請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/69373
標題: | 機器學習在衛星圖片道路擷取上的應用 Road Extraction from Satellite Images Using Machine Learning Techniques |
作者: | Alexandre Constantin 康亞力 |
指導教授: | 丁建均 |
關鍵字: | 衛星圖,圖像分類,計算機視覺,道路提取,機器學習,神 經網絡, Satellite Images,Image classification,Computer Vision,Road Extraction,Machine Learning,Neural Network, |
出版年 : | 2018 |
學位: | 碩士 |
摘要: | 在計算機視覺中,圖像分類由於其廣泛應用在許多領域上,如圖像壓縮、目標跟踪、圖像分析和地形開發追蹤,一直佔有重要地位。在這個特定的區域,本文我們將探討圖像分類在衛星圖像中提取道路網絡的應用上。其中包括給每個像素做標籤,區分它是否屬於或不屬於道路。這個應用在地圖繪製或監控安全上一直被非常廣泛使用。
為了準確地提取道路,我們提出使用一種基於機器學習的技術,並附加圖像分類算法作為後處理。我們的程序描述如下:首先,將輸入的圖像切割成小方塊,可以更好地適用於簡化後的神經網絡,以便減少計算時間和提供更詳細的信息。除此之外,我們還對顏色進行預處理來增加輸入數據域並且利用顏色通道和附加的梯度通道作為神經網絡的入口。最後通過圖像分類算法對檢測到的道路進行驗證,也更進一步地利用後處理完善缺失的道路,建立最終的衛星道路圖。 測試結果表明,我們提出的方法能夠檢測出大部分道路,並優於目前最先進的方法。 In computer vision, image classification has always had an important place due to its widespread applications such as image compression, object tracking, image analysis and ground use evolution. In this particular area we will use classification to extract road networks from satellites images. It consists of giving a label to each pixel if it belongs or not to a road, the applications are very wide from Network mapping to monitor territories. To extract the roads accurately we propose a machine learning based technique with additional classification algorithm as post-processing. Our procedure is described as follows: first the input image is cutted in small squares, it will fit better to a reduced neural network and will decrease computation time and provide more detailed information, we also increase the range of input data using color’s pre-processing algorithm. Then we exploit color channels and additional gradient channels to put inside the neural network that is trained to classify road and non-road pixels. Finally we verify with additional classification algorithm the detected roads and with post-processing complete the missing road parts to construct the final road map. Simulations show that our proposed method detect most of the road pixels and outperforms state-of-the-art methods. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/69373 |
DOI: | 10.6342/NTU201801173 |
全文授權: | 有償授權 |
顯示於系所單位: | 電信工程學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-107-1.pdf 目前未授權公開取用 | 13.13 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。