Combination of Classifier Cascades and Training Sample Selection for Robust Face Detection

Face detection is one of the most fundamental tasks in human-computer-interaction, surveillance, and, more recently, image retrieval. Determining the location and size of faces in input images is a prerequisite for many other applications, including face recognition. In recent years several breakthroughs have been made in this field. These days, face detectors deliver high detection rates, low false alarm rates and run in real-time. Despite the efforts and publicly available tools, training high-performance face detectors from scratch remains a challenge. Mostly, because training time for a single cascade can be in the order of days and various training parameters have to be chosen carefully. Usually, training involves acquiring heuristics and a feeling for the intricacies of the training process and the influence of training parameters. A substantial amount of time is spent training classifiers iteratively and modifying parameters, while usually discarding intermediate results. The goal of this work is to overcome some of the problems of training cascade classifiers and to promote the use of custom-trained classifiers. Specifically, two problems are addressed in this work. First, an approach to combine several trained cascade classifiers into a single cascade is presented and evaluated. Second, a technique to optimize the training set is explored. A major challenge during cascade training is the choice of training parameters. There is no ideal way to choose these parameters and optimization is not feasible. Usually, the process involves several attempts or guesses at the right parameters and, finally, the best performing classifier is selected. Instead of discarding intermediate results, several of these classifiers are combined into a single new classifier. Unlike previous work, the base classifiers are not run in parallel but a fixed number of individual classifier stages are optimized, selected and combined into a new classifier without added run-time overhead. Experiments have shown the importance of a proper choice of training samples. Classifiers trained with a reduced amount of well-chosen samples can outperform a classifier that was trained on a far larger training set. The use of less training samples to achieve the same performance decreases the required training time, especially with large training sets, where results cannot be cached. Additionally, forcing the classifier to focus on difficult training examples has shown to increase classification performance. Therefore, a method to select an optimized set of training samples from a large set with the help of support vector machines is explored. The results of both presented approaches have been evaluated on the widely used, publicly available CMU+MIT database. Both the SVM-based training sample selection and the cascade combination approaches are shown to improve the performance over the base classifiers. Cascade combination allows to generate a classifier within a single day that performs nearly as well as a single, high-performance classifier trained in more than ten days. Additionally, classifiers generated by cascade combination outperform the orignal base cascades. Face detectors trained with SVM-based training set selection perform better than equally trained base classifiers with a random choice of training samples. Both presented approaches were able to produce cascade classifiers that clearly outperform the publicly available OpenCV face detectors.