激光与光电子学进展, 2019, 56 (15): 151502, 网络出版: 2019-08-05
基于深度残差网络与边缘监督学习的显著性检测 下载: 926次
Salient Object Detection Based on Deep Residual Networks and Edge Supervised Learning
机器视觉 显著性检测 边缘残差块 三分类模型 多尺度空洞卷积 machine vision saliency detection edge residual block three-classification model multi-scale atrous convolution
摘要
针对复杂背景下,图像显著区域显著值低和目标边缘表现模糊的问题,提出了基于深度残差网络和多尺度边缘残差学习的显著性检测方法。一方面提出了边缘残差块,使用边缘残差块在深度残差网络的基础上构建边缘监督网络,用于显著图边缘监督学习;另一方面,通过构建基于背景、前景和边缘的三分类模型,训练网络学习边缘特征,使目标边缘更加准确,同时输出采用空洞卷积构建多尺度空洞卷积单元,多尺度地对全局信息进行特征整合提取。最后,将提出的算法在数据集SED2和ECSSD上进行模型简化测试,使用公认评价指标对所提算法和当前多种算法进行评价。实验结果表明,该方法的准确率和召回率更高,对显著目标保持了良好的完整性,且在边缘轮廓区域更好地区分了显著目标与背景。
Abstract
This paper proposes a saliency detection method based on deep residual networks and multiscale edge residual learning to address the problems of low salient values and blurred edges in images having complex backgrounds. Further, an edge residual block is proposed, and an edge residual network is constructed based on the deep residual network using the edge residual block for the salient graph edge supervised learning. In addition, the edge features are learned while training the network by constructing a three-category model based on the background, foreground, and edge, which can make the target edge more accurate. The output uses atrous convolutions to construct a multiscale atrous convolution unit for integrating and extracting the multiscale features based on the global information. Finally, the proposed algorithm is tested in an ablation study using two datasets (SED2 and ECSSD) and compared with various existing algorithms based on the common evaluation indicators. The experimental results demonstrate that the proposed method exhibits high accuracy and recall rate, maintains good integrity for the significant target, and distinguishes the significant target and background from the edge contour regions.
时斐斐, 张松龙, 彭力. 基于深度残差网络与边缘监督学习的显著性检测[J]. 激光与光电子学进展, 2019, 56(15): 151502. Feifei Shi, Songlong Zhang, Li Peng. Salient Object Detection Based on Deep Residual Networks and Edge Supervised Learning[J]. Laser & Optoelectronics Progress, 2019, 56(15): 151502.