Research On Image Dehazing Based On The Method Of Convolutional Neural Networks

Images may be greatly degraded by fog and haze.Therefore,it is urgently necessary and also widely applicable to further study image defogging technologies for improving the quality of images with fog and haze.In this thesis,the mechanism of image degradation in foggy weather is analyzed,and the physical model of image defogging is introduced;the mapping relationship of foggy images and its transmission maps is established based on deep learning method;finally,the fog-free image will be reconstructed from the transmission map based on the atmospheric scattering model.The detailed research issues are as follows:1.Image defogging method based on multi-scales convolutional neural networks.In order to improve the feature learning capability of convolutional neural networks(CNN)in image processing,three CNNs with different scales are adopted to build the deep learning model for the transmission map prediction.Firstly,the first CNN is obtained based on the AlexNet model,and two CNNs with different scales are appended.Then the more accurate transmission map is got with the gradual optimization of the three CNNs.Finally,the fog-free image is obtained based on atmospheric scattering model.2.Image defogging method defogging based on an improved convolutional neural network model with dark channel prior.Inspired by the image defogging algorithm with dark channel prior,a CNN model is designed based on the dark channel features.The main idea is using Maxout nonlinear active function to simulate an extreme filter,so as to extract the dark channel features from the input images.And,then three scales of convolutional layers are applied in ensuring the scale invariance;subsequently,the spatial invariance of the network model is achieved by operating a pooling layer.The transmission map of the foggy image can be predicted by the improved CNN model,and then the fog-free image can be restored according to the atmospheric scattering model.3.The design and implementation of a prototype system.A prototype system is developed based on the above two defogging methods.The test's result shows that the system can effectively defog images.