Yere Nüfuz Eden Radar Verilerinde Tel Tespiti İçin Aktarım ve Çok Görevli Öğrenme Yöntemleri
xmlui.mirage2.itemSummaryView.MetaDataShow full item record
Ground Penetrating Radar (GPR) is a commonly used tool for buried target detection, classification and identification applications. GPR produces output based on the electrical conductivity differences of buried targets relative to the ground. Classification of targets with same shape but made up of different materials is a difficult problem, especially for targets with similar electrical conductivity. Reflections from objects except the target exponentially increase the difficulty of this problem. In such a classification problem, deep learning has been used since the extraction of features is very difficult and requires expert knowledge. Deep learning is a method that has demonstrated state-of-art performance over the last five years thanks to automatic learning of features in data and classifier. The deep learning model needs a lot of data to learn the features automatically. Since real data collection is costly and difficult, in this thesis synthetic GPR data is generated with the gprMax program, which simulates GPR. This dataset consist of plastic wire targets, wires which are made of other materials and clutter objects placed in dry / damp / wet soils. Generated data has been input into three classification algorithms; namely; (i) standard deep convolutional neural networks, (ii) a deep learning model that uses transfer learning and (iii) a deep learning model that uses multitask learning. Transfer learning is a process of transferring information acquired from a deep learning model previously trained with a large number of data to the current model. In this way, general features (color and edge perception) of the pre-trained model acquired from a large number data are transferred to the existing deep learning model and target detection performance has been increased. However, it has been observed that detecting targets in different soil types is difficult. For this reason, multitask learning method has been used to increase performance of the target detection. Multitask learning is an approach that improves learning about a task using the training knowledge of other related tasks. In this thesis, the primary task of multitask learning is to separate target and non-target objects. The second task is to determinate the soil type. Learning from each task can provide a better learning process for the other task. With multitask learning, a customized classifier has been trained to detect targets according to soil type. Application of the multitask learning method has shown that soil type task contributes positively to the achievement of target detection. The purpose of this thesis is to correctly classify buried targets with high performance. For this purpose, deep learning, transfer learning and multitask learning methods have been proposed and their contribution of achievement has been examined.