Bangla Image Caption Generation Using Vision Transformer (ViT) Based Model

In the era of digital content and visual communication, Bangla image captioning has emerged as a crucial technology for enhancing accessibility, improving content discoverability, and bridging the language gap for millions of Bangla speakers worldwide. Our work proposes a novel approach combining a...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:2025 International Conference on Electrical, Computer and Communication Engineering (ECCE) S. 1 - 6
Hauptverfasser: Sarker, Arpita, Das, Udoy, Murad, Hasan
Format: Tagungsbericht
Sprache:Englisch
Veröffentlicht: IEEE 13.02.2025
Schlagworte:
Bangla Image Captioning
Bangla NLP
Computational modeling
Computer architecture
Computer vision
Encoder-Decoder Transformer
Feature extraction
Image edge detection
Measurement
Meteors
Residual neural networks
Transformers
Vision Transformer
Visual communication
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:In the era of digital content and visual communication, Bangla image captioning has emerged as a crucial technology for enhancing accessibility, improving content discoverability, and bridging the language gap for millions of Bangla speakers worldwide. Our work proposes a novel approach combining a vision transformer as a feature extractor with a customized encoder-decoder architecture for Bangla language generation. We use a wide range of metrics, such as BLEU, ROUGE-L, and METEOR, that have been specifically tailored for Bangla to assess the effectiveness of our model. The proposed model performs at the cutting edge with a BLEU score of 0.6572, a ROUGE-L score of 0.6218, and a METEOR score of 0.4513. Comparative analysis with other architectures, such as Xception, ResNet101, ResNet50, and InceptionV3 combined with encoder-decoder models, gives information about both the advantages and drawbacks of several methods for captioning images in Bangla.
DOI:10.1109/ECCE64574.2025.11013210