基于注意力机制的视觉位置识别方法  

Visual Place Recognition Method Based on Attention Mechanism

作　　者：戴天虹[1] 杨晓云 宋洁绮 DAI Tian-hong;YANG Xiao-yun;SONG Jie-qi(School of Mechanical and Electrical Engineering, Northeast Forestry University, Harbin Heilongjiang 150040, China)

机构地区：[1]东北林业大学机电工程学院,哈尔滨150040

出　　处：《哈尔滨理工大学学报》2022年第2期63-68,共6页Journal of Harbin University of Science and Technology

基　　金：黑龙江省自然科学基金(C201414);哈尔滨市科技创新人才项目(2014RFXXJ086);中央高校基本科研业务费专项资金(2572019CP17).

摘　　要：针对现有的视觉位置识别方法在图像外观变化和视角变化时准确性和鲁棒性表现不佳的问题,提出了一个与注意力机制结合的视觉位置识别方法。首先,采用在大型位置数据集上预训练的卷积神经网络HybridNet提取特征。然后,运用上下文注意力机制对图像不同区域分配权重值,构建基于多层卷积特征的注意力掩码。最后,将掩码与卷积特征结合,构建融合注意力机制的图像特征描述符,从而提高特征的鲁棒性。在两个典型位置识别数据集上做测试实验,结果表明结合注意力机制的方法可以有效区分图像中与位置识别有关的区域和无关的区域,提高在外观变化和视角变化场景中识别的准确性和鲁棒性。Aiming at the problem of poor accuracy and robustness of the existing visual place recognition methods when the image appearance changes and the viewing angle changes,a visual place recognition method combined with the attention mechanism is proposed.Firstly,we use the convolutional neural network HybridNet pre-trained on a large location dataset to extract features.Then,we use the context attention mechanism to assign weight values to different regions of the image to construct an attention mask based on multi-layer convolution features.Finally,we combine the mask with the convolution feature to construct the image feature descriptor fused with the attention mechanism so as to improve the robustness of the feature.Testing experiments on two typical place recognition datasets show that the method combined with the attention mechanism can effectively distinguish between the regions related to place recognition and the unrelated regions in the image,and it can improve the accuracy and robustness of recognition in scenes with changes in appearance and viewpoints.

关 键 词：图像处理 位置识别 注意力机制 卷积神经网络 深度学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]