25-08-2017, 09:32 PM
AN EDGE-BASED FACE DETECTION ALGORITHM ROBUST AGAINST
ILLUMINATION, FOCUS, AND SCALE VARIATIONS
ABSTRACT
A face detection algorithm very robust against illumination,focus and scale variations in input images has been developedbased on the edge-based image representation. Themultiple-clue face detection algorithm developed in our previouswork [1] has been employed in conjunction with a newdecision criterion called .density rule,. where only high densityclusters of detected face candidates are retained as faces.As a result, the occurrence of false negatives has been greatlyreduced. The robustness of the algorithm against circumstancevariations has been demonstrated.
1. INTRODUCTION
The development of robust image recognition systems isquite essential in a variety of applications such as intelligenthuman-computer interfaces, robotics, security systems, andso forth. Real-time face recognition, in particular, plays animportant role in establishing user-friendly human-computerinterfacing. In realizing real-time automatic face recognitionsystems, the issues are twofold [2]. In the _rst place, locationsof faces in an input image must be detected robustlyindependent of circumstance conditions such as illumination,focus, size of faces, and so on. The most important point atthis stage is to detect all faces without loss due to false negativesdespite some false positives included in the detectionresults. The decision that a partial image belongs to facesor non faces needs be done very fast because such decisionmust be repeated for all locations in the entire image.The detected face candidates are then identi_ed as individualsby matching with stored samples in the system. Thecomputational cost at this stage is quite reduced as comparedto that in the _rst stage since only a small number of candidatesare involved in the recognition process. This allowsus to execute more complex computation for veri_cation andidenti_cation. Therefore, false positives detected in the _rststage can be eliminated by this operation.A number of face detection algorithms such as thoseusing eigenfaces [3] and neural networks [4], for instance,have been developed. In these algorithms, however, a largeamount of numerical computation is required, making theprocessing extremely time-consuming. Therefore, it is notfeasible to build real-time responding systems by softwarerunning on general-purpose computers. In this regard, thedevelopment of hardware-friendly algorithms compatible todedicated VLSI chips is quite essential.For this purpose, we have developed search engine VLSIchips in both CMOS digital [5] and analog [6] technologies.The projected principal-edge distribution (PPED) algorithmhas been developed as an image representation scheme compatibleto the search engine VLSI's and its robust nature hasbeen proven in applications to hand-written pattern recognitionand medical X-ray analysis [7][8][9]. The multiplecluepattern classi_cation algorithm has been developed asan extension of the PPED and successfully applied to facedetection [1]. However, we found a severe degradation in theperformance when the number of face templates is increasedaiming at enhancing its robustness against the variations ininput images.The purpose of this paper is to analyze the cause of thedegradation and to develop a face detection algorithm that isvery robust against illumination, focus, and scale variations.We have introduced a new decision criterion called .densityrule,. where only high density clusters of detected face candidatesare retained as faces. And the enhanced robustness ofthe multiple-clue face detection algorithm has been demonstratedfor a number of sample images.
2. FACE DETECTION ALGORITHM
2.1 Edge-Based Feature Maps
Edge-based feature maps are the very bases of our image representationalgorithm [1][9]. The feature maps represent thedistribution of four-direction edges extracted from a 64_64-pixel input image. The input image is _rst subjected to pixelby-pixel spatial _ltering operations using kernels of 5_5-pixel size to detect edges in four directions, i.e. horizontal,+45 degree, vertical, or -45 degree. The threshold foredge detection is determined taking the local variance of luminancedata into account. Namely, the median of the 40values of neighboring pixel intensity differences in a 5_5-pixel kernel is adopted as the threshold. This is quite importantto retain all essential features in an input image inthe feature maps. Fig. 1 shows an example of feature mapsgenerated from the same person under different illuminationconditions. The edge information is very well extracted fromboth bright and dark images
.2.2 Feature Vectors
64-dimension feature vectors are generated from featuremaps by taking the spatial distribution histograms of edge_ags. In this work, three types of feature vectors, twogeneral-purpose vectors and a face-speci_c vector generatedfrom the same set of feature maps were employed to performmultiple-clue face detection algorithm [1]. Fig. 2 illustratesthe feature vector generation procedure in the projectedprincipal-edge distribution (PPED) [9]. This provides a generalpurpose vector. In the horizontal edge map, for example,edge _ags in every four rows are accumulated and the spatialdistribution of edge _ags along the vertical axis is representedby a histogram.
Download full report
http://www.eurasipProceedings/Eusipco/Eu...cr1828.pdf