Fire Detection and Semantic Fire Image Segmentation using Deep Learning

국문 초록 (Abstract)

최근 딥 러닝은 다양한 분야에서 가장 중요하고 강력한 주제이다. 딥러닝은 이미지 분류에서 뛰어난 성능을 보였으며, 이후 컴퓨터 비전의 이지미에서 객체 감지 및 시맨틱 분할에도 적용되었다. 본 논문에서는 뛰어난 성능을 가진 딥 러닝을 사용하여 화재 이미지 감지 및 분할 작업에 적합한 네트워크를 제안한다. 또한 딥러닝 압축 기법을 사용하여 화재 이미지 분할 딥러닝 모델에 적용하여 소규모 네트워크를 제안하였고 이를 임베디드 장치에 적용했다. 여러가지 광범위한 실험을 통해 화재감지와 화재 이미지 분할에서 기존의 기법보다 좋다는 점을 보였다.

다국어 초록 (Multilingual Abstract)

Recently, deep learning has become the most important and powerful topic in various research fields. It has shown excellent performance in image classification and has been applied to the fields of object detection and semantic image segmentation of computer vision. In this thesis, we proposed deep neural networks suitable for fire image detection and segmentation tasks with excellent performance. In addition, we proposed a small-sized network for fire image segmentation based on squeezed deep-learning techniques and applied it to an embedded device. Several extensive experiments are presented to demonstrate its better performance compared with the existing methods for fire detection and image segmentation.

목차 (Table of Contents)

• 1 Introduction 1
• 2 Preliminaries 4
• 2.1 Image classification 4
• 2.2 Object detection 9
• 2.3 Semantic image segmentation 12
• 2.4 Compressed deep learning 15
• 3 Fire Detection and Localization 17
• 3.1 Related work 17
• 3.2 Proposed method 19
• 3.3 Experiments 21
• 3.3.1 Fire area localization 28
• 3.3.2 Fire localization results 30
• 3.4 Conclusion 32
• 4 Semantic Segmentation using Deep Learning for Fire Images 34
• 4.1 Related work 34
• 4.2 Proposed architecture 38
• 4.2.1 Comparison with FusionNet 40
• 4.3 Experimental results 41
• 4.3.1 Experimental results using FiSmo dataset 42
• 4.3.2 Experimental results using Corsican Fire Database 45
• 4.4 Conclusion 48
• 5 Squeezed Semantic Segmentation for Fire Images 49
• 5.1 Related work 49
• 5.2 Squeezed Fire Binary Segmentation Networks(SFBSNet) 50
• 5.2.1 SFBSNet architecture 50
• 5.2.2 Implementation details 54
• 5.3 Experiments 56
• 5.3.1 Ablation study 58
• 5.3.2 Experiments on FiSmo dataset 58
• 5.3.3 Experiments on Corsican Fire Database 60
• 5.3.4 Additional experiments on Still dataset 60
• 5.4 Conclusion 62
• 6 Conclusion and Future Works 64
• Abstract (in Korean) 73