Fine-Grained Food Image Recognition: A Study on Optimising Convolutional Neural Networks for Improved Performance.

Area of Science:

• Computer Vision
• Machine Learning
• Food Science

Background:

• Food waste is a significant environmental and resource issue.
• Existing computer vision datasets lack fine-grained food item categorization.
• There is a need for precise food identification to enable targeted waste reduction strategies.

Purpose of the Study:

• To develop a model for identifying individual food items for a mobile application.
• To bridge the gap in fine-grained food datasets for waste reduction research.
• To integrate computer vision technology for practical food waste mitigation.

Main Methods:

• Evaluated seven convolutional neural network architectures for multi-class food image classification.
• Utilized a custom dataset of 41,949 images across 20 food item classes.
• Performed extensive parameter tuning, including optimizers, image sizes, dropout rates, and activation functions.

Main Results:

• DenseNet architecture established a baseline performance with 74% training accuracy and 68% validation accuracy.
• Optimized DenseNet significantly improved performance, achieving 99% training accuracy and 79% validation accuracy.
• The optimized model demonstrated superior generalization capabilities and reduced loss metrics.

Conclusions:

• Optimized DenseNet architecture is highly effective for fine-grained food item identification.
• The developed model, integrated into the FridgeSnap app, aids users in recognizing food and reducing waste.
• This research contributes to environmental sustainability through advanced technological solutions for food waste reduction.