Image captioning in Bengali language using visual attention.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Natural Language Processing

Background:

• Automatic image captioning is a complex AI task, especially for intricate languages like Bengali.
• There is a significant lack of Bengali-captioned image datasets, hindering research and development.

Purpose of the Study:

• To curate a human-annotated Bengali image caption dataset.
• To develop an innovative end-to-end deep learning architecture for Bengali image captioning.

Main Methods:

• A new, meticulously human-annotated Bengali image caption dataset was created.
• An attention-driven decoder architecture was employed, integrating Gated Recurrent Units (GRUs) for visual features.
• An attention mechanism facilitated the interrelation between visual and linguistic representations.

Main Results:

• The model achieved a BLEU-4 score of 43%, METEOR score of 39%, and ROUGE score of 47%.
• These scores represent the highest performance to date for Bengali image captioning tasks.

Conclusions:

• The developed model and dataset significantly advance the field of Bengali image captioning.
• The attention-based architecture effectively generates coherent Bengali image descriptions.