基于金字塔特征与边缘优化的显著性对象检测-《郑州大学学报(工学版)》

文章信息/Info

Title:: Salient Object Detection Based on Pyramid Features and Edge Optimization

作者:: 田旭¹; 彭飞¹; 刘飞¹; 陈庆文²; 闫馨宇³; 4; 国网青海省电力公司经济技术研究院;中国电建集团西北勘测设计研究院有限公司;天津大学智能与计算学部;天津大学天津机器学习重点实验室;

Author(s):: TIAN Xu¹; PENG Fei¹; LIU Fei¹; CHEN Qingwen²; YAN Xinyu³; 4; 1.State Grid Qinghai Electric Power Company Economic and Technical Research Institute, Xining 810000, China;
2.Northwest Engineering Corporation Limited, Xi′an 710065, China;
3.College of Intelligence and Computing, Tianjin University, Tianjin 300350, China;
4.Tianjin Key Lab of Machine Learning, Tianjin University, Tianjin 300350, China

Keywords:: salient object detection; multi-scale feature extraction; fully convolutional networks; edge feature extraction; deep learning

摘要:: 针对图像显著性对象检测领域中多尺度特征提取不充分’对象边缘模糊等问题!提出了一个端到端的基于注意力嵌入的金字塔特征以及渐进边缘优化的显著性对象检测模型" 首先!设计了由多个扩张卷积构成的注意力嵌入的密集空洞金字塔模块%DPJDH:& !在不减小特征分辨率的前提下!得到丰富且有效的多级多尺度特征#其次!为了解决显著性对象边缘模糊的问题!提出了渐进边缘优化模块% 0PI:& !在特征恢复分辨率的过程中逐步补充空间细节信息!使模型检测出的显著对象能够拥有清晰的边缘轮廓" 在J[R0)RP’P+00J’J[R)I:2IK’Yd[)Q0’HD0+DZ)0 $ 个显著性领域公开的数据集上与其他"% 种已有的先进方法在’ 个常用指标下进行了比较!结果表明$所提方法能够得到更加准确’边缘更加清晰的显著性结果" 此外!自对比实验也充分证明了提出的注意力嵌入的密集空洞金字塔模块和渐进边缘优化模块的有效性"

Abstract:: To solve the problems of insufficient multiscale feature extraction and object edge blur in image-based salient object detection, an end-to-end salient object detection model was proposed based on attention embedding pyramid feature and stepped edge optimization. Firstly, the attention embedded dense atrous Pyramid Module (AEDAPM) composed of multiple dilated convolutions was designed to obtain rich and effective multi-level multi-scale features without reducing the feature resolution； Secondly, in order to solve the problem of blurring the edges of salient objects, a stepped edge optimization module (SEOM) is proposed, which gradually supplements spatial detail information during the process of feature restoration resolution, so that the salient objects detected by the model could have clear edge contours. The method in this paper was compared with 12 state-of-the-art saliency methods under 3 common indicators on 5 public datasets, such as DUTS-TE, ECSSD, DUT-OMRON, HKU-IS, and PASCAL-S. The experimental results show that the method proposed in this paper can obtain more accurate and clearer saliency results. In addition, the ablation study also fully proved the effectiveness of the AEDAPM and the SEOM proposed in this study.