Simultaneously exploiting spectral similarity and spatial distribution for patterned minefield detection

摘要

In this paper we investigate how shape/spectral similarity of the mine signature and the minefield like spatial distribution can be exploited simultaneously to improve the performance for patterned minefield detection. The minefield decision is based on the detected targets obtained by an anomaly detector, such as the RX algorithm in the image of a given field segment. Spectral, shape or texture features at the target locations are used to model the likelihood of the targets to be potential mines. The spatial characteristic of the patterned minefield is captured by the expected distribution of nearest neighbor distances of the detected mine locations. The false alarms in the minefield are assumed to constitute a Poisson point process. The overall minefield detection problem for a given segment is formulated as a Markov marked point process (MMPP). Minefield decision is formulated under binary hypothesis testing using maximum log-likelihood ratio. A quadratic complexity algorithm is developed and used to maximize the log-likelihood ratio. A procedure based on expectation maximization is evaluated for estimating unknown parameters like mine-level probability of detection and mine-to-mine separation. The patterned minefield detection performance under this MMPP formulation is compared to baseline algorithms using simulated data.