基於非監督域適應之衛星圖資切割優化

Abstract

在國防領域中，衛星圖資因其高海拔尺度、獨特的視角以及廣大覆蓋率的特性扮演了重要的角色，同時也被應用在各種情境中，包括農作物預測、土地調查等繪圖應用。而在本論文中，我們使用深度學習框架於「影像切割與色彩調配」任務，其中前者目的在於準確地對衛星影像中的每個像素進行不同分類的預測；後者旨在開發相關技術以穩健地在具不均勻分布的不同來源影像中進行影像切割。\n\n深度學習本質上是資料驅動的，十分仰賴訓練資料的數量以及品質，且訓練好的模型往往無法泛化到其他資料上，而這也是應用在衛星圖資分析常常會被觀察到的現象。針對此一議題，我們採用了非監督域適應(UDA) 技術，嘗試在不同來源資料中轉移域知識，進而讓模型具更強能力來處理異質性資料。對此，我們首先研究及實驗目前現有的Source-Only模型以及 UDA 演算法，並獲得綜合性的結果。接著我們進一步提出創新的模型架構，稱為域轉換與強化網路(Domain Transfer and Enhancement Network, DTEN)，當中包含Skip Connection、Mixed Pooling Module以及域轉換模組，並實驗中達到46.4% MIOU，超越當今SOTA約3% 的成績。

Satellite imagery plays an important role in national defense due to its high attitude, unique view point and large coverage. It has also been utilized in areas such as crop prediction, land surveying as mapping. In this thesis, we focus on the task of image segmentation and color matching using deep learning framework. The former is concerned with accurate classification of pixels in satellite images into different categories, and the latter is to develop techniques to perform robust semantic segmentation when satellite images are obtained from different sources, causing non-uniformity in color distribution.\n\nDeep neural network is inherently data-driven, relying heavily on the quantity and quality of the training data. Models trained on one dataset might not generalize well to the other, which is often observed in the analysis of satellite images. To address the above issue, we employ unsupervised domain adaptation (UDA) techniques to transfer domain knowledge between different sources of satellite images, hoping to strengthen the model’s ability to cope with heterogeneous data. We start by experimenting with existing UDA algorithms and performing comparative analysis. We then propose a novel architecture named domain transfer and enhancement network(DTEN) which incorporates skip connection, mixed pooling module and domain transfer module, and design some ablation studies. The resulting model supersedes the state-of-the-art methods in satellite image segmentation, achieving a highest MIOU of 46.4%, an improvement of nearly 3% over existing approaches.