• Informações do Trabalho

Titulo Development of an Artificial Intelligence-Aided Software for Annotating Image Datasets Subtítulo Autor PAULO VICTOR DE MAGALHÃES ROZATTO Orientador LUIZ MAURILIO DA SILVA MACIEL Resumo Deep learning is a highly successful class of methods in the artificial intelligence (AI) field that has a variety of applications. To perform well, deep learning models require a large amount of high-quality annotated data. Data annotation is a time-consuming and laborious task that requires a significant amount of human labor, which makes it expensive. This work aims to reduce the time required to annotate image datasets by building an easy-to-use software tool that has semi-automated annotation powered by an artificial intelligence model. We developed a web-based tool and employed HQ-SAM, a deep neural network for image segmentation based on Vision Transformers, to generate polygon annotations based on the user’s prompts. Although HQ-SAM has a good zero- shot generalizability, we fine-tuned it on the Bean Leaf Dataset to evaluate how well the network adapts to specific tasks. We observed an increase in accuracy of the fine-tuned model in comparison with the pre-trained one. We tested our tool with 20 participants, all of whom are from the computer vision and graphics fields. We asked them to annotate the same two images both manually and AI-aided, and we recorded the annotation times. Lastly, we asked the participants to fill out a usability form about their user experience. In our evaluation, we registered a median speedup of 1.5× regarding the AI-aided annotation compared to manual annotation and overly positive answers regarding our tool’s ease of use and usefulness. Palavras-Chave Deep learning, web application, semi-automated annotation, image datasets, vision transformers