Ship detection and classification using type of convolution neural networks

Ship classification and detection systems are used in many areas that may cause problems such as national and local defense in countries with intense coastal and strait crossings, ship traffic control, illegal fishing, pirate, human traffickers and the global trade chain. In order to meet these kinds of needs easily, solutions with many different methods have been developed related to ship classification and detection. Research on these methods is divided into two groups intensely. These are the methods in used in Satellite and SAR images. Although the number of studies on these sources is high, the images that are come are not public. In addition, their resolution is not enough for ship classification and detection. As an alternative solution to these methods, CNNs with deep learning technology, have become quite popular. Over the past 30 years, computer vision technologies have had a hard time helping people in visual tasks. However, major advances in deep learning technology today have enabled computers to process images as successfully and even better as a person. One of the reasons why this thesis is based on CNN architectures is that the number of classes can be easily expanded according to the scenarios in the classification and detection problems of CNN models. In this thesis, the most popular deep learning CNN architectures in the literature such as VGG16, VGG 19, DenseNet121 and Xception are used for ship classification. Also, CNN architecture YOLOv3 with high performance is preferred for ship detection. The results of these models are compared in this thesis. Our data set of 6 different ship types, created by us, is used for training and testing. Since CNN models require a large number of training data sets, the transfer learning technology is applied in these models in order to eliminate the problem of lack of data. The contribution of our thesis to literature can be listed as comparing CNN model architectures for ship classification problems, investigating the effects of layer numbers in CNN model architecture and choosing the appropriate hyper parameter for this problem. For ship detection, we can evaluate the number of labeled data as an impact on performance.