Image Dehazing Algorithm Based On Multi-Level Hourglass Structure

With the rapid development of computer vision technology in industrial control,target detection and other fields,more and more technologies have been applied to remote sensing satellite imaging,video surveillance detection and other outdoor scenes,and achieved good results.However,when fog,dust and other bad weather occurs,the image quality of such outdoor vision system will be affected,and it is difficult for the algorithm to complete the established task.Therefore,the defogging of degraded images has important research significance.In order to improve the image restoration ability of single image defogging algorithm,this thesis mainly completes the following work:(1)In order to solve the problem of inadequate estimation of atmospheric scattering coefficient in the process of building synthetic training data set,scholars proposed to analyze the relationship between human visual contrast and atmospheric scattering coefficient in atmospheric scattering model,calculate the range of the parameters,and construct standard "fog-free" data set based on the calculation results and multiple basic data sets.(2)Aiming at the problem that the existing learning-based method has smaller receptive field and the network can not obtain global information,which leads to the difficulty of image restoration in heavy foggy environment and poor recovery of vision details,a transmission graph estimation algorithm based on multi-level hourglass structure is proposed.Through multi-level up-and-down sampling structure,the network receptive field is enhanced,the network's perception of global information is improved,and the transmission graph is better estimated.On this basis,a multi-level Hourglass network cascade method is proposed,which can restore image details more accurately through multiple up-and-down sampling.At the same time,this thesis compares the influence of different levels of Hourglass on the ability of feature extraction,and finally decides to use the third-order and second-order Hourglass cascade as the infrastructure of the outgoing graph estimation sub-network.(3)To solve the problem that the existing atmospheric light estimation algorithms rely on prior assumptions,lack of overall image observation,have poor accuracy and are difficult to integrate into the network,a method of estimating atmospheric light using convolutional neural network is proposed.Two strategies of classification and regression are used in this thesis.Aiming at the problem that Softmax Loss can not accurately describe the relationship between predicted value and real value,Euclidean Loss gradient disappears and predicted value is in the middle,a new Bell Loss function is proposed.This function combines the advantages of the two loss functions mentioned above,and makes use of bell curve to describe the deviation between predicted value and real value smoothly and accurately,and the appropriate gradient is provided to update the network parameters and guide the convergence of the model,so as to improve the prediction accuracy of atmospheric light.(4)Aiming at the problem that the multi-level Hourglass network structure is too deep and gradient dispersion exists in training,the method of relay supervision is proposed for gradient optimization.At the same time,in order to improve the generalization ability of the network in the real foggy environment,the traditional characteristics were introduced in the way of dark channel prior to improve the robustness of the network,and the effects of different optimization modules were compared and analyzed.Finally,this thesis fuses the above network to build an end-to-end single image defogging model.In order to verify the effectiveness of the algorithm,this thesis uses synthetic data sets and real outdoor data sets to compare the proposed defogging algorithm based on multi-level hourglass structure with six more advanced algorithms in both objective and subjective evaluation.The image restoration ability of the algorithm in different foggy environments has been greatly improved.It proves that the algorithm can better restore the details of the near and long-range image,more realistic color restoration,and more natural color transition.