Landmine detection using semi-supervised learning.

Abstract

Landmine detection is imperative for the preservation of both military and civilian lives. While landmines are easy to place, they are relatively difficult to remove. The classic method of detecting landmines was by using metal-detectors. However, many present-day landmines are composed of little to no metal, necessitating the use of additional technologies. One of the most successful and widely employed technologies is Ground Penetrating Radar (GPR). In order to maximize efficiency of GPR-based landmine detection and minimize wasted effort caused by false alarms, intelligent detection methods such as machine learning are used. Many sophisticated algorithms are developed and employed to accomplish this. One such successful algorithm is K Nearest Neighbors (KNN) classification. Most of these algorithms, including KNN, are based on supervised learning, which requires labeling of known data. This process can be tedious. Semi-supervised learning leverages both labeled and unlabeled data in the training process, alleviating over-dependency on labeling. Semi-supervised learning has several advantages over supervised learning. For example, it applies well to large datasets because it uses the topology of unlabeled data to classify test data. Also, by allowing unlabeled data to influence classification, one set of training data can be adopted into varying test environments. In this thesis, we explore a graph-based learning method known as Label Propagation as an alternative classifier to KNN classification, and validate its use on vehicle-mounted and handheld GPR systems.