计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (22): 259-267.DOI: 10.3778/j.issn.1002-8331.2206-0392

• 图形图像处理 • 上一篇    下一篇

复杂场景下显著性目标检测注意力金字塔网络

方金生,陶余昊,朱古沛,陈彦佑   

  1. 1.闽南师范大学 计算机学院,福建 漳州 363000
    2.数据科学与智能应用福建省高校重点实验室(闽南师范大学),福建 漳州 363000
  • 出版日期:2023-11-15 发布日期:2023-11-15

Saliency Detection from Complex Background with Attention-Based Boundary-Aware Pyramid Pooling Network

FANG Jinsheng, TAO Yuhao, ZHU Gupei, CHEN Yanyou   

  1. 1.School of Computer Science and Engineering, Minnan Normal University, Zhangzhou, Fujian 363000, China
    2.Fujian Province Key Laboratory of Data Science and Intelligence Application(Minnan Normal University), Zhangzhou, Fujian 363000, China
  • Online:2023-11-15 Published:2023-11-15

摘要: 近年来,深度卷积神经网络在显著性目标检测中得到广泛关注和研究,并取得了重要进展。但当显著性目标处于复杂背景中时,当前算法的性能仍有待提高。提出一种利用边界感知和注意力机制的金字塔池化网络(attention-based boundary-aware pyramid pooling network,ABAPNet),用于复杂场景下的显著性目标检测。ABAPNet通过引入级联式通道注意力和空间注意力机制,采用特征金字塔网络架构获取多层次的语义特征,以高层特征信息来辅助浅层特征;再通过融合二进制交叉熵、结构相似性和联合交集的混合损失函数,可增强获取重要语义特征并且关注目标边界特征,从而引导网络从复杂背景中更好地检测目标。在5个公开数据集上的实验表明,ABAPNet在多个评价指标上均优于比较算法,达到最优性能。

关键词: 显著性目标检测, 注意力机制, 深度学习, 金字塔池化, 边界感知

Abstract: Deep convolution networks have been widely adopted in saliency detection in recent years and achieved state-of-the-art performance. A core issue for previous methods is the limited capability of detecting objects that are overwhelmed by complex background. Aiming to address this problem, a novel attention-based boundary-aware pyramid pooling network is proposed, namely ABAPNet, which consists of feature aggregation module, pyramid pooling module and cascading dual attention module. The proposed ABAPNet method transits high-level information to shallow layers to keep transferring important features between layers and obtains more semantic features. Hybrid loss by fusing binary cross entropy, structural similarity and intersection-over-union is employed to reduce the gradient of mixed loss of network output features, which guides the network to better segment boundaries from complex background. Experimental results on 5 public datasets show that the proposed ABAPNet outperforms the competitors.

Key words: saliency object detection, attention mechanism, deep learning, pyramid pooling, boundary-aware