基于卷积神经网络的建筑构件图像识别

Image Recognition for Building Components Based on Convolutional Neural Network

对于现存的大量既有建筑,利用三维激光扫描所得到的点云数据生成BIM模型,需要将点云数据转换成建筑RGB-D图像,并对图像进行分类处理。传统图像识别技术无论是识别准确度还是面对复杂场景的模型泛化能力等,都难以满足现在的需求。文中基于深度学习算法,针对室内建筑门窗构件图像的分类问题,提出了一种运用卷积神经网络模型进行建筑构件图像识别的方法。该方法首先将收集的数据集进行数据增强处理以增加数据丰富度,并使用在ImageNet上已经训练好权重的VGG16作为识别网络,随后对网络进行优化,包括增加Dropout层、L2正则化以及采用Fine-tune操作来提升网络的识别精度。实验结果表明,进行了Fine-tune等优化后的模型的平均识别准确率达到95.4%,相比于未经过优化的模型的准确率提高了大约5.1%。

It is necessary to convert the point cloud data into the RGB-D images of building and classify the images,when using the point cloud data obtained by 3 D laser scanner to generate BIM model for a large number of existing buildings.In this paper,based on the deep learning algorithm,a method of building components image recognition employing the convolution neural network is proposed by using the transfer learning theory to dealing with the classification problem of interior building components image such as doors and windows.First of all,the VGG16 with weight parameters trained in Imagenet is used as the image recognition neural network.In addition,the network is optimized by adding Dropout layer,L2 regularization and using Fine-tune operation to improve the recognition accuracy of the network.The experimental results show that the average recognition accuracy of the model optimized by Fine-tune is 95.4%,which is about 5.1%higher than that of the model without optimization.