Improved Sparse Representation of Open Set HRRP Recognition Method

Traditional target recognition methods classify targets within a closed set, and the methods cannot effectively identify unknown categories outside the library. Research on open set recognition (OSR) has been developed in computer vision, yet rare research is conducted in radar target recognition. First the HRRP recognition performance of several state-of-the-art OSR algorithms are compared in open set, including SROSR, WSVM, and 1-vs Set. Among them the SROSR algorithm performances better than the others. However, manual setting of tail and weight parameters is required in the algorithm, which may difficult to be obtained in practical applications. Focused on this issue, an modified SROSR method based on the mean ratio of reconstruction errors is presented. The mean ratio of reconstruction errors is used as a new metrics and its statistical distribution is modeled with extreme value theory (EVT). Compared with the classical SR-OSR, manual setting of the tail of reconstruction errors and weight parameters are avoided. Experiments were conducted on HRRP data generated from MSTAR SAR chips, and the results show that the recognition performance of the proposed method is better than the classical SROSR algorithm.