Image Dehazing And Object Detection Methods Based On Binocular Reinforcement And Adaptive Features

Image dehazing and object detection are key technologies of intelligent driving and other applications.Intelligent driving systems including assisted driving and unmanned driving require clear images and accurate object detection.Image dehazing aims to deal with the degradation of image quality caused by fog and other factors.It is utilized in the assisted driving system to improve the driver’s field of vision and reduce the probability of traffic accidents in foggy days.Object detection aims to locate objects and classify them.It is mainly used in the unmanned driving system to detect objects such as vehicles,pedestrians,and traffic signs,so as to lay the foundation for subsequent decision-making and control tasks such as path planning.Furthermore,image dehazing methods can provide high-quality inputs for visual perception tasks such as object detection,and reduce the negative impacts of foggy weather on their performance.Based on mutual reinforcement of binocular images and adaptive feature learning,this thesis carries out researches on the visual perception of intelligent driving in adverse weathers from the following two aspects: For assisted driving,this thesis studies the image dehazing method based on mutual reinforcement of binocular images,and fully excavates their coupling relationships by using the multi-view information to improve the performance of image dehazing.For unmanned driving,the object detection method with adaptive feature learning is studied.Specifically,contextual information of small objects are selectively extracted to improve the accuracy of small object detection,and features are adaptively aligned to improve the location accuracy for the cascade detector.The main works and innovations are as follows:(1)In terms of the research on image dehazing methods for assisted driving,two methods are proposed,which are based on multi-view information provided by binocular images.This thesis proposes a stereo correlation mechanism based dehazing method.It learns the horizontal one-dimensional correlation between left and right views to predict transmission maps related with depth,and restores images efficiently according to the atmospheric scattering model.Moreover,this thesis proposes a progressive dehazing method with feature decoupling.It is a cascaded network directly dehazing images and enhancing images in a coarse-to-fine manner.Experiments demonstrate that the above two dehazing methods can enhance the quality of binocular images and obtain clear and natural visual results.At the same time,they can improve the detection accuracy of various detection methods including the following proposed 2D object detection method and a stereo-based 3D object detection method in foggy scenes.(2)In the aspect of small object detection,small objects occupy fewer pixels in images and have less discriminative information,which leads to low accuracy for small object detection.The context of small objects plays an important role in inferring the existence of small objects.It is the key of improving the performance of small object detection to extract efficient context.However,existing methods introduce information irrelevant of objects and noises when extracting contextual information.To solve the problem,this thesis proposes a selective context encoded object detection approach to selectively extract the context.Specially,a multi-branch context encoding module is proposed and extracts multi-scale contextual information to enhance the feature discrimination of small objects.A triple attention module is constructed and corporately learns global-level,channel-level,and spatial-level attentions to conduct selective context fusion,which makes full use of the context and filters out noisy information.Experimental results on the large-scale challenging MS COCO dataset demonstrate that the proposed object detection approach can significantly improve the accuracy of small objects.(3)In the localization issue of object detection,existing single-stage object detection methods improve the location accuracy by repeatedly regressing the positions of objects using cascade structures.However,the inconsistency between the feature sampling positions and the object positions limits the further improvement of location accuracy in cascade prediction.To solve the above problem,this thesis proposes an object detection method with feature refinement and adaptive alignment.An adaptive feature alignment module is designed.It learns objectness maps to restrict the positions of existing objects and utilizes deformable convolutions to realize the adaptive alignment of features with the object locations.A multi-branch shallow feature module is proposed and introduces shallow refined information of objects to improve the localization power of features.Experiments on the MS COCO dataset show that the proposed method effectively improves the location accuracy with negligible overheads.