Research On Object Detection And Image Restoration Algorithms In Hazy Weather

Deep learning-based object detection technology has recently made significant progress and has a wide range of applications in robotics,autonomous driving,traffic surveillance,etc.However,in inclement weather,especially in haze weather,the scattering of light by particles such as water droplets and dust in the atmosphere not only reduces the quality of the image captured by the camera,but also affects the performance of the object detection task.Thus,there have been growing demands for recovering clear images from hazy images and improving the accuracy of object detection in haze weather.This thesis reviews the research progress of object detection in hazy weather,introduces the imaging principles of haze images,discusses the theories related to the atmospheric scattering model,convolutional neural networks,and learning-based object detection algorithms,and proposes two methods to improve object detection performance in hazy weather.The main contributions are as follows:Firstly,this thesis proposes a domain adaptation method for object detection in the frequency domain.Under unsupervised learning in the frequency domain,the image features between different domains are aligned to reduce the performance gap between hazy and clear weathers.The proposed method consists of two stages.In the first stage,it translates the annotated training data from the source domain to the target domain using unsupervised image-to-image translation.An adversarial domain adaptation is then applied to the object detection model to align the features of the translated data and the real data in the target domain.In light of the energy concentration property of the discrete cosine transform,the proposed algorithm conducts domain adaptation for object detection by processing only a few most significant frequency coefficients,reduces memory and computing resource consumptions,and alleviates the domain shift problem,which effectively improves the object detection performance under haze weather.Secondly,this thesis designs a holistic framework for image dehazing and object detection.By domain alignment and feature sharing,the framework can handle the two tasks at the same time.The framework consists of a detection branch,a dehazing branch,and a domain classifier for feature alignment.The detection branch learns the domain-invariant features of hazy and clear images in a semi-supervised manner by competing with the domain classifier.By virtue of feature alignment,we use a lightweight dehazing branch with only a few dilated convolutions to recover clear-scene images by re-using the intermediate features generated by the detector.The dehazing branch and the detection network benefit each other during training,and the detection performance under hazy weather has also been significantly improved.We also build a new outdoor dataset in the real scenes to evaluate the effectiveness of joint object detection and haze removal.Experimental results on synthetic images in RESIDE(REalistic Single Image DEhazing)dataset,the Foggy Cityscapes dataset,and the dataset constructed in this thesis,as well as real images in wild scenarios,demonstrate that the proposed algorithm has higher quality of dehazing images.The peak signal-to-noise ratio reaches 34 db,which exceeds second place by 13.3%;object detection accuracy in hazy weather is improved by more than 40%,and the performance is significantly better than the comparison algorithms.