Analysis of the Pattern Recognition Efficiency On Non-Optical Images
Nikita Andreevich Andriyanov, Vitaly Dementiev, Anatoly A. Gladkikh · 2021 Ural Symposium on Biomedical Engineering, Radioelectronics and Information Technology (USBEREIT) · 2021
The article examines algorithms for processing non optical images using convolutional neural networks (CNN). Particular attention is paid to the recognition of objects on radar images of the Earth's surface and on X-ray images of baggage inspection. A brief overview of research in this topic is presented. The indicators of recognition efficiency on limited datasets using data augmentation are considered. In addition, the article describes in detail the processes of building neural networks and their training. A comparative analysis of CNNs is carried out for different training parameters and different network architectures. It has been shown that CNNs provide satisfactory results when processing non-standard types of images, such as Xray images of baggage or radar images from Synthetic Aperture Radar (SAR) systems. Furthermore, the text deals with various metrics such as accuracy, precision and recall. The suggested algorithms provides about 90-100% for recognition of SAR images and about 90% for X-ray images.