Abstract:
Object detection, such as face detection using supervised learning, often requires extensive training for the computer, which results in high execution times. If the trained system needs re-training in order to accommodate a missed detection, waiting several hours or days before the system is ready may be unacceptable in practical implementations. This dissertation presents a generalized object detection framework whereby the system can efficiently adapt to misclassified data and be re-trained within a few minutes. Our developed methodology is based on the popular AdaBoost algorithm for object detection. AdaBoost functions by iteratively selecting the best among weak classifiers, and then combining several weak classifiers in order to obtain a stronger classifier. Even though AdaBoost has proven to be very effective, its learning execution time can be high depending upon the application. For example, in face detection, learning can take several days. In our dissertation, we present two techniques that contribute to reducing to the learning execution time within the AdaBoost algorithm. Our first technique utilizes a highly parallel and distributed AdaBoost algorithm that exploits the multiple cores in a CPU via lightweight threads. In addition, our technique uses multiple machines in a web service similar to a map-reduce architecture in order to achieve a high scalability, which results in a training execution time of a few minutes rather than several days. Our second technique is a methodology to create an optimal training subset to further reduce the training execution time. We obtained this subset through a novel score-keeping of the weight distribution within the AdaBoost algorithm, and then removed the images that had a minimal effect on the overall trained classifier. Finally, we incorporated our parallel and distributed AdaBoost algorithm, along with the optimized training subset, into a generalized object detection framework that efficiently adapts and makes corrections when it encounters misclassified data. We demonstrated the usefulness of our adaptive framework by providing detailed testing on face and car detection, and explained how our framework applies to developing any other object detection task.
