Research On Image Defogging Algorithm Based On AOD-NET

Many tasks such as video monitoring and target detection and tracking need to be performed in outdoor scenes.When encountering haze weather,images collected by image acquisition devices may encounter issues such as lack of detail,dim colors,and reduced brightness,making it difficult to complete the required tasks.Therefore,image defogging is of great significance in practical work.In order to restore higher quality fog free images,the research content of this article includes the following points:Firstly,based on the atmospheric scattering model,in order to solve the problem of error accumulation caused by the need to estimate the two parameters of transmittance and atmospheric coefficient in the atmospheric scattering model,this paper applies an improved atmospheric scattering model formula to convert the two parameters in the scattering model into a unified parameter through formula conversion,reducing the accumulation error.Secondly,in view of the limitations of traditional defogging algorithms,this paper proposes two deep learning network models based on the unified theory of parameters.The first defogging network model is a Multi Scale All in one(MSAOD)model.This model is divided into three modules: The first module is a preprocessing module,which divides the input image into two blocks,performs convolution processing respectively,and then performs image merging;The second module is the backbone module,which is mainly used for feature extraction.The output of the first part is used to extract image features through multi-scale encoder decoder;The third module is the post-processing module,which is mainly to reduce the details of the image lost through the convolution calculation of the backbone network.The obtained feature map is fused with the original image and then further trained.This can reduce the loss of details after the network module,so that the fog free image with less deviation can be obtained,and the image distortion will be less.A unified parameter model in the image defogging formula is obtained through training.Finally,the parameters are substituted into the atmospheric scattering model to obtain the defogged image.Another defogging network model is the Dual Multi Scale All in one(DMSAOD)neural network model.This network model follows the multi scale module and post processing module of the MSAOD network model.A multi scale module is added between the two modules to form a dual multi scale module for the feature extraction function module,and the post processing module is used to preserve the original details of the image.The experimental results show that the defogging results of the two network models proposed in this paper are compared with eight traditional defogging algorithms on seven sets of images,and are superior to mainstream depth learning and traditional methods in terms of PSNR,SSIM,and subjective vision.After defogging,the image is optimized in terms of detail retention,color,brightness,and other aspects.In addition,ablation experiments have proved that each module in the two network models proposed in this article is necessary.Thirdly,in order to solve the problem of color deviation caused by color projection that cannot be solved in the defogging algorithm,this paper incorporates a white balance algorithm after the defogging algorithm.This algorithm solves the color deviation problem by maximizing the overlapping area of the R,G,and B channel histograms through histogram matching.However,which of the three channels can be used as the matching reference channel cannot be automatically selected.To solve this problem,this paper proposes a white balance algorithm that can independently select the best matching channel,making the histogram matching white balance algorithm more intelligent and more applicable.The experimental results show that the final image obtained by the image defogging system in this paper has more realistic colors,less color bias,and is more consistent with human visual characteristics.