摘要
针对水下距离感知任务真实训练数据缺乏,水下目标感知任务目标模糊、密集、多尺度的问题,提出一种基于颜色校正和深度信息去雾的水下视觉感知系统。设计了一种改进的融合增强方法,并建立了一个水下单目图像数据集,以解决距离感知任务数据不足的难点。设计了一种基于深度信息的去雾方法,结合水下成像模型对图像进行去雾处理,提升图像质量。设计了一种基于中心点检测的通道重排网络,将卷积神经网络中浅层的详细特征完全集成到深层中,且无需锚框,增强对小目标、密集目标的特征提取能力。实验表明,该系统可从水下图像中恢复真实陆地色彩,准确感知水下场景相对距离,并实现域内和跨域高精度目标感知,在URPC数据集上取得了78.2%的域内目标检测精度,比基准CenterNet高出4.6%,在UTTS数据集上取得81.5%跨域目标检测精度,证明了该系统的有效性。
The core of an Autonomous Underwater Vehicle(AUV)lies in its ability to accurately perceive objects and the surrounding environment.With advancements in underwater optical vision sensor technology,optical imaging for environment perception is now feasible.Despite progress in object detection,underwater images'inherent degradation poses challenges.High underwater pressure complicates distance information acquisition,leading to limited training datasets.Moreover,the degradation and blurriness of underwater images often obscure object features.To enhance AUVs'capabilities in distance perception and scene reconstruction,research is increasingly focusing on precise localization and depth scene construction in underwater scenarios.To this end,this paper introduces an underwater visual perception system which incorporates color correction and depth information dehazing to overcome these challenges.Specifically,we propose an improved color correction method that combines white balance and adaptive histogram equalization for effective white balance and histogram adjustments to original images.This approach effectively mitigates the common issue of red artifacts in underwater images,thus rendering the images more realistic.Additionally,our method leverages white balance adjustments to enhance overall image contrast,thereby improving feature clarity.Moreover,to address the challenge of data insufficiency in underwater distance perception tasks,we have developed an improved fusion enhancement method.Through this approach,we establish an underwater monocular image dataset.Specifically,we collected a large number of underwater images from the Internet and enhanced them using the aforementioned image enhancement method.Building upon this,we integrated a monocular depth estimation network into our framework,where the depth estimation network is trained on the collected underwater images in an unsupervised manner.This approach provides depth map information,which is essential for subsequent image dehazing within the framework.Fur
作者
毛昭勇
刘楠
陈刚琦
侯冬冬
沈钧戈
MAO Zhaoyong;LIU Nan;CHEN Gangqi;HOU Dongdong;SHEN Junge(Unmanned System Research Institute,Northwestern Polytechnical University,Xi′an 710072,China;China Aerospace Science and Technology Innovation Academy,Beijing 100088,China;School of Marine Science and Technology,Northwestern Polytechnical University,Xi′an 710072,China;Henan Key Laboratory of Underwater Intelligent Equipment,713 Research Institute of China Shipbuilding Industry Corporation,Zhengzhou 450015,China)
出处
《光子学报》
EI
CAS
CSCD
北大核心
2024年第6期183-198,共16页
Acta Photonica Sinica
基金
中央基本科研业务费(No.5000220192)
陕西省自然科学基金(No.2022JM-206)
陕西省秦创原“科学家+工程师”(No.2022KXJ-006)
西安市科技计划-人工智能示范项目(No.21RGZN0008)。
关键词
目标检测
去雾
深度估计
颜色校正
水下图像
Object detection
Dehazing
Depth estimation
Color correction
Underwater image
