基於頭部偵測的膚色辨識方式

Skin Color Detection Based on Head Detection

羅群翰(Qun-Han Luo)

指導教授 ： 涂世雄

中原大學/工學院/電機工程研究所/碩士(2013年)

https://doi.org/10.6840/cycu201300928

摘要

在本篇論文中，我們提出一個基於頭部偵測的膚色偵測方法。我們主要的目的是，即使在一個與膚色相近的光線下我們依然可以辨識出屬於人體的膚色部份，以增加膚色的辨識率。
在本論文中，我們會介紹如何去找出我們所想要的膚色區域。首先我們會先利用 Prewitt 濾波器去找出輸入圖片的所有邊界。之後我們再利用自己設計的濾波器來去加強頭部兩側邊界的部份。其次，根據我們找出的頭部兩側邊界，我們取兩邊界的中心文參考點，並計算出兩中心參考點的歐基里德距離，以此距離當作我們的ROI(Region of Interesting)的寬度。然後我們再使用人臉的黃金比例來定義出ROI的長度，由此為基準，我們就可以定位出我們完整的頭部位置。第三，根據我們所訂的 ROI 在 YCbCr 色彩空間中做Gaussian分布的取樣。我們針對取樣出來的數值去找出它們的平均值，之後再給這個平均值一個誤差範圍，並以這個閥值區間作為我們的膚色分布範圍。第四，根據我們用所取出來的閥值在 YCbCr 色彩空間中來針對我們的輸入影像作分析，之後我們可以找出符合這個閥值區間的影像部分。最後藉由符合閥值區間的影像與原始影像作重疊處理，如此一來我們就能清楚分辨出在原始影像中可能的膚色範圍。最後，根據實驗出來的結果表示這個分析的方法是可行的。就算整個環境的光線顏色趨近於膚色，這個分析方式確實提升了膚色的辨識率。
在這篇論文中，我們的研究結果如下：
(1) 提升了頭部偵測的正確率：
在本篇論文中，我們解決了在頭部偵測中頭部歪斜的問題。結果顯示克服了歪斜的問題後，我們的頭部偵測正確率為90.27%
(2) 降低了膚色偵測的錯誤率：
由於我們是針對每一張不同的輸入影像來做分析，因此我們膚色偵測的正確率為93.77%

關鍵字

頭部 ； 膚色 ； 偵測 ； 定位

並列摘要

In this thesis, we propose a skin color detect method based on head detection, the main propose in this thesis is to increase the identification rate of skin color even the environment light is close to skin color or others.

In this thesis, we will introduce how to define our skin region. First, we use Prewitt filter to find the edge of input image. After that, we use the self-designed image filter to enhance the edge of the head on both sides. Second, according to the head edge, we locate its center position and calculate its Euclidean distance to be the width of ROI (Region of Interesting). We use the gold ratio of human face to define the length of our ROI so we can locate the head position. Third, focusing on the ROI, we will sample the pixel value by using Gaussian distribution in YCbCr color space. The average pixel value of the sample is calculated. We give an error rate for the average value let it be our skin color threshold. Fourth, we use the threshold to analyze the input image in YCbCr color space and find out the region which matches the threshold. By overlap the region which is matched the threshold with original image; we can distinguish clearly the possible skin colors of original image. Finally, the experimental result shows that the analysis method in this thesis is feasible. It actually increases the identify rate of skin color even the environment light is close to skin color.

In this thesis, the results of our research are as follows:
(1) Increase the head detection rate:
In our thesis, we solve the crooked problem of head detection. The result shows that the identify rate oh head detection is 90.27%.
(2) Reduce the error rate of Skin color detection rate:
We focused on every different condition to do the color sampling so our skin color detection rate is 93.77%.

並列關鍵字

location ； detection ； skin color ； head

參考文獻

[1] LUO Tao, Zaifeng Shi, GAO Ying, GAO Jing, “A Circuit of Configurable Skin Tone Adjusting method base on Exact Skin Color Region Detection”, 2011 IEEE

[5] M. Menke, M.S. Atkinst, and K.R. Buckley, “Compensation Methods for Head Motion Detected During PET Scans”, 1995 IEEE

[6] Carolina Moreno Salcedo, Cesar Augusto Pe˜ na, J´ es de Jesus Cerqueira, Antonio Marcus Nogueira Lima, “Designing a real time artificial vision system for human interaction with an omnidirectional mobile platform”, 2012 Brazilian Robotics Symposium and Latin American Robotics Symposium

[7] Rafael C. Gonzalez, Richard E. Woods, “Digital Image Processing 2/e”, 繆紹剛譯, 台灣培生教育出版

[8] Yohei Ishii, Hitoshi Hongo, Kazuhiko Yamamoto and Yoshinori Niwa, “Face and Head Detection for a Real-Time Surveillance System” Proceedings of the 17th International Conference on Pattern Recognition