Weapon detection with unmanned air vehicle images through deep learning algorithms

The proliferation of deep learning algorithms today has led to an increase in object detection and recognition applications in images and videos. Object detection and recognition applications have found solutions to many problems in the fields of security, defense, natural disasters (flood, earthquake and fire etc.), health (prevention of the spread of outbreaks etc.), agriculture, forestry in recent years. Regional Based Convolutional Neural Networks (R-CNN) are among the most widely used algorithms in object detection and recognition applications. Region based Convolutional Networks (Fast R-CNN) and faster region based convolutional neural networks (Faster R-CNN) algorithms have been developed in order to assist the detection applications of R-CNN. Another Convolutional Neural Networks (CNN) algorithm used to further increase the success of object detection applications is the ResNet101 algorithm. Especially, ResNet101, which is widely used in image detection, has been preferred to minimize differences such as object detection accuracy rate, object detection time of R-CNN, Fast R-CNN and Faster R-CNN algorithms. In this study, it is aimed to detect objects (weapons) from aerial images taken by unmanned aerial vehicle. In the images obtained, R-CNN were preferred because the correct prediction rate was higher than other R-CNN types in object (weapon) detection. In addition to R-CNN algorithms, the use of ResNet101 algorithm has been tried in this study in order to see its contribution to the correct prediction rate. In this context, training and test data sets were created using 200 images taken from the air with a drone. As a result of the training, the image based result was obtained with R-CNN architecture and ResNet101 architecture with 99% accuracy rate on the data set. With this study, it has been demonstrated how successful the R-CNN architecture and ResNet101 architecture are in the detection of objects (weapons) in unmanned aerial images.