摘要
为了高效地实现对边缘、边界和目标轮廓的同时检测,提出了一种基于新型网络的多层次边缘检测器(Multi-level edge detector, MLED)。与一般的多任务网络不同,MLED使用特征转移模块(Feature transfer module, FTM)和空间交集机制(Spatial intersection, SI)来取代传统的多解码器设计,在网络的不同层进行不同层次的边缘检测,有效地降低了模型复杂度。FTM实现了高级特征的自顶向下的传播,有利于提高低层次边缘检测的稳定性。SI能够提高高层次边缘检测的空间精度,以弥补下采样带来的精度损失。另外,Canny算法被用来辅助低级边缘检测的训练,解决了低级边缘标签缺失的问题。实验表明,MLED能够在较低的参数量和运算量的情况下实现高质量的多层次边缘检测,有更强的泛化能力。
In order to efficiently detect edges,boundari In order to efficiently detect edges,boundaries and target contours at the same time,a multi-level edge detector(MLED)based on a new network is proposed.Different from general multi-task networks,MLED uses Feature transfer module(FTM)and spatial intersection(SI)to replace the traditional multi-decoder design,and performs edge detection at different levels at different layers of the network,effectively reducing the complexity of the model.FTM realizes top-down propagation of high-level features,which is conducive to improving the stability of low-level edge detection.SI can improve the spatial accuracy of high-level edge detection to compensate for the accuracy loss caused by downsampling.In addition,the Canny algorithm is used to assist the training of low-level edge detection,which solves the problem of missing low-level edge labels.Experiment results show that MLED can achieve high quality multi-level edge detection with lower parameter number and computational cost,and has stronger generalization ability.
作者
李正
贺赛先
LI Zheng;HE Saixian(Electronic Information School,Wuhan University,Wuhan 430079,China)
出处
《激光杂志》
CAS
北大核心
2023年第3期62-68,共7页
Laser Journal
关键词
边缘检测
边界检测
目标轮廓检测
全卷积网络
edge detection
boundary detection
object contour detection
fully convolutional network