臺灣博碩士論文加值系統

直方圖均化（Histogram Equalization, HEQ）是一種處理曝光不足或曝光過度的影像方法，它能參考目標的累積分佈函數（Cumulative Distribution Function, CDF）進行分佈的轉換，此方法直觀且不需繁複的細節，近年來有許多研究將此方法應用在不同的領域上以及解決不匹配的問題，例如強健語音辨識的噪音改善問題，或是自然語言處理的跨語料庫問題。

本論文針對HEQ進行幾項基本的實驗，並利用分類問題來評估該方法對於轉換問題是否有效。結果顯示，在資料量不足的情況所得到的粗糙CDF分佈曲線，HEQ反而會造成錯誤的映射而影響轉換的結果。HEQ直接轉換與間接轉換的作法對於線性轉換與非線性轉換皆有接近的效果，但間接轉換的作法在不同分類器的情況顯得不穩定。在資料量充足的前提下，HEQ和標準差倍數正規化法在線性轉換下達到相近的效果，但對於非線性轉換，HEQ則有比標準差倍數正規化法更好的效能。

Histogram equalization (HEQ) is a technology for improving the darkness and the brightness of the image by adjusting the gray levels based on the cumulative distribution function (CDF). In recent years, this method has been applied to different issues, including robust speech recognition for solving the mismatch between the noisy speech and the clean speech, and natural language processing for the cross-database problem.

This paper analyzed how histogram equalization may influence a simple classification problem by simulation. The results showed the rough curve of CDF caused by insufficient data would lead to the poor mapping between training and test data and degrade the performance. Direct and indirect operations of histogram equalization achieve similar performance for linear or non-linear transformation, while the performance of the indirect one is more sensitive to type of classifiers. With sufficient amount of training data, HEQ and mean-standard deviation weight (MSW) can achieve compatible performances for linear transformation, while HEQ appears superior for nonlinear transformation.

第1章 緒論 1
1.1 研究動機 1
1.2 背景簡介 2
1.3 成果簡介 3
1.4 討論組織與架構 4
第2章 文獻探討與相關技術 5
2.1 直方圖均化 5
2.2 標準差倍數正規化法 8
2.3 高斯混合模型分類器 10
2.4 多層次倒傳遞類神經網路簡介 11
2.5 本文摘要 15
第3章 基礎實驗 16
3.1 模擬資料的產生 16
3.2 分類器訓練 17
3.2.1 高斯混合模型分類器 17
3.2.2 多層次倒傳遞類神經網路分類器 17
3.3 訓練資料的多寡對直方圖均化的影響 18
3.4 直方圖均化對於資料轉換的效果 21
3.4.1 對於平移量變化的分類效果 23
3.4.2 對於伸縮變化的分類效果 26
3.4.3 對於雙彎曲函數的分類效果 27
3.5 直方圖均化與標準差倍數正規化法的比較 31
3.5.1 對於平移量變化的結果比較 31
3.5.2 對於伸縮變化的結果比較 33
3.5.3 對於雙彎曲函式轉換的結果比較 35
3.6 本章摘要 39
第4章 結論與未來研究方向 40
4.1 結論 40
4.2 未來研究方向 41
參考文獻 42

[1]Robert Hummel,” Image enhancement by histogram transformation”, Comp. Graph. Image Process , vol. 6, pp. 184-195, 1977.
[2]John B. Zimmerman, Stephen M. Pizer, Edward V. Staab, J. Randolph Perry, William Mccartney, and Brenton, “An Evaluation of the Effectiveness of Adaptive Histogram Equalization for Contrast Enhancement,” IEEE TRANSACTIONS ON MEDICAL IMAGING, vol. 7, pp. 304-312, December 1988.
[3]Stephen M. Pizer, E. Philip Amburn, John D. Austin, Robert Cromartie, Ari Geselowitz, Trey Greer, Bartter Haar Romeny, John B. Zimmerman and Karel Zuiderveld, “Adaptive histogram equalization and its variations,” Computer Vision, Graphics, and Image Processing, vol. 39, pp. 355-368, September 1987.
[4]Rajesh Balehandran and Richard J. Mammone, “NON-PARAMETRIC ESTIMATION AND CORRECTION OF NON-LINEAR DISTORTION IN SPEECH SYSTEMS,” in Proc. ICASSP, 1998.
[5]S. Dharanipargda and M. Padmanabhan, “A Nonlinear Unsupervised Adaptation Technique for Speech Recognition,” In Proc. Int. Conf. on Spoken Language Processing (ICSLP), pp. 556-559, October 2000.
[6]C. Y. Wan and L.S. Lee, “Joint Uncertainty Decoding (JUD) with Histogram-Based Quantization (HQ) for Robust and/or Distributed Speech Recognition,” in Proc. ICASSP, 2006.
[7]C.Y. Wan and L.S. Lee, “Histogram-based quantization (HQ) for robust and scalable distributed speech recognition,” in Proc. EUROSPEECH, 2005.
[8]Shih-Hsiang Lin, Yao-Ming Yeh, and Berlin Chen, “A Comparative Study of Histogram Equalization (HEQ) for Robust Speech Recognition,” Computational Linguistics and Chinese Language Processing, vol. 12, pp. 217-238, June 2007.
[9]陳弈璁，「應用多特徵與正規化於統計式未知詞萃取之研究」，碩士論文，國立台灣科技大學資訊管理研究所，台北（2010）
[10]J. A. Bilmes, “A Gentle Tutorial of the EM Algorithm and its Application to Parameter Estimation for Gaussian Mixed and Hidden Markov Models,” University of Berkeley, Tech. Rep. ICSI-TR-97-021, 1997.
[11]David Rumelhart, James McClelland, eds., “Parallel Distributed Processing : Explorations in the Microstructure of Cognition,” Cambridge, Ma: MIT Press, Vol.1, 1986.