Scholarlink

<p>As e-commerce continues to expand, the demand for automated fashion image captioning has surged, particularly in enhancing customer discovery and boosting business revenue. Traditional methods for generating fashion image captions often fall short in capturing the nuances and dynamics of fashion trends. To address this, we introduce MTFIC, a novel multi-transformer model specifically designed for fashion image captioning. MTFIC leverages Contrastive Language-Image Pretraining (CLIP) to extract robust image features and Bidirectional Encoder Representations from Transformers (BERT) to encode caption features. These features are then fed into a new feedforward neural network decoder, employing a multi-head self-attention mechanism, to generate natural and semantically rich descriptions. Notably, the decoder is randomly initialized, allowing for high customization and performance optimization. Our extensive experiments on the FACAD and Fashion-Gen datasets, evaluated using metrics such as ROUGE, BLEU, METEOR, and CIDEr, demonstrate that MTFIC significantly outperforms previous approaches, underscoring its effectiveness in fashion image captioning. This work presents a substantial step toward automating fashion image description, with potential applications in e-commerce, social media, and beyond. To facilitate readers in replicating our experiments and evaluating our results, we have published the source code at the link: <a href="https://github.com/huyquoc22/ImageCaption">https://github.com/huyquoc22/ImageCaption</a>.</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

MTFIC: enhanced fashion image captioning via multi-transformer architecture with contrastive and bidirectional encodings

  • 期刊论文

  • 发表日期：

    2025年7月1日

查看全文
  • Bui Thanh Hung,
  • Vo Quoc Huy

摘要

As e-commerce continues to expand, the demand for automated fashion image captioning has surged, particularly in enhancing customer discovery and boosting business revenue. Traditional methods for generating fashion image captions often fall short in capturing the nuances and dynamics of fashion trends. To address this, we introduce MTFIC, a novel multi-transformer model specifically designed for fashion image captioning. MTFIC leverages Contrastive Language-Image Pretraining (CLIP) to extract robust image features and Bidirectional Encoder Representations from Transformers (BERT) to encode caption features. These features are then fed into a new feedforward neural network decoder, employing a multi-head self-attention mechanism, to generate natural and semantically rich descriptions. Notably, the decoder is randomly initialized, allowing for high customization and performance optimization. Our extensive experiments on the FACAD and Fashion-Gen datasets, evaluated using metrics such as ROUGE, BLEU, METEOR, and CIDEr, demonstrate that MTFIC significantly outperforms previous approaches, underscoring its effectiveness in fashion image captioning. This work presents a substantial step toward automating fashion image description, with potential applications in e-commerce, social media, and beyond. To facilitate readers in replicating our experiments and evaluating our results, we have published the source code at the link: https://github.com/huyquoc22/ImageCaption.