Analysis of GPR B scan images with deep learning methods

In this thesis, GPR B scan images are presented with architectures that efficiently analyze buried structures using deep learning methods. GprMax program was used to obtain two-dimensional GPR B scan images. Two data sets with different characteristics were generated by using simulation programs. A total of 180 GPR B Scan images are available in the first data set. These images in this data set have buried structures with different shapes and materials. There are patterns in the images that cause a false alarm condition due to the thickness of the material and the proximity of buried structures to each other. In the second data set, there are 4280 GPR B scan images in total by using data augmentation methods. Soil and materials with different electrical properties were used in the generation of these images. At the same time, different types of shape were used. Thirdly, a total of 9000 GPR B-G images obtained from GPR devices were analyzed. Within the scope of this analysis, it has been tried to determine type of GPR device, scanning frequencies and soil type in the surveyed area. The fourth data set contains 39 real GPR B scan images. These images were obtained from different concrete blocks with cavities. In the analysis of the data sets, several methods were used by considering the data set structure. Deep dictionary learning structure was used in order to realize the less number of data and analysis in the first data set. At the same time, different number of layers and classifiers were used to perform a comparative analysis. Training on Convolutional Neural Network (CNN) was carried out by performing on the second data set. CNN structures were used as AlexNet, VGG-16, GoogloeNet, ResNet-50 and SquezeeNet. During the training phase, trainings were carried out by using scratch, pre-trained and transfer learning structures of these models. At the same time, the sensitivity of training data on the model has been tested. In addition, Convolutional Support Vector Machines structures were trained. These structures are composed of three different models as small, medium and large type. Trained network structures have been tested with both simulation data and real data. In the third data set, different CNN pre-trained models were trained by using transfer learning. In the context of the proposed method for the specified data set, AlexNet, VGG-19, GoogloeNet, ResNet-50 and InceptionNet models were used. At the same time, two different CNN architectures were designed and comparisons were made with the performance of pre-trained CNNstructures. In the fourth data set, YOLO, Fast Regional-based Convolutional Neural Networks (Fast R-CNN) and Faster Regional-based Convolutional Neural Networks (Faster R-CNN) methods were used to detect cavities regions in the GPR B scan images. The obtained results were found to be very successful.