多尺度特征融合技术的建筑寿命分析方法  

Building Life Analysis Method Based on Multi-scale Feature Fusion Technology

作　　者：王振东 刘思航 WANG Zhen-dong;LIU Si-hang(School of Architecture and Urban Planning,Beijing University of Civil Engineering and Architecture,Beijing 100044,China;School of Architecture and Art,Dalian University of Technology,Dalian 116000,Liaoning,China)

机构地区：[1]北京建筑大学建筑与城市规划学院,北京100044 [2]大连理工大学建筑与艺术学院,辽宁大连116000

出　　处：《建筑节能（中英文）》2021年第12期126-131,共6页Building Energy Efficiency

摘　　要：针对传统分析建筑寿命的深度学习目标检测算法计算效率差、检测精度低,提出了一种基于密集卷积网络(Dense convolutional network, DenseNet)的新型多尺度特征融合单点检测算法(Multi-scale Feature fusion One point Detection, MFSOD)。该算法将DenseNet与单激发多盒探测器(Single Shot MultiBox Detector, SSD)算法框架相结合。在骨干网后增加卷积层,实现多尺度特征检测。并设计了特征融合模块,对不同层次的多尺度特征进行融合,在目标检测中引入上下文信息。为了减小噪声的影响,针对空间位置和特征信道的特点,提出了一种基于去噪的多尺度特征融合机制,嵌入到各种特征融合模块结构中,在整个编解码器框架中起到编码器模型的作用。实验结果表明,在数据量为2TB时,相对于支持向量机(Support Vector Machine, SVM)算法模型MFSOD模型的检测精度更高,处理数据效率提高了43%,相关性能更好。Aiming at the low computational efficiency and low detection accuracy of traditional deep learning target detection algorithm for building life analysis, a Multi-scale Feature fusion One point Detection(MFSOD) algorithm based on dense convolution network(DenseNet) is proposed. The algorithm combines DenseNet with SSD algorithm frame work. Firstly, a convolution layer is added after the backbone network to realize multi-scale feature detection. In addition, feature fusion module is designed to fuse multi-scale features of different levels, and context information is introduced into target detection. In order to reduce the influence of noise, a multi-scale feature fusion mechanism based on de-noising is proposed according to the characteristics of spatial location and feature channel. It is embedded in various feature fusion module structures and plays the role of encoder model in the whole codec framework. The experimental results show that, compared with the support vector machine(SVM) algorithm model, the MFSOD model has higher detection accuracy, higher data processing efficiency and better correlation performance when the amount of data is 2 TB.

关 键 词：多尺度 特征融合 目标检测 密集卷积网络 编解码器 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]