請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/73426
標題: | 基於動態點擊轉換之互動式物件切割演算法及其高效率嵌入式系統實作 Interactive Object Segmentation with Dynamic Click Transform Algorithm and Efficient Embedded System Implementation |
作者: | Chun-Tse Lin 林均澤 |
指導教授: | 簡韶逸(Shao-Yi Chien) |
關鍵字: | 互動式分割,物件分割,深度學習, interactive segmentation,object segmentation,deep learning, |
出版年 : | 2020 |
學位: | 碩士 |
摘要: | 互動式物件切割(interactive object segmentation)會依照使用者給予的指示將目標對象切割出來,並且透過互動持續修正標示錯誤的區域,達到更加精確的切割結果。這項技術中最主要的挑戰來自於使用者給予的指令與目標物之間的不確定性, 如何明確表示使用者意象並以最少的互動次數達到精確的切割結果一直都是這項技術中熱門的研究對象。傳統演算法需要大量的使用者標示來估計前景和背景的分佈。近年來隨著深度學習在電腦視覺上的成功應用,基於深度學習的互動式物件切割算法將使用者互動轉換為圖片,並通過卷積神經網絡(Convolutional Neural Network)預測切割物件。這些方法在展示優異結果的同時也提高了計算複雜度, 不利於嵌入式系統上的實作。 在本篇論文中,我們提出了動態點擊轉換演算法來更好地表示使用者給予的指令與互動,並同時考慮空間幾何與特徵分佈,善加利用使用者傳達的資訊。與現有演算法相比,動態點擊轉換演算法展現了良好的表現,證實了提出方法的有效性。 除此之外,我們透過一連串的優化在Nvidia Jetson TX2開發版上實現了高效率的演算法實作,達到與使用者的即時互動,提供精確的切割結果和低延遲的使用者體驗。 In the interactive segmentation, a user initially indicates the target object to segment the main body and then provides corrections on mislabeled regions to iteratively refine the segmentation mask. The main challenge of this task originates from the ambiguity in the correlation between user annotations and the target object. Researchers have been finding ways to represent user interactions and segment a precise mask with the least interactions. Traditional algorithms require substantial user annotation to estimate the distribution of foreground and background. With the success of deep learning techniques on computer vision tasks, learning-based interactive segmentation algorithms have become popular in recent years. These methods convert user annotations into interaction maps and predict the mask from a convolutional neural network (CNN). However, these methods demonstrate superior results while increasing the computational complexity, which is not suitable for deploying on an embedded system. In this thesis, we propose a Dynamic Click Transform algorithm to better represent user interactions and take both spatial geometry and feature distribution into consideration. We demonstrate the effectiveness of our proposed method and achieve favorable performance compared to the state-of-the-arts. Furthermore, we accelerate our algorithm and implement it on an embedded system, NVIDIA Jetson TX2. This system performs a real-time interactive segmentation that provides high-quality results and a low-latency user experience. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/73426 |
DOI: | 10.6342/NTU202004397 |
全文授權: | 有償授權 |
顯示於系所單位: | 電子工程學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
U0001-0712202000292400.pdf 目前未授權公開取用 | 24.94 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。