Performance analyses of deep learning techniques in aircrafts classification

With the developing technologies in recent years, many attack and defense systems are needed. The most prominent area of this need is the aviation sector. Many aircraft operate national or international flights during the day. However, there may be possible attacks. Detection of aircraft is of great importance in order to prevent this situation and take the necessary precautions. In this study, it is aimed to analyze the performance of deep learning techniques on aircraft classification. In this study, GoogleNet and ResNet18 are used in image classification techniques. Random test and learning classes with different percentages are created for approximately 5000 aircraft images in the ready-made dataset, which includes 5 separate classes: Passenger Aircraft, Military Aircraft, Helicopter, Drone and Rocket. According to the results of these test and learning classes, it is decided to continue with 90% learning and 10% test classes, as a generally higher performance rate is observed in the results with more training data. While the data input size of ResNet18 and GoogleNet architectures are 224x224x3, the data inputs are changed and training is performed 5 times with 90% training class of 210x210x3, 230x230x3 and 250x250x3 dimensions and the average values of the results are taken. These values have shown us that; better performances can be achieved by changing the data sizes, compared to the results given by the original methods. With Googlenet 210x210x3 data entry, a better percentage of performance has been achieved for criticality situations.It has been observed that with RestNet18 230x230x3 data entry, an acceptable level of lower performance was achieved in case of time criticality, but less than half of the time was saved.