Attention-Based Convolutional Neural Network for Ingredients Identification

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Smart catering relies on accurate ingredient identification for efficiency and cost reduction.
• Existing ingredient classification methods suffer from low accuracy and flexibility.
• Automating ingredient identification is crucial for reducing labor costs in food service.

Purpose of the Study:

• To develop a highly accurate and flexible ingredient identification system for smart catering.
• To address the limitations of current methods in recognizing a wide variety of fresh ingredients.
• To implement an open-set recognition module for handling previously unseen ingredients.

Main Methods:

• Construction of a large-scale fresh ingredients database.
• Design of an end-to-end multi-attention-based convolutional neural network (CNN) model.
• Integration of an open-set recognition module for unknown category prediction.

Main Results:

• The proposed CNN model achieved 95.90% accuracy in classifying 170 types of ingredients.
• The open-set recognition module attained 74.6% accuracy in identifying unknown samples.
• Successful deployment in smart catering systems resulted in 92% average accuracy and 60% time savings.

Conclusions:

• The developed AI model represents a state-of-the-art solution for automatic ingredient identification.
• The system demonstrates practical effectiveness and significant operational benefits in real-world smart catering applications.
• The inclusion of open-set recognition enhances the model's adaptability to dynamic inventory changes.