Evaluation Of Image Defogging Effect Based On Convolution Neural Network

Image quality assessment is very important for the development of image processing technology.As the research focus in the field of image processing,image quality assessment has been attracting much attention from researchers.In recent years,the fogging algorithm in the field of image processing has developed rapidly,and many excellent algorithms have been put forward one after another,However,the scientific quantitative evaluation method for the effect of fog removal still needs to be studied.Therefore,the objective evaluation of the effectiveness of image de fog algorithm is of great significance for the development of fog removal algorithm.The existing evaluation method of fog image is mainly based on subjective evaluation method,which requires a large number of participants,and the grading process is tedious and not timeliness,so the applicability is not high.Objective evaluation method,no reference quality assessment is the research focus of the objective evaluation method of fog removal at the present stage.The non reference image quality evaluation method is evaluated by the extracted image features,both of the prior knowledge of feature selection and extraction process and the degree of fitting between features and vision decided the effect of the fog image evaluation algorithm.In order to solve the problem of difficult feature extraction and low fitting degree of image feature and visual perception in objective evaluation of fogging image,An evaluation model of image de haze effect based on convolutional neural network is proposed in this paper.By improving the AlexNet network,it transforms it from a network for image classification to an image quality evaluation network model.This paper proposes to fog image convolution neural network,differential image convolution neural network,feature fusion convolutional neural network three network models to evaluate the effect of fog image.In this paper,180 fog images are treated with histogram equalization fogging algorithm,Retinex fog removal algorithm,dark channel priori fog removal algorithm,Tarel and Fattal fogging algorithm,and 900 restoration images with fog effect score are obtained by subjective evaluation of the fog removal effect after fog removal.And these 900 images are used as data sets for training and testing this algorithm.In order to verify the effectiveness of the algorithm,we carried out three network models of image de haze effect evaluation simulation experiment.On the training set,the optimal network structure of the deconvolution neural network and the differential image convolution neural network is determined by designing the different quantity convolution layer and the number of convolution kernel.Then the final network structure of feature fusion convolution neural network is determined.Further,the evaluation model of the optimal structure is used to compare the data on the test set,and the analysis shows that the correlation coefficient and Spielman coefficient of the evaluation results of the model and the subjective score are more than 0.8,and the mean square error between the score and the subjective score is less than 0.5.It shows that the correlation between the evaluation result and the subjective score is higher,and the difference between the result of the model and the subjective score is small.The correlation coefficient and Spielman coefficient of the feature fusion neural network are higher and the mean square error is less than 0.95,and the mean square error is less than 0.2.It shows that the performance of the feature fusion convolution neural network is the best,which proves the effectiveness of the algorithm in the evaluation of the image defogging effect.