Method for generating infrared big data for deep learning algorithm training by using small sample data

Abstract

With regard to precision guided weapon, a large amount of feature data is required for the training of non-cooperative target deep learning recognition models. In this paper, based on the remote sensing data and target temperature information, the variable parameters of big data generation are proposed by analyzing the formation mechanism of infrared radiation characteristics. And then, the 3D temperature field of the target is constructed based on small sample information. Finally, based on the 3D temperature field, multiple sets of infrared characteristic data are generated with different observation angles and sun positions. By analyzing the data results, it can be seen that when the observation angle is changed, the variation of the tank barrel radiation characteristics are up to 80%, and when the position of the sun is changed, the variation of the tank hatch radiation characteristics is up to 100%. The big data generation method proposed in this paper has the characteristics of diverse characteristics, high data scalability and high feature continuity, and can be used as training data for deep learning models.