A Method for Creating Realistic Synthetic Images Using a Generative Deep Learning Model for Classifying Anomalies in Panoramas

This paper describes a method for generating realistic synthetic images using a generative deep learning model. When generating training examples, the features of climate, relief, architecture, and landscape of a particular area are considered. The generative model is represented by a variational autoencoder consisting of an encoder, a latent space, and a decoder. To train the autoencoder, a new dataset CSCDroneCL_v1 was created, consisting of manually labeled panoramic images and training examples selected from the open VisDrone2022 dataset. Based on the SOTA-ConvNet template, a discriminative neural network model was designed and trained to classify anomalies obtained from multitemporal panoramas. Architectural compatibility between generative and discriminative models is ensured by using a single design pattern. The unified architecture of neural network models made it possible to apply the transfer learning method. The use of pretrained parameters in the variational autoencoder model makes it possible to compensate for the small size of the CSCDroneCL_v1 dataset. The generated synthetic training examples of panoramas were added to the VisDrone2022 dataset. Using the new dataset, a discriminative model was trained, resulting in an increase in anomaly classification accuracy by 21.2% for the selected class.