Scholarlink

Natural language production and visual comprehension are connected through image captioning. Gujarati and other low-resource languages are still underrepresented, despite the fact that state-of-the-art models perform well in high-resource languages. Using an encoder-decoder system with a GRU-based decoder and ResNet18 as visual feature extractor, this study suggests a lightweight image captioning model for Gujarati. The Flickr8k and Flickr30k datasets were used to manually translate and validate Gujarati captions. Morphological issues are addressed with SentencePiece tokenization. For generalization, minor image enhancement techniques are used. With BLEU-4 scores of 0.2255 for Flickr8k and 0.1925 for Flickr30k, the model performs competitively compared to current captioning systems of other regional languages. This study provides a scalable foundation for multilingual captioning in resource-constrained scenarios.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep Learning-Based Gujarati Image Caption Generator Using ResNet18 and GRU

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Dhrumi Patel,
  • Palak J. Parmar,
  • Het Bhatt,
  • Desai Tanvish,
  • Riya Patel

摘要

Natural language production and visual comprehension are connected through image captioning. Gujarati and other low-resource languages are still underrepresented, despite the fact that state-of-the-art models perform well in high-resource languages. Using an encoder-decoder system with a GRU-based decoder and ResNet18 as visual feature extractor, this study suggests a lightweight image captioning model for Gujarati. The Flickr8k and Flickr30k datasets were used to manually translate and validate Gujarati captions. Morphological issues are addressed with SentencePiece tokenization. For generalization, minor image enhancement techniques are used. With BLEU-4 scores of 0.2255 for Flickr8k and 0.1925 for Flickr30k, the model performs competitively compared to current captioning systems of other regional languages. This study provides a scalable foundation for multilingual captioning in resource-constrained scenarios.