Landmine Detection Using Multispectral Images

The conditions in which images are obtained to perform multispectral detection of landmines have a direct influence on the methods that are used to perform automatic detection of landmines. In this paper, two methodologies are proposed: one using traditional classifiers and the other using deep learning, namely, a convolutional neuronal network (CNN). In the first methodology, classifier fusion techniques are also used. The performance of the first methodology was evaluated as a function of the number of landmine features, the environment, and the depth of the mine. In deep learning, a study was carried out based on the feature map, the type of landmine, and the environment. A quantitative analysis shows that traditional classifiers achieved an overall accuracy of above 97% in indoor and outdoor environments for the detection of landmines. The adopted deep learning methodology presented an increase in the performance for larger mines and a decrease for smaller ones. These experimental results shed light on the factors that influence the detection of mines and into the advantages and disadvantages of CNN compared with the classical classifier methods.