Dataset Expansion by Generative Adversarial Networks for Detectors Quality Improvement

Alexander Kostin, Vadim Gorbachev

Abstract: Modern neural network algorithms for object detection tasks require large labelled dataset for training. In a number of practical applications creation and annotation of large data collections requires considerable resources which are not always available. One of the solutions to this problem is creation of artificial images containing the object of interest. In this work the use of generative adversarial networks (GAN) for generation of images of target objects is proposed. It is demonstrated experimentally that GAN’s allows to create new images on the basis of the initial collection of real images on background images (not containing objects), which simulate real images accurately enough. Due to this, it is possible to create a new training collection containing a greater variety of training examples, which allows to achieve higher precision for detection algorithm. In our setting, GAN training does not require more data than is required for direct detector training. The proposed method has been tested to teach a network for detecting unmanned aerial vehicles (UAVs).