Classification with transfer learning

Learning an object category from only a few samples is generally not adequate for correct classification. One needs many training samples to obtain a classifier that generalizes well and that has sufficient success rates. However, in several applications, including target detection from Ground Penetrated Radar (GPR) data, collecting a lot of annotated data is not always possible. In a GPR data collection, the images formed show a nonlinear dependence on the soil properties such as the permeability and permittivity. Therefore, even if enough training data were available to train a good classifier for one soil type (such as dry sand); the success of this classifier does not translate well if the soil type is changed (say, to wet sand). In this work, multi-model knowledge transfer (KT) is proposed to detect a picnic tube buried in different media using GPR. In the proposed method, scale invariant feature transform (SIFT) features are extracted from GPR data. Then, least-square support vector machine (LS-SVM) classifiers are trained for three soil types (i.e. dry sand, dry sandstone, and wet sandstone) where there is ample data available. Then, adaptive LS-SVM is used to train a classifier that detects the target picnic tube in wet sand from where there is only scarcely available training data. We show that (i) knowledge transfer from multiple sources (i.e. multiple types of sand) generates better results than single source transfer; and (ii) as little as 3 training data from the unknown source increases the detection rates by 25% for single source KT, and 32% for multiple source KT