摘要
材质识别旨在识别自然材质图像中的主要对象及其所属材料类别。针对材质图像数据集通常数据量少、人工标注局部纹理区域困难所导致的材质识别准确率低的问题,提出了一种基于注意力机制和深度卷积神经网络的材质识别方法,该方法的核心是材质识别深度卷积神经网络(MaterialNet)。MaterialNet利用深度残差网络对图像进行特征提取,采用所提出的级联空洞空间金字塔池化的方式引入注意力机制,使网络可以通过端到端训练自适应地关注包含纹理特征的关键区域,从而有效识别材质的局部纹理特征。在FMD材质数据集上进行实验,结果表明,MaterialNet的总体识别准确率可达到82.3%,比当前主流的B-CNN和CNN+FV材质识别方法分别提高了7.2%和4.5%,对多种材质的识别准确率较高且具有参数量少、计算量小等优点。
The purpose of material recognition is to identify the main objects and their material categories in natural material images.Aiming at the problem of low recognition accuracy caused by the lack of data in material image data sets and the difficulty of manually labeling local texture regions,a material recognition method based on attention mechanism and deep convolutional neural network is proposed.The core of the method is material recognition deep convolutional neural network(MaterialNet).MaterialNet uses the deep residual network to extract the features of the image,and introduces the attention mechanism by the proposed cascaded atrous spatial pyramid pooling method,so that the network can adaptively focus on the key areas containing texture features through end-to-end training,so as to effectively identify the local texture features of materials.Based on the FMD material datasets,the experimental results show that the overall identification accuracy of MaterialNet is 82.3%,which is 7.2%and 4.5%higher than the current mainstream B-CNN and CNN+FV material identification methods,respectively.The recognition accuracy of MaterialNet is high for a variety of materials,and it has the advantages of less parameters and less calculation.
作者
许华杰
杨洋
李桂兰
XU Hua-jie;YANG Yang;LI Gui-lan(College of Computer and Electronic Information,Guangxi University,Nanning 530004,China;Guangxi Key Laboratory of Multimedia Communications and Network Technology,Nanning 530004,China;Guangxi Institute of Product Quality Inspection,Nanning 530007,China)
出处
《计算机科学》
CSCD
北大核心
2021年第10期220-225,共6页
Computer Science
基金
广西壮族自治区科技计划项目(2017AB15008)
崇左市科技计划项目(FB2018001)。
关键词
注意力机制
深度卷积神经网络
空洞卷积
空间金字塔池化
Attention mechanism
Deep convolutional neural network
Atrous convolution
Spatial pyramid pooling
