臺灣博碩士論文加值系統

人臉偵測的技術目前已經融入在我們的生活中，並且廣泛的應用在手機及數位相機上，因此快速且精準地找到影像中的人臉是最重要的目標。本論文利用成熟的人臉偵測技術來完成一個船舶偵測器。

Viola與Jones臉部偵測器於2001年被提出，藉由蒐集人臉的樣本來分析其中的規律性，可以快速地偵測出影像中的人臉。但由於樣本都是採用正面且固定框架的人臉進行訓練，因此對於不同角度的人臉，Viola與Jones偵測器的表現並不理想；所以後人提出一個正規像素差特徵(Normalized Pixel Difference)的特徵值來改善這個問題，此特徵值是表示兩像素之間的關係，不具有固定的模板，因此可以偵測多元的影像；而後使用二次式來做為決策樹的判斷，這樣有效的將圖片歸類，進而完成分類器。

本論文是利用臉部偵測的技術，使用船隻的圖庫做為訓練樣本，再取出樣本的正規像素差特徵送入自適應增強演算法(AdaBoost)與級聯演算法訓練出船舶偵測器。本論文提出了三種作法，首先是使用船隻圖片做為樣本所訓練出的偵測器；其次是將樣本的船隻圖片去除背景後，送入演算法訓練出的偵測器；最後是先歸類出不同角度的船隻樣本，各自建立偵測器，讓各種角度的船隻可以進入該角度偵測器進行偵測。測試的過程中，我們對船隻圖片進行偵測，根據偵測結果計算其偵測率；我們也使用海面的圖片送入偵測器偵測，並且評估偵測器的錯誤率。最後我們發現，將船隻圖片去背景做為樣本所訓練出的偵測器擁有不錯的偵測率，並且具備最低的錯誤率。

Face detection is popular in mobile phones and digital cameras and has become a part of our life. For these applications, fast and accurate face detection is important. In this thesis, we apply these face detection techniques in ship detection.

Viola and Jones face detector was proposed in 2001. It analyzes the regularity among face samples and can detect face very quickly. However, because all face samples are frontal, Viola and Jones face detector is not good for face from different angles. Therefore, a new feature called “normalized pixel difference” (NPD) is proposed to solve this deficiency. NPD represents the relationship between two pixels. With this new feature and a deep quadratic tree classification, it can effectively detect face from different angles.

This thesis applies the face detector with NPD in ship detection. We propose three different approaches in data training. The first uses various ship images as training samples. The second removes the ocean background from the ship images in the training samples. The third divides the training images according to the ship’s orientation angles and, then, trains each classes separately. In the experiments, we uses both images with and without ships to test the detector. We find that the second approach can achieve the highest detection rate as well as the lowest failure rate.

誌謝 i
摘要 ii
Abstract iii
目次 iv
圖目次 v
表目次 vii
第一章 緒論 1
1.1 研究背景 1
1.2 方法簡述 1
1.3 論文架構 2
第二章 背景介紹 3
2.1 哈爾特徵(Haar-like Feature) 3
2.2 AdaBoost 演算法 6
2.3 級聯分析(Attentional Cascade) 8
2.4 Normalized Pixel Difference特徵值 10
2.5 Deep Quadratic Tree 12
2.6 總結 13
第三章 研究方法 14
3.1 訓練偵測器 14
3.2 作法一 15
3.3 作法二 17
3.4 作法三 18
3.5 總結 19
第四章 實驗結果 20
4.1 實驗概觀 20
4.2 偵測結果 21
4.3 偵測結果比較 27
4.4 總結 29
第五章 結論 30
參考文獻 31

[1]P. Viola and M. Jones, “Rapid Object Detection using a Boosted Cascade of Simple Features,” in Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recog., 2001, pp. I-511-I-518.
[2]Y. Q. Wang, “An Analysis of the Viola-Jones Face Detection Algorithm,” Image Processing On Line, vol.4, pp. 128-148, Jun. 2014.
[3]S. Liao, A. K. Jain, S. Z. Li, “A Fast and Accurate Unconstrained Face Detector,” IEEE Transactions on Image Processing. vol. 38, no. 2, pp. 211-223, Feb. 2016.
[4]J. Friedman, T. Hastie, and R. Tibshirani, “Additive Logistic Regression: A Statistical View of Boosting,” The Ann. Statist. vol. 28, no. 2, pp. 337–374, Apr. 2000.
[5]L. Bourdev and J. Brandt, “Robust Object Detection Via Soft Cascade,” in Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recog., 2005, vol. 2, pp. 236–243.