Weather Recognition And Non-Homogeneous Image Degazing Based On Deep Learning

With the development of deep neural network,its excellent performance in visual field has attracted the attention of scholars at home and abroad.The convolutional neural network method has become the most important tool to solve the related tasks in visual field.For the recognit ion and restoration of weather degradation images,it has been a hot topic of research at home and abroad,especially in the field of automatic driving,which is helpful to realize all-weather automatic driving.Tradit ional weather recognition algorithm only considers the recognition of static weather images such as sunny and cloudy days.For the field of automatic driving,the environment of automobile driving often focuses on the dynamic weather conditions.In the field of defogging,many researches assume that the image distribution of fog is uniform,but the actual image distribution is not uniform,which also leads to the poor restoration effect of many algorithms in nonuniform fog.Based on the method of deep neural network,this paper proposes an image recognition algorithm under dynamic weather conditions,and open up a four kinds of dynamic weather image data set.In order to study the field of dynamic weather recognition in the future;For the restoration of non-uniform fog images,this paper proposes a neural network method based on migration learning.Aiming at the characteristics of non-uniform fog images,the branch network of migration learning is introduced to make the algorithm better in the restoration effect of nonuniform fog images,especially in the area of thick fog cover.Firstly,most of the data sets in the field of weather classificat ion are collected under static weather.For example,in image2 weather data set,there are five weather images,including snow,sunny,cloudy,rainy and hazy.But most of the images are collected after rain or snow.For the weather images which are raining or snowing,there is no relevant open source data set.In this paper,we search the real scene dynamic weather capture image on the Internet,and construct a four kinds of weather image data set(FWID)including rainy day,snow day,hazy day and sand dust weather.And a lightweight weather recognit ion network is constructed by using the data set proposed in this paper.Secondly,a two stream dehazing algorithm based on transfer learning and attention mechanism is designed for the uneven distribution of fog in non-uniform fog images.Considering that the current defogging field is trained and tested on synthetic data sets,and all assume that fog is uniform in image distribution,this paper extracts the essential features of images on large-scale data set Image Net by means of transfer learning according to the characteristics of non-uniform fog images,and then uses the attention mechanism to pay more attention to the dense fog area on the non-uniform images,The results show that the proposed algorithm is superior to the previous algorithm in the quality of the restored image.Then,in view of the characteristics of generative adversarial network,the generator and discriminator need mutual supervision and training,and the generator can not be too strong and the discriminator can not be too weak.This paper proposes a discriminator network which combines the high and low frequency information of the image,and combines with the two stream image defogging neural network proposed in this paper,which further reduces the color distortion in the restoration of non-uniform fog images The problem of detail ambiguit y and fog residue.Traditional discriminator only discriminates the true and false of single input prediction image,but the high and low frequency informat ion of the image can distinguish the true and false from the degree of the sharp change of pixels.Therefore,the high and low frequency components of the image are sent to the discriminator together with the predict ion image itself for discrimination.The experimental verification is carried out on the image data set of non-uniform fog,After the original algorithm is added with the improved discriminator,the image qualit y is better.Finally,the algorithm proposed in this paper is ablation learning,which is divided into three parts: verifying the effectiveness of migration learning method,verifying attention mechanism and mult i-scale enhancement module.Firstly,the effectiveness and rationality of the proposed method are fully proved by the evaluation of peak signal to noise ratio(PSNR)and structural similarity metric(SSIM),and the local recovery evaluation of the concentrated non-uniform image.Finally,the verification is carried out on the I-HAZE,O-HAZE and RESIDE image data set,and the recovery effect is good,which proves the effectiveness of the proposed method.