An l2-norm Regularized Underwater Target Classifier with Improved Generalization Capability

Improving the generalization capability of a target classifier has become one of the primary challenges in underwater target recognition systems. This paper addresses the task of classification in the framework of ill-posed inverse problems, and discusses the problem of overfitting, the solution to which has been formulated using the technique of regularization. l2 norm regularization on a logistic regression classifier has been implemented utilizing Newton's method to minimize the cost function for parameter optimization. Evaluation results with the help of Receiver Operating Characteristics and classification accuracy reveal the performance improvement of the classifier while making predictions on unseen samples.