Research On Image Dehazing Algorithm Based On Deep Learning

Fog and haze are particulate matter suspended in the air.When light propagates in fog and haze,a significant portion of the light will be absorbed,scattered,or reflected,significantly reducing the visibility of the landscape,and seriously degrading the quality of the collected images.Often,phenomena such as blurred contours,low contrast,dim brightness,and unclear object features will occur,increasing the difficulty of advanced computer vision research.Therefore,how to restore images taken under foggy conditions to corresponding images under sunny weather conditions has become an extremely important research direction.In recent years,more and more image defogging methods have emerged in this field.From the initial image enhancement method based on non-physical model to the method based on physical model,until the great development of artificial intelligence technology,the defogging method based on deep learning has become the mainstream of current research.The machine learning method based on big data has a strong feature perception and the ability to obtain context information,and has a very obvious processing advantage for the restoration of atmospheric degradation images and the enhancement of detail information.It can more accurately establish the mapping relationship between degraded fogged images and clear images,so that the images after the removal of fog are more close to the real image.This type of algorithm has high accuracy and good effect,and the restored fog free image is closest to the real image.However,existing defogging algorithms based on deep learning also have some drawbacks,such as incomplete defogging,color oversaturation,large parameter quantities,and difficulty in collecting paired datasets.To this end,this thesis makes the following innovative work to address the above issues:(1)To solve the problem of insufficient feature extraction in existing defogging algorithms,an end-to-end image defogging algorithm based on multi-scale residual and attention mechanism is proposed.Firstly,using three small-scale convolution kernels to extract the shallow features of the fog image can obtain a larger receptive field while reducing the amount of parameters.Then,it is input into multiple network modules composed of a multiscale residual hole convolution feature extraction module and a multiscale attention mechanism module in series.The multi-scale hole convolution residual feature extraction module can extract fog map features of different Receptive field and fuse features of different dimensions,effectively solving the problem of single feature scale.The multi-scale attention mechanism module can reasonably allocate the weight of different features and suppress irrelevant redundant information.Finally,the fog features in the fog map are filtered to obtain the feature map of the defog map,and then the fog free image is restored by convolution operations.Through a series of comparative experiments,it is proved that the network can effectively remove the haze existing in the image.(2)An unsupervised defogging network based on Cycle GAN is proposed.To solve the problems of large network parameters,slow training,and difficulty in obtaining paired datasets,the Cycle GAN network originally used for domain migration was improved to make it better applicable to the field of image defogging.Firstly,the structure of the codec is improved,using convolutions with a convolution core of 3 for image feature extraction and image reconstruction,and adding multiple residual connections to reduce the amount of parameters.Next,in the feature conversion section,the scheme of residual block plus attention mechanism is adopted,and multiple modules are connected in series for feature conversion from fog map to non-fog map.The attention mechanism adopts an attention mechanism that combines space and channel,which is a simple and effective lightweight attention mechanism that can effectively focus on channel features and better extract spatial features,making the defogging effect improved.In addition,perceptual loss is added on the basis of the original adversarial loss,feature loss and cyclic consistency loss,which uses the output of the fourth pooling layer of VGG19 network to extract the combination of high-level and low-level features,thus preserving the original structure of the image.Through experimental analysis,the proposed defogging method can effectively remove fog.(3)An image defogging algorithm based on Transformer architecture is proposed.After extensive research,it has been found that convolutional neural networks(CNN)have many advantages in the field of deep learning image defogging,but its shortcomings cannot be ignored,such as the generally small receptive field of CNN and its high adaptability to input content.In recent years,another type of neural network architecture,Transformer,has shown significant performance improvements in natural language processing and advanced visual tasks,which can effectively compensate for the shortcomings of CNN.Therefore,the thesis combines the advantages of both to design a new image defogging algorithm.Through a series of key designs in constructing multiple self-attention and feedforward network blocks,it proposes an efficient Transformer model that can capture remote pixel interactions,that is,establish connections between distant pixels,and is also suitable for large size images.The overall network architecture is in the form of an encoder decoder,which is composed of several Transformer blocks.The features extracted by the encoder through the Transformer are input to the corresponding decoding stage,enabling effective fusion of low-level features and highlevel features,and efficient restoration of image detail information.Experimental analysis shows that the proposed defogging method has better defogging effect.