Sh.F. Mukhamedova, H.B. Sharipov

Annotation:

This paper presents a novel method for semantic data augmentation based on probabilistic diffusion models. A software module has been developed based on the Django framework and the Diffusers library, implementing the Image-to- Image technique using the Stable Diffusion v1.5 model. A key feature of the method is the mathematical optimization of the "strength" parameter (transformation strength) and the utilization of Markov chains to generate photorealistic variations of students' faces while preserving their biometric identity. Experimental validation conducted on a dataset of 500 students demonstrated the high effectiveness of the proposed approach. Training the ResNet-50 model using the augmented data improved recognition accuracy in low-light conditions from 72.3% to 89.4%. The study confirms that generative artificial intelligence can effectively address the problem of data scarcity, thereby saving resources on the collection of real photographs and enhancing the overall robustness of security systems.
Keywords: computer vision, Stable Diffusion, data augmentation, face recognition, Markov chain, latent space, artificial intelligence.