Image Dehazing Using Two-phase Features And Attention Refinement Module Based Fully Convolutional Networks

Haze is a very common phenomenon in nature scenes,which leads to poor image quality.In the field of image processing,images or videos obtained through image capture equipment are often noisy,and haze is a type of noise.It will make the image more blurry and lose some image information.There are countless situations similar to this kind of image degradation,such as outdoor scene images taken by cameras,remote sensing maps taken by remote sensing satellites,medical images taken by medical photography equipment in vivo,and pictures taken underwater by underwater cameras.Therefore,image defogging has important research significance in practical applications such as machine vision,intelligent transportation,medical images,remote sensing satellites,and underwater surveys.Image defogging is also an image enhancement technology.After defogging the image,you can obtain better visual effects.In graphics calculation or post-processing,the image after defogging can get more details.Image dehazing provides more information for scientific research such as feature extraction and pattern recognition.At present,image dehazing algorithms in the scientific research community are divided into two categories: traditional a priori and learning-based.The traditional apriori algorithm needs the advantage of having strong theoretical derivation and mathematical induction to support the algorithm,but its actual effect is in many scenarios Ineffective,some scenes with high fog density are not applicable,and some scenes with white areas are not effective;learning-based algorithms are not accurate enough in feature selection,and some algorithms add features that are not related to dehazing to learning In the model,the feature redundancy is caused,and the learning-based algorithm needs a large amount of data to support its model,which leads to the improvement of the algorithm efficiency.Based on the above unresolved problems,this article proposes two defogging methods.The two innovations are asfollows.First,for the problem that the traditional algorithm is not applicable in scenes with many white areas or dense fog,this paper proposes a two-stage feature image dehazing algorithm.After studying the characteristics of existing algorithms and synthesizing their advantages and disadvantages,the algorithm in this paper selects the one-stage features of foggy images,and filters them on this basis to obtain the two-stage features.By observation,in the foggy image and the fogless image,the former has a lower saturation than the latter.In RGB space,the value of one channel of the fogless image is lower than the value of the other two channels.This minimum value for the fog map is much larger than the fog-free map.And according to human visual habits,it is often easy to distinguish the fog in the image,and the Gabor feature is consistent with the image characteristics of human visual observation.Based on the above observations,this paper extracts the saturation,minimum channel,maximum channel of the input image,and the Gabor features of the figure at 5 different wavelengths and 8 different directions.The 43-dimensional features are used as the first stage features.Secondly,based on the feature map of the first stage,the maximum,minimum,mean,variance,skewness,kurtosis,Gauss and other features of each window are extracted in the form of filtering as the features of the second stage.These features Contains(43 * 7)dimensions.Among these features,such as mean and variance,the numerical values??shown on images with larger fog density are also very different,and higher-order features such as skewness and kurtosis are more sensitive to changes in fog.With the change of fog density,these characteristics change significantly,which is very conducive to the model learning in the direction of solving.These feature map groups are mapped to transmittance maps,and mathematical models are used to model transmittance prediction models.Finally,use the predicted transmittance map to perform the final operation to obtain a fog-free map.Second,for the learned defogging algorithm,most of the network models used forimage defogging are relatively single,and the network structure is deeper.The more layers of the network means the more parameters are required,which will lead to The load of the processing machine is larger,and the cost of image defogging is higher,which results in less improvement in defogging efficiency than traditional algorithms.Most of the network frameworks used for defogging do not combine the global information of feature maps,making the transmission map mapping inaccurate.Therefore,this paper proposes a dehazing algorithm based on full convolutional network learning of attention mechanism.Because the full convolutional network is suitable for image segmentation,in this paper,the prediction of the transmittance map is also a scene depth map segmentation problem,so the structure of the full convolutional network is suitable for the research work in this paper.In addition,in order to make the global information of the convolutional layer play a role in model training,an attention mechanism module is added to the full convolutional network structure proposed in this paper to extract each convolutional layer Global information.After researching and analyzing the characteristics of the fog map,the features selected by the first method can well characterize the relationship between transmittance and them.In the experiment,the scene maps in various occasions were compared.The effect is very good.By improving the original full convolutional network,the full convolutional network model designed for defogging in this paper can effectively solve the problems caused by deep networks.In the experiments,by comparing with other deep learning defogging algorithms,it is concluded that the algorithm is mostly better than other algorithms in the defogging effect.It has small parameters,low hardware requirements,Fast speed and other advantages.