Scholarlink

<p>In this study, we utilized CLIP (Contrastive Language-Image Pretraining) as a recommendation model for retrieving images of Thai food recipes relevant to a query ingredient text in the Thai language. We scraped a total of 22,500 food recipes online and used the recipe image and ingredient list as an image-text pair for training. After preprocessing the data, we trained CLIP using four variations of image and text encoders, employing distilbert-base-th-cased and wangchanberta-base-att-spm-uncased as text encoders, and ResNet50 and EfficientNet-B2 as image encoders. The model trained with distilbert-base-th-cased and ResNet50 achieved the best top-10 accuracy of 17.1%. CLIP applications for the Thai language are rare due to the language’s complexity and difficulty in embedding. However, our study demonstrates that CLIP can comprehend the relationship between image and text in the Thai language, particularly for ingredients with unique shapes or colors and popular recipes in the dataset.</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

Using contrastive language-image pre-training for Thai recipe recommendation

  • 期刊论文

  • 发表日期：

    2025年3月21日

查看全文
  • Thanatkorn Chuenbanluesuk,
  • Voramate Plodprong,
  • Weerasak Karoon,
  • Kotchakorn Rueangsri,
  • Suthasinee Pojam,
  • Thitirat Siriborvornratanakul

摘要

In this study, we utilized CLIP (Contrastive Language-Image Pretraining) as a recommendation model for retrieving images of Thai food recipes relevant to a query ingredient text in the Thai language. We scraped a total of 22,500 food recipes online and used the recipe image and ingredient list as an image-text pair for training. After preprocessing the data, we trained CLIP using four variations of image and text encoders, employing distilbert-base-th-cased and wangchanberta-base-att-spm-uncased as text encoders, and ResNet50 and EfficientNet-B2 as image encoders. The model trained with distilbert-base-th-cased and ResNet50 achieved the best top-10 accuracy of 17.1%. CLIP applications for the Thai language are rare due to the language’s complexity and difficulty in embedding. However, our study demonstrates that CLIP can comprehend the relationship between image and text in the Thai language, particularly for ingredients with unique shapes or colors and popular recipes in the dataset.