以物件偵測方法建立台灣建物結構辨識模型;Taiwan Building Recognition Model Based on Object Detection Method

NCUIR > College of Electrical Engineering & Computer Science > Executive Master of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/86510

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/86510

Title:	以物件偵測方法建立台灣建物結構辨識模型;Taiwan Building Recognition Model Based on Object Detection Method
Authors:	吳少凱;Wu, Shao-Kai
Contributors:	資訊工程學系在職專班
Keywords:	卷積神經網路;物件偵測;建物分類;街景圖片;CNN;Object detection;Building instance classification;Street view images
Date:	2021-10-26
Issue Date:	2021-12-07 12:55:05 (UTC+8)
Publisher:	國立中央大學
Abstract:	在過去的數年中，由於卷積神經網絡(CNN)技術的發展，以街景圖資分類建築物結構或使用類型，在方法上已有一定的成果。這種分類方式通常是對單張街景圖片中的單一建物類型進行標記，由CNN分類架構訓練辨識模型，並在其後獲得建物結構類型的地理分布，也就是災害風險中的暴露度地圖。然而，這種由單一圖片進行分類的方式，在許多基礎場景的應用上成功率有限。比如台灣這樣壅擠的都市區域，一張街景圖片中通常包含多個類別的建物，導致單純CNN分類架構所訓練出的模型無法良好的運作。在本文中，我們使用物件偵測方法對單個建築物的建造材質進行分類，此方法可根本性的優化整個分類辨識的流程，也妥善處理了從前無法在擁擠都市進行應用的問題。此方法在台灣的資料集中獲得了75\%的辨識成功率(mAP)，同時我們也比較了北美的資料集，在物件偵測方法上可獲得約5\%的正確率提升。本文所使用的物件偵測方法是YOLOv4的物件偵測架構，此方法可以從街景圖像（例如Google街景）中對單一建物的立面結構進行分類，並將分類完成的目標與建物的地理資訊進行關聯處理，進而獲得一個區域的暴露度地圖。我們同時創建了一個針對台灣建物的基礎數據集，用於訓練和評估CNN物件偵測或分類模型。此外，我們還針對臺中市中區，應用本研究所訓練的模型，實際製作城市尺度的建物結構分類圖，並且針對圖片擷取方法以及流程進行修正，獲得一個可做為後續災害暴露度應用之圖資。;In the past few years, with the development of Convolutional Neural Network (CNN) technology, there are extensively researched on classifying the structure or using types of buildings based on street view images. These kinds of classification methods usually use a typical CNN classification framework, training by the single label of each image. This recognition framework is used to obtain an exposure model for earthquake hazard analysis in an entire region. The exposure model here represents the geographical distribution of building structure types in an area. However, classifying by a single image is not well fitted to many basic scenarios. In crowded urban areas such as Taiwan, one street view image usually contains multiple types of buildings, which causes the weights trained by a simple CNN classification model is not able to make reasonable recognition. In this paper, we proposed the object detection method to classify the construction materials of a single building. This method fundamentally optimizes the entire classification and identification process and also handles the problem that could not be applied in crowded areas properly. This method has achieved a mean Average Precision(mAP) of 75\% in the Taiwan dataset. At the same time, we also compared the North American dataset, and we can obtain an accuracy rate improvement of about 6\% on our method. The method used in this article is an object detection architecture based on CNN. This method can classify the facade structure of a single building from street view images (such as Google Street View). After correlating with the geographic information of the building, we obtain a regional exposure model. We create a benchmark dataset for buildings in Taiwan that could be used on training and validating the CNN object classification model. In addition, we have also produced a city-level structure classification map for Taichung City.
Appears in Collections:	[Executive Master of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	133	View/Open

社群 sharing

Loading...