive normalization module and residual channel enhancement module complement each other to effectively improve the network dehazing effect.Based on these two ideas,another novel dehazing network is proposed.The results of training and evaluation on multiple data sets show that the model has a high efficiency in dehazing compared with classical dehazing methods in the field.Finally,we will propose two dehazing models in multiple synthetic data sets as well as real data sets to conduct detailed experiments,including comparison and ablation experiments,to prove the effectiveness and advancement of the proposed method.In addition,as an enhanced algorithm,the haze removal model was applied to the production scenario of grape yield prediction.We made a haze map grape data set to fine-tune the dehazing model,and combined the fine-tuned dehazing model with the grape string detection algorithm to solve the problem of inaccurate yield estimation algorithm results caused by haze environment.