Research On Single Image Deraining And Network Lightweight Based On Deep Learning

Computer vision algorithms such as target detection and tracking,scene semantic segmentation,and image understanding have made great development and are widely used in monitoring links of intelligent systems such as smart city,automatic driving and public safety early warning.However,the performance of these visual algorithms is often severely degraded when the quality of acquired image is degraded due to complex weather conditions such as rain and snow.Therefore,how to improve the image quality,reduce the negative effects of rain weather,and enhance the adaptability of existing intelligent systems,has important research value.The performance and operation efficiency of the existing image deraining methods can be greatly improved.In this dissertation,the improvement of image deraining performance,the reduction of deep neural network parameters,and the physical model of the rain image are studied in depth.The main research work and contributions are as follows:(1)As rain streaks seriously damage the texture details of the image background,we propose a recurrent attention dense network for single image deraining.A region-level direction-aware attention module is first utilized to detect rain streaks and generate an attention map to guide subsequent modules to accurately focus on the rain areas.Then the dilated convolution is used to increase the perception field,the squeeze-and-excitation modules combine the spatial features and the channel features,the densely connect convolutional further enhance the representational ability of the network in high dimension.What's more,the gate recurrent unit transmits the rain removal information from the previous step to the latter step to enhance the long-term memory of the network.Finally,the edge loss is added to the objective function,which is helpful for the network to retain more edge details.The proposed network can remove the rain streaks and preserve the background texture.(2)Due to the large number of parameters in the existing deep convolutional neural network and the cumbersome model,it is difficult to deploy them in the platform with limited storage space and computing resources.Therefore,we propose lightweight deraining networks based on knowledge distillation and fine design.Firstly,the knowledge distillation method is applied to the problem of image deraining,and the proposed recurrent attention dense network is taken as the teacher network.The relationship matrix between feature maps of different depths is adopted as the objective function of the recurrent module,the square of the amplitude of the feature map is used as the objective function of the attention module.The network parameters are greatly reduced while the performance is reserved.And then we combine the refined network design with knowledge distillation,the convolution separation is used to reduce the complexity of convolution operation and channel exchange is used to ensure the communication between feature channels.Convolution kernels of different sizes are used in different feature channels to obtain more feature information.The designed compact network is trained under the supervision of the teacher network,the network parameters are further reduced.(3)Most of the existing rain datasets ignore the "shield" effect caused by fog,and the ability to represent the real rain image is limited.To solve the problem,the influence of scene depth on the visual effect of rain image is analyzed,and a rain image model based on the atmospheric scattering model is established,which is more consistent with reality.Then,an unsupervised monocular image depth estimation method based on uncertainty is proposed,which takes monocular sequence image as input,combines uncertainty,structure-from-motion,and image reprojection to realize monocular image scene depth map estimation.At last,we utilize the estimated depth map to synthesize the realistic rain image based on the atmospheric scattering model for the training of deraining to improve the performance under fog.(4)For the atmospheric scattering rain image model,we propose a depth based generative adversarial network for single image deraining.Firstly,we utilize a multiscale aggregation attention module to detect rain streaks,and apply a covariance based second order attention module to estimate transmission map,then remove rain streaks according to the atmospheric scattering model preliminary,and improve the results by the refined module.At last,the relative discriminator containing depth information is used to distinguish the input is a ground truth or a fake image to further improve the performance.The proposed method can remove rain streaks effectively and restore the texture even with a shield from fog.