Scholarlink

<p>Using onboard vision collection and recognition technologies, the visual assistance driving system for civil aviation aircraft captures information about the surrounding threat scenario to assist pilots in making driving judgments. However, nighttime scene images deteriorate, leading to low visual clarity and decreased scene object detection accuracy. Current object detection algorithms are unable to simultaneously enhance nighttime visual clarity and accurately detect scene objects. To address this problem, a convolutional neural network named VnDark based on iterative feature enhancement learning is proposed. A feature enhancement sub-network is designed based on an iterative light enhancement curve mechanism, which employs a U-shaped encoder-decoder structure to produce sharp edge features from nighttime images, enhancing scene visibility. A multi-task feature collaborative optimization strategy is designed, which shares the enhanced features generated by the feature enhancement sub-network with the object detection sub-network and employs a multi-task weight learning algorithm to optimize the weights for nighttime image enhancement, object classification, and localization tasks. This strategy improves the accuracy of scene object detection. A nighttime scene object dataset containing real-world and simulation data is established to validate VnDark’s performance. Experimental results indicate that the network can accurately detect scene objects while simultaneously enhancing the visibility of dark images. In the enhanced images, the original dark area objects are clearly visible, making them easier for human observation. The average precision of object detection reached 78.5%. Additionally, the proposed algorithm achieved an average precision of 69.3% on the public dataset Exclusively Dark, surpassing typical general object detectors such as Faster R-CNN, SSD, YOLOv7, and YOLOv5s. Both precision and speed of the proposed algorithm surpass that of typical nighttime object detection networks DK-YOLOv5 and DENet.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

VnDark: iterative enhanced feature learning network for nighttime airfield object detection

  • 期刊论文

  • 发表日期：

    2025年5月12日

查看全文
  • Qijun Luo,
  • Xuerui Zhang,
  • Xigao Fang,
  • Zhiming Zheng,
  • Tao Zhang

摘要

Using onboard vision collection and recognition technologies, the visual assistance driving system for civil aviation aircraft captures information about the surrounding threat scenario to assist pilots in making driving judgments. However, nighttime scene images deteriorate, leading to low visual clarity and decreased scene object detection accuracy. Current object detection algorithms are unable to simultaneously enhance nighttime visual clarity and accurately detect scene objects. To address this problem, a convolutional neural network named VnDark based on iterative feature enhancement learning is proposed. A feature enhancement sub-network is designed based on an iterative light enhancement curve mechanism, which employs a U-shaped encoder-decoder structure to produce sharp edge features from nighttime images, enhancing scene visibility. A multi-task feature collaborative optimization strategy is designed, which shares the enhanced features generated by the feature enhancement sub-network with the object detection sub-network and employs a multi-task weight learning algorithm to optimize the weights for nighttime image enhancement, object classification, and localization tasks. This strategy improves the accuracy of scene object detection. A nighttime scene object dataset containing real-world and simulation data is established to validate VnDark’s performance. Experimental results indicate that the network can accurately detect scene objects while simultaneously enhancing the visibility of dark images. In the enhanced images, the original dark area objects are clearly visible, making them easier for human observation. The average precision of object detection reached 78.5%. Additionally, the proposed algorithm achieved an average precision of 69.3% on the public dataset Exclusively Dark, surpassing typical general object detectors such as Faster R-CNN, SSD, YOLOv7, and YOLOv5s. Both precision and speed of the proposed algorithm surpass that of typical nighttime object detection networks DK-YOLOv5 and DENet.