MCE training based continuous density HMM landmine detection system

Landmines that scattered all over the world have caused the global landmine crisis. The detection task is very difficult due to the large variety of landmines and environmental conditions. Therefore the de-mining technology is acting a critical rule in reducing the damage of landmines.;Hidden Markov Model (HMM) was proposed in this field in recent years. The baseline HMM based landmine detection system is based on the Maximum Likelihood (ML) estimation. Although ML approach is efficient and the model convergence is guaranteed, its performance is limited due to the unknown true data distribution and insufficient training data.;In this thesis, the Minimum Classification Error (MCE) training is applied to the CDHMM based landmine detection system. The generalized probabilistic descent (GPD) algorithm is used to update the system's parameters in the MCE approach, where the training procedure is proved to converge with probability one. From the parameter analysis, we found that the convergence factor did not affect the system performance significantly, and for different CDHMM parameter, different iteration step size should be assigned. We also found that the parameters of the sigmoid function have big effect on utilizing training data. Through a set of experiments of GPR landmine data, we show that the MCE trained HMMs can achieve better performance than the ML trained HMMs. The MCE approach can increase the detection rate by about 2--4% over the baseline system with the same HMM structure. With different HMM structures, system performance improvements due to MCE training are also significant.