用于手绘检索的伪监督边缘检测算法

PSEUDO-SUPERVISED EDGE DETECTION FOR SKETCH-BASEDIMAGE RETRIEVAL

基于深度学习的边缘检测算法需要大量的标注,这阻碍了边缘检测的推广应用。因此提出一种伪监督边缘检测算法,能够在无标注的手绘图像数据集上提取图像边缘。算法分成三部分,包括伪监督标签生成、多尺度边缘检测网络和特征增强模块。伪监督标签为多尺度监督的边缘检测网络生成监督信息,特征增强模块可以弥补伪监督带来的信息丢失。该算法比现有边缘检测算法提取的边缘更完整,在手绘数据集QMUL-Shoe和QMUL-Chair上可以提高1%~6%的检索精度,对需要边缘检测的所有领域都有启发性意义。

Edge detection algorithm based on deep learning needs large amounts of annotation,which hinders the popularization and application of edge detection.This paper proposes a pseudo-supervised edge detection algorithm,which can extract edges from unannotated sketch datasets.The algorithm was divided into three parts,including pseudo-supervised label generation,multi-scale edge detection network and feature enhancement module.Pseudo-supervised label generated supervision information for multi-scale supervised edge detection network,and the feature enhancement module could make up for the information loss caused by pseudo supervision.Compared with the existing edge detection algorithms,the extracted edges were more complete and the retrieval accuracy could be improved by 3%~5%on sketch datasets QMUL-Shoe and QMUL-Chair.As far as we know,this is the first proposed pseudo-supervised edge detection algorithm,which is instructive for all fields that need edge detection.