ConvNet Fine-Tuning Investigation for GPR Images Classification

Abstract

Deep learning has been widely implemented as a new classification platform during the past few years. One of the main problems facing deep learning is the problem of data dependency as it requires a very large amount of data for training. Therefore, transfer learning (TL) has been introduced as a solution to this problem. This study focuses on the fine-tuning strategy of the transfer learning and how it can be implemented to classify the ground-penetrating radar (GPR) images. The GPR data has been collected and processed using the matched filter algorithm and further clutter reduction techniques. The resultant GPR images compromises of a limited number of samples, therefore, the deployed convolutional neural network (ConvNet) has been trained first using another larger dataset, then fine-tuned using the GPR dataset. The obtained results are promising and show a high degree of precision and accuracy compared to previously conducted researches.