计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (24): 238-247.DOI: 10.3778/j.issn.1002-8331.2306-0341

• 图形图像处理 • 上一篇    下一篇

基于全局响应的多级融合监督显著性目标检测

陈慧,彭力   

  1. 1.江南大学 物联网工程学院,江苏 无锡 214122
    2.无锡职业技术学院 物联网技术学院 江苏 无锡 214121
  • 出版日期:2023-12-15 发布日期:2023-12-15

Multi-Level Fusion Supervised Saliency Object Detection Based on Global Response

CHEN Hui, PENG Li   

  1. 1.School of Internet of Things Engineering, Jiangnan University, Wuxi, Jiangsu 214122, China
    2.Internet of Things Technology College, Wuxi Institute of Technology, Wuxi, Jiangsu 214121, China
  • Online:2023-12-15 Published:2023-12-15

摘要: 显著性目标检测是自主寻找图像视频中最具显著性物体的过程。针对目前常见的尺度变化和背景误判等问题,现有的显著性目标检测方法主要从特征融合、注意力机制和深度监督等角度进行优化以提高网络的检测能力。所提的基于全局响应的多级融合算法基于以上优化方向,主要通过轮廓信息与高级语义交融学习、补充目标结构特征以及抑制预测噪声来增强网络对显著目标的表征能力,同时提高了网络对目标尺度变化的感知能力以及对显著特征的敏感度。全局响应模块的构建强调了图像的全局特性,计算和判断了图像在空间和通道的响应值和不同位置的显著性,此举有效滤除了浅层背景噪声,使网络更快地锁定显著区域,提高学习效率。在通用指标上,实验数据表明了所提算法的优越性和高效性。

关键词: 显著性目标检测, 深度学习, 卷积神经网络, 特征融合, 注意力机制

Abstract: Saliency object detection is the process of autonomously locating the most salient objects in images and videos. In response to common challenges such as scale variations and background misjudgment, existing saliency object detection methods have primarily focused on feature fusion, attention mechanisms, and deep supervision to enhance the detection capabilities of networks. This paper proposes a global response-based multi-level fusion algorithm that builds upon these optimization directions. It enhances the network’s representation capabilities of salient objects by integrating contour information with high-level semantic feature, supplementing target structural features, and suppressing prediction errors. Moreover, it improves the network’s sensitivity to scale variations and salient region. The global response module emphasizes the global characteristics of images, computing and evaluating the responsiveness of images in spatial and channel dimension as well as the saliency at different locations. This effectively filters out shallow background noise, allowing the network to quickly identify salient regions and improve learning efficiency. Experimental results using general metrics demonstrate the superiority and efficiency of the proposed algorithm.

Key words: salient object detection, deep learning, convolutional neural network, feature fusion, attention mechanism