spit-iccict 2012

2 downloads 0 Views 3MB Size Report
Ycbcr skin color model. The third, paper uses well known opening and closing morphological operations to exclude the impact of noise on color segmentation.
2012 International Conference on Communication, Information & Computing Technology (ICCICT), Oct. 19-20, Mumbai, India

Reliable Face Detection in Varying Illumination and Complex Background Varsha Powar

Aditi Jahagirdar

PG Student, Department of IT, MITCOE Pune MITCOE, Pune Pune, India [email protected]

Assistant Professor, Department of IT, MITCOE Pune MITCOE Pune Pune, India [email protected] If we can find out some practical algorithms to filter out the noise generated by the above mentioned cases, it will be possible to construct a face detection and authentication system which has practical application value.

20 12

Face detection according to various approaches are classified into following four categories: Knowledge-based methods[2]. These techniques derived from heuristics, which are commonsense rules derived from human knowledge. It is usually the relationships between facial features. For example a face appears with two eyes that are symmetric to each other, a nose and a mouth and relationships between them can be represented by their relative distance and position. Problems of these methods consist in the difficulty to transfer the human knowledge into well-defined rules.

CT

Abstract— In this paper we propose a novel face detection approach. In general, lighting in images can have negative impact on the performance of face detection systems. Many current face detection systems are only reliable under controlled condition such as indoors under stable lighting conditions. So there is need to devise a solution to overcome illumination problems faced by face detection system. As one of the solution proposed approach firstly adopts a ‘reference white’ to realize light compensation; Then it realizes the detection of candidate face region using Ycbcr skin color model. The third, paper uses well known opening and closing morphological operations to exclude the impact of noise on color segmentation. Last Adaboost algorithm is improved and used to accurate face detection. The system uses java to achieve and is verified in self designed photo library, and is particularly suitable for illumination variations and complex background.

INTRODUCTION

IC

I.

T-

In the past decade, numerous face detection applications have been developed in order to enhance the human computer interaction with new non-traditional interfaces.

S.

P. I.

Face detection is a complex and challenging pattern detection problem which involves difficulties like varying illumination, complex background and other factors such as geometry and appearance of the face. The focus of our work is on illumination variation and complex background. Illumination variation can be caused by lighting source, impact of light such as brightness, contrast changes, shadows of the image and imaging conditions such as focal length of the camera, imaging distance, incorrect camera settings and so on. Furthermore, the human face has rather complicated detail changes such as shape of the face, skin color etc, in addition to different expressions such as opening and closing of eyes and mouth. External factors such as glasses, hair and head accessories as well as change in external conditions increase the difficulty of face detection. Different gestures of the faces which are caused by different imaging angles such as plane rotation, deep rotation (which has a greater impact); up and down spin etc also create obstacles in the face detection process.

978-1-4577-2078-9/12/$26.00©2011 IEEE

Feature invariant approaches[4][5]. These methods are based on the observation that humans can detect faces and objects in different poses and lighting conditions, so there must be some invariant properties over these variable scenes. A statistical model is built over the extracted features to describe their relationships. In these methods, however, image features can be corrupted due to illumination, occlusion and noise. Skin color methods belong to this group.

CI

Keywords- Light Compensation, Reference white, Skin color model, YCbCr, Haar features, Adaboost, Improved Adaboost

Template matching methods[3]. A standard face pattern, usually a frontal face, is manually predefined or parameterized by a function. To determine if a face is present or not in a scene, correlation values with the standard patterns are evaluated. These methods have the advantage of being simple to implement but they are inadequate since they don’t deal with scale, pose and shape. Appearance-based methods[4][5]. Machine learning based models (or templates) are learned from a set of training images which capture the representative variability of facial appearance. One popular appearance based approach which is also used in our work is to use of wavelet-based features followed by the Adaboost algorithm. As the human face is an extremely complex non- rigid model, the detection method with integration of multiple feature information has given more attention. In order to study issue like variant illumination and complex background, it is needed to build a reliable and robust face detection system. In the following section system architecture of face detection is presented.

1

2012 International Conference on Communication, Information & Computing Technology (ICCICT), Oct. 19-20, Mumbai, India of the skin region lies in chrominance space. Y is luminance component which carries no information. So, only consideration of chrominance components increases the speed of skin detection. In this system, program checks if values of Cb, Cr pixels satisfy condition given in algorithm, then this pixel considered as skin pixel. The algorithm is described below: Temp t if (Cr > 137 && Cr < 177 { if (Cb < 127 && Cb > 77) {

II.

System Architecture of Face Detection System

t = Cb + 0.6 * Cr; if (t > 190 && t < 215)

FACE DETECTION FOR VARYING ILLUMINATION

A. Light Compensation Poor illumination conditions may mean that color information is degraded and so techniques that rely upon color information may fail. One of the most popular techniques for correcting poor lighting condition is Light compensation. The brightest regions of the image are defined to be the reference white, and all other color channels are then scaled so that these regions become true white [1].

} }

return false;

Thus the accuracy of given algorithm is quite good but skin detection alone is not good to detect faces from any given image. It significantly drops its performance when image is having complex background. To overcome this problem we have used combined approach of light compensation, skin color model and Adaboost together. Adaboost is discussed in later section.

CT

We have proposed a simple and effective method for light compensation. We lighten the whole image through the reference white approach with proper brightness. The experiments show that our method can make it possible to enhance the brightness of the dark regions and meanwhile can preserve the details of the bright regions. However, the detail enhancing method usually enhances the noise too but it can be removed in the later using noise removal techniques.

return true;

20 12

Figure 1.

IC

CI

D. Skin Binarization After color segmentation the color image in complex background turned to a binarization image. First of all we have to convert skin colored image to binary image means the pixel which are containing skin color will be converted into white pixel and rest of the pixel will be converted to black. Then it further refined through binary morphological operations, opening and closing operations to reduce the background noise and remove holes within faces.

P. I.

T-

B. Color Space In order to segment human skin regions from non-skin regions, a reliable skin model is needed that is adaptable to different colors and light conditions. The common RGB representation of color images is not suitable for characterizing skin-color. The color spaces that are frequently used in studies are HIS, HSV, YCbCr. YCbCr color space has been defined in response to increasing demands for digital algorithms, and has since become a widely used model in a digital video/image.

S.

The YCbCr color space is commonly used in image processing as it separates the luminance, in Y component, form the chrominance described through Cb and Cr components. The Cb and Cr components are used to characterize the skin color information. The formulae for converting from RGB to YCbCr are given below. Y = 0.299R + 0.587G + 0.114B Cb = -0.169R - 0.332G + 0.500B

(1)

Cr = 0.500R - 0.419G - 0.081B C. Skin Detection Algorithm In his system skin color model is based on chrominance component Cb and Cr. It is based on the assumption that most

978-1-4577-2078-9/12/$26.00©2011 IEEE

III.

ADABOOST FOR FACE DETECTION

After the skin regions obtained, we applied Adaboost training and classification algorithm. In 2001, Paul Viola and Michael Jones have proposed Adaboost algorithm for object detection [2]. The classifiers used in early work on face detection, for example, neural networks [4] and SVM [5], were complex and computationally expensive. In Adaboost, most non-face patches were quickly rejected by the early nodes. Cascade detectors have demonstrated impressive detection speed and high detection rates This algorithm is very fast and effective for object detection. The framework of this method includes following steps[2]: •

Use of haar like features and computation of haar features using integral image.

2

2012 International Conference on Communication, Information & Computing Technology (ICCICT), Oct. 19-20, Mumbai, India •

Selection of rectangle feature(weak classifier) to form a strong classifier.



Cascade framework and use of adaboost to train cascade nodes.

3) Programming in Java offers the interesting opportunity to easily port applications to the Web, through the mechanism of applets. 4)

The algorithm of Adaboost works as follows[2]: Adaboost starts with uniform distribution of “weights” over training examples. The weight tells the learning algorithm the importance of the example.

2.

Obtain a weak classifier from the weak learning algorithm ,hj(x).

3.

Increase the weights on the training examples that were misclassified.

4.

Repeat.

5.

At the end, carefully make a linear combination of the weak classifiers obtained at all iterations. IV.

IMPROVED ADABOOST FOR FACE DETECTION

Viola and Jones, first, proposed an “Integral Images” for each image to speed up the calculation of Haar Features. Second, they proposed an Adaboost algorithm for determining weak classifiers Finally, they formed cascade classifier by combining strong classifiers.

VI.

EXPERIMENTAL RESULTS

In this paper, our face detection system was carried out on an Intel Pentium processor with 2 GHZ processor, RAM of 1 DDR GB. System is windows XP. For our part we have chosen java. Front end of our system is designed with Swing, Development kit and IDE used is JDK 1.6 and Net beans 6.9.1 respectively. In this experiment, Standard formulas have been used to compute the precision and recall for query images is shown below. The evaluation metrics consists of Precision and recall rate of the detection. Table III shows that proposed approach gives higher precision and recall rate as compared to skin detection and adaboost algorithm. The test images are digital photos from life, personal photo and collection from the internet (indoor and outdoor).

V.

Figure 2.

Test Result of Image1

Figure 3.

Test Result of Image2

P. I.

T-

IC

CI

CT

Theoretically more the training samples, better the result is, but updating training samples is the procedure of reusing training samples which makes the training process overfitting and time consuming. For avoiding the phenomena of overfitting in training process, this paper proposes a new method to update weight[7], its basic idea is to set a weight threshold focusing on each loop, when the sample is sorted by error and its current weight is smaller than the threshold, the weight of this sample is increased, otherwise the weight is decreased. At one time, the process of constructing cascade classifier is added to training process, so we do not need to construct cascade classifier solely[8]. Thus even if the difficult samples are misclassified in each round, their weight will not be increased too much. So it can avoid the degradation of the classifier to some extent.

5) It has also other advantages in terms of computation speed and modularity.

20 12

1.

The system is quite stable and appears to be robust against the student’s programming errors.

IMPLEMENTATION OF FACE DETECTION SYSTEM

S.

Most of the face detection systems are developed in traditional way either through MATLAB or VC++.We developed our proposed system using JAVA as a programming language. The image processing parts of Java are buried within the java.awt.image package. For experimental simulation we have chosen java because of the following arguments. 1) Java is a platform independent. Java is free and network ready. 2) Java is robust with a good handling of errors and garbage collection; this eliminates the main source of bugs and crashes.

978-1-4577-2078-9/12/$26.00©2011 IEEE

3

2012 International Conference on Communication, Information & Computing Technology (ICCICT), Oct. 19-20, Mumbai, India TABLE I.

COMPARISON OF DETECTION SPEED Image1

Face Detection Methods

FUTURE WORK

Image2

Skin Detection

~250 ms

~295 ms

Adaboost

~200 ms

~235 ms

Improved Aaboost

~110 ms

~150 ms

This system can be implemented in future for online face detection and user authentication. Simulated system can also be extended on a wide range of small low power devices, including hand-held’s and embedded processors. REFERENCES

TABLE II.

EXPERIMENTAL RESULT OF DIFFERENT VARIATION IN AN IMAGE

1

Single Face Image

150

Correct Faces Detected 150

2

Single face with complex background

150

150

3

Illumination Variation

120

115

4

Multiple faces with poor illumination

120

110

Image Type

QUANTITATIVE EVALUATION

Face Detection Methods

Precision

Recall

Skin Detection

86.66 %

87.79 %

Adaboost

86.13 %

76.12 %

Improved Adaboost

92.3%

93.10 %

CONCLUSION

CT

TABLE III.

No. of Images in the Gallery

Jinjun Wang, Huiyuan Wang, A Survey Preprocessing Methods for Illumination Insensitive Face Recognition. Journal of Electrical & Electronic Education, Vol 30, pp.43-46, December 2008 (In Chinese). [2] Paul Viola, Michael Jones, "Robust Real-time Object Detection", Second International Workshop on Statistical and Computational Theories of Vision -Modeling, Learning,Computing, and Sampling, 2001. [3] Ming-Hsuan Yang, David J. Kriegman, and Narendra Ahuja, “Detecting faces in images: A survey”,IEEE transactions on pattern analysis and machine intelligence, vol. 24, no. 1, january 2002. [4] H.A. Rowley, S. Baluja, and T. Kanade, “Neural Network-Based Face Detection,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 20, no. 1, pp. 23-38, Jan. 1998. [5] E. Osuna, R. Freund, and F. Girosi, “Training Support Vector Machines: An Application to Face Detection,” Proc. Int’l Conf. Computer Vision and Pattern Recognition, pp. 130-136, 1997. [6] Raman Maini & Dr. Himanshu Aggarwal “Study and Comparison of Various Image Edge Detection Techniques “International Journal of Image Processing (IJIP), Jan-Feb 2009, Volume (3) Issue (1)pp.1-11. [7] XIONG Sheng-wu,ZONG Xin-lu and ZHU Guo-feng,”Improved face detection method based on AdaBoost algorithm”,Application Research of Computers,Vol.24,No.11,Nov.2007,298-300. [8] Lin Peng,”study and implementation of face detection based on adaboost algorithm”,University of Technology,2007,pp.24-28. [9] R. C. Gonzales and R. E. Woods, “Digital Image Processing”, 2nd , Prentice Hall, Upper Saddle River, NJ, 2002. [10] Wilhelm Burger, Mark J. Burge,“Digital Image Processing An algorithmic Introduction Using Java.

20 12

Sr.No

[1]

S.

P. I.

T-

IC

CI

Skin detection is very simple technique for finding faces. Use of haar features and Adaboost made face detection practical. It also gives high speed and detection rate compared to previous face detection methods. Proposed face detection approach which combines skin color model and Adaboost together reduces false detection rate in complex background and in varying illumination of image.

978-1-4577-2078-9/12/$26.00©2011 IEEE

4