Using A Neural Network to Generate Images When Teaching Students to Develop an Alternative Text
| Autor: | Yekaterina A. Kosova, Kirill I. Redkokosh, Pavel O. Mikheyev |
|---|---|
| Jazyk: | English<br />Russian |
| Rok vydání: | 2024 |
| Předmět: | |
| Zdroj: | Открытое образование (Москва), Vol 28, Iss 1, Pp 9-20 (2024) |
| Druh dokumentu: | article |
| ISSN: | 1818-4243 2079-5939 |
| DOI: | 10.21686/1818-4243-2024-1-9-20 |
| Popis: | The purpose of research. The purpose of the study is to develop and test an approach to training digital content compilers in creating alternative text that accurately describes the original image, using a neural network to generate reference images reconstructed from the text. The lack of textual descriptions of visual content in a web resource limits digital accessibility, especially for users with visual disorders. To ensure accessibility, each informative image should be accompanied by the alternative text. Text alternatives generated by means of automated tools are known to be lower in quality to human-generated descriptions. Therefore, a digital content compiler must be able to develop the alternative text for images. It has been suggested that a neural network for generating images from text descriptions can act as a tool for checking the relevance of the developed text alternatives.Materials and methods. The study was carried out in April-May 2023. 17 undergraduate students studied the requirements for developing text alternatives, completed initial text descriptions for three proposed photographs, and then corrected the text using the Kandinsky 2.1 neural network according to the algorithm: generating an image from the description; visual comparison of the resulting image with the original; returning to editing the description or ending the process. Based on the initial and final descriptions, the researchers reconstructed the images using the same neural network. Further work consisted of assessing the quality of all text descriptions and the similarity of all generated images to the original ones. The results of the study (text descriptions; expert evaluations; links to generated images) were published as a data set in the Mendeley Data repository. The t-test, Pearson correlation and multivariate regression were used to analyze the data (at the specified significance level p = 0,05).Results. It was found that the quality scores of the initial and final text descriptions were not significantly different (p > 0,05), and also there were no significant differences for the length of the text (p > 0,05). At the same time, the similarity of the generated images and original photographs after students used the neural network has increased considerably (p < 0,05). Therefore, training in the neural network contributed to improving the quality (similarity to the original) of images generated from modified text descriptions, without losing the descriptions’ quality. It was also shown that the quality of the final text alternatives was higher the larger their size within the allotted limit, the better and shorter the initial descriptions (p < 0,05). Thus, concise and accurate alternative descriptions for images after training students in a neural network can be converted into equally high-quality text alternatives, the relevance of which is increased by adding plot details to the description.Conclusion. Neural networks generating images can be applied as a software tool to encourage potential content authors to create more accurate and complete alternative text while keeping it concise. It seems important to continue the research by extending it to other types of images and using a variety of neural networks. |
| Databáze: | Directory of Open Access Journals |
| Externí odkaz: |
|