Classification of Biscuit Quality With Deep Learning Algorithms.

Area of Science:

• Food Science and Technology
• Computer Science
• Artificial Intelligence

Background:

• Traditional quality control in food production is labor-intensive and prone to human error.
• Automated defect detection systems are needed to enhance efficiency and reduce costs in biscuit manufacturing.

Purpose of the Study:

• To develop and evaluate deep learning models for automated detection of defective biscuits.
• To compare the performance of different deep learning architectures for binary and multi-class defect classification.

Main Methods:

• Creation of two datasets: one for binary classification (defect/no defect) and one for multi-class classification (overcooked, texture defect, incomplete).
• Training and evaluation of deep learning models including EfficientNet, ResNet, XceptionNet, and MobileNet.
• Utilizing Grad-CAM for model interpretability to visualize attention on defect regions.

Main Results:

• EfficientNet demonstrated superior performance with 93.89% accuracy in binary classification and 95.03% in multi-class classification.
• ResNet achieved comparable results, while XceptionNet and MobileNet showed competitive F1 scores, especially for texture defects.
• Grad-CAM analysis confirmed EfficientNet's effective focus on critical defect areas.

Conclusions:

• Deep learning models, particularly EfficientNet, offer a viable solution for efficient and precise automated quality control in biscuit production.
• The study highlights the potential of AI-driven inspection systems to reduce errors, time, and costs in industrial food manufacturing.
• These findings support the integration of advanced machine learning techniques for enhanced food product quality assurance.