机器自监督学习的建筑物面要素几何形状度量

Geometry shape measurement of building surface elements based on self-supervised machine learning

针对传统方法在度量建筑物面要素几何形状时,未能考虑形状认知的视觉特征因素且形状特征需要人为定义等问题,该文提出一种建筑物几何形状度量方法。首先,利用深度卷积神经网络的图像特征学习特性,结合自动编码机的自监督学习能力,构建基于机器自监督学习的建筑物面要素几何形状度量神经网络;其次,利用建筑物图像形状大数据对网络进行训练;最后,利用训练完成的神经网络识别并提取建筑物形状特征集并作为形状度量的结果。实验表明,该方法形状度量结果区分度高,一定程度上克服了人为定义形状特征的缺点,且与视觉感知结果基本一致。

As the visual cognition of shape features is not considered and the shape features are manually extracted by human when measuring the geometry building shapes in many traditional algorithms,a new method to solve these problems was proposed.Firstly,the neural network model based on machine self-supervised learning was constructed with the convolution neural network having the ability to learn the visual features of shape images and with the auto encoder model having the self-supervised learning characteristic.Secondly,the features of shape image were learned by the model itself by feeding it a set of big data of building shape.Finally,the model was used to extract the shape feature set of the building and the result of the shape measurement was taken as the result of the shape measurement.Experiments showed that the measurement for the building shape was distinctive and were consistent with visual perception results.