Search

In this paper, the performance influencing class analysis (PI-CA) framework is proposed for performance analysis of pattern recognition systems dealing with data with great variety and diversity. Through the PICA procedure, the population of data is divided into subsets on which the system achieves different performances by means of statistical methods. On basis of the division, performance assessment and analysis are conducted to estimate the system performance on the whole data population. The PICA framework can predict true performance in real application and facilitate comparison of different systems without the same test set. The PICA framework is applied to the analysis of a broadcast news segmentation system. The procedure is presented and experimental results were given, which verified the effectiveness of PICA. [ABSTRACT FROM AUTHOR]