A New CNN-Based Single-Ingredient Classification Model and Its Application in Food Image Segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Food Science

Background:

• Pixel-level annotation for food ingredient segmentation is labor-intensive and time-consuming.
• Existing methods struggle with the complexity and cost of creating detailed food datasets.

Purpose of the Study:

• To develop an efficient framework for food ingredient segmentation using image-level annotations.
• To reduce the dependency on extensive pixel-level data for training segmentation models.

Main Methods:

• A standardized biological-based hierarchical ingredient structure was introduced.
• A single-ingredient classification model was trained on an image-level annotated dataset.
• Feature maps from the classification model were utilized for ingredient segmentation.

Main Results:

• The proposed framework achieved significant results on the FoodSeg103 dataset.
• Key metrics include mIoU of 0.65, mDice of 0.77, mPurity of 0.83, mEntirety of 0.80, and mLoGTs of 0.06.
• The method demonstrates effectiveness in segmenting ingredients from food images.

Conclusions:

• The developed framework offers a viable alternative to pixel-level annotation for food ingredient segmentation.
• This approach provides a foundation for enhanced food recognition systems.
• The study highlights the potential of leveraging image-level data for complex segmentation tasks.