Abstract
Auto-focus (AF) has been a key feature inconsumer level digital and cell-phone cameras allowing usersto focus automatically on a particular plane in a scene inorder to get a sharp image. Face priority AF has become ofinterest lately due to the fact that most pictures captured byconsumers are of human faces. In face-priority AF, thefocusing decision is made based on a detected face area in theimage, thus capturing a sharp picture of the face. While manyface detection algorithms exist in the literature, very few ofthem are actually suitable for real-time software deploymenton resource limited digital or cell-phone camera processors.In this paper, a fast face-detection algorithm is introduced bycombining a skin color model cluster with a computationallyefficient shape processing scheme. Comparison results with astandard algorithm in terms of robustness, speed andaccuracy are provided. This face detection algorithm isincorporated into our previously developed rule-based AFmethod to achieve real-time face-priority AF on an actualdigital camera platform1
.Index Terms — Face-priority auto focus, real-time facedetection, digital cameras, cell-phone cameras.
I. INTRODUCTION
Nowadays, most digital and cell-phone cameras have anauto focus (AF) feature allowing users to capture focusedimages without any user intervention. Auto focusing involvesthe adjustment of the distance between the lens and the imagesensor in order to obtain a focused image. Various passive AFtechniques, e.g. [1], [2] have been developed to replace thetedious process of manual focusing by merely using thecaptured image information and not using any additionaldistance sensor. In passive AF, a measure of image sharpnessor a sharpness function is extracted from captured images atdifferent lens positions and then the in-focus position isobtained by locating the peak of the sharpness function.Noting that the great majority of pictures taken byconsumers are of friends’ and family’s faces, cameramanufacturers have started introducing an AF feature whichuses an area around a face for extracting the sharpnessfunction. Due to the computational complexity of existing facedetection algorithms, a hardware solution or a dedicated co processor is normally used for this purpose. In this paper, wehave come up with a real-time or computationally efficientsoftware solution which can be deployed on resource limitedembedded processors of digital or cell-phone cameras withoutusing any additional hardware.Face detection has been extensively examined in the imageprocessing literature, e.g. [3]-[10]. In [3], Yang et al.presented a survey on various face detection techniques.While most of these algorithms have been reported to haveachieved high detection rates, very few of them are suitablefor real-time deployment on a digital or cell-phone cameraprocessor due to their high computational and memoryresource demands. Multi-resolution face detection methodshave been shown to provide fast detection, but they do so atthe expense of relatively low detection rates [4]. In [5],Kotropoulos et al. developed a fast approach to face detectionutilizing horizontal and vertical image profiles throughlocating facial candidates followed by post processing rules tovalidate the candidate regions. Although the method is fast, itsuffers from low detection rates when the background iscomplex or there are multiple faces in the scene. Facedetection based on the relationship amongst key facialfeatures, such as eyes, nostrils and mouth, has also beenintroduced [6]. Such features seem to be effective only whenthe entire frontal face area is captured at high resolution.Moreover, using eyes, nose or mouth features may causefailure when cameras are rotated during photography,requiring additional post-processing and thus computationtime to obtain high detection rates.Viola and Jones [7] have introduced a rapid facedetection scheme based on a boosted cascade of simplefeatures. Recently, this method has drawn much attentiondue to its speed and accuracy. Though this method is muchfaster than previous feature based face detection methods,when used for auto focusing, it increases the focusing timebeyond an acceptable focusing time without using adedicated hardware or co-processor. Another issue with thismethod is its robustness. For frontal faces, its detection rateis quite high but it drops significantly for profile, rotated,tilted, or partially covered faces. Although modificationshave been proposed to increase its robustness, e.g. [8],these modifications increase the computational time further.Our intension in this work has thus been to develop arobust real-time software solution for face-priority AFwithout utilizing any dedicated face detection hardware andby retaining more or less the same AF time that we wereable to achieve in our previously introduced fast AFapproach

Download full report
http://ieeexplore.ieeeiel5/30/4711185/04...er=4711194