基於 YOLOX-CS 與插圖資料增強的台灣建物結構 辨識模型之實作與方法;The Implementation and Method of Taiwan Building Recognition Model based on YOLOX-CS and Illustration Data Augmentation

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/89757

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/89757

Title:	基於 YOLOX-CS 與插圖資料增強的台灣建物結構辨識模型之實作與方法;The Implementation and Method of Taiwan Building Recognition Model based on YOLOX-CS and Illustration Data Augmentation
Authors:	陳偉翔;CHEN, WEI-HSIANG
Contributors:	資訊工程學系
Keywords:	物件偵測, 資料增強, 地震風險評估, 建物結構辨識,YOLOX-S;Object Detection, Data Augmentation , Earthquake Hazad Analysis,YOLOX-S
Date:	2022-07-13
Issue Date:	2022-10-04 11:58:34 (UTC+8)
Publisher:	國立中央大學
Abstract:	地震工程一直以來都是研究中很重要的問題，為了要快速了解城市的建物結構分布，以及以可靠的方式估計相關區域的建物結構分佈情況，隨著深度學習 CNN 快速的發展，已有許多學者利用相關深度學習技術來製作建物辨識模型，相較於傳統辨識建築物方法，傳統的建築物辨識模型使用人工設計，基於建築輪廓的目標特徵，這不僅需要大量的時間和人力來處理資料，且手動的不一致性也會導致 mAP 降低。基於深度學習的物件偵測模型具有較高的檢測 mAP 且可以省下大量時間以及經濟成本; 然而目前台灣建物結構辨識模型精準度及速度是有待改善的，因為台灣建築物資料集中，存在背景複雜、建築物常遭到雜物遮擋...等等問題導致模型辨識困難。本研究為了提高建物結構辨識模型的平均準確率 (mAP) 和速度，在 YOLOX-S 網路中加入卷積注意力模塊(CBAM)，增加有用特徵的權重，同時抑制無效特徵的權重，提高檢測 mAP，本研究將此模型命名為 YOLOX-CS，並使用了本研究提出的插圖資料增強法改善了台灣建物中，背景複雜、建物遭遮擋等問題，增加了台灣模型辨識能力以及透過實驗結果表明本研究的提出的方法比原 YOLOX-S 具有更好的 mAP 及速度，對小目標、多目標都取得了更好的偵測效果，插圖增強法也更勝於常見的資料增強方法，與 YOLOv4 物件偵測方法建立台灣建物結構辨識模型相比，mAP 提升了 21.74% ，訓練模型時間縮短了約 37.5%，最終使YOLOX-CS 辨識模型的 mAP 達到了 76.32% ，相較 YOLOX-S 模型提升了 2.45% mAP。 ;Earthquake engineering has always been a very important issue in re- search. In order to quickly understand the distribution of buildings in the city and estimate the distribution of buildings in relevant areas in a reli- able way, with the rapid development of deep learning CNN , many scholars have used related deep learning techniques to make building recognition models. Compared with traditional building recognition methods, tradi- tional building recognition models use artificial design and target features based on building outlines, which not only requires a lot of time and man- power to process data, and manual inconsistencies can also lead to lower mAP. The object detection model based on deep learning has a high de- tection mAP and can save a lot of time and economic cost; however, the accuracy and speed of the current Taiwanese building structure recognition model need to be improved, because Taiwanese building data is concen- trated and the background is complex , buildings are often blocked by debris... and so on, which makes the model identification diﬀicult. In this study, in order to improve the mAP and speed of the build- ing recognition model, a convolutional attention mechanism (CBAM) was added to the YOLOX-S network to increase the weight of useful features, while suppressing the weight of invalid features to improve detection mAP. The model is named YOLOX-CS, and the illustration Data Augmentation method proposed in this study is used to improve the problems of com- plex backgrounds and occlusion of buildings in Taiwan, and to increase the recognition ability of the Taiwan model. The experimental results show that the method proposed in this study is Compared with the original YOLOX-S, it has better mAP and speed, and has achieved better detec- tion results for small targets and multiple targets. The illustration Data Augmentation method is also better than the common Data Augmentation method and YOLOv4 object detection method. Compared with the estab- lishment of the Taiwan building structure identification model, the mAP is increased by 21.74%, and the training time of the model is shortened by about 37.5%. Finally, the mAP of the YOLOX-CS identification model reaches 76.32%, which is 2.45% higher than that of the YOLOX-S model.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	55	View/Open

社群 sharing

Loading...