Research On Haze Classification Based On Knowledge-Driven And Domain Adaptation

Due to the influence of weather and human factors,the phenomenon of smog has become more common and seriously affects people’s lives.Haze weather will degrade the image quality obtained by image acquisition equipment,which will affect the safety and accuracy of vision computing systems such as unmanned driving.Different concentrations of haze bring different degrees of difficulty to computer vision tasks in the field of deep learning such as object detection and image segmentation on images.Therefore,it is particularly important to enable deep learning models to learn different degrees of haze features in images through training.At present,there are two types of haze classification methods.One class of methods is the haze classification based on traditional numerical statistics.This kind of method uses image numerical features and mathematical analysis tools to make the haze classification process relatively simple,but the deep learning model cannot learn the haze features of the image,so this kind of method cannot be applied to computer vision tasks in the field of deep learning.One class of methods is the haze classification by using support vector machines.This kind of method relies heavily on a large number of high-quality labeled data,so it cannot accurately identify the haze features of the image in the real use process.In view of the above problems existing in the existing methods,this thesis proposes two new ideas for haze image classification and constructs an image dataset for research in the field of haze.Therefore,the research content of this thesis is mainly composed of the following three aspects:(1)A knowledge-driven haze classification method is proposed.The most direct and effective way to improve the performance of haze classification model is to acquire enough images with real haze features.Due to the uncertainty of haze weather,the real haze image data is scarce.A large amount of artificial data can be obtained by adding different degrees of haze to real images,but artificial data and real data have different distribution domains,so they cannot be directly applied to real image classification training.This method uses the ability of the style transfer network to transfer the haze features of real images to artificial data to obtain a large amount of data with real haze features for the training of the haze classification model.This method provides a new solution for the small sample classification task.(2)A domain-adaptive-based haze classification method is proposed.The haze classification method based on domain adaptation takes artificial data and real data as the source domain and target domain,respectively,and maps the feature knowledge of the two domains to the latent style space through feature extraction.The objective function obtained by training artificial data in latent style space is applied to real images,thereby improving the classification performance of haze classification model for real haze images.(3)A multi-domain foggy image dataset(MDFIDA)is constructed for related research in the field of haze images.Affected by the uncertainty of haze weather,the cost of obtaining data both geographically and temporally is huge,so the number of real haze images is scarce and there is a problem of uneven distribution among various haze images.The multi-domain foggy image dataset consists of three parts,namely the foggy road image dataset(FRIDA)generated by Tarel et al.,the dataset obtained by artificially adding different degrees of haze to clear images(SIM),and a dataset consisting of a small number of real images(REAL).