A neural network-based system for face detection in low quality web camera images

The rapid and successful detection and localization of human faces in images is a prerequisite to a fully automated face image analysis system. In this paper, we present a neural network-based face detection system which arises from the outcome of a comparative study of two neural network models of different architecture and complexity. The fundamental difference in the construction of the two models lies in approaching the face detection problem either by seeking a general solution based on the full-face image or by composing the solution through the resolution of specific portions/characteristics of the face. The proposed system is based on the brightness contrasts between specific regions of the human face. We show that the second approach, even though more complicated, exhibits better performance in terms of detection and false-positive rates. We tested our system with low quality face images acquired with web cameras. The image test set includes both front and side view images of faces forming either a neutral or one of the "smile", "surprise", "disgust", "scream", "bored-sleepy", "angry", and "sad" expressions. The system achieved high face detection rates, regardless of facial expression or face view.