A METHOD FOR SYNTHESIZING THERMAL IMAGES USING GAN MULTI-LAYERED APPROACH

Abstract. The active development of neural network technologies and optoelectronic systems has led to the introduction computer vision in various fields science technology. Deep learning made it possible solve complex problems that a person had not been able before. use multi-spectral optical significantly expanded field application video systems. Tasks such as image recognition, object re-identification, surveillance require high accuracy, speed reliability. These qualities are provided by algorithms based on deep convolutional networks. However, they have large databases images objects achieve state-of-the-art results. While color different widely available public domain, then similar thermal either available, or represent small number types objects. quality three-dimensional modeling for imaging spectral range remains at an insufficient level solving important tasks, which precision realistic synthesis is especially due complexity cost obtaining real data. This paper focused method synthesizing generative adversarial We developed algorithm image-to-image translation. changed original GAN architecture converted loss function. presented new approach. For this, we prepared special training dataset including about 2000 tensors. evaluation results obtained showed proposed can be used expand images.