Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (16): 187-195.DOI: 10.3778/j.issn.1002-8331.2211-0132

• Graphics and Image Processing • Previous Articles     Next Articles

Cross-Domain Adaptive Object Detection Based on CNN Image Enhancement in Foggy Conditions

GUO Ying, LIANG Ruilin, WANG Runmin   

  1. The Joint Laboratory for Internet of Vehicles of Ministry of Education-China Mobile Communications Corporation, Chang’an University, Xi’an 710018, China
  • Online:2023-08-15 Published:2023-08-15

基于CNN图像增强的雾天跨域自适应目标检测

郭迎,梁睿琳,王润民   

  1. 长安大学 车联网教育部-中国移动联合实验室,西安 710018

Abstract: This paper proposes a cross-domain adaptive object detection method based on CNN(convolutional neural networks) image enhancement, which addresses the problem that the accuracy of target detection algorithm decreases due to the low quality of images captured by the visual perception system of autonomous vehicles in foggy conditions. An end-to-end target detection network is constructed, which integrates DIP(digital image processing) and CNN adaptive image enhancement module, to improve the image quality in foggy weather through a small CNN parameter predictor that learns enhancement parameters adaptively. Furthermore, a multi-scale DA(domain adaptive) module is connected to YOLOv4 backbone, which through adversarial training, reduces the domain differences caused by foggy conditions and increases the accuracy of target detection in foggy weather. In the stage of training, CNN, DA and YOLOv4 are learned in an end-to-end manner. In the stage of detection, both CNN and DA modules are removed, only using the images that pre-training weights have adaptively detected in normal and foggy weather, which will not increase the complexity of the original network and thus satisfy the timing requirement of autonomous vehicles. An experiment based on the open dataset Foggy Cityscapes indicates that the proposed method can significantly enhance the image quality in foggy weather, increasing the average accuracy of target detection by 10.4%, which effectively enhances the target detection ability of autonomous vehicles in foggy conditions.

Key words: autopilot, object detection, image enhancement, convolutional neural networks(CNN), domain adaptation, YOLOv4

摘要: 针对自动驾驶车辆视觉感知系统在雾天条件下捕获图像质量较低,造成目标检测算法精度下降的问题,提出一种基于卷积神经网络(convolutional neural networks,CNN)图像增强的跨域自适应雾天目标检测方法。构建一个端到端目标检测网络,融合数字图像处理技术(digital image processing,DIP)和CNN的自适应图像增强模块,通过小型CNN参数预测器自适应学习增强参数,提升雾天图像质量;进一步地,将多尺度领域自适应(domain adaptive,DA)模块与YOLOv4主干网络相连,通过对抗训练减少由雾天造成的领域差异,提高雾天目标检测精度。在训练阶段,所提方法以端到端的方式学习CNN、DA模块以及YOLOv4,而在目标检测阶段将移除CNN及DA模块,仅使用预训练权重在正常天气和雾天天气自适应地检测图像,不会增加原有网络复杂性,从而保证自动驾驶车辆的实时性要求。在公开数据集Foggy Cityscapes上的实验表明,采用所提方法使雾天图像质量显著增强,目标检测平均精度提升了10.4%,有效提升了雾天条件下自动驾驶车辆对目标的识别能力。

关键词: 自动驾驶, 目标检测, 图像增强, 卷积神经网络(CNN), 领域自适应, YOLOv4