Face Detection Using Single Cascade of Customized Features Discriminators

Face detection has become an important and helpful tool for camera and video processing. Useful human-computer interaction (HCI) applications such as drivers assistant system that prevents accidents and saves pedestrian lives when drivers attention is absent, needs a head pose estimator. A head pose estimator cannot function without face detector. There has been a considerable amount of literature to address the problem. The most significant results obtained on uptight frontal face detection which is a sub-problem of a larger problem of face detection. There are other types of sub-problems that has been studied with least significant advancements that the upright frontal face detection had accomplished. The problem of multi-pose detection is still under study and it remains hard. A solution to this large scale of the problem (multi-pose face detection) is critical in head pose accuracy. This thesis suggests a multi-pose face detection algorithm for uncontrolled environments. The detector is designed to be used in building head pose estimator for a human-computer interaction application. The observed design of the detector has to implement a cascade of classifiers. Each classifier has to address at least one certain area of the problem. The design have to maintain speed and an acceptable detection rate. These requirements can be satisfied by constructing the cascade to implement fast and simple classifiers at first stages of the cascade. A novel use of the integral image as a fast filter was invented to be placed at the start of the detection process. Included in the cascade, classifiers that are trained on special designed features aimed to solve part of the problem. One special unique classifier is a data mining based classifier that uses a modified version of the Maximal Frequent Itemset Algorithm (MAFIA) [2] for feature extraction. Special features classifiers use the extracted facial features information extracted from a new knowledge-based classifier/filter that was created with the capacity to locate to an acceptable ac- curacy the location of eyes, mouth and nose using a suite of approaches including discreet local minima and geometric measures. The extracted facial features were used to estimate head pose and extract classifier features accordingly to enhance detection rates. A cascade of classifiers based on fast and simple contrast features was used to refine and speed up the detection process. To further improve speed some components were parallelized. As an attempt to overcome some of the fundamental challenges of face detection, lighting correction and noise reduction were implemented based on the information extracted from images. Results are reported on the FDDB [12] benchmark showed 5.22% detection rate with 2000 false positives while OpenCV implementation of Viola-Jones [19] face detector showed 65.92 detection rate with 2010 false positives. This comparison is flawed; because Viola-Jones is an upright face detector and even though FDDB [12] includes a number on non-frontal faces and profiles the majority of the faces are frontal. The two solutions address two different problems that reflect large differences in difficulty. A standard benchmark testset and evaluation system as FDDB [12] benchmark and com- parable results from the same class of the problem at the time of writing this document was not available. The key points to building good face detector in general are; (1) resolving speed issues using fast techniques (e.g. integral image) at the start of the cascade and a powerful design, (2) using a huge number of different strong and weak features, and (3) eliminating variations (i.e. pose , noise and lighting variations). The algorithm was also tested on MIT+CMU upfront faces testset and reported 43.56% detection rate with 504 false positives