深度学习融合局部聚合向量的增强现实标志物检索研究

Research on Augmented Reality Marker Retrieval Based on Deep Learning and Vector of Locally Aggregated Descriptors

为了降低计算量,增强现实(augmented reality,AR)经常在真实场景中设置人工标志物,用以协助相机对场景的跟踪,因而对人工标志物的检索与跟踪技术十分关键,是增强现实融入大众生活的重要前提。针对标志物检索问题,设计了一个端到端的神经网络模型,该模型由特征提取和局部特征聚合两大核心模块组成:特征提取模块使用卷积神经网络提取图像高维特征;局部特征聚合模块对高维特征进行降维与压缩,为图像编码全局特征。该网络在文中提出的一种新的损失函数的指导下进行训练。为了解决真实场景数据收集成本较高的问题,提出了一种模拟数据集,网络在模拟数据集上训练及测试,取得了较高的检索准确率,证明了深度学习与局部特征聚合结合的方法在图像检索领域的可行性。

In order to reduce the amount of computation,Augmented Reality(AR)often sets artificial markers in the real scene to assist the camera in tracking the scene.Therefore,the retrieval and tracking technology of artificial markers will become an important premise for augmented reality to integrate into our public life.For the problem of marker retrieval,this paper designs an end-to-end neural network model,which is composed of two core modules:feature extraction and local feature aggregation.The feature extraction module uses convolution neural network to extract high-dimensional features of the image;The local feature aggregation module compresses the high-dimensional features to code global features for images.The network is trained under the guidance of a new loss function proposed in this paper.Meanwhile,in order to solve the problem of high cost of real scene data collection,we propose a simulation dataset.The network is trained and tested on the dataset and has achieved high accuracy.Our experiment proves the feasibility of the combination of deep learning and local feature aggregation in the field of image retrieval.