注意力机制与卷积神经网络融合的军队标号识别方法

Recognition Method for Military Symbology Based on Fusion of Attention Mechanism and Convolutional Neural Network

手绘军队标号识别是基于草图的战场态势表达的重要组成部分。针对现有的草图特征手工提取方法费时费力,以及依靠数据驱动的深度学习方法泛化能力受到训练数据多样化制约的问题,提出一种注意力机制与卷积神经网络融合的军队标号识别方法。首先构建了特征提取模型Sketch-Net,实现手绘军队标号特征的初步提取;然后引入注意力机制模块,分别捕获特征在通道和空间位置中的依赖关系,增强模型提取特征信息的有效性。结果表明,提出的方法在建立的军队标号数据集上的识别正确率达到95.75%,能有效应用于手绘军队标号的识别。

Sketch recogition for military symbology is a significant component of sketch-based military battlefield situation expression.Existing sketch feature extraction methods are time-consuming and laborious,and the generalization ability of data-driven deep learning methods is restricted by the diversity of training data.A sketch recognition method based on the fusion of attention mechanism and convolution neural network is proposed in the paper.Firstly,the preliminary features of sketch military symbology are extracted by the Sketch-Net model.Then,the attention mechanism module is introduced to capture the dependency of features in channel and spatial location respectively,to enhance the validity of extracted feature information of the model.The experimental results show that the accuracy of the proposed method in the SF-Sketch dataset can reach 95.75%,which proves that the method can be effectively applied to sketch recognition of military symbology.