Estimation of effectiveness image generation using diffusion algorithms to increase the training sample for classifications tasks on geospasial data
Keywords: Diffusion Algorithms, LoRA, Image lassifications, Synthetic Data, Geospasial Data
Abstract. The article presents a comparative analysis of learning approaches based on two sampling methods. The training is conducted on samples obtained when shooting real objects, and on several samples in which the images obtained when shooting real objects are supplemented with data generated using diffusion algorithms. The article also discusses the method of obtaining the generated data. The data generation method is based on diffusion algorithms. The initial model of the selected generation and fine-tun is the FLUX, fine-tun method - LoRA (Low-Rank Adaptation). The fine-tun method is being considered for use on a data set of no more than one thousand elements. An open source geospasial dataset was used in training and testing.