Real-time face location on gray-scale static images

Abstract

This work presents a new approach to automatic face location on gray-scale static images with complex backgrounds. In a first stage our technique approximately detects the image positions where the probability of finding a face is high; during the second stage the location accuracy of the candidate faces is improved and their existence is verified. The experimentation shows that the algorithm performs very well both in terms of detection rate (just one missed detection on 70 images) and of efficiency (about 13 images/s can be processed on Hardware Intel Pentium II 266 MHz).

Introduction

Automatic face location is a very important task which constitutes the first step of a large area of applications: face recognition, face retrieval by similarity, face tracking, surveillance, etc. (e.g. Ref. [1]). In the opinion of many researchers, face location is the most critical step towards the development of practical face-based biometric systems, since its accuracy and efficiency have a direct impact on the system usability. Several factors contribute to making this task very complex, especially in the case of applications requiring to operate in real-time on gray-scale static images. Complex backgrounds, illumination changes, pose and expression changes, head rotation in the 3D space and different distances between the subject and the camera are the main sources of difficulty.

Many face-location approaches have been proposed in the literature, depending on the type of images (gray-scale images, color images or image sequences) and on the constraints considered (simple or complex background, scale and rotation changes, different illuminations, etc.). Giving a brief summary of the conspicuous number of works requires a pre-classification; unfortunately, due to the large amount of different techniques used by researchers this task is not so easy. While we are aware of the unavoidable inaccuracies, we have tried to make a tentative classification:

• Methods based on template matching with static masks and heuristic algorithms which use images taken at different resolutions (multiresolution approaches) [2], [3].

• Computational approaches based on deformable templates which characterize the human face [4] or internal features [5], [6], [7], [8]: eyes, nose, mouth. These methods can be conceived as an evolution of the previous class, since the templates can be adapted to the different shapes characterizing the searched objects. The templates are generally defined in terms of geometric primitives like lines, polygons, circles and arcs; a fitness criterion is employed to determine the degree of matching.

• Face and facial parts detection using dynamic contours or snakes [6], [9], [10], [11]. These techniques involve a constrained global optimization, which usually gives very accurate results but at the same time is computationally expensive.

• Methods based on elliptical approximation and on face searching via least-squares minimization [12], incremental ellipse fitting [13] and elliptic region growing [14].

• Approaches based on the Hough transform [7] and the adaptive Hough transform [15].

• Methods based on the search for a significant group of features [triplets, constellations, etc.] in the context considered: for example, two eyes and a mouth suitably located constitute a significant group in the context of a face [7], [16], [17], [18], [19].

• Face search on the eigenspace determined via PCA [20] and face location approaches based on the information theory [21], [22].

• Neural Network approaches [23], [24], [25], [26], [27], [28], [29], [30]. The best results have been obtained by using feed forward networks to classify image portions normalized with respect to scale and illumination. During the training, examples of face objects and non-face objects are presented to the network. The high computational cost, induced by the need to process at different resolutions all the possible positions of a face in the image, is the main drawback of these methods.

• Face location on color images through segmentation in a color-space: YIQ, YES, HSI, HSV, Farnsworth, etc [27], [31], [32], [33], [34], [35], [36]. Generally, color information greatly simplifies the localization task: a simple spectrographic analysis shows that the face skin pixels are usually clustered in a color space, and then an ad hoc segmentation allows the face to be isolated from the background or, at least, to drastically reduce the amount of information which must be processed during the successive stages.

• Face detection on image sequences using motion information: optical flow, spatio-temporal gradient, etc. [27], [33], [37].

Since in several applications it is mandatory (or preferable) to deal with static gray-scale images we believe it is important to develop a method which does not exploit additional information like color and motion. For example, most of the surveillance cameras nowadays installed in shops, banks and airports are still gray-scale cameras (due to the lower cost), and the electronic processing of mug-shot or identity card databases could require to detect faces from static gray-scale pictures printed on paper. Unfortunately, if we discard color and motion-based approaches, the most robust methods are generally time consuming and cannot be used in real-time applications.

The aim of this work is to provide a new method which is capable of processing gray-scale static images in real time. The algorithm must operate with structured backgrounds and must tolerate illumination changes, scale variations and small head rotations.

Our approach (Fig. 1(a)) is based on a location technique which starts by approximately detecting the image positions (or candidate positions) where the probability to find a face is high (module AL) and then for each of them improves the location accuracy and verifies the presence of a true face (module FLFV). Actually, most of the applications in the field of biometric systems require detection of just one object in the image (i.e. the foreground object): under this hypothesis, a more efficient implementation of our method is reported in Fig. 1(b), where at each step the module AL passes only the most likely position to FLFV and FLFV continues to require a new position until a valid face is detected or no more candidates are available. It should be noted that, even in this case, the system could be used to detect more faces in an image, assuming that the iterative process is not prematurely interrupted.

Although AL and FLFV have been implemented in a very different manner, both these modules work on the same kind of data: that is the directional image extracted by the starting gray-scale image.

In Section 2 the directional image is defined and some comments about its computation are reported. Section 3 describes the module AL which is based on the search of elliptical blobs in the directional image by means of the generalized Hough transform. In Section 4 we present the dynamic-mask-based technique used for fine location and face verification (module FLFV) and in Section 5 we discuss how to practically combine AL and FLVF in order to implement the functional schema of Fig. 1(b). Section 6 reports the results of our experimentation over a 70 image database; finally, in Section 7, we present our conclusions and discuss future research.