基于卷积神经网络的地图相似度匹配方法研究

Research on map similarity matching method based on convolutional neural network

针对目前基于深度学习的图像分类方法难以提取地图子类图像的高级语义信息,该文采用一种孪生网络结构下利用卷积神经网络提取图像特征计算相似度的方法进行地图子类的识别,将地图样本输入到孪生网络模型中进行训练及测试,并与直方图、灰度共生矩阵两种传统图像相似度计算方法进行对比分析。结果表明,基于孪生网络结构利用卷积神经网络提取图像特征计算地图相似度的方法准确率为93%,比两种传统方法分别高出了48%、43%;F1测度值为93.2%,比两种传统方法分别高出61.1%、26.53%,每张图像的运行速度也可达毫秒级。得出结论:该相似度匹配方法的高准确度和高效性为地图子类图像识别提供了技术方法,为互联网地图监管提供了新的思路。

In view of the present image classification method based on the deep learning is difficult to extract high-level semantic information of map subclass,this paper adopts a convolution neural network is utilized to extract image features under the Siamese network structure map subclass of recognition method of similarity calculation,to map sample input to the Siamese network model for training and testing,compared and analyzed with two traditional image similarity calculation methods:histogram and gray co-occurrence matrix.The results show that the accuracy of convolutional neural network based on Siamese network is 93%,which is 48%and 43%higher than the two traditional methods respectively.The F1 measure value is 93.2%,which is 61.1%and 26.53%higher than the two traditional methods,respectively.The running speed of each image can also reach millisecond level.It is concluded that the high accuracy and high efficiency of the similarity matching method provide a technical method for map subclass image recognition and a new idea for Internet map supervision.