Face detection is the task of determining the locations and sizes of human faces in arbitrary digital images, while ignoring any other objects to the greatest possible extent. A fundamental problem in computer vision, it has important applications in fields ranging from surveillance-based security to autonomous vehicle navigation. Although face detection has been studied for almost a decade, the results are not satisfactory for a variety of practical applications, and the topic continues to receive attention.

A commonly used approach for detecting faces is based on the techniques of "boosting" and "cascading", which allow for real-time face detection. However, systems based on boosted cascades have been shown to suffer from low detection rates in the later stages of the cascade. Yet, such face detectors are preferable to other methods due to their extreme computational efficiency.

In this thesis we introduce a novel variation of the boosting process that uses features extracted by Independent Component Analysis (ICA), which is a statistical technique that reveals the hidden factors that underlie sets of random variables or signals. The information describing a face may be contained in both linear as well as high-order dependencies among the image pixels. These high-order dependencies can be captured effectively by representation in ICA space. Moreover, it has been argued that the metric induced by lCA is superior to other methods in the sense that it may provide a representation that is more robust to the effect of noise such as variations in lightening. We propose that features extracted from such a representation may be boosted better in the later stages of the cascade, thus leading to improved detection rates while maintaining comparable speed. We present the results of our face detector, as well as comparisons with existing systems.


Computer Sciences