东北大学学报(自然科学版) ›› 2024, Vol. 45 ›› Issue (7): 936-943.DOI: 10.12068/j.issn.1005-3026.2024.07.004

• 信息与控制 • 上一篇    下一篇

一种基于编解码结构的多尺度边缘检测方法

田岸霖1, 雷为民1(), 张鹏1,2, 张伟1   

  1. 1.东北大学 计算机科学与工程学院,辽宁 沈阳 110169
    2.沈阳二一三电子科技有限公司,辽宁 沈阳 110027
  • 收稿日期:2023-03-07 出版日期:2024-07-15 发布日期:2024-10-29
  • 通讯作者: 雷为民
  • 作者简介:田岸霖(1998-),男,山西吕梁人,东北大学硕士研究生
  • 基金资助:
    2022年辽宁省“揭榜挂帅”科技重大专项(2022JH1/10400025);国家重点研发计划项目(2018YFB1702000);中央高校基本科研业务费专项资金资助项目(N2216010)

A Multi-scale Edge Detection Method Based on Encoder-Decoder

An-lin TIAN1, Wei-min LEI1(), Peng ZHANG1,2, Wei ZHANG1   

  1. 1.School of Computer Science & Engineering,Northeastern University,Shenyang 110169,China
    2.Shenyang 213 Electronic Technology Co. ,Ltd. ,Shenyang 110027,China.
  • Received:2023-03-07 Online:2024-07-15 Published:2024-10-29
  • Contact: Wei-min LEI
  • About author:LEI Wei-minE-mail:leiweimin@ise.neu.edu.cn.

摘要:

针对传统边缘检测算法易出现边缘线断裂、不连续的情况以及基于深度学习的方法无法保证边缘清晰性和准确性且容易丢失边缘细节信息的问题,提出了一种基于编解码结构的轻量级的边缘检测方法.通过编码器提取图像的边缘特征,在解码端恢复编码器下采样时丢失的边缘信息,编解码器之间采用跳跃连接方法实现低层特征和高层特征之间的融合,采用具有注意力机制的深监督模块,进一步学习多尺度多层次的边缘特征生成精细的图像边缘.该网络模型在BSDS500S数据集上进行训练,实验结果表明,本文方法ODS(optimal dataset scale)与OIS(optimal image scale)分别达到0.808和0.830,在 GTX 1060机器上帧率达到60帧/s,超过了基于卷积神经网络的主流边缘检测方法,具有较好的效果.

关键词: 边缘检测, 编解码, 跳跃连接, 注意力机制, 深监督

Abstract:

Traditional edge detection algorithms often suffer from fractured and discontinuous edge lines, while deep learning?based approaches fail to ensure edge clarity and accuracy, often leading to the loss of edge details. To address these issues, a lightweight edge detection method based on encoder?decoder structure is proposed. The encoder extracts the edge features of the image, recovers the edge information lost when sampling under the encoder at the decoding end, uses the jump connection method between the codecs to achieve the fusion between low?level features and high?level features, and uses the deep supervision module with attention mechanism to further learn multi?scale and multi?level edge features to generate fine image edges. The network model is trained on the BSDS500S dataset. Experimental results show that the ODS and OIS of the proposed method reach 0.808 and 0.830 respectively, and the frame rate on the GTX 1060 machine reaches 60 frames/second, which exceeds the mainstream edge detection methods based on convolutional neural networks, thus showing its effectiveness.

Key words: edge detection, encoder?decoder, skip connection, attention mechanism, deep supervision

中图分类号: