Scholarlink

<p>Image Captioning (IC), a critical juncture of computer vision and Natural Language Processing (NLP), seeks to generate meaningful textual descriptions for visual content. A recurring challenge in this domain is the loss evaluation mismatch, where standard training losses do not always correlate well with human-evaluated metrics, leading to suboptimal model performance in real-world applications. To overcome this divergence, we introduce an innovative dual-objective framework for image captioning that synergistically combines conventional training loss with the BLEU score, a widely accepted metric for evaluating caption quality. Specifically, our optimization strategy aims to simultaneously minimize training loss while maximizing the BLEU score, ensuring tighter alignment between model training and human-centric evaluation metrics. Extensive experiments exhibit that the proposed approach mitigates the loss evaluation mismatch and yields captions that are contextually relevant and linguistically rich. This work underscores the potential of reframing optimization objectives for bridging the gap betwixt algorithmic training procedures and human-perceived quality, setting a new paradigm for future research in IC. The source code and dataset for this study are publicly available on GitHub(<a href="https://github.com/ranjithjms/image_captioning_dual_optimization">https://github.com/ranjithjms/image_captioning_dual_optimization</a>) and COCO (<a href="https://cocodataset.org/">https://cocodataset.org/</a>), respectively, to facilitate further research and benchmarking.</p>

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

Innovative Dual-Objective Framework for Image Captioning: Harmonizing Visual Analysis and Linguistic Precision

  • 期刊论文

  • 发表日期：

    2025年6月24日

查看全文
  • Ranjith Gnana Suthakar Alphonse Raj,
  • B. J. Sandesh

摘要

Image Captioning (IC), a critical juncture of computer vision and Natural Language Processing (NLP), seeks to generate meaningful textual descriptions for visual content. A recurring challenge in this domain is the loss evaluation mismatch, where standard training losses do not always correlate well with human-evaluated metrics, leading to suboptimal model performance in real-world applications. To overcome this divergence, we introduce an innovative dual-objective framework for image captioning that synergistically combines conventional training loss with the BLEU score, a widely accepted metric for evaluating caption quality. Specifically, our optimization strategy aims to simultaneously minimize training loss while maximizing the BLEU score, ensuring tighter alignment between model training and human-centric evaluation metrics. Extensive experiments exhibit that the proposed approach mitigates the loss evaluation mismatch and yields captions that are contextually relevant and linguistically rich. This work underscores the potential of reframing optimization objectives for bridging the gap betwixt algorithmic training procedures and human-perceived quality, setting a new paradigm for future research in IC. The source code and dataset for this study are publicly available on GitHub(https://github.com/ranjithjms/image_captioning_dual_optimization) and COCO (https://cocodataset.org/), respectively, to facilitate further research and benchmarking.