טכניון מכון טכנולוגי לישראל
הטכניון מכון טכנולוגי לישראל - בית הספר ללימודי מוסמכים  
M.Sc Thesis
M.Sc StudentBraunstain Eyal
SubjectColor Space Methods as an Aid in Face Detection
DepartmentDepartment of Biomedical Engineering
Supervisors Professor Emeritus Gath Isak
Professor Gur Moshe
Full Thesis textFull thesis text - English Version


Abstract

Most contemporary object and face detection methods rely solely on intensity images.

Color space information is very often ignored, as it usually exhibits large variations due to e.g. illumination changes and shadows, and due to the lower spatial resolution in color channels than in the intensity image. In our work, we addressed the problem of utilizing color information to detect faces in images. Color is represented in CIE-Lab space and the first stage towards face detection is omprised of skin detection. Skin detection is carried out by applying a Gaussian model constructed from skin chroma histogram. This is followed by discarding non-face skin regions, using a fuzzy clustering algorithm. To further utilize color information, we design a new color descriptor, based on a variant of Local Binary Patterns (LBP), which is composed of histograms of encoded chroma texture similarities of local patches. The designed operator couples information of CIE-Lab chroma channels, based on inter-channel correlation. The descriptor is designed to achieve invariance to monotonic changes in both luminance and chroma channels. We demonstrate experimentally that the descriptor is highly invariant to photometric and geometric (e.g. pose, camera viewpoint) variations, and exhibits high discriminative power in a face detection setting. We implement a face detection system, combining intensity-based and the new color descriptor, and it is found to produce substantial contribution to detection performance. The face detector is based on the Support Vector Machine (SVM) classifier. A fuzzy variant for SVM is examined, with an algorithm for the computation of the fuzzy memberships matrix in a face detection setting.

This fuzzy SVM classifier formulation leads to further improvement in face detection

rates.