FMLED:细粒度级多尺度特征表示的轻量级边缘检测方法  

Fine-grained Multi-scale Feature Representation Lightweight Network for Edge Detection

作　　者：杨红菊[1,2] 王昱蓉 YANG Hong-ju;WANG Yu-rong(School of Computer and Information,Shanxi University,Taiyuan 030006,China;Computational Intelligence and Chinese Information Processing of Ministry of Education,Shanxi University,Taiyuan 030006,China)

机构地区：[1]山西大学计算机与信息技术学院,太原030006 [2]山西大学计算智能与中文信息处理教育部重点实验室,太原030006

出　　处：《小型微型计算机系统》2023年第4期812-817,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61976128)资助;山西省高等学校科技创新项目(2019L0103)资助;山西省1331工程项目资助。

摘　　要：针对现有基于深度学习的轻量级边缘检测方法存在检测效果不佳的问题,本文提出一种细粒度级多尺度特征表示的轻量级边缘检测方法——FMLED.首先,以细粒度方式提取图像多尺度特征,得到更细致并具有全局性的特征图;其次,对特征图进行上采样,得到不同阶段的中间边缘图;最后,使用学习滤波器为每个像素分配不同权重并融合中间边缘图特征,使生成的边缘更加清晰准确.本文方法可用于任何边缘检测任务,而无需进行预训练或微调;并以较小的模型复杂度产生优质边缘.在BIPED、BSDS500和NYUD 3个数据集上的实验结果表明,FMLED框架在不增加计算负载的情况下,可以得到更高质量的边缘.Aiming at the problem that the existing lightweight edge detection methods based on deep learning have poor edge detection results,this paper proposes a lightweight edge detection method(FMLED)based on finegrained multi-scale feature representation.First,by extracting the multi-scale features of the image in a finegrained way,more detailed and global feature maps are obtained;secondly,the feature map is up-sampled to obtain the middle edge map at different stages;finally,the learning filter is used to assign different weights to each pixel and fuse the features of the middle edge maps,which makes the generated edges more clear and accurate.The method can be used for any edge detection task without pre-training or fine-tuning;and it can produce high-quality edges with less model complexity.The experimental results on the three datasets of BIPED,BSDS500 and NYUD show that the FMLED framework can detect higher quality edges without increasing the computational load.

关 键 词：深度学习 轻量级 多尺度表示 细粒度级 边缘检测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]