Research On Dehazing Method Of Real Scene Fog Image Based On Deep Learning

Image defogging aims to remove the fog elements in the image to the greatest extent and restore the original information of the image.It plays an important role in the fields of video surveillance,automatic driving,drone imaging,and intelligent security.It is also the basis of high-level image processing tasks such as object tracking,object detection,object segmentation.Most of the existing image dehazing methods use synthetic fog datasets for training and achieve good results on synthetic fog test sets.However,the actual environment is complex and changeable.In this environment,there is a big difference between the real fog image and the synthetic fog image generated by the algorithm,which leads to poor performance of algorithms trained on synthetic fog datasets when dealing with realworld fog images.In view of the above problems,this thesis launched the research on the dehazing method for real fog images,aiming at dealing with real fog images in different environments.The specific research contents and contributions are as follows:(1)Research on Dehazing Method of Real-scene Foggy Image Based on Cycle Generative Adversarial Network and Feature Fusion.Based on the Cycle generative adversarial network,this thesis forms the generator via using residual blocks,and designs image-level and feature-level discriminators,which realizes the conversion of synthetic fog images to real fog images,and bridges the feature gap between synthetic fog and real fog to a certain extent.In addition,since the features of different scales carry information of different levels of the image,this thesis introduces the residual block to designs the multi-scale feature fusion network,which realizes the comprehensive perception of the local information and global information of the fog image.In addition,this thesis introduces an unsupervised dark channel loss for real fog images to optimize the training of the network.The experimental results show that the method proposed in this thesis is able to effectively remove the fog elements in the fog image and maintain the overall harmony of the output image.(2)Research on Dehazing Method of Real-scene Foggy Image Based on Attention Mechanism and Contrastive Learning.Aiming at dealing with the phenomenon of nonuniform fog density in different areas in the real fog image,and most image defogging methods only consider narrowing the gap between the output image and the fog-free image from the positive direction and ignore the constraints of the reverse direction,this thesis proposes a dehazing method based on attention mechanism and contrastive learning is proposed,using two-stage pooling to obtain channel attention,and using convolution to obtain pixel attention.In addition,the constraints on the weight update of the image dehazing network are realized from the positive and negative directions of the fog-free image and the fog image in this thesis.Compared with methods based on feature fusion,it can better realize the differential processing of different regions of uneven fog images.Comprehensive experiments show that the method in this thesis shows excellent performance in the processing of real fog images.In summary,this thesis proposes two different dehazing methods for real-world foggy images.Quantitative experiments and qualitative evaluations on synthetic fog and real fog datasets demonstrate the effectiveness of the proposed method.