A new face detection method based on shape information

https://doi.org/10.1016/S0167-8655(00)00008-8Get rights and content

Abstract

Automatic detection of human faces is one of the most difficult problems in pattern recognition. In many practical applications (e.g. personal identification using photo-IDs), face images often have a simple background. In this paper, we propose a new method based on shape information. Our system can be very efficient for images with simple background. Input image is first enhanced by means of histogram equalization, followed by edge detection based on a multiple-scale filter. The extracted edges are then linked using a method based on an energy function. The face contour is finally extracted using the direction information of the linked edges. Experimental results are included to verify the effectiveness of this method.

Introduction

Face detection is the important first step of face recognition. It is basically an image segmentation problem as the image is to be segmented into two parts: one containing faces and the other representing non-face regions. The performance of face detection influences the performance of the entire recognition system.

A number of methods have been proposed, and some have been used in practical applications (Miller, 1994). These methods can be broadly classified into two groups: the first group is based on separate facial features, which means the final decision comes from the integration of several detection results. Among these, Yang and Huang have described a system that uses a hierarchical knowledge-based method; Li and co-workers (Chow and Li, 1993, Li and Roeder, 1995) use several deformable templates to locate the face contour; facial feature detection and verification based on geometrical relations have been studied in (Jeng et al., 1996). The second group considers a face as a single pattern and features are extracted from the entire face region. Existing methods have employed techniques ranging from neural networks (Rowley et al., 1998a, Rowley et al., 1998b; Juell and Marsh, 1996; Lin and Kung, 1997; Ranganathan and Arun, 1997), color information (Saber and Tekalp, 1996; Sobottka and Pitas, 1996) to spatial gray-level dependence matrix (SGLD) (Dai and Nakano, 1996) etc. In most cases, multiple techniques are used together to achieve better performance.

We realize that face images with simple background are seen in many practical applications, such as photo-IDs for personal identification and security work. So far several different methods have been proposed toward such specific images. Among these, the method of template matching (Yang and Huang, 1994; Chow and Li, 1993) is most prevalent. In this method, the face contour is extracted using six templates–Two eye templates and one mouth template are used to verify a face and locate its main features; then two cheek templates and one chin template to extract the face contour. The performance is favorable except that it cannot detect faces with shadow, rotation and bad lighting conditions.

In most cases, the overall shape of the face or the whole head is very similar to an ellipse. Several methods have been proposed using this shape information (Saber and Tekalp, 1996; Tang et al., 1998; Sobottka and Pitas, 1996). However, face or head contour cannot easily be modeled using a single ellipse. In this paper, we attempt to extend previous shape-based face detection methods by applying a special template containing directional information of edges. Extensive experiments show that the new system is very efficient when processing images with a simple background regardless of variations on size, head pose (moderate head rotation) and lighting condition.

Details of this system are described in the remainder of this paper. Section 2 covers the basic principles of the system. In Section 3, we give the experimental results and the detailed comparison with other systems. Section 4 concludes the paper with suggestions for possible extension and improvements.

Section snippets

Description of the system

The basic flowchart of our system is given in Fig. 1. For the first three steps, we only give the basic principles for they can be easily found in many image processing textbooks. The last two steps are the kernel of our method and we discuss them in greater details.

Experimental results and comparison

In experiments, a set of 20 images is used as the training set. All thresholds and parameters used in our system are obtained from the training set.

For clarity, we define the following terms:

Correct detection: refers to the detection results which can be used in applications such as recognition, tracking etc.

False detection: for detection results not classified as correct detection.

First our algorithm is tested on MIT face database. Each of 16 people is digitized 27 times in the office, varying

Extensions and conclusions

This paper has demonstrated the effectiveness of a new face detection algorithm in the images with simple or complex background. The algorithm is able to correctly detect all faces in the images with simple backgrounds. Compared with other similar algorithms, the new algorithm appears to be more robust to noise and shape variations.

We intend to extend the algorithm for multi-face detection. First, the number of faces should be known before the detection. This is similar to Govindaraju (1996)

References (20)

There are more references available in the full text version of this article.

Cited by (87)

View all citing articles on Scopus
View full text