基於語意分割之電纜線偵測

• Main Authors: 游晉瑞, YOU, CHIN-JUI
• Other Authors: 廖文宏, Liao, Wen-Hung
• Format: Thesis
• Language: unknown
• Published: 2021
• Subjects: 電腦視覺; 語意分割; 深度學習; Computer vision; Semantic segmentation; Deep learning; Arctic
• Online Access: https://nccur.lib.nccu.edu.tw//handle/140.119/135981; https://doi.org/10.6814/NCCU202100504; https://nccur.lib.nccu.edu.tw/bitstream/140.119/135981/1/304301.pdf

碩士 國立政治大學 資訊科學系 107753043 電腦視覺的領域中，語意分割是一項非常重要的技術，目前被廣泛應用於無人載具、自動駕駛、場景理解與避障等任務。隨著深度學習技術的進步，語意分割已取得一定的成果，然而針對電纜線檢測，仍有許多待改善的空間。現今電塔及電纜線的開源資料有限，本研究主要對目前兩大開源的電纜線資料集進行電纜線檢查的語意分割實驗，分別為Ground Truth of Powerline Dataset及PLD-UVA資料集，並將兩資料集之Ground Truth重新標記，使模型能夠精準的獲得Ground Truth資訊。\n近年來學者已注意到將不同感測器的資料融合，如熱感測器或深度感測器的資料，可以協助提升光學影像模型的準確度，其中RTFNet利用兩個編碼器將熱影像特徵融合到光學影像中，但此架構並未考慮到也可以將光學影像的特徵融合到熱影像中，達到互相幫助的效果。藉此，本研究以RTFNet為基礎提出Dual Segmentation Model，希望透過邊緣強化來加強電纜線的光學影像模型，讓語意層面的資訊得以互補，進而提升分割模型的準確度，使其超越現今電纜線檢測表現相當優異的LS-Net之結果。本論文提出之dual segmentation 模型，其Precision以0.7919與LS-Net的0.8004並駕齊驅，Recall以0.7710超越LS-Net之0.5368，總結以上兩者所表現的F-score得出0.7753，超越LS-Net的0.5940將近0.2之多，體現出本研究所提出模型之可靠性。 In the field of computer vision, semantic segmentation is a technique widely employed in tasks such as autonomous driving, scene understanding and obstacle avoidance. With the advancement of deep learning technology, the performance of semantic segmentation has progressed rapidly. Yet there is still much room for improvement in the detection of power lines. Currently, open-source datasets of electric towers and power lines are quite limited. This research mainly conducts the semantic segmentation experiment of power lines inspection using two open-source datasets, namely, the Ground Truth of Power lines dataset and the PLD-UVA dataset. We re-labeled the above two datasets to better locate the region occupied by power lines.\nResearchers have reported the benefits of fusing data from different sensors, such as thermal sensors or depth sensors, to enhance the accuracy of optical image models. Among them, RTFNet utilizes two encoders to fuse thermal image features. However, this architecture fails to consider the characteristics of the optical image to promote mutual assistance. This research proposes the dual segmentation model (DSM) based on RTFNet. We hope to strengthen the optical image model of the power lines through edge enhancement, so that the semantic-level information can be complemented, and the accuracy of the segmentation model can ...