Scholarlink

<p>With people’s pursuit of healthy diet, many tasks about food have emerged. Food recognition is the most important. However, in Chinese food recognition, there are two challenges. Firstly, the presence of multiple foods in a Chinese meal results in each food item appearing in localized areas of the image. Additionally, the cutting and mixing of ingredients in Chinese dishes make it challenging that the global features of the food can’t represent accurately about the food. Secondly, different foods may share similar ingredients, leading to comparable visual features. In response to these challenges, our paper introduces a solution named Learning Multi-Scale Features Automatically from food and ingredients (LMSFA). LMSFA uses multi-scale fusion block to extract local features of different sizes and use a late fusion block to align image and ingredient information. During the training phase, LMSFA adopts a strategic approach by selecting hard negative pairs for each image and use a loss function on these pairs. This involves choosing pairs of images that contain similar foods but should be embedded separately in the feature space. Because of the lack of multi-label Chinese food image dataset, a public dataset named ChineseFood-200 is proposed. After experiments, LMSFA exhibits a 7.76% increase in F1 score in food recognition compared with ResNet50.</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

Learning multi-scale features automatically from food and ingredients

  • 期刊论文

  • 发表日期：

    2025年4月29日

查看全文
  • Ruoxuan Zhang,
  • Dantong Ouyang,
  • Ximing Li,
  • Hongtao Bai,
  • Chenming Zhang,
  • Lili He

摘要

With people’s pursuit of healthy diet, many tasks about food have emerged. Food recognition is the most important. However, in Chinese food recognition, there are two challenges. Firstly, the presence of multiple foods in a Chinese meal results in each food item appearing in localized areas of the image. Additionally, the cutting and mixing of ingredients in Chinese dishes make it challenging that the global features of the food can’t represent accurately about the food. Secondly, different foods may share similar ingredients, leading to comparable visual features. In response to these challenges, our paper introduces a solution named Learning Multi-Scale Features Automatically from food and ingredients (LMSFA). LMSFA uses multi-scale fusion block to extract local features of different sizes and use a late fusion block to align image and ingredient information. During the training phase, LMSFA adopts a strategic approach by selecting hard negative pairs for each image and use a loss function on these pairs. This involves choosing pairs of images that contain similar foods but should be embedded separately in the feature space. Because of the lack of multi-label Chinese food image dataset, a public dataset named ChineseFood-200 is proposed. After experiments, LMSFA exhibits a 7.76% increase in F1 score in food recognition compared with ResNet50.