Compressive classification based on autoregressive features
Matteo Testa, Enrico Magli
2016 International Conference on Communications (COMM), Bucharest, 2016, pp. 433-438
Compressed Sensing (CS) allows to efficiently acquire and compress a signal within a single operation. However, reconstructing the original signal is typically expensive. Hence, being able to perform signal processing operations in the compressed domain is extremely important. In this paper we propose a new technique to perform classification tasks in the compressed domain. In order to perform compressive classification we employ feature vectors which can estimated directly in the compressed domain. The feature vectors we propose are the autoregressive (AR) parameters since they are well suited to approximate signals in a more general fashion than specifically suited feature vectors. We validate the proposed technique with two kind of natural signals: texture and speech showing the effectiveness of the proposed technique which can achieve excellent classification performance.