| Under hazy conditions,there are a large number of particles in the air.The luminous intensity reaching the imaging system is weakened due to the scattering effects of the medium.The quality of scattered images is decreased,which brings difficulties to the target detection and recognition of the visual system.The image restoration of scattering imaging has important research value.Deep learning methods provide data-driven solutions for image restoration of scatter imaging,but it usually requires a large number of training data to extract features.And there are still some problems,such as poor global connection and lack of processing focus.In addition,the interpretability of deep learning method is also an important research field.The lack of interpretability limits the development of theory and the application in practice.In this paper,we address the problem of image degradation caused by hazy scattering and study the reconstruction of images using end-to-end networks.For the problems of neural networks lacking interpretability and not focusing on key information,this paper designs and implements two deep network structures with improved results in both qualitative and quantitative metrics.The main research content of this paper are as follows:(1)Unrolled deep network for scatter imaging of image restoration.In order to solve the problem that the neural network method needs a large number of training data and lack of interpretability,an interpretable unrolled network is designed for scatter imaging of image restoration.Firstly,an optimization model is designed based on the atmospheric scattering model with the addition of a sparse regularization prior.And the optimization problem is decomposed into multiple subproblems and an iterative framework is constructed using the alternating miniaturization method.Secondly,the iterative algorithm is unrolled into a deep network using algorithm unrolling,and the learnable module is embedded in the network.The unrolled network uses back propagation to learn parameters that are difficult to design in traditional iterative algorithms,and the parameters can take different values in each network layer.The parameters no longer remain fixed and the model is more flexible.In addition,the unrolled network is designed specifically according to the target,which reduces the requirement for training data and has better generalization performance.The unrolled network is tested by two datasets,and the experimental results show that the unrolled network can improve the visual perception and achieve better results in quantitative metrics.(2)Attention network for scatter imaging of image restoration.In order to solve the problem that the global information extraction of convolution neural network method is insufficient and the focus of feature extraction is not prominen,an attention network is designed to enhance the performance of feature exaction.On the basis of convolutional neural networks,the squeeze-and-excitation attention module is added,which can adaptively adjust the feature weights according to the importance of channels and improve the network performance.The attention network contains two main parts: an encoder decoder module to extract features and a pyramid pooling module to enhance the recovered image details.The attention network can extract interesting information faster,quickly locate the hazy area and dig for hidden information.The effect of the network is tested with two datasets,and the results show that the method can effectively restore the detail information of scattered images taken under severe hazy condition,and provides a more significant improvement in both subjective perception and objective evaluation criteria. |
