Non-faces principled way to choose some important free parameters such as the number of clus- ters it uses. Similarly,Rowley and his colleagues have used problem information in the design ofa retinally connected neural network trained to classify face and nonface patterns.Their ap- proach relies on training several neural net- works emphasizing sets of the training data to obtain different sets of weights.Then,their approach uses different schemes of arbitra- 国 tion between them to reach a final answer. Our SVM approach to the face-detection system uses no prior information to obtain the decision surface.this being an interest- ing property that can be exploited in using the same approach for detecting other ob- 四 jects in digital images. The SVM face-detection system This system detects faces by exhaus- tively scanning an image for face-like pat- Faces terns at many possible scales,by dividing the original image into overlapping subim- ages and classifying them using an SVM to Figure 6.Geometrical interpretation of how the SVM separates the face and nonface classes.The patters are real determine the appropriate class-face or support vectors obtained after training the system.Notice the small number of total support vectors and the fact that a nonface.The system handles multiple higher proportion of them correspond to nonfaces. scales by examining windows taken from scaled versions of the original image. Clearly,the major use of SVMs is in the classification step,which is the most criti- cal part of this work.Figure 6 gives a geo- metrical interpretation of the way SVMs work in the context of face detection. More specifically,this system works as follows.We train on a database of face and nonface 19x19 pixel patterns,assigned to classes +1 and-1,respectively,using the support vector algorithm.This process uses a second-degree homogeneous polynomial kernel function and an upper bound C= 200 to obtain a perfect training error. To compensate for certain sources of 293 image variation,we perform some prepro- cessing of the data: Figure 7.False detections obtained with the first version of the system.These false positives later served asnonface Masking:A binary pixel mask removes examples in the training process. some pixels close to the window-pattern boundary,allowing a reduction in the dimensionality of the input space from Tomaso Poggio5,7 and Henry Rowley2re- their result is the clustering and use of 19x 19 361 to 283.This step reduces flect systems with very high detection rates combined Mahalanobis and Euclidean met- background patterns that introduce un- and low false-positive detection rates. rics to measure the distance from a new necessary noise in the training process. Sung and Poggio use clustering and dis- pattern and the clusters.Other important Illumination gradient correction:The tance metrics to model the distribution of features of their approach are the use of process subtracts a best-fit brightness the face and nonface manifold and a neural nonface clusters and a bootstrapping tech- plane from the unmasked window pixel network to classify a new pattern given the nique to collect important nonface patterns values,allowing reduction of light and measurements.The key to the quality of However,this approach does not provide a heavy shadows. 24 IEEE INTELLIGENT SYSTEMSTomaso Poggio6,7 and Henry Rowley2 re￾flect systems with very high detection rates and low false-positive detection rates. Sung and Poggio use clustering and dis￾tance metrics to model the distribution of the face and nonface manifold and a neural network to classify a new pattern given the measurements. The key to the quality of their result is the clustering and use of combined Mahalanobis and Euclidean met￾rics to measure the distance from a new pattern and the clusters. Other important features of their approach are the use of nonface clusters and a bootstrapping tech￾nique to collect important nonface patterns. However, this approach does not provide a principled way to choose some important free parameters such as the number of clus￾ters it uses. Similarly, Rowley and his colleagues have used problem information in the design of a retinally connected neural network trained to classify face and nonface patterns. Their ap￾proach relies on training several neural net￾works emphasizing sets of the training data to obtain different sets of weights. Then, their approach uses different schemes of arbitra￾tion between them to reach a final answer. Our SVM approach to the face-detection system uses no prior information to obtain the decision surface, this being an interest￾ing property that can be exploited in using the same approach for detecting other ob￾jects in digital images. The SVM face-detection system This system detects faces by exhaus￾tively scanning an image for face-like pat￾terns at many possible scales, by dividing the original image into overlapping subim￾ages and classifying them using an SVM to determine the appropriate class—face or nonface. The system handles multiple scales by examining windows taken from scaled versions of the original image. Clearly, the major use of SVMs is in the classification step, which is the most criti￾cal part of this work. Figure 6 gives a geo￾metrical interpretation of the way SVMs work in the context of face detection. More specifically, this system works as follows. We train on a database of face and nonface 19×19 pixel patterns, assigned to classes +1 and –1, respectively, using the support vector algorithm. This process uses a second-degree homogeneous polynomial kernel function and an upper bound C = 200 to obtain a perfect training error. To compensate for certain sources of image variation, we perform some prepro￾cessing of the data: • Masking: A binary pixel mask removes some pixels close to the window-pattern boundary, allowing a reduction in the dimensionality of the input space from 19 × 19 = 361 to 283. This step reduces background patterns that introduce un￾necessary noise in the training process. • Illumination gradient correction: The process subtracts a best-fit brightness plane from the unmasked window pixel values, allowing reduction of light and heavy shadows. 24 IEEE INTELLIGENT SYSTEMS Figure 6. Geometrical interpretation of how the SVM separates the face and nonface classes. The patterns are real support vectors obtained after training the system. Notice the small number of total support vectors and the fact that a higher proportion of them correspond to nonfaces. Figure 7. False detections obtained with the first version of the system. These false positives later served as nonface examples in the training process. Non-faces Faces