Hybrid approach of video indexing and machine learning for rapid indexing and highly precise object recognition

Video database and object recognition have been treated as separate problems in the past. Previous retrieval applications achieved rapid indexing and robust retrieval capabilities, but the precision of recognizing objects in video images is lower than object recognition using machine learning. In contrast with video retrieval, machine learning needs vast computational training time in advance and cannot handle similarity easily. To solve this problem, we present an image-based video recognition framework combined with video retrieval and object recognition. To develop an effective combination, we evaluated several retrieval methods and the support vector machine (SVM), which is one of the most popular supervised learning techniques. From experimental results, we found that the combination of extended color-pair retrieval and SVM using color location is the most effective pair for high precision and rapid indexing of a video recognition system.