1509.04887.pdf论文翻译
An Improved Algorithm for Eye Corner Detection
Anirban Dasgupta, Anshit Mandloi, Anjith George and Aurobinda Routray
Department of Electrical Engineering
Indian Institute of Technology
Kharagpur, India 721302
[email protected], [email protected], [email protected], [email protected]
译文：
文摘：本文提出了一种改进的检测算法。给出了鼻和颞眼角的问题。该算法是一个对桑托斯和普卡法的修改。在第一步中,我们使用基于Haarlike的分类器检测人脸和眼睛特性。然后我们将巩膜，从检测到的眼睛区域。从分段的巩膜，我们分离出一个近似眼睑轮廓。获得眼角的候选人使用哈里斯和史蒂芬斯角探测器。我们引入一个postpruning在眼角的余光中找到了眼角，最后。该算法已经在耶鲁大学，JAFFE数据库中进行了测试以及我们创建的数据库。
————————————————————————————————————————————
摄像机——面部检测——ROI选择——眼部检测——巩膜分割——候选眼角检测——后剪枝（后处理）——最终眼角检测
—————————————————————————————————————————
the next step is segmentation of the sclera region. The sclera being the most unsaturated region in the eye image, an HSV transformation yields the lowest magnitudes on the saturation plane. The iris and sclera being segmented out, the next stage constitutes of the approximation of the eyelids contour. This is achieved using morphological dilation of the iris segmentation mask with a horizontal structuring element is carried out. This expands the iris regions horizontally. Finally, point-bypoint multiplication between the dilated and the enhanced data provides a good approximation to the eyelids contour. The subsequent step comprises of the generation of a set of candidate eye corner positions, which was performed using the Harris and Stephens corner detector. This method works well for an image having proper resolution and clarity. However, for inferior quality images, such as those obtained using a standard webcam, the performance of such methods is limited.
下一步是巩膜区域的分割。巩膜作为眼睛图像中最不饱和的区域，HSV转化产生的饱和度最低飞机。虹膜和巩膜被分割，下一阶段构成了眼睑轮廓的近似。这是利用虹膜分割的形态扩张采用水平结构单元进行掩模。这将横向扩展虹膜区域。最后,point-bypoint放大和增强之间的乘法数据提供了一个很好的近似于眼睑轮廓。接下来的步骤包括生成一组候选眼角位置，使用的是哈里斯和斯蒂芬斯角探测器。这种方法对于具有适当分辨率的图像效果很好。和清晰。但是，对于质量低劣的图像，比如那些使用一个标准的网络摄像头，这样的性能方法是有限的。
————————————————————————————————————————————
Our proposed scheme is a real-time camera-based system depicted as a schematic in Fig. 1. For image acquisition, thescheme uses a standard webcam of resolution 640480 pixels.The algorithm has been implemented in real-time using a SonyPS3 Eye webcam at 30 fps. First, face detection is carried outin a given frame, followed by eye detection. Eye detectionis confined to a Region of Interest (ROI) in the detected facearea. The RoI is selected based on our previous work [6]. Thisstep finds each eye separately. The algorithm has been maderobust using some preprocessing techniques for illuminationcorrection.
我们提出的方案是一个基于摄像头的实时系统图1描绘的示意图。图像采集的scheme使用标准的分辨率为640x480像素的网络摄像头。该算法是通过索尼实时实现的PS3的眼睛摄像头，30 fps。首先，进行人脸检测在一个特定的坐标系中，然后是眼睛检测。眼睛检测被限制在检测到的区域内的一个感兴趣的区域（ROI）区域。RoI是根据我们以前的工作6来选择的。这一步一步地找到每只眼睛。
————————————————————————————————————————————
A. Image Enhancement
Preprocessing of the image is necessary for filtering out the noise and preserving relevant information. The filtering is particularly essential as the eye corner is sometimes not visible clearly because of illumination conditions. The pre-processing begins by equalizing the histogram of the image. We have used Contrast Limited Adaptive Histogram Equalization (CLAHE) technique [7] for this purpose. In this method, the image is fractioned into small blocks of size 88. Each of these blocks undergoes histogram equalization. This confines the histogram to a small region. An issue of this method is the amplification of noise if present. This issue is avoided by contrast limiting. After the CLAHE operation, bi-linear interpolation is applied to remove artifacts in tile borders.
A.图像增强
图像的预处理对于滤除噪声和保留相关信息是必要的。过滤特别重要，因为由于照明条件，有时不能清楚地看到眼角。预处理开始于均衡图像的直方图。为此，我们使用了对比度有限自适应直方图均衡（CLAHE）技术[7]。在该方法中，图像被分成大小为88的小块。这些块中的每一个都经历直方图均衡。这将直方图限制在一个小区域。该方法的问题是如果存在噪声的放大。对比度限制可以避免这个问题。在CLAHE操作之后，应用双线性插值来移除图块边界中的伪像。
————————————————————————————————————————————
B. Face and Eye Detection
The face and eye detection forms the first step in the algorithm for the localization of the eye corners. A classifier based on Haar-like features was selected for this stage. For optimal use, we have used parameters based on the earlier work [8]. Once the face is detected, an ROI is selected from the facial region. The ROI selection scheme has been reported in the previous work [9]. Now, the search for eyes is confined to a reduced area, which also reduces computation and improves the speed.
B.面部和眼睛检测
面部和眼睛检测形成了用于眼角定位的算法的第一步。在此阶段选择了基于类Haar特征的分类器。为了最佳使用，我们使用了基于早期工作的参数[8]。一旦检测到面部，就从面部区域中选择ROI。 ROI选择方案已在之前的工作中报道[9]。现在，对眼睛的搜索仅限于缩小的区域，这也减少了计算并提高了速度。
————————————————————————————————————————————
C. Eye Corner Detection
The corner detection operates on the detected eye region.
C.眼角检测
角点检测对检测到的眼睛区域进行操作。
1) Sclera Segmentation: As proposed in [3], the sclera is segmented out by converting the RGB image to HSV and subsequently thresholding the saturation plane. For grayscale images, the color space conversion is not required, and the thresholding operation can be applied directly. The idea behind this lies in the fact that the sclera is the most unsaturated segment in the eye image. There remain certain noisy pixels because of the blood vessels in the sclera. This issue is addressed using the morphological opening of the sclera portion using an elliptical mask.
1）巩膜分割：如[3]中所提出的，通过将RGB图像转换为HSV并随后对饱和平面进行阈值处理来分割出巩膜。对于灰度图像，不需要颜色空间转换，并且可以直接应用阈值处理操作。这背后的想法在于巩膜是眼睛图像中最不饱和的部分。由于巩膜中的血管，仍然存在某些噪声像素。使用椭圆形掩模使用巩膜部分的形态开口来解决该问题。
2) Eyelid Contour Approximation: The sclera region is hence segmented out with the morphological post-processing. The boundary of the mask is overlaid on the eye image to obtain the eyelid contour approximation.
2）眼睑轮廓近似：因此，巩膜区域通过形态学后处理被分割出来。掩模的边界覆盖在眼睛图像上以获得眼睑轮廓近似。
3) Eye Corner Candidate selection: The corner score is obtained by the sum of squared differences (SSD), S(x;y) between the corresponding pixels of two patches in the eyelid contour image I(x;y). The differential of the corner score is considered for finding out the corner candidates. A circular window w(u;v) is used, to make the response isotropic.
3）眼角候选者选择：通过眼睑轮廓图像I（x; y）中的两个斑块的相应像素之间的平方差（SSD），S（x; y）之和来获得角分数。角点分数的差异被认为是用于找出角落候选者。使用圆形窗口w（u; v），以使响应各向同性。
With Taylor series expansion and proper approximation, we have
The Harris matrix, H is obtained as
4) Post Pruning: Since the actual eye corners will lie at the extreme ends of the eyelid contour, the eye corner pair having the farthest distance are selected as the eye corners. In cases of a tie, the mean corner is chosen as the correct corner estimate, among the corner candidates bearing equal distances.
4）后修剪：由于实际眼角位于眼睑轮廓的末端，因此选择具有最远距离的眼角对作为眼角。在平局的情况下，在具有相等距离的角落候选者中，选择平均角作为正确的角估计。
Fig. 6. Detection Results for images taken (a) Yale Database (b) JAFFE Database
Fig. 7. Sample images from our created database IV. RESULTS
————————————————————————————————————————————
IV. RESULTS
A face database of 30 subjects was prepared using a standard Sony PS3 web camera for testing the algorithm. Some sample images of the database is shown in Fig. 7. The algorithm has also been tested on Yale Face Database [10] and JAFFE Database [11]. A sample set of 200 images total were randomly selected from the databases. The percentage of mean-squared pixel error in eye localization has been provided in Table II for the different databases. The error is eye corner localization is based on the manual marking of the endpoints. The processing-rate of the algorithm with online testing was found to be 16.2 fps while on offline database, it was 20.4 fps. The speed may be improved by using GPU based implementations, and employing parallel schemes.
IV.结果
使用标准Sony PS3网络摄像机准备30名受试者的面部数据库以测试该算法。数据库的一些样本图像如图7所示。该算法也在耶鲁人脸数据库[10]和JAFFE数据库[11]上进行了测试。从数据库中随机选择总共200个图像的样本集。表II中针对不同的数据库提供了眼睛定位中的均方像素误差的百分比。错误是眼角定位是基于端点的手动标记。在线测试的算法处理速度为16.2 fps，而在离线数据库中，处理速率为20.4 fps。可以通过使用基于GPU的实现并采用并行方案来提高速度。
————————————————————————————————————————————
V. CONCLUSION
In this paper, we propose an algorithm that uses a standard web camera to localize effectively the eye corner. This is an improvisation over the Santos and Proenca Method. The significant modification lies in the post-pruning the eye corner candidates. The method has less than 10% localization errors in all the three tested databases. A future scope of the work may be testing of the algorithm of infrared and near infra-red images [12], and make appropriate modifications so that the applications of the algorithm can be extended to areas where night vision is preferable.
V.结论
在本文中，我们提出了一种算法，该算法使用标准网络摄像头有效地定位眼角。这是对Santos和Proenca方法的即兴创作。重大修改在于修剪眼角候选者。该方法在所有三个测试数据库中的定位误差小于10％。未来的工作范围可能是测试红外和近红外图像的算法[12]，并进行适当的修改，以便算法的应用可以扩展到夜视最好的区域。
————————————————————————————————————————————
ACKNOWLEDGMENT
The authors would like to acknowledge the subjects for participation in the experiment.
致谢
作者要感谢参与实验的主题。