FACE DETECTION IS one of the most popular topics

Transcription

1 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION 1 Fast and Robust Face Detection Using Evolutionary Pruning Jun-Su Jang and Jong-Hwan Kim, Senior Member, IEEE Abstract Face detection task can be considered as a classifier training problem. Finding the parameters of the classifier model by using training data is a complex process. To solve such a complex problem, evolutionary algorithms can be employed in cascade structure of classifiers. This paper proposes evolutionary pruning to reduce the number of weak classifiers in AdaBoost-based cascade detector, while maintaining the detection accuracy. The computation time is proportional to the number of weak classifiers and, therefore, a reduction in the number of weak classifiers results in an increased detection speed. Three kinds of cascade structures are compared by the number of weak classifiers. The efficiency in computation time of the proposed cascade structure is shown experimentally. It is also compared with the state-of-the-art face detectors, and the results show that the proposed method outperforms the previous studies. A multiview face detector is constructed by incorporating the three face detectors: frontal, left profile, and right profile. Index Terms AdaBoost learning, constrained optimization, evolutionary computer vision, face detection, pattern recognition. I. INTRODUCTION FACE DETECTION IS one of the most popular topics of research in the computer vision field. It has many applications including face recognition, crowd surveillance, and human-computer interaction. These applications use face detection as a fundamental first step. Many successful methods have been developed [1], however, there is still a gap between the capabilities of humans and those of state-of-the-art face detectors. To make a robust real-time face detector, Viola and Jones proposed a method based on AdaBoost learning [2]. They obtained both fast computation and high detection rates by introducing rectangle features, integral images, and cascade structures of classifiers. Subsequently, many improved methods were researched. Li and Zhang proposed FloatBoost method for training the classifier [3]. Backtracking scheme was employed for removing unfavorable classifiers from the existing classifiers. Wu et al. carried out multiview face detection using nested structure and real AdaBoost [4]. In view-based face detection approach, face detection task is regarded as a classifier training problem. Classifier training is the process of finding the parameters of the classifier model by using training data. The standard approach is to specify a Manuscript received December 13, 2006; revised April 14, 2007 and July 17, This work was supported in part by the Korea Research Foundation Grant funded by the Korean Government (MOEHRD) (KRF D00145). The authors are with the Department of Electrical Engineering and Computer Science, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Korea ( jsjang@rit.kaist.ac.kr; johkim@rit.kaist.ac.kr). Digital Object Identifier /TEVC model having many parameters and then estimate their values from training data. When the models are simple, it is possible to find the optimal parameters by solving equations explicitly. However, it is very difficult to find the optimal parameters if the models become more complex. Therefore, stochastic approaches can be a good method to find the parameters. Evolutionary algorithms (EAs) are one of the most powerful stochastic search methods. They have robust performance without recourse to domain-specific heuristics. EAs have been applied in many classifier training tasks such as face detection [5] [7], face recognition [8], and car detection [9]. Some of the above studies applied EAs to AdaBoostbased classifier training problems. They tried to improve either making ensemble classifiers or selecting features. AdaBoost is a greedy search algorithm which provides an ensemble of classifiers [10]. It iteratively chooses a local optimal classifier at every round of boosting. The classifier selected in the current round has an effect on choosing the classifier of later rounds by updating the weights of training examples. An ensemble of classifiers is given by a weighted linear combination of many classifiers. AdaBoost can provide us with a greedy approach to generate an ensemble of classifiers, however, global search may make a different ensemble result. Wang and Wang utilized evolutionary approach to update the weights of training examples [11]. Replacing the weight update rule in AdaBoost learning, they tried to find classifiers which had large diversity. Treptow and Zell proposed efficient feature selection method by employing EAs to an extended Haar-like feature set [7]. EAs were used to search over a larger feature set instead of using an exhaustive search over a limited feature set. The exhaustive search in much bigger feature sets may be impossible in current computing power. Abramson et al. employed EAs to enable feature selection process since the exhaustive search was impossible for their huge feature set [9]. In this paper, a fast and robust face detection system is proposed by introducing EAs to optimize ensembles of classifiers. AdaBoost and cascade structure are employed as a basic framework. Our method is focused on the weight of each component classifier rather than the weight of training data [11]. We aim to find the minimal set of classifiers for a given detection rate and false positive rate. To solve this problem, a pruning step is added after AdaBoost learning. EAs are employed as a global search method in the pruning step. They can be expected to have a better chance to find a smaller set of classifiers than AdaBoost method. A similar approach in an ensemble of neural networks was taken by Zhou et al. [12]. In their study, genetic algorithms were employed to make a subset of neural network classifiers instead of using all neural networks. And they showed that it may be better to ensemble many instead of all. Our method X/$ IEEE

2 2 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION starts based on this idea and extends it to find the minimal set of weak classifiers. We propose evolutionary pruning method to minimize the number of classifiers without degrading the detection accuracy. The reduced number of classifiers provides faster computation time. This paper is organized as follows. Section II presents AdaBoost learning method and cascade structure of classifiers. In Section III, evolutionary pruning method is proposed to construct an efficient cascade structure. Section IV shows the experimental results of frontal and multiview face detection, and conclusions follow in Section V. II. ADABOOST LEARNING AND CASCADE STRUCTURE OF CLASSIFIERS A. AdaBoost Learning With Rectangle Features AdaBoost is a general method to obtain an ensemble of weak classifiers whose accuracy may be poor, but slightly better than random guess. Given a set of weak classifiers, a strong classifier is obtained in terms of weighted linear combination of weak classifiers as follows: if otherwise where is an input image, is a weak classifier, is a corresponding weight, and represents a threshold value. Weak classifiers are selected from the huge number of rectangle features. Each weak classifier consists of a rectangle feature, a threshold, and a polarity indicating the direction of the inequality if otherwise To find the best weak classifier in every boosting round,an exhaustive search is employed the same as [2]. Fig. 1 shows six types of rectangle features used in this study. Types (a) (e) are similar to basic Haar-like features proposed by Viola and Jones. These features are computed by subtracting the sum of pixel values in dark rectangle from the sum of pixel values in bright rectangle. Type (f) indicates variance feature. It is calculated by taking the variance of pixel values in the rectangle region. The variance feature can extract second-order statistics in a region, and it was experimentally proved to be an informative feature [13]. In implementation, the feature values can be calculated very quickly using integral image technique [2], [14]. After calculating the feature values, they are normalized to minimize the effect of illumination conditions except type (f). Normalization is simply performed by dividing the variance of the region (dark and bright together). Given a base window size of pixels, the possible positions and scales of six types of features are very large. Our feature pool consists of rectangle features. (1) (2) Fig. 1. Six types of rectangle features. B. Cascade Structure of Classifiers Window scanning techniques are used in the most view-based detection methods. Among the millions of possible subwindows, only very few subwindows are classified as a face. The potential frequency of faces and nonfaces should be considered for real-time performance. The cascade structure of a classifier is a good framework to implement a fast detector. This approach was also used in the other face detection systems. Rowley [15] constructed two-stage-cascade of neural network for fast version of his detector. Similarly, SVM face detector with cascade structure was proposed by Heisele et al. [16]. At each stage, a strong classifier is trained to pass almost all face training data, while discarding a certain portion (typically between 0.2 and 0.5) of nonface training data. The learning goal for th stage is to satisfy detection rate ( ) and false positive rate ( ), i.e., detection rate in each stage should be greater than or equal to, and false positive rate in each stage should be less than or equal to. Feature selection is performed until the strong classifier satisfies the learning goal. After a stage classifier is trained, a new negative training data set is collected for the next stage and AdaBoost algorithm is performed the same way. To enhance the efficiency of cascade, nested structure [4], [17] is considered. The notion of nested structure is that the previous stage classifier can be a good starting point for the current stage training. For, where indicates current stage number, the current stage classifier can be organized by adding the previous stage classifier as follows: if otherwise (3)

3 JANG AND KIM: FAST AND ROBUST FACE DETECTION USING EVOLUTIONARY PRUNING 3 where is the previous stage classifier with new threshold and is a corresponding coefficient found by the AdaBoost algorithm. By changing the threshold value from to, the previous stage classifier is employed with no additional computation cost. Since every feature value is already computed at the previous stage, only the new threshold needs to be compared. III. EVOLUTIONARY PRUNING FOR CASCADE STRUCTURE A. Evolutionary Algorithms (EAs) EAs are principally stochastic search and optimization methods based on the principles of natural biological evolution [18]. Compared with traditional optimization method, EAs have robust performance and global search characteristics. Applying EAs to the classifier optimization problem, the representation of an individual is considered as a real-valued string. Individuals can be directly mapped to parameters to be optimized. The typical mutation and selection methods used in this study are self-adaptation and -tournament, respectively. Selfadaptation is introduced by Schwefel for controlling the stepsize of mutation [19]. A real-valued individual consists of object variables and strategy parameters. For each object variable and parameter,, the mutation operator works by following equations: Fig. 2. Procedure of evolutionary pruning. (4) where and. The notation means that the random variable is sampled anew for each value of the index. -tournament selection is a kind of probabilistic selection method [20]. After creating offspring from parent, the fitness value of each of the individuals are compared with those of individuals which are chosen randomly from the whole population. Then, each individual is ranked according to the number of wins, and the best individuals survive. As the tournament size is increased, the selection pressure becomes high. Since the best individual is always top-ranked with score, it is guaranteed to survive. B. Evolutionary Pruning In this section, the evolutionary pruning method is proposed for pruning the cascade structure. Each cascade stage consists of many weak classifiers for satisfying the learning goal, detection rate, and false positive rate. The evolutionary pruning is applied to reduce the number of weak classifiers, while the learning goal is also satisfied. The procedure of evolutionary pruning is presented in Fig. 2. Let a set of weak classifiers be an empty set in the initialization step. Features are added to the set by applying the AdaBoost process. After constructing a set of weak classifiers, which satisfies the learning goal by using AdaBoost, evolutionary pruning is applied to the set. The idea is that some dependency may exist among the weak classifiers. AdaBoost algorithm produces a strong decision by merging the decisions Fig. 3. Evolutionary rearrangement for weak classifiers. of weak classifiers. If there are some dependencies among many weak classifiers, a set of reduced number of weak classifiers can make a decision as good as the original set do. Evolutionary pruning is a stochastic search method for discarding the redundant weak classifiers, while maintaining the quality of the final strong decision. The first step of evolutionary pruning is called evolutionary rearrangement. It is a step to vary the weight of each decision of weak classifier. Generally, the weight decreases as round increases in AdaBoost process. Fig. 3 shows a typical distribution of in the stage classifier. The cross marks present the weights of corresponding features. The feature index zero means the previous stage classifier which forms a nested structure. There are 36 weak classifiers indexed from 1 to 36. Therefore, a total of 37 weights are involved in the current stage decision. The weight of previous stage classifier is relatively larger than that of the other weak classifiers. It is obvious that the

4 4 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION Fig. 4. Structure of the chromosome for evolutionary pruning. previous stage classifier can make a better decision than any single weak classifier, because it consists of a few weak classifiers selected at the previous stage. Except for the first two or three weights, the other weights have similar value. It implies that every single feature has similar influence to final strong decision. If one feature is discarded while the others keep their weights, then the learning goal is not satisfied. EAs are applied to step 3(a) and 3(d) in Fig. 2. In each step, EAs are expected to solve a constrained optimization problem. Let an estimation of the weight be. Then, the target parameter,, forms a chromosome. Fig. 4 shows the structure of the chromosome for evolutionary pruning. The number of weak classifiers,, decides the length of chromosome at each stage of training. Since the first stage does not have a previous stage classifier, is not activated in the first stage. The boundary of each parameter is given by using maximum value of as follows: where is the maximum value among the weights produced by AdaBoost algorithm. The fitness function is employed by scores and penalty function as follows: Two kinds of scores are evaluated to define appropriate fitness value. and indicate scores for face training data and nonface training data, respectively. Basically, it is added by for every correct classification. The stage learning goal should be satisfied. These constraints can be satisfied by employing penalty functions as follows: if otherwise if (8) otherwise where and indicate detection rate and false positive rate for th stage, which should be satisfied by stage classifier. and mean the detection rate and the false positive rate computed by applying a set of to training data. and are proper positive constants. The penalty functions are applied if an individual does not satisfy the learning goal. Otherwise, the penalty values are set to zero for feasible solutions. By applying the linear penalty function, the EAs can search feasible solutions that maximize the fitness function. In the experiments, population size was set to 100. The mutation method was self-adaptation presented by (4). Initial individuals were generated by random number satisfying the boundary condition in (5). A solution from AdaBoost was already known, but it did not join in population. It was experimentally verified that the solution from AdaBoost was not helpful to find diverse (5) (6) (7) with large variance. It is explained that a well-fitted solution in the early generation restricts the exploration of evolution. The -tournament selection method was employed for the next generation. After five random tournaments for each individual, 50 parents were selected for reproducing. Fifty parents and 50 new offsprings made up the next generation. As shown in Fig. 3, the EAs find a set of weights which satisfy the learning goal, but a solution set is different from the result of AdaBoost learning. The set of weights has more diverse value than that from AdaBoost. Thus, it is easier to select one weak classifier, which is likely to be pruned. By discarding one weak classifier which has the minimum weight, the number of weak classifiers reduces by one. Evolutionary search is followed to verify that there exists a feasible solution under a new set of weak classifiers. If the feasible solution is found, then the set is stored as a possible solution set, and one weak classifier, which has the minimum weight, is discarded again. This procedure is continued until evolutionary search fails to find the feasible solution. Finally, the strong classifier is constructed by weighted linear combination of surviving weak classifiers. As evolutionary search processes by reducing the number of weak classifiers, the length of string, shown in Fig. 4, also decreases. Evolutionary pruning can be seen as an iterative approach of Zhou s work [12]. Our method prunes weak classifiers iteratively, while Zhou s method selects weak classifiers by comparing with a preset threshold. It should be noted that there is a possibility to improve detection accuracy although evolutionary pruning fails to find a reduced set of weak classifiers. Because the evolutionary search is basically employed to maximize the fitness function which drives the detection rate higher and false positive rate lower. One feasible solution from AdaBoost is already obtained and, therefore, the evolutionary rearrangement starts with the certainty of existence of feasible solutions. The fitness function is not designed to reduce the number of weak classifiers directly, but it works for satisfying the learning goal whenever a weak classifier is discarded. IV. EXPERIMENTS A. Frontal Face Detection The face training set consisted of 6000 hand labeled faces aligned to a base window size of pixels with 256 gray levels per pixel. For detecting the frontal faces, face training data were collected in the range of in-plane rotation and out-of-plane rotation from the exact upright face. The face training set consisted of various face images including the training data used by Viola and Jones. The examples of the face training data are shown in Fig. 5(a). The nonface training set consisted of 6000 images for the training of the first stage. The examples of the nonface training data are shown in Fig. 5(b). After training the stage classifier, nonface training set is updated. A portion of nonface examples, which is classified to nonface by current stage classifier, is removed from the training set. False positive examples produced by the stage classifier substitute for removed examples. False positive examples can be obtained by applying the stage classifier to images which do not contain faces.

5 JANG AND KIM: FAST AND ROBUST FACE DETECTION USING EVOLUTIONARY PRUNING 5 TABLE I COMPARISON OF THE THREE CASCADE STRUCTURES. STANDARD CASCADE WITH (SC_AB), NESTED CASCADE WITH ADABOOST (NC_AB), AND NESTED CASCADE WITH EVOLUTIONARY PRUNING (NC_EP) ARE COMPARED IN TERMS OF THE NUMBER OF WEAK CLASSIFIERS (WC), DETECTION RATE p, AND FALSE POSITIVE RATE q FOR EACH STAGE Fig. 5. Examples of face training data. (a) Examples of face training data. (b) Examples of nonface training data. The learning goal for each stage was set to satisfy a detection rate, and a false positive rate. It means that the training process is performed until satisfying greater than or equal to, and less than or equal to on given training data. Exceptionally, a false positive rate of 0.1 was assigned for the final stage (15th stage). Three kinds of cascade structures were constructed for comparative experiments as follows: Standard cascade with AdaBoost (SC_AB). Cascade structure was constructed without nest. It was the same way of Viola and Jones approach [2]. AdaBoost was employed to select weak classifiers for each stage. Nested cascade with AdaBoost (NC_AB). Cascade structure was constructed with nest. AdaBoost was employed to select weak classifiers for each stage. Nested cascade with evolutionary pruning (NC_EP). Cascade structure was constructed with nest. Evolutionary pruning was employed to select weak classifiers for each stage. The nested cascade structures were constructed by reusing the previous stage classifier. The training results for the three cascade structures are summarized in Table I. At the first stage in Table I, six weak classifiers are boosted for both SC_AB and NC_AB. They are exactly the same for both cascade structures, since the training data are identical and the nested structure is not available at the first stage. Starting from the set of six weak classifiers, evolutionary pruning method finds a set of five weak classifiers, while satisfying the learning goal. Even the false positive rate is less than others. From the second stage to the final stage, the training data between SC_AB and NC_EP are different because the first stage classifiers are not identical, and then they produce different

6 6 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION TABLE II DETECTION RATES FOR VARIOUS NUMBERS OF FALSE POSITIVES ON CMU+MIT DATABASE training data for the next stage. To demonstrate the performance of NC_EP on the same training data, the experimental results of NC_AB were derived from the same training data in NC_EP. It means that the training data for NC_AB and NC_EP are exactly the same through the whole stage. In total, 1706 weak classifiers were employed for SC_AB, 1228 for NC_AB, and 1002 for NC_EP. The number of weak classifiers of NC_EP was reduced to 58.7% of SC_AB, and 81.6% of NC_AB. The number of weak classifiers is related to computation time. Especially, the early stages are closely related since the scan windows are applied to former stage classifier more than latter. For example, the first-stage classifier is performed for every scan window, while the second-stage classifier may treat about 40% of scan windows. The ratio of 40% is referred from the false positive rate in training step. It varies on input images because the nature of input images may be somewhat different from the training data. The computation time of NC_EP was approximately 15 ms for a pixels input image on Pentium-IV 2.4 GHz with MS Windows XP. Visual C++ was used for implementation. In the same environment, the computation time of SC_AB was about 26 ms. The ratio of the computation time is similar to the ratio of the number of weak classifiers. To show the effectiveness of pruning, NC_EP was compared with the forward feature selection (FFS) method [21] proposed by Wu et al. In their recent study [22], the linear asymmetric classifier (LAC) was applied after feature selection. We compared both methods in the same constraints, learning goal, and the number of weak classifiers in each stage. The same training data were used for single stage training. Comparison results for fifth and tenth stages are presented in Fig. 6. As shown in the ROC curves, NC_EP has higher detection rate on the same false positive rate. The similar results were found in all the other stages. In Fig. 6(a), NC_EP can achieve the learning goal (, ), while FFS+LAC cannot satisfy the learning goal given the number of weak classifiers. It means that FFS+LAC need to add additional weak classifiers to achieve the learning goal. It is obvious that the larger number of weak classifiers requires more runtime computation. However, FFS+LAC has faster training time than NC_EP. Because FFS+LAC trains weak classifiers only once per stage, while NC_EP trains at every round of boosting the same as the Viola-Jones approach. The proposed detector NC_EP was tested on the CMU+MIT database [23]. It consists of 130 images with 507 frontal faces Fig. 6. Comparisons of the ROC curves between NC_EP and FFS+LAC. (a) ROC curves for fifth stage. (b) ROC curves for tenth stage. with various conditions such as facial expression, occlusion, pose, and scale variations, etc. Most images contain more than one face on various backgrounds. Face detection was performed by the window scanning technique, applying the classifier to every subwindow in an input image. This was an exhaustive search for possible face locations at all scales. The input image was scaled down by a factor

7 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. JANG AND KIM: FAST AND ROBUST FACE DETECTION USING EVOLUTIONARY PRUNING Fig. 7. Some face detection results on CMU+MIT database. Fig. 8. Examples of profile face training data. of 1.25 until it became a smaller size than the base window size of pixels. The corresponding integral image was prepared for each scale of image. The detector was applied to Fig. 9. Profile detector result on CMU profile database. 7

8 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 8 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION Fig. 10. Some face detection results on CMU profile database. each scale of image using a shift step size of 1 pixel. Table II shows the evaluation results of the three cascade detectors and also gives comparison with previous approaches. By adjusting the threshold value of the final stage classifier, various number of false positives and corresponding detection rate can be obtained. As shown in the evaluation results, SC_AB, NC_AB, and NC_EP have similar detection rates. Since the learning goal and initial training data sets for three detectors are the same, their final detection rates are similar. Even though SC_AB is trained by the Viola-Jones approach, selection of training data can vary the detection rate. The Rowley-Baluja-Kanade detector [15] is based on neural network and the evaluation result is given by Viola and Jones. The Schneiderman-Kanade detector [24] is based on naive Bayes classifier which estimated the joint probability of local appearance. Their detection rate is a little high, however, the computation time takes about 600 times more than the Viola-Jones detector. The proposed detector has a similar or slightly better detection rate than the Viola-Jones detector. Their detector was constructed by a total of 38 stages with 6061 weak classifiers. It has about six times more weak classifiers than NC_EP. This is caused by three reasons. The first is the extended feature pool. Variance features are added to basic Haar-like features and they reduce the total number of weak classifiers. The second is the

9 JANG AND KIM: FAST AND ROBUST FACE DETECTION USING EVOLUTIONARY PRUNING 9 nested structure. It reuses the previous stage classifier so that a good weak classifier is employed without additional computation cost. Generally, the previous stage classifier has relatively high weight, therefore, a smaller number of weak classifiers is needed to construct the stage classifier. It is verified in Table I by comparing NC_AB with SC_AB. The third aspect is the evolutionary pruning. Evolutionary pruning can reduce the number of classifiers by considering the dependency among the weak classifiers. It helps that a set of has large variance, so that the minimum has little effect on the stage classifier performance. It should be noted that our method is not proposed for improving the detection rate. The detection rate mainly depends on the relation between training data and test data. Therefore, selection of training data can have an effect on the detection rate. There are many factors concerning the detection rate, such as face alignment in training data, illumination, number of subdetectors considering face pose angles, etc. We provide Table II to show that the proposed method has comparable detection accuracy, while reducing the computation time. Fig. 7 shows some results on CMU+MIT database. The overlapping detection results merge into single detection by postprocessing. The label in the upper left corner of each image ( ) indicates the number of faces detected ( ), the total number of faces in the image ( ), and the number of false positives ( ). The label in the lower right corner of each image presents its resolution. B. Multiview Face Detection In the real environment, faces appear with various ranges of in-plane rotation and out-of-plane rotation. It is very difficult to detect faces that have a large range of rotations by one detector. In-plane rotated faces can be detected by rotating input images because the face pattern is just rotated by certain angle. However, out-of-plane rotations can make large differences, which cannot be overcome by rotating input images. Therefore, profile faces are treated as a different class from frontal faces. The multiview face detection system was implemented by employing frontal and profile face detectors. The profile face detector was constructed the same way as the frontal face training. The face training data were profile faces instead of frontal faces. The profile face training data consisted of faces which have out-of-plane rotation between A total of 6000 profile faces were cropped from various images in world-wide-web. The examples of the profile face training data are shown in Fig. 8. The initial set of nonface training data was the same as the set used in the training of frontal face detector. The trained profile detector was tested on CMU profile database. It consists of images containing the various frontal and profile faces. The experimental result was compared with the Viola-Jones profile detector [25]. For verifying the performance of the profile detector, frontal faces in the database were not counted the same as Viola-Jones. Fig. 9 shows a comparative result including detection rates for various false positives. The computation time of the NC_EP profile detector was approximately 80 ms for a pixels input image on Pentium-IV 2.4 GHz. Compared with the Viola-Jones detector that reported 120 ms, the proposed method is faster, while maintaining higher detection rate. A multiview face detection system was constructed by incorporating the three face detectors, frontal, left profile, and right profile. The face locations from each detector were merged into the final result. Some results of the multiview face detector on the CMU profile database are shown in Fig. 10. V. CONCLUSION The main contribution of this paper was to propose a method to construct effective cascade structure of classifiers. Evolutionary pruning was proposed to reduce the number of weak classifiers in each stage of cascade. The computation time is proportional to the number of weak classifiers, therefore the reduction process directly affects detection speed. It was shown that evolutionary search could find a smaller set of weak classifiers than AdaBoost. The number of weak classifiers was successfully reduced by pruning the weak classifier which has minimum weight among the ensemble of classifiers. The total number of weak classifiers in the proposed structure was reduced to 58.7% of that constructed from the AdaBoost method. The proposed method detected between 90.1% and 94.7% of the faces on the CMU+MIT database under acceptable number of false positives. It was also compared with state-of-the-art face detectors. The experimental results showed that the detection accuracy of the proposed detector was similar or better than the Viola-Jones detector, while the proposed detector had less computational cost. The computation time of the proposed detector was about 15 ms for a pixels input image on Pentium-IV 2.4 GHz with MS Windows XP. A multiview face detector was constructed by incorporating the three face detectors: frontal, left profile, and right profile. The profile face detector was trained the same way in frontal face detector, just changing the face training data. REFERENCES [1] M.-H. Yang, D. Kriegman, and N. Ahuja, Detecting faces in images: A survey, IEEE Trans. Pattern Anal. Mach. Intell., vol. 24, no. 1, pp , [2] P. Viola and M. Jones, Robust real-time face detection, Int. J. Computer Vision, vol. 57, no. 2, pp , [3] S. Z. Li and Z. Q. Zhang, Floatboost learning and statistical face detection, IEEE Trans. Pattern Anal. Mach. Intell., vol. 26, no. 9, pp , Sep [4] B. Wu, H. Ai, C. Huang, and S. Lao, Fast rotation invariant multi-view face detection based on real Adaboost, in Proc. 6th Int. Conf. Autom. Face Gesture Recogn., 2004, pp [5] J.-S. Jang, K.-H. Han, and J.-H. Kim, Evolutionary algorithm-based face verification, Pattern Recogn. Lett., vol. 25, pp , [6] J.-S. Jang and J.-H. Kim, Evolutionary pruning for fast and robust face detection, in Proc. IEEE Congr. Evol. Comput., 2006, pp [7] A. Treptow and A. Zell, Combining Adaboost learning and evolutionary search to select features for real-time object detection, in Proc. IEEE Congr. Evol. Comput., 2004, pp [8] C. Liu and H. Wechsler, Evolutionary pursuit and its application to face recognition, IEEE Trans. Pattern Anal. Mach. Intell., vol. 22, no. 6, pp , [9] Y. Abramson, F. Moutarde, B. Stanciulescu, and B. Steux, Combining adaboost with a hill-climbing evolutionary feature search for efficient training of performant visual object detectors, in Proc. 7th Int. FLINS Conf. Appl. Artif. Intell., 2006, pp [10] Y. Freund and R. E. Schapire, A decision-theoretic generalization of on-line learning and an application to boosting, J. Comput. Syst. Sci., vol. 55, no. 1, pp , 1997.

10 10 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION [11] X. Wang and H. Wang, Classification by evolutionary ensembles, Pattern Recogn., vol. 39, pp , [12] Z.-H. Zhou, J. Wu, and W. Tang, Ensembling neural networks: Many could be better than all, Artif. Intell., vol. 137, no. 1-2, pp , [13] R. Xiao, M.-J. Li, and H.-J. Zhang, Robust multipose face detection in images, IEEE Trans. Circuits Syst. Video Technol., vol. 14, no. 1, pp , Jan [14] F. Crow, Summed-area tables for texture mapping, in Proc. SIG- GRAPH, 1984, vol. 18, pp [15] H. A. Rowley, S. Baluja, and T. Kanade, Neural network-based face detection, IEEE Trans. Pattern Anal. Mach. Intell., vol. 20, no. 1, pp , [16] B. Heisele, T. Serre, S. Mukherjee, and T. Poggio, Feature reduction and hierarchy of classifiers for fast object detection in video images, in Proc. IEEE Conf. Comput. Vision Pattern Recogn., 2001, pp [17] J. Šochman and J. Matas, Inter-stage feature propagation in cascade building with AdaBoost, in Proc. Int. Conf. Pattern Recogn., 2004, pp [18] T. Bäck, U. Hammel, and H.-P. Schwefel, Evolutionary computation: Comments on the history and current state, IEEE Trans. Evol. Comput., vol. 1, no. 1, pp. 3 17, Apr [19] H.-P. Schwefel, Evolution and Optimum Seeking. New York: Wiley, [20] T. Bäck, Evolutionary Algorithms in Theory and Practice: Evolution Strategies, Evolutionary Programming, Genetic Algorithms. New York: Oxford Univ. Press, [21] J. Wu, J. M. Rehg, and M. D. Mullin, Learning a rare event detection cascade by direct feature selection, Advances in Neural Information Processing Systems, vol. 16, [22] J. Wu, M. D. Mullin, and J. M. Rehg, Linear asymmetric classifier for face detectors, in Proc. Int. Conf. Mach. Learn., 2005, pp [23] H. A. Rowley, Neural network-based face detection, Ph.D., Carnegie Mellon Univ., Pittsburgh, PA, [24] H. Schneiderman and T. Kanade, A statistical method for 3D object detection applied to faces and cars, in Proc. IEEE Int. Conf. Comput. Vision, Pattern Recogn., 2000, vol. 1, pp [25] P. Viola and M. Jones, Fast multi-view face detection Mitsubishi Electric Research Laboratories, Tech. Rep., TR , Jun-Su Jang received the B.S., M.S., and Ph.D. degrees in electrical engineering and computer science from the Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Korea, in 1999, 2001, and 2006, respectively. He is currently a Postdoctoral Fellow at the Robotics Institute, Carnegie Mellon University, Pittsburgh, PA. His research interests are in the areas of statistical pattern classification, evolutionary algorithms, robot localization, and ubiquitous robotics. Dr. Jang was the recipient of the bronze prizes at the Samsung HumanTech Thesis Prize Awards in Jong-Hwan Kim (S 85 M 88 SM 03) received the B.S., M.S., and Ph.D. degrees in electronics engineering from Seoul National University, Seoul, Korea, in 1981, 1983 and 1987, respectively. Since 1988, he has been with the Department of Electrical Engineering and Computer Science, Korea Advanced Institute of Science and Technology (KAIST), where he is currently a Professor. His current research interests are in the areas of ubiquitous and genetic robotics. Dr. Kim currently serves as an Associate Editor of the IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION and of the IEEE Computational Intelligence Magazine. He was one of the co-founders of the International Conference on Simulated Evolution and Learning (SEAL). He was General Chair for the IEEE Congress on Evolutionary Computation in Seoul, Korea, His name was included in the Barons 500: Leaders for the New Century in 2000 as the Father of Robot Football. He is the Founder of The Federation of International Robosoccer Association (FIRA) and The International Robot Olympiad Committee (IROC). He is currently serving FIRA and IROC as President.