A method for improving the quality of image annotation in semantic monitoring gis of business processes
R. M. Pasichnyk, Lyudmila Babala, M.V. Machuliak · INFORMATICS AND MATHEMATICAL METHODS IN SIMULATION · 2024
This article addresses the pressing issue of automating the image labeling process for computer vision systems in agriculture.The authors investigate methods for creating image datasets and configuring parameters for image classification models using neural networks based on the TensorFlow framework.The scientific significance of the work lies in developing new approaches to automated collection of thematic image collections and formalizing the methodology for parametric training of classification models.The practical value of the research is expressed in improving the efficiency of the image labeling process for geoinformation systems in the agricultural sector.The research methodology includes analyzing existing approaches to image labeling, developing an algorithm for automated formation of thematic image collections, formalizing a method for parametric training of the classification model, and experimental verification of the proposed approaches.Main results of the work: 1.An algorithm for automated formation of thematic image collections has been developed.2. A method for parametric training of the image classification model using the TensorFlow framework has been formalized.3. The dependence of classification accuracy on the size of the training sample and image augmentation parameters has been experimentally established.The study showed that with optimal selection of augmentation parameters and using 48 images per label in the training sample, it is possible to reduce the classification error to an acceptable level of 8%.The work makes a significant contribution to the development of automated image processing methods for agricultural geoinformation systems.The practical significance of the results lies in improving the efficiency of monitoring and management processes in the agricultural sector.