计算机科学与探索

• 学术研究 •    

融合卷积特征的清晰边缘检测研究

王兵,黄刚,张兴鹏   

  1. 西南石油大学 计算机科学学院,成都 610500

Research on crisp edge detection based on fusion of convolutional features

WANG Bing, HUANG Gang, ZHANG Xingpeng#br#   

  1. School of Computer Science, Southwest Petroleum University, Chengdu 610500, China

摘要: 受益于卷积神经网络(Convolution Neural Networks, CNNs),边缘检测在多个基准数据集上都已经超过人类水平。但这类算法无法保证边缘的清晰性和定位的准确性。为获取细化清晰、有效抑制背景纹理、定位准确的目标边缘图,本文提出了一种融合卷积特征(Fuse Convolutional Features, FCF)的清晰边缘检测算法。该算法以VGG16作为卷积特征提取主干网络,将不同阶段的卷积特征上采样后特征融合,并通过本文所设计的细化融合模块(Refine Fusion Block, RFB)来获得清晰的边缘图。RFB使用多个归一化细化块(GroupNorm Refine Block, GRB)来细化得到的边缘图。此外,为平衡边缘像素和非边缘像素,本文还提出一个细化骰子损失函数(Refine Dice Loss, RD)。在BSDS500数据集上,本文所提的方法将HED、RCF等深度边缘检测器的F-score(ODS)分别提高了2.8%和2.1%;当不使用非极大值抑制(non-maximal suppression, NMS)进行边缘检测评估时,F-score(ODS)、F-score(OIS)分别达到0.801和0.816,超过了其他算法。

关键词: 清晰边缘检测, 融合卷积特征, 细化骰子损失, 卷积神经网络

Abstract: Benefiting from Convolution Neural Networks (CNNs), edge detection has surpassed human performance on several benchmark datasets. However, such algorithms cannot guarantee the crispness of edges and the accuracy of positioning. In order to obtain the target edge map that is refined and clear, effectively suppresses background texture, and locates accurately, a crisp edge detection algorithm based on Fusionof Convolutional Features (FCF) is proposed in this paper. The algorithm uses VGG16 as the backbone network for convolutional feature extraction, and fuses the convolutional features at different stages after upsampling, and obtains a crisp edge map through the Refine Fusion Block (RFB) designed in this paper. RFB uses multiple GroupNorm Refine Blocks (GRB) to refine the resulting edge map. In addition, to balance edge pixels and non-edge pixels, this paper also proposes a Refine Dice Loss (RD) function. On the BSDS500 dataset, the method proposed in this paper improves the F-score (ODS) of deep edge detectors such as HED and RCF by 2.8% and 2.1%, respectively; When edge detection evaluation is performed without non-maximal suppression (NMS), the F-score (ODS) and F-score (OIS) reach 0.801 and 0.816, respectively, outperforming other algorithms.

Key words: crisp edge detection, fuse convolutional features, refine dice loss, CNNs