Research On Incremental Image Dehazing Algorithm Based On Deep Learning

In recent years,the air quality has deteriorated year by year due to industrial production and vehicle exhaust emissions,and haze weather has become very frequent.When hazy weather occurs,various particles floating in the atmosphere will have a greater impact on the propagation of light,which in turn will affect the clarity of the image captured by the camera device.However,the quality of the image is the basic guarantee for subsequent computer vision tasks.Therefore,the study of haze image sharpening is an important part of the field of computer vision.In recent years,the development of deep learning has been very rapid,and the image dehazing algorithm based on deep learning has gradually replaced other dehazing algorithms.Most deep learning dehazing algorithms adopt an end-to-end structure,which avoids the problem of inaccurate estimation of intermediate parameters when solving physical models for dehazing.At the same time,the algorithm has stronger ability to handle fog images in different scenes and more thorough dehazing..However,in the process of image dehazing using deep neural networks,the training of the neural network and the augmentation of the dataset are usually performed simultaneously.One of the challenging problems is how to extend the trained neural network to supplementary datasets while ensuring the dehazing capability for both datasets.Secondly,in the process of deep neural network dehazing,the features in the image are usually extracted in the form of convolution kernel weight sharing,which cannot reflect the importance of the features,thus reducing the ability to reconstruct clear images.In response to the above problems,this paper proposes two end-to-end incremental dehazing algorithms based on the deep learning method.The main tasks are as follows:(1)A network structure with feature enhancement and multi-scale loss constraints is proposed,and incremental training is used to generalize the dehazing ability.The network consists of teacher network and student network.The teacher network adopts an auto-encoder structure,and the student network enhances the features by learning the attention information of the labeled samples extracted by the teacher network;the multi-scale semantic features of the labeled samples are used as the soft labels of the student network,and the multi-scale semantic feature loss measurement mechanism is established.The global pixel difference loss is cascaded to construct a loss function oriented to two levels of features and pixels;in addition,the incremental training method is adopted,the teacher network guides the student network to balance the relationship between the old and new knowledge,so that the network retains the original knowledge and rapidly improve the generalization ability to supplementary datasets.In addition,the effectiveness of each module of the algorithm is verified by ablation experiments.Finally,the algorithm verifies the dehazing performance of the algorithm from the perspectives of subjective vision and objective evaluation.Experiments show that the proposed algorithm achieves good results in both indicators.(2)A dehazing structure and incremental training method for deep neural networks based on multiple transfer attention are proposed.The network consists of three parts.Among them,the attention generation network adopts the form of auto-encoder to extract the multiple attention of clear labels and haze distribution,and form the original attention of the dehazing network.The transfer medium network is between the attention generation network and the dehazing network.It uses the original attention as the soft target of the network,extracts the features of the haze image,and forms the transfer medium attention,thereby transferring the original attention to the dehazing network.The dehazing network achieves differential reconstruction of images under the guidance of transfer media attention.The incremental training method of neural network is based on some samples of the original data set,and a small number of new samples are added to construct an incremental data set,so that the neural network can increase the processing ability of new samples without forgetting the knowledge of the original data set..The effectiveness of the algorithm is verified on the ITS and OTS images in the fog image dataset RESIDE.There are 29 figures,6 tables,and 75 references in the full text.