| Object detection in real foggy scenes is an important research topic in the field of computer vision.The main challenge is that the images collected in real foggy scenes have some problems such as object occlusion,blur and color distortion,which lead to the loss of object details.At present,object detection in complex scenes has a wide range of application prospects in the fields of video surveillance,unmanned driving,military applications and smart cities,etc.Therefore,this research is of great scientific significance and very challenging work.Although the general object detection has made great progress,the existing research still lacks the discussion on object detection methods in the complex hazy scene.And there are urgent problems to be solved:scarcity of dataset and singularity of researching method.This thesis focuses on these two issues:firstly,to address the problem of scarcity of dataset of object detection in real-world hazy condition,this thesis designs a systematic hazy image synthesis method.Images in existing MS COCO are synthesized into hazy images in a low-cost way,and the object detection dataset in hazy scene,that is S-COCO,is established.Then,aiming at the difficulty of object detection in real-world hazy scene,this thesis discusses the solution from three research directions.The research contents and main contributions of this thesis include the following aspects:Firstly,we propose object detection model based on joint optimization of model.Most of the current object detection methods in the hazy scene take image dehazing as a pre-processing operation before model training,but artifacts,detail missing and color distortion generated in the process of image restoration seriously affect the accuracy of the downstream detection task.For this problem,the cascade optimization framework of image dehazing and object detection joint learning can effectively learn the structural details and color features recovered in image dehazing through the joint optimization learning of dehazing model and object detection model,so as to improve the accuracy of object detection in real fog scenes.Secondly,we propose object detection model based on weighted normalization attention mechanism.This thesis propose weighted normalization attention mechanism to deal with the problem of the style of hazy scene with the representation learning.In addition,the appearance invariant feature is introduced to make the network filter the interference information irrelevant to the object and learn the appearance invariant feature representation.Weighted normalization mechanism by using instance normalization to delete the statistics information of style which affect the appearance of image.Moreover,it extracts the object-specific features from the deleted residual information after instance normalization,and integrates them with the features after instance normalization to make the network pay attention to the discriminative features related to the object,thus improving the generalization performance of the object detection network in hazy scene.Thirdly,we propose Dual Complementary Learning between Image-Instance Alignment for Domain Adaptive Object Detection.Because most of the current object detection research in hazy scene is carried out on the training set and test set from different sources,the learning problem of inconsistent feature distribution is easy to occur in the process of model training.Aiming at this problem,this thesis put forward Dual Complementary Learning between Image-Instance Alignment for Adaptive Object Detection(DCLIA),and through the dual complementary image-level alignment to generate adaptation process of network.In the meantime,we introduce cross domain prototype alignment to establish the tight relationship between categories and domains and enhance intra-class compactness and inter-class separability.The experimental results show that this method can effectively improve the adaptive ability between similar domains,that is,the domain adaptive from haze-free scene to hazy scene.In addition,experiments on Pascal VOC,CLIPART and SIM 10K show that the domain adaptive object detection method based on Dual Complementary Learning between Image-Instance Alignment also improves the adaptability between dissimilar domains(from natural images to unnatural images). |
