Study On Vehicle Detection In Foggy Environment Based On Generative Adversarial Network And Domain Adaptation

The object detection assignment based on the camera will face huge challenges under the hazy environment.The reduced visibility will affect the image quality and cause the decline of the accuracy in the object detection algorithm,which will affect the subsequent decision-making control system.In the hazy environment,the current object detection algorithms have the following problems: they are trained under normal weather conditions,and the high cost of collecting and labeling large-scale hazy data makes it hard to directly use the large-scale hazy data for training.To overcome those challenges,this paper proposes a feature-supervised algorithm based on generative adversarial networks for image dehazing and an object detection algorithm based on the dehazing network and domain adaptation to improve the accuracy of object detection under hazy environment.The main contents of this paper are as follows:A feature-supervised dehazing algorithm based on the generation adversarial network is designed,and the accuracy is improved through a two-stage detection scheme of image restoration.A feature-supervision mechanism is introduced in the dehazing algorithm.Specifically,paired clean and hazy images are fed into the network and the prior knowledge contained in the clean image is used to constrain the restoration process.At the same time,to further strengthen the constraints on the training process,introducing adversarial loss,perception loss,style loss and feature regularization loss to constrain the training process from different levels.The experimental results show that the proposed algorithm achieves better performance on both synthetic public datasets and real-world hazy datasets.The application experiment of object detection shows that the two-stage way can effectively improve the detection accuracy.To overcome the problems of unstable,and longtime consumption of two-stage object detection scheme,this paper proposes a one-stage algorithm based on the featuresupervised dehazing network and domain adaptation,which includes image translation module,feature extraction module and detection module.The feature supervised dehazing algorithm and its inverse transformation are used to the style transfer between two domains,which can reduce the distribution difference between normal weather and hazy weather.In addition,the domain classifier is introduced in the high-dimensional and low-dimensional of the feature extraction module,and the feature difference between domains is narrowed by adversarial training.The experimental results show that the detection accuracy of the proposed method is significantly improved compared to the method without domain adaptation,and the accuracy is almost the same as the two-stage way,but the operating efficiency of the model is greatly improved.Furthermore,the detection model adapted to hazy weather is also suitable for normal weather conditions.