以防災救治為導向之高通用性物體偵測與分割模型

Abstract

本研究基於現有State-of-art檢測模型更著重在檢測時運算效能之提升，而欠缺針對未知複雜場景通用性(Generalization Capability)之探討的背景之下，結合防救災領域針對災難現場細部場景的數據集嚴重不足的現狀，在防災救治願景下，實現高通用性物體偵測與分割模型建置。為此，本研究旨在開發一種高通用性的物體偵測與分割模型，以提升災後快速評估建築損傷和制定援救計畫的效率。首先，本研究開發了一種新型窗戶偵測模型用於災難救援，名為FOpen-YOLO。該模型在來自巴黎和臺灣的非災害街景圖像上進行訓練，並在包括無人機和災後圖像在內的各種未知複雜場景的數據集上評估其通用能力。結果表明，所提出的FOpen-YOLO在多個測試數據集上超過了基線模型，全球街景圖像的準確率提高了15.1%，無人機拍攝的圖像增強了26.6%，災後損壞建築的窗戶檢測提升了21.2%。其次，本研究開發了一種新型裂縫分割模型用於災損評估，GCUnet，利用時間域和頻率域對特徵圖進行處理，感知裂縫在圖像中的分佈，學習如何區分裂縫和干擾,實現通過像素級的裂縫分割數據進行訓練，完成在災害室內場景中高效的裂縫檢測。結果表明，所提出的GCUnet在災害室內場景上超過了基線模型Unet，MIoU提升了約17.6%，Dice的表現增強了13.1%。最後，本研究通過有效感受野分析解釋了高通用性模型的性能差異，發現高通用性模型在特徵提取層和目標輸出層的有效感受野具有規律特點。這些發現為未來提升防災救治系統的效能提供了重要依據。

This study addresses the limited exploration of the generalization capability of current state-of-the-art detection models in unknown complex scenarios. Due to the scarcity of detailed disaster site datasets, this study aims to create a highly generalizable object detection and segmentation model for disaster mitigation. The objective is to improve the efficiency of rapid post-disaster building damage assessment and rescue planning. The study introduced a new window detection model named FOpen-YOLO to achieve this. This model underwent training using non-disaster street view images from Paris and Taiwan and was subsequently tested in various complex scenarios, including UAV and post-disaster images. The results revealed that FOpen-YOLO outperformed the baseline model across multiple test datasets, demonstrating a 15.1% improvement in accuracy on global street view images, a 26.6% enhancement on UAV-captured images, and a 21.2% increase in window detection for damaged buildings post-disaster. Furthermore, the study also developed a novel crack segmentation model, GCUnet, which utilizes feature maps in both time and frequency domains to detect cracks in images and differentiate them from noise. Training the model using pixel-level crack segmentation data resulted in efficient crack detection in indoor disaster scenarios. In comparison to the baseline model Unet, results indicated that GCUnet surpassed it with a 17.6% improvement in MIoU and a 13.1% increase in the Dice coefficient. Finally, the study explains the performance variances of the highly generalizable models through practical receptive field analysis, unveiling distinctive characteristics in the feature extraction and target output layers. These findings establish an essential foundation for future enhancements in the effectiveness of disaster prevention and rescue systems.