Research On End-to-End Image Dehazing Algorithm Based On Deep Learning

With the advancement of industrialization in China,environmental problems are becoming more and more serious.They make the hazy weather appear more and more frequently.In hazy weather,many fine particles are suspended in the air.The light transmitted from the object to the lens changes.Thus,this leads to a series of problems,such as a lack of color and loss of contrast in the captured images.These problems can seriously affect the proper functioning of many systems and devices that use images,and hazy images can also affect other advanced vision tasks,such as target detection and object recognition.Most scholars use deep learning methods for image dehazing and have gained specific results.However,there are still the following problems.First,the dehazing models are very dependent on the atmospheric scattering model,and the details of the dehazed image are seriously lost.Secondly,the dehazing models often fail to completely dehaze or excessive haze removal when processing dense haze images.Third,the processing time of the dehazing models becomes longer and longer as the dehazing effect improves.Accordingly,how to address these issues has become a crucial challenge in current research on image dehazing.In response to the above problems,the main research contents of this thesis are listed as follows:1)A dense residual convolutional neural network for single image dehazing is proposed.The network uses an end-to-end dehazing model.The network inputs a hazy image and obtains the corresponding clear dehazed image directly.It is divided into three main parts: encoder,feature extraction,and decoder.The encoder and decoder increase the depth of the network and avoid the occurrence of gradient explosion and gradient disappearance by using the Residual Group 3(RG3)module.Meanwhile,the Dense Residual(DR)module guides local feature learning and local feature fusion.The feature extraction uses the Residual Group 16(RG16)module to process the encoder output features further and help the decoder better process the valid feature information.The encoder and decoder use a add operation to fuse features,compensating for the loss of detailed information during up-sampling and down-sampling.The network uses L1 loss and MS-SSIM loss.The result indicates that the proposed method surpasses the current state-of-the-art single image dehazing.2)A multi-scale attentive feature fusion network for single image dehazing is proposed.The network is mainly divided into the encoder and the decoder.The Pixel and Channel Attention(PCA)module can assign different weight values to different channels and pixels,thus directing the network to focus on more important information.The Boost Information Connection(BIC)module is used to obtain detailed features.The Feature Fusion(FF)module can fuse feature information from different layers.The network uses L1 loss and gradient loss.The result indicates that the proposed method surpasses the current state-of-the-art single image dehazing.3)An end-to-end image dehazing using depthwise separable convolution network is proposed.The network uses a lightweight dehazing model,which is mainly divided into shallow feature extraction,deep feature extraction,and reconstruction.The shallow feature extraction uses depthwise separable convolution layer to extract features from the hazy input image.The deep feature extraction consists of three Group modules,a Pixel and Channel Attention(PCA)module,and two depthwise separable convolution layers,which help the network focus on more important features and effectively reduce the computation and number of parameters.The reconstruction uses a convolution layer,which is used to obtain the dehazed images.The L1 loss function is used for the training process.The result indicates that the proposed method surpasses the current state-of-the-art single image dehazing.In conclusion,the dense residual convolutional neural network for single image dehazing can effectively solve the problem of severe loss of detail in the dehazed image.The multi-scale attentive feature fusion network for single image dehazing can effectively eliminate the inability to complete dehazing or excessive haze removal.The end-to-end image dehazing using depthwise separable convolution network can significantly improve the efficiency of dehazing.The three methods are aimed at addressing specific problems in specific environments and have made beneficial attempts for different application scenarios to meet their respective needs.