臺灣博碩士論文加值系統

將陳舊竣工圖 (HDAD maps) 做圖片二值化是一項必要的工作，因為後續的操作如數位化、 編輯、檢索和壓縮都需要二值化做前置處理。植基於端到端全卷積網絡 (EEFCN)，本文提出 了一種用於 HDAD maps 的高效率二值化方法，簡稱為 DeepBinarization。基於所選擇的訓練 HDAD maps，所提出的 DeepBinarization 方法首先產生 Groundtruth foreground（GF）模型， 然後在訓練步驟中使用 GF 模型來訓練 EEFCN 的學習權重。其次，對於每個輸入的 HDAD 區塊，我們提出了一種基於混合的方法來生成該塊的 Estimated foreground（EF）模型，然後 在測試步驟中使用 EF 模型來修復由 EEFCN 生成的粗糙二值化 HDAD 區塊，生成高質量的 二值化HDAD map。基於大量測試HDAD maps，本文提出的DeepBinarization方法於recall、 specicity、precision、F-measure 和可視化效果等特點上大大優於傳統的二值化方法和基於卷
積神經網絡的二值化方法。

Binarizing historically degraded as-built drawing (HDAD) maps is a very important and necessary task because after that, the subsequent manipulations, such as the digitization, editing, retrieval, and compression, can be followed. Based on end-to-end fully convolutional networks (EEFCNs), in this paper, we propose a coarse-to-fine based binarization for HDAD maps. For convenience, the proposed deep learning-based binarization method is called the DeepBinarization method. Based on the selected training HDAD maps, the proposed DeepBinarization method first produces the groundtruth foreground (GF) models, and then the GF models are used in the training step to train the learned weights of EEFCNs. Secondly, for each input HDAD block, we propose a hybrid-based approach to generate the estimated foreground (EF) model of that block, and then the EF model is used in the
testing step to refine the rough binarized HDAD block generated from EEFCNs, producing a highquality binarized HDAD map. Based on a large amount of test HDAD maps, thorough experiments have been carried to show that in terms of recall, specificity, precision, F-measure, and visual effect, the proposed DeepBinarization method substantially outperforms the traditional binarization methods and the state-of-the-art convolutional neural networks-based binarization methods.

指導教授推薦書i
論文口試委員審定書ii
中文摘要iii
Abstract in English iv
Contents v
List of Figures vii
List of Tables viii
1 Introduction 1
1.1 Related works and motivation 1
1.2 Contributions 4
2 The proposed method for building up the estimated foreground (EF) models 6
2.1 The proposed MLT binarization method 6
2.2 The iterative histogram equalization-based global thresholding method by Kavallieratou:IHEGT 7
2.3 The hybrid-based approach to build up the EF and GF models 8
3 The proposed DeepBinarization method on EEFCNs 11
3.1 The determination of 5-EEFCNs 11
3.1.1 The determination of encoder configuration 11
3.1.2 The determination of decoder configuration 15
3.2 Coarse-to-fine training step 16
3.3 Coarse-to-fine testing step 17
4 Experimental result 19
4.1 Performance comparison metrics 19
4.2 Binarization accuracy and Visual effect merit 20
5 Conclusion 24

[1] V. Badrinarayanan, A. Kendall, and R. Cipolla
Segnet: a deep convolutional encoder-decoder architecture for image segmentation
IEEE Transactions on Pattern Analysis and Machine, 39 (12) (2017), pp. 2481-2495.
[2] J. Calvo-Zaragoza, G. Vigliensoni, I. Fujinaga
Pixel-wise binarization of musical documents with convolutional neural networks
IAPR International Conference on Machine Vision Applications, (2017), pp. 362-365.
[3] Q. Chen, Q.S. Sun, P.A. Heng, and D.S. Xi
A double-threshold image binarization method based on edge detector
Pattern Recognition, 41 (4) (2008), pp. 1254-1267
[4] Y.H. Chiu, K.L. Chung, W.N. Yang, Y.H. Huang, and C.H. Liao
Parameter-free based two-stage method for binarizing degraded document images
Pattern Recognition, 45 (12) (2012), pp. 4250-4262.
[5] V. Dumoulin, F. Visin
A guide to convolution arithmetic for deep learning
arXiv: 1603.07285, (2016).
[6] ftp://140.118.175.164/HDAD maps.
[7] B. Gatos, I. Pratikakis, and S.J. Perantonis
Adaptive degraded document image binarization
Pattern Recognition, 39 (3) (2006), pp. 317-327.
[8] R. C. Gonzalez and R. E. Woods
Digital Image Processing
4th Edition, Pearson, NJ, USA, 2017.
[9] N.R. Howe
Document binarization with automatic parameter tuning
International Journal on Document Analysis and Recognition, 16 (3) (2012), pp. 247-258.
25
[10] E. Kavallieratou
A Binarization Algorithm Specialized on Document Images and Photos
in: Proceedings of the 8th International Conference on Document Analysis and Recognition,
(2005), pp. 463-467.
[11] J.N. Kapur, P.K. Sahoo, A.K.C. Wong
A new method for gray level picture thresholding using the entropy of histogram
Computer Vision, Graphics, and Image Processing, 29 (3) (1985), pp. 273-285.
[12] J. Long, E. Shelhamer, and T. Darrell
Fully convolutional networks for semantic segmentation
in: Proceedings of the CVPR, (2015), pp. 3431-3440.
[13] Y. LeCun, Y. Bengio, and G. Hinton
Deep learning
Nature, 521 (7553) (2015), pp. 436-444.
[14] R.F. Moghaddam and M. Cheriet
A multi-scale framework for adaptive binarization of degraded document images
Pattern Recognition, 43 (6) (2010), pp. 2186-2198.
[15] W. Niblack
An introduction to digital image processing
Prentice-Hall, Englewood Cliffs, NJ (1986), pp. 115-116.
[16] N. Otsu
A threshold selection method from gray level histograms
IEEE Trans. Syst. Man Cybern, 9 (1) (1979), pp. 62-66.
[17] J. Sauvola and M. Pietikäinen
Adaptive document image binarization
Pattern Recognition, 33 (2) (2000), pp. 225-236.
[18] C. Tensmeyer and T. Martinez
Document image binarization with fully convolutional neural networks
in: 14th IAPR International Conference on Document Analysis and Recognition (ICDAR), 1
(2017), pp. 99-104.26
[19] Q.N. Vo, S.H. Kim, H.J. Yang, and G. Lee
Binarization of degraded document images based on hierarchical deep supervised network
Pattern Recognition, 74 (2018), pp. 568-586.
[20] M. Zhao,Y. Yang,H. Yan
An adaptive thresholding method for binarization of blueprint images
Pattern Recognition Letters, 21 (10) (2000), pp. 927-943.
[21] M.D. Zeiler
Adadelta: an adaptive learning rate method
arXiv: 1212.5701, (2012).