Research On Deep Learning Method Of Image Dehazing Based On Visual Attention Model

With the advent of the information age,computer vision system has played a huge role in the fields of visual navigation,video monitoring,automatic driving and so on.However,the performance of the current vision system is closely related to the image quality received by the imaging interface.The visual system is not good at processing the images taken in the harsh environment,especially in the fog,which affects the normal operation of the computer vision system due to the loss of more details and poor visibility.Therefore,it is necessary to remove haze from hazy images.Based on the convolution neural network,this thesis introduces the attention mechanism into the task of image dehazing,mainly from the following three aspects:1.We propose an image dehazing method based on selective attention mechanism.Firstly,channel attention and spatial attention are used in image dehazing task for the first time,and spatial attention is improved according to the characteristics of image dehazing task.In addition,this thesis proposes a fusion module based on selective attention,which can better integrate channel attention module and spatial attention module.In order to verify the ubiquity of this algorithm,we also do experiments on the image deraining task.A large number of experimental results show the effectiveness and ubiquity of this algorithm.2.We propose an image dehazing method based on Non-local network.This method applies self-attention to image dehazing task for the first time.In this thesis,non-local operations are used to capture global features in spatial dimension and channel dimension respectively.In this thesis,the idea of block is used to reduce the computation of nonlocal operations in the spatial dimension,and the fusion module based on selective attention is used to fuse two kinds of non-local modules.In addition,this thesis improves the residual connection structure in the previous chapter of the network,converges the shallow information and deep information,and extracts features from different layers of the network for fusion as the output of the network.Experimental results and ablation experiments show the effectiveness of the proposed algorithm.3.An image dehazing system is designed and implemented.The users can select the image from the folder directly as the input of the system.The system can dehaze with one button and get the image after dehazing quickly.