Research On Image Dehazing Algorithm Based On Model Optimization And Feature Learning

Outdoor vision systems are usually sensitive to weather factors,especially haze and dust,etc.In bad weather such as haze and dust,the turbid medium and suspended aerosol particles in the air greatly hinder the normal transmission of light between the target scene and the imaging device,causing a series of undesirable phenomena in the images or videos acquired by the imaging device,such as impaired contrast,loss of image details,scene color shift and visibility degradation,resulting in a serious reduction in the working effectiveness of outdoor vision systems.Therefore,with the development and demand of intelligent and vision systems,achieving visibility recovery for degraded scenes under haze and dusty weather has become one of the important research tasks in computer vision.To address the shortcomings of existing methods for haze image recovery,this paper implements some optimizations and improvements in terms of both physical models and deep learning frameworks,respectively.The root cause of foggy image formation is the absorption and scattering of light between the target scene and the imaging device,it is described by atmospheric scattering models.Therefore,atmospheric scattering model is widely used in image dehazing tasks.Based on the existing research,five optimized de-fogging algorithms using atmospheric scattering models are proposed for problems such as insufficient estimation of transmittance and atmospheric light,and the specific work is as follows.(1)In order to solve the problems such as insufficient transmission estimation and color cast of results of dehazing algorithms,an image restoration algorithm based on m inimum channel interval estimation and transmission adaptive constraint model is proposed.Firstly,bright channel of hazy image is obtained by using maximum operation of different sizes,and average value processing and frequency domain filtering are combined to get the atmospheric light estimation.Secondly,starting from the atmospheric imaging theory,minimum channel of hazy image is regarded as a constraint,then upper and lower boundaries of minimum channel of hazy image are fitted by plane model and adaptive mapping model respectively,and minimum channel of dehazed image and initial transmission estimation are obtained.Finally,the initial transmittance can be refined by filter smoothing and adaptive boundary constraints to obtain the optimized transmittance,and according to atmospheric scattering model,restoration results are obtained.Experiments show that the restoration results of the proposed algorithm have natural colors,appropriate brightness,thorough degree of dehazing,rich detailed information and low time complexity,which effectively solves the problems of insufficient transmittance estimation and color cast.(2)Aiming at problem of scene degradation in haze and sandy weather,a degraded scene restoration algorithm based on Gaussian convex optimization and double constraints of light curtains is proposed.Firstly,according to correlation between depth of field and scene brightness and saturation,Gaussian model and convex optimization are used to estimate depth of field.Secondly,through in-depth analysis of the relationship between atmospheric light curtain and scene,combining minimum channel smoothing and depth attenuation dual constraints to obtain degraded scenes atmospheric light curtain.In addition,atmospheric light value is obtained through improvement of bright channel priori and local atmospheric light.Finally,degraded scene is restored based on the restoration model,and color of sand and dust scene is corrected to realize scene restoration.Experiments show that proposed algorithm restores scenes with suitable brightness,natural colors,and rich detailed information,which also achieves ideal scores in quantitative indicators,effectively solve the problems of color cast and loss of details in degraded scenes.(3)Aiming at the problem of color cast in dehazing problem and incomplete dehazing of dense haze images,an image dehazing algorithm based on haze density distribution and adaptive linear attenuation is proposed.According to the characteristics of haze,a haze distribution model based on brightness,saturation and texture characteristic is established.From the nature of image degradation in hazy weather,the characteristic of negative correlation between haze density and transmission is used to propose an adaptive linear attenuation model.This model can complete the estimation of the dark channel of clear image,and then obtain transmission.According to the characteristic that atmospheric light can only reflect brightness information,haze distribution can be used to improve the local atmospheric light.And the dehazing result can be obtained by atmospheric scattering model.Experiments show that the proposed algorithm has thorough dehazing,natural color and suitable brightness,and has achieved satisfactory results in both subjective and objective evaluation.(4)Most of existing dehazing methods have high computational complexity and poor dehazing quality,therefore,a fast haze removal method is proposed based on haze density prior(HDP).HDP is a kind of statistics of real haze image.Haze density is reflected by the difference between maximum channel and minimum channel.The HDP enables effective distinction between thin haze images and dense haze images,as well as fast estimation of atmospheric light veil.In addition,an optimized method for estimating atmospheric light with mid-channel of haze image is proposed,which overcomes limitations of global atmospheric light.Our method can fast and efficiently recover clear images because there is no estimation of transmission map.Experiments show that the proposed method outperforms existing methods in haze removal,especially for dense haze images.Comparison of objective evaluation and running time show effectiveness and real-time of the proposed method.(5)Visibility restoration of images under haze and dust weather is essential in computer vision tasks.In this work,an algorithm for image visibility restoration based on color correction and composite channel prior(CCP)is proposed.Firstly,the color of a dust image is corrected by color compensation and white balance for blue and red channels.Haze and dust images are effectively distinguished by channel differences.Secondly,the composite channels are defined by simple multiplication and subtraction,and the composite channels of haze image and clear image have a very close pixel distribution.Then,according to atmospheric imaging rules,haze is the main factor that causes brightness difference of each composite channel.To eliminate the brightness difference,an adaptive gamma correction function based on haze density is proposed.In addition,as another important parameter of image restoration task,mean inequality and morphological operations are used to obtain more accurate mid-channel atmospheric light.Finally,a practical transmission map and high-quality clear image are obtained by atmospheric scattering models.Experimental results show that the proposed method is usable and practical.Our results have rich details and edges,especially for dust images.Model-based approaches require accurate estimation of transmittance and atmospheric light,while deep learning methods enable end-to-end restoration of haze images without using physical models.Therefore,aiming at the problems of error accumulation and insufficient extraction of related features introduced by existing learning methods,two end-to-end restoration structures are designed in this paper,and the specific work is as follows.(1)Convolutional neural network is developing rapidly in image processing.Most image dehazing algorithms only focus on dehazing and neglect the overall quality of dehazing image.This leads to problems such as loss of information and blurred texture.We p ropose a dehazing and enhancement convolutional neural network.Hazy image and clear image are obtained by encoding and decoding.Enhancement network is used to restore the texture and details of dehazing image.Experiments show that our method has excelle nt results in subjective evaluation and quality indexes.Haze can be removed more thoroughly,and images with clearer details and texture can be obtained.Proposed network solves the problems of information loss and texture blur.(2)Dehazing methods based on convolutional neural networks have made great progress in recent years,but there are still problems such as loss of detail,color distortion,and incomplete defogging.In response to the above problems,this paper proposes an end-to-end image dehazing algorithm based on ladder network and attention cross fusion.The network model includes three modules: feature extraction,feature fusion and image reconstruction.Feature extraction includes extraction of hazy image detail and c ontour features,which are extracted from different steps of the ladder network.Feature fusion module is intersection of attention mechanism fusion is realized,combined with adaptive residual processing to obtain the final fusion feature.Finally,in the image reconstruction module,dehazing image is obtained through non-linear mapping.Experiments show that the proposed method dehazing thoroughly,and the dehazing image has rich details,which effectively solves the problems of color distortion and loss of details,at the same time,ladder network overcomes the time-consuming problem of deep network training process.