A Top-down Attention Based Approach for Printed Mathematical
發(fā)布時間:2022-07-19 17:14
研究背景和意義隨著互聯(lián)網(wǎng)的發(fā)展,如今絕大多數(shù)材料以電子文檔的形式存儲在計算機上。印刷文件也可以通過掃描,拍照等轉(zhuǎn)換成電子文檔?蛇x字符識別(OCR)是將文檔內(nèi)容轉(zhuǎn)換為計算機文本的關(guān)鍵技術(shù)。經(jīng)過多年的發(fā)展,OCR技術(shù)逐漸成熟,其應(yīng)用范圍也越來越廣泛。目前,OCR不僅可以識別文獻中的常用詞,還可以識別數(shù)學(xué)表達式(ME)。通過使用OCR來識別打印的數(shù)學(xué)表達式,可以實現(xiàn)數(shù)學(xué)表達式的重用,F(xiàn)有的OCR系統(tǒng)能夠準(zhǔn)確有效地識別文檔中的字符,但仍然無法很好地處理數(shù)學(xué)表達式。對于某些沒有特殊數(shù)學(xué)符號的一維數(shù)學(xué)表達式,可以識別它,但對于諸如積分符號和根符號的數(shù)學(xué)符號,它不能很好地工作。數(shù)學(xué)表達式仍保存為圖像,無法識別,無法編輯和重復(fù)使用。這使得一些以數(shù)學(xué)表達為中心的文章難以編輯,并且圖像占用大量存儲空間以影響傳輸速度。因此,擴展OCR系統(tǒng)的應(yīng)用以識別文本中的數(shù)學(xué)表達式具有重要意義。數(shù)學(xué)表達符號與普通文本不同,它們的布局主要以二維結(jié)構(gòu)呈現(xiàn),這導(dǎo)致字符在各種情況下被卡住并且分割的復(fù)雜性。使用傳統(tǒng)的字符粘附分割方法很難獲得滿意的結(jié)果。大多數(shù)方法只能解決一兩種特定情況,例如簡單的水平或垂直關(guān)系等。那些擁有復(fù)雜符號...
【文章頁數(shù)】:61 頁
【學(xué)位級別】:碩士
【文章目錄】:
Acknowledgements
Abstract
Chapter 1 Introduction
1.1 Motivation
1.2 Research Status
1.3 Objective
1.4 Thesis Structure
Chapter 2 Related Work
2.1 Artificial Neural Network
2.1.1 Feedforward Neural Network
2.1.2 Convolutional Neural Network
2.1.3 Recurrent Neural Network——Gated Recurrent Unit
2.2 Encoder-Decoder Framework
2.3 Attention Mechanism
2.3.1 Why Introduce Attention Mechanism
2.3.2 Classification of Attention Mechanism
2.4 Word Embedding
Chapter 3 Proposed Mathematical Expression Recognition System
3.1 Encoder
3.1.1 Feature Extraction
3.1.2 Context Representation
3.2 Decoder
3.2.1 Language Model
3.2.2 Decoder with Attention Mechanism
3.2.3 Attention Visualization
Chapter 4 Experiments
4.1 Model Architecture
4.1.1 Deep Convnets Architecture
4.1.2 Bi-RNN Architecture
4.2 Training Procedure
4.2.1 Data Preprocessing
4.2.2 Word Embedding
4.2.3 Experimental Environment
4.2.4 Experimental Parameter Setting
4.2.5 Testing Stage
4.3 Evaluation Metrics
4.3.1 Match Score
4.3.2 BELU
4.4 Comparison of Experimental Results
4.4.1 Experiment Ⅰ: with Or without Attention Mechanism
4.4.2 Experiment Ⅱ: with Or without Bi-RNN
4.4.3 Comparison with Other Systems
4.4.4 Experiment in Handwritten Mathematical Expression Recognition
Chapter 5 Summary and Future Work
5.1 Summary
5.2 Future Work
References
Appendix A
【參考文獻】:
期刊論文
[1]基于改進遺傳算法的下采樣圖像水印算法研究[J]. 魏志成,李建雄,戴居豐. 光電子.激光. 2007(07)
本文編號:3663818
【文章頁數(shù)】:61 頁
【學(xué)位級別】:碩士
【文章目錄】:
Acknowledgements
Abstract
Chapter 1 Introduction
1.1 Motivation
1.2 Research Status
1.3 Objective
1.4 Thesis Structure
Chapter 2 Related Work
2.1 Artificial Neural Network
2.1.1 Feedforward Neural Network
2.1.2 Convolutional Neural Network
2.1.3 Recurrent Neural Network——Gated Recurrent Unit
2.2 Encoder-Decoder Framework
2.3 Attention Mechanism
2.3.1 Why Introduce Attention Mechanism
2.3.2 Classification of Attention Mechanism
2.4 Word Embedding
Chapter 3 Proposed Mathematical Expression Recognition System
3.1 Encoder
3.1.1 Feature Extraction
3.1.2 Context Representation
3.2 Decoder
3.2.1 Language Model
3.2.2 Decoder with Attention Mechanism
3.2.3 Attention Visualization
Chapter 4 Experiments
4.1 Model Architecture
4.1.1 Deep Convnets Architecture
4.1.2 Bi-RNN Architecture
4.2 Training Procedure
4.2.1 Data Preprocessing
4.2.2 Word Embedding
4.2.3 Experimental Environment
4.2.4 Experimental Parameter Setting
4.2.5 Testing Stage
4.3 Evaluation Metrics
4.3.1 Match Score
4.3.2 BELU
4.4 Comparison of Experimental Results
4.4.1 Experiment Ⅰ: with Or without Attention Mechanism
4.4.2 Experiment Ⅱ: with Or without Bi-RNN
4.4.3 Comparison with Other Systems
4.4.4 Experiment in Handwritten Mathematical Expression Recognition
Chapter 5 Summary and Future Work
5.1 Summary
5.2 Future Work
References
Appendix A
【參考文獻】:
期刊論文
[1]基于改進遺傳算法的下采樣圖像水印算法研究[J]. 魏志成,李建雄,戴居豐. 光電子.激光. 2007(07)
本文編號:3663818
本文鏈接:http://sikaile.net/kejilunwen/shengwushengchang/3663818.html
最近更新
教材專著