MACHINE LEARNING BASED FALSE ALARMS REDUCTION FOR SMALL INFRARED TARGETS

There are many techniques for feature detection in Small Infrared Images and to minimize false alarm rates up-gradation is a must. This paper gives the idea of machine learning dependent target detection where feature extraction is done using Discrete Wavelet Transformation and feature selection is done using Nature Inspire Algorithm that is Particle Swarm Optimization at the end classification is done using classifier Random Forest with filter-based target detection method. Modified Top- Hat filter is used for the filtered database. DWT is used for extraction of features that breaks a target into small parts termed as wavelets obtained from mother wavelet via shifting and dilation. Before feature selection each feature is analyzed using two different approaches first is intensity distribution and other one is probability density function and from both the observation median, mean, entropy, variance shows the best result because overlapping of probability distribution is lease in these four. PSO works on the principle of bird flocking together after calculating gbest and pbest of each feature, Variance, Mode, Kurtosis, Zernike moment, Entropy, Skewness are the highest-ranking feature values these features are used further for the classification purpose. Out of all the classifiers logistic regression and random forest classifiers have shown best results for the false alarm reduction rate. Earlier many wrapper-based feature selection algorithms have been used but PSO and random forest classifiers have combinedly reduced false alarm rate by 7.9 whereas without classifiers the false alarm rate was 30.6. FAR without classifier and with classifier improved by 3.7 factor.