Visual Food Ingredient Prediction Using Deep Learning with Direct F-Score Optimization.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Food ingredient prediction from images is a complex multi-label classification problem.
• Real-world datasets exhibit severe class imbalance, complicating model training and evaluation.
• The F-score is crucial for evaluating performance in imbalanced classification tasks.

Purpose of the Study:

• To develop a computationally efficient method for direct F-score optimization in food ingredient prediction.
• To address the challenges posed by class imbalance in multi-label classification tasks.
• To improve the state-of-the-art performance in food ingredient recognition.

Main Methods:

• Reformulated direct F-score optimization as a cost-sensitive classifier optimization problem.
• Developed an efficient algorithm for estimating optimal relative cost parameters.
• Evaluated the proposed method on the Recipe1M dataset.

Main Results:

• Achieved a micro F1 score of 0.5616 on the Recipe1M dataset.
• Demonstrated a substantial improvement over the previous state-of-the-art score of 0.4927.
• The proposed framework offers a principled and generalizable solution for class imbalance.

Conclusions:

• The novel F-score optimization framework provides an efficient and effective solution for imbalanced multi-label classification.
• This approach significantly enhances food ingredient prediction accuracy.
• The method is generalizable to other domains facing similar class imbalance challenges.