Face detection in images by transform classifiers

Abstract

The thesis describes operation of classifiers based on Haar and LBP features and evaluates their performance on selected data sets. Haar and LBP based classifiers are popular in machine learning due to their good performance and fast operation after training. These classifiers were developed over the years and their structure is now quite complicated due to the Adaboost cascade which is used to adaptive optimize the classification. The classifier problem is studied in the thesis for the face detection in images. Classifiers are trained on image sets with faces in different positions and image sets with no faces.Two type of image sets are utilized during detection. First set is made by real faces from available database. Second set is made by faces which were generated by 3D graphics software called Facegen. Both data type were used in experiments separately and in mixture. We found that the Haar classifier performs better than LBP but this is at the cost of increased computation time. For real faces the performance is very good, over 90% correct face detection. For 3D generated faces the performance is much weaker, 65% correct face detection. This indicates that 3D faces are much more difficult to detect by the classifiers.