查詢結果分析
來源資料
頁籤選單縮合
| 題 名 | 應用深度學習與街景影像探討臺灣都市廣告招牌街道空間特性:以新北市板橋區為例=Investigating the Spatial Characteristics of Urban Street Signboards in Taiwan Using Deep Learning and Street View Imagery: A Case Study of Banqiao District, New Taipei City |
|---|---|
| 作 者 | 羅章秀; 林柏丞; | 書刊名 | 地理學報 |
| 卷 期 | 111 2025.08[民114.08] |
| 頁 次 | 頁55-97 |
| 分類號 | 312.831 |
| 關鍵詞 | 廣告招牌; 街景影像; 深度學習; 地理人工智慧; 空間分析; Signboard; Street view imagery; Deep learning; GeoAI; Spatial analysis; |
| 語 文 | 中文(Chinese) |
| DOI | 10.6161/jgs.202508_(111).0003 |
| 中文摘要 | 由於街景影像涵蓋範圍廣泛與具有地理坐標特性,以及深度學習(Deep Learning)之卷積神經網路(Convolutional Neural Network, CNN)於影像處理層面有顯著準確度與時間效率改善,近年來許多研究結合街景影像與深度學習於都市街道空間大規模調查。臺灣街道廣告招牌多元且密集,若能全面取得廣告招牌街道量化成果,可為商業、都市規劃重要參考依據,亦可探討廣告招牌與實際商業活動之空間異質性(Spatial Heterogeneity),剖析更深入商業現象,惟缺乏相關研究探討。故此,本研究首先以Deeplab v3+預訓練模型,透過遷移學習(Transfer Learning)方式,建置臺灣都市廣告招牌辨識模型與建構廣告招牌街道空間聚集度(Signboard Density Index, SDI)量化指標。接著以空間自相關(Spatial Autocorrelation)探討研究區域廣告招牌街道空間聚集度之空間型態(Spatial Pattern)。最後則以地理加權迴歸(Geographically Weighted Regression, GWR)與K-means分群方式,探討村里空間尺度之土地利用、人潮活動以及消費熱度與廣告招牌街道空間聚集度關聯性,並進一步區分不同區域商業特性。成果顯示,本研究廣告招牌辨識模型的MIoU值達83%,有一定辨識廣告招牌能力,而局域(Local)空間自相關可以發現廣告招牌分布有地區差異,亦符合實際分布情形,而地理加權迴歸分析成果,除發現廣告招牌分布與商業活動有空間異質現象,根據迴歸係數的分群成果亦可將實驗區分為「商業街道、傳統市集」、「商圈、百貨公司」以及「日常消費商業、鐵皮建築形態之商業」三種商業特性區域。本研究可作為後續應用參考與研究開端。 |
| 英文摘要 | Because of the extensive coverage and spatial characteristics of street view imagery, and the advancements in deep learning, specifically Convolutional Neural Networks (CNNs), have significantly enhanced the accuracy and efficiency of image processing. Consequently, the integration of street view imagery and deep learning has gained prominence in large-scale surveys of urban street space, and also a significant aspect of GeoAI development in recent years. In Taiwan, the streets are characterized by a diverse array of dense signboards. However, there is a research gap on the quantification and spatial heterogeneity of these signboards, which are crucial for business and urban planning. Hence, this study proposes a methodology that leverages the pre-trained model Deeplab v3+ and transfer learning techniques to develop a recognition model for urban signboards in Taiwan. Furthermore, this study introduces a Signboard Density Index (SDI) to assess the spatial clustering of signboards on streets. Also, spatial autocorrelation analysis is applied to explore the spatial patterns of signboard density within the study area. Additionally, Geographically Weighted Regression (GWR) and K-means clustering techniques are employed to investigate the relationships between signboard density, land use, human activities, and consumption intensity at the village scale in order to identify different regional commercial characteristics. The results demonstrate that the model for recognizing urban signboards achieves a MIoU value of 83%, indicating a significant level of signboard recognition capability. By utilizing local spatial autocorrelation analysis, this study reveals spatial variations in signboard distribution that align with the actual distribution patterns observed. Furthermore, the findings from the geographically weighted regression and K-means analysis not only unveil spatial heterogeneity in signboard distribution and commercial activities but also categorize the study areas into three commercial characteristic regions based on the regression coefficients: "commercial streets and traditional markets," "commercial districts and department stores," and "daily consumption businesses and commercial areas with tin buildings." The proposed methodology presents interdisciplinary opportunities for applications in surveys, urban design, and planning, as well as policy-making. Overall, this study serves as a valuable reference and a starting point for future applications and research. |
本系統中英文摘要資訊取自各篇刊載內容。