计算机工程与应用 ›› 2024, Vol. 60 ›› Issue (18): 256-265.DOI: 10.3778/j.issn.1002-8331.2306-0298

• 图形图像处理 • 上一篇    下一篇

采用特征优化和引导的显著目标检测研究

吴文介,王丰   

  1. 广东工业大学 信息工程学院,广州 510006
  • 出版日期:2024-09-15 发布日期:2024-09-13

Research on Salient Object Detection Using Feature Optimization and Guidance

WU Wenjie, WANG Feng   

  1. School of Information Engineering, Guangdong University of Technology, Guangzhou 510006, China
  • Online:2024-09-15 Published:2024-09-13

摘要: 针对目前深度图存在对比度不明显和预测图边界模糊等问题,提出了一种新型显著目标检测网络模型。该模型包括特征优化模块和特征引导模块。为了降低低质量深度图的负面影响,并精确地突出显著目标,在特征优化模块对深度图的各层特征进行混合注意力计算并进行双向融合。为解决边界模糊问题,在特征引导模块利用引导融合的方式引入低层特征来精细化目标边界。在解码阶段,引入不增加模型参数的权值计算方法,计算RGB特征和深度特征对最终预测的贡献比重。通过与近年来十二种先进方法进行的对比实验表明,所提算法模型在NJU2K、NLPR、DES、SIP、STERE和LFSD测试数据集上具有更优秀的检测性能,其中在SIP数据集上,提出的模型与第二名相比,最大F值提升了1.3%,平均F值提升了1%,E-measure提升了1.7%,S-measure提升了1.5%,消融实验证明了所提模块的有效性。

关键词: 深度图, 显著目标检测, 混合注意力, 特征融合

Abstract: Aiming at the problems of weak contrast in depth map and fuzzy boundary in prediction map, this paper proposes a new salient object detection model. It includes feature optimization module and feature guidance module. In order to reduce the negative effects of low quality depth maps and precisely highlight salient objects, it applies mixed attention to compute the feature of each layer, and then the bidirectional fusion is used to fuse in the feature optimization module. In the feature guidance module, it introduces low-level features to refine the object boundary by means of guidance and fusion. In the decoding stage, it measures the contribution of RGB features and depth features by the weight calculation method without increasing the model parameters. Experiments compared with twelve advanced methods in recent years show that the proposed algorithm model has better detection performance on NJU2K, NLPR, DES, SIP, STERE and LFSD test datasets. On the SIP dataset, the performance of the model is improved by 1.3% in maximum F-value, 1% in average F-value, 1.7% in E-measure, 1.5% in S-measure compared with the second place, and ablation experiments show the effectiveness of the proposed module.

Key words: depth map, salient object detection, mixed attention, feature fusion