激光与光电子学进展, 2019, 56 (7): 071501, 网络出版: 2019-07-30
基于级联全卷积神经网络的显著性检测 下载: 1104次
Salient Detection Based on Cascaded Convolutional Neural Network
机器视觉 显著性检测 级联全卷积神经网络 金字塔池化模块 边缘检测网络 machine vision saliency detection cascaded full convolution neural network pyramid pooling module edge detection network
摘要
提出了一种级联全卷积神经网络的显著性检测方法。网络主要由两层级联的全卷积神经网络组成,第一阶段构建了一个带金字塔池化模块编码-解码架构的全卷积神经网络,金字塔池化模块有效抑制了背景噪声的干扰。第二阶段设计了边缘检测网络,学习显著区域的边缘信息,通过融合两个阶段显著图得到边界精确的显著图。实验结果表明,所提方法在图像显著性检测数据集ECSSD和SED2上均具有较高的准确率、召回率和较低的平均绝对误差,为目标识别、机器视觉等提供了可靠的预处理结果。
Abstract
A saliency detection method is proposed based on a cascaded full convolutional neural network. This network is mainly composed of two full convolutional neural networks. In the first stage, a full-convolutional neural network with a pyramid pooling module encoding and decoding architecture is constructed, and the pyramid pooling module can be used to effectively suppress the interference of background noises. In the second stage, an edge detection network is designed to learn the edge information of a salient region, and the accurate boundary saliency map is obtained by the fusion of two-stage saliency maps. The experimental results show that the proposed method has high accuracy, high recall rate, and low average absolute error in image significance detection dataset ECSSD and SED2, which provides the reliable pretreatment results for target recognition, machine vision and other applications.
张松龙, 谢林柏. 基于级联全卷积神经网络的显著性检测[J]. 激光与光电子学进展, 2019, 56(7): 071501. Songlong Zhang, Linbo Xie. Salient Detection Based on Cascaded Convolutional Neural Network[J]. Laser & Optoelectronics Progress, 2019, 56(7): 071501.