Year: 2024
Type: conference-paper
Abstract: Automatic picture captioning is a prominent re-search area of artificial intelligence technology (AI). Its ability to enhance AI models by translating observed data into human language opens up a wide range of real-time applications. In this study, we explore picture captioning in the Bengali language using the global attention mechanism. Given the limited prior research in this area, we provide a comprehensive assessment of two distinct global attention approaches: the general approach and the concatenation approach. Additionally, we evaluate the performance of two CNN encoders, VGG19 and InceptionV3,within these models. The models are trained on a secondary dataset consisting of 4,849 nature-based images, each annotated with a single caption, enabling the models to gain a broad understanding of related categorical information. To achieve our research objectives, we developed and trained four separate models using this new dataset. Our analysis, both qualitative and quantitative, demonstrates that these algorithms are capable of generating human-like captions for similar images. The results indicate that models using the concatenation approach, particu-larly with the InceptionV3encoder, performed best, achieving a BLEU-1 score of 84.85. In contrast, the model using the general approach with VGG19 underperformed in generating satisfactory captions.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4404037696