Scholarlink

Image captioning is an important problem of artificial intelligence that involves computer vision and natural language generation to provide textual descriptions of images. This paper proposes a new method of using EfficientNet-B0 for image encoding and a Transformer for decoding to enhance the captioning process. Unlike other models which depend on ResNet or VGG structures, EfficientNet-B0 is a lightweight yet very powerful feature extraction system that also increases the computational efficiency. The proposed model was trained using the Flickr8k dataset with the pre-processing to enhance the variability of the data. The performance evaluation based on the BLEU scores indicates the proposed approach is better than the baseline models with BLEU-1 of 0.5924 as compared to ResNet-50 with 0.4800, and VGG16 with 0.5680. Nevertheless, there are some issues which can be associated with creating detailed captions for intricate pictures. Future work will look into using larger datasets and also extend the use of vision-language pretrained models for better context. The experiment results also show that incorporating the EfficientNet-B0 and transformers provides a highly efficient and feasible solution for real-world image captioning.

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

Vision-to-Words: A Resource-Efficient Transformer-Based Approach for Image Captioning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Nagendra B. Hanchinale,
  • Parikshith Patil,
  • Shridhar B. Anigolkar,
  • Karthik K. Noolvi,
  • Raghavendra V. Vadavadagi,
  • Uday Kulkarni,
  • Shashank Hegde

摘要

Image captioning is an important problem of artificial intelligence that involves computer vision and natural language generation to provide textual descriptions of images. This paper proposes a new method of using EfficientNet-B0 for image encoding and a Transformer for decoding to enhance the captioning process. Unlike other models which depend on ResNet or VGG structures, EfficientNet-B0 is a lightweight yet very powerful feature extraction system that also increases the computational efficiency. The proposed model was trained using the Flickr8k dataset with the pre-processing to enhance the variability of the data. The performance evaluation based on the BLEU scores indicates the proposed approach is better than the baseline models with BLEU-1 of 0.5924 as compared to ResNet-50 with 0.4800, and VGG16 with 0.5680. Nevertheless, there are some issues which can be associated with creating detailed captions for intricate pictures. Future work will look into using larger datasets and also extend the use of vision-language pretrained models for better context. The experiment results also show that incorporating the EfficientNet-B0 and transformers provides a highly efficient and feasible solution for real-world image captioning.