This paper presents an example-based learning approach for detecting vertical frontal views of human faces in complex scenes. The method models the distribution of human face patterns using a few view-based "face" and "nonface" model clusters. At each image location, a difference feature vector is computed between the local image pattern and the distribution-based model. A trained classifier determines whether or not a human face exists at the current image location. The distance metric used for computing difference feature vectors and the inclusion of "nonface" clusters are critical for the system's success. The paper discusses the challenges of face detection, including variations in facial appearance, expression, skin color, and lighting conditions. It also reviews existing face detection approaches, such as correlation templates, deformable templates, and image invariance schemes. The authors propose a distribution-based face model using Gaussian clusters to represent canonical face patterns and nonface patterns. This model is trained on a large database of face and nonface examples, allowing the system to generalize and detect faces in various conditions. The system uses a two-value distance metric to measure the difference between a test pattern and model clusters. This metric combines a normalized Mahalanobis distance (D1) and a Euclidean distance (D2) to capture both the major directions of data distribution and the smaller eigenvector directions. The system is trained using a multilayer perceptron (MLP) classifier, which combines the 12 distance measurements into a single similarity measure for classification. The paper evaluates the system's performance on two test databases, showing high detection rates with low false positives. It also analyzes the impact of different classifier architectures, distance metrics, and "nonface" model clusters on the system's performance. The results indicate that the two-value distance metric and the inclusion of "nonface" clusters significantly improve the system's ability to distinguish between face and nonface patterns. The system is robust to variations in lighting, expression, and pose, and can be trained to detect faces across a wide range of scales and orientations.This paper presents an example-based learning approach for detecting vertical frontal views of human faces in complex scenes. The method models the distribution of human face patterns using a few view-based "face" and "nonface" model clusters. At each image location, a difference feature vector is computed between the local image pattern and the distribution-based model. A trained classifier determines whether or not a human face exists at the current image location. The distance metric used for computing difference feature vectors and the inclusion of "nonface" clusters are critical for the system's success. The paper discusses the challenges of face detection, including variations in facial appearance, expression, skin color, and lighting conditions. It also reviews existing face detection approaches, such as correlation templates, deformable templates, and image invariance schemes. The authors propose a distribution-based face model using Gaussian clusters to represent canonical face patterns and nonface patterns. This model is trained on a large database of face and nonface examples, allowing the system to generalize and detect faces in various conditions. The system uses a two-value distance metric to measure the difference between a test pattern and model clusters. This metric combines a normalized Mahalanobis distance (D1) and a Euclidean distance (D2) to capture both the major directions of data distribution and the smaller eigenvector directions. The system is trained using a multilayer perceptron (MLP) classifier, which combines the 12 distance measurements into a single similarity measure for classification. The paper evaluates the system's performance on two test databases, showing high detection rates with low false positives. It also analyzes the impact of different classifier architectures, distance metrics, and "nonface" model clusters on the system's performance. The results indicate that the two-value distance metric and the inclusion of "nonface" clusters significantly improve the system's ability to distinguish between face and nonface patterns. The system is robust to variations in lighting, expression, and pose, and can be trained to detect faces across a wide range of scales and orientations.