Deep Learning-Based Prediction of Fish Freshness and Purchasability Using Multi-Angle Image Data.

Area of Science:

• Food Science
• Computer Science
• Artificial Intelligence

Background:

• Assessing fish freshness is crucial for food safety and quality control.
• Traditional methods for evaluating fish freshness can be subjective and time-consuming.
• Objective, automated methods are needed to complement existing quality assessments.

Purpose of the Study:

• To develop and evaluate deep learning models for predicting sea bass (Dicentrarchus labrax) freshness using image data.
• To identify the most effective deep learning algorithms and architectures for freshness prediction.
• To assess the performance of models in predicting various quality parameters, including purchasability.

Main Methods:

• Daily imaging of sea bass from purchase to spoilage across multiple angles.
• Evaluation of 22 quality parameters: 10 categorical (sensory) and 12 numerical (color-based).
• Training 2464 predictive models using seven transfer learning algorithms (EfficientNetB0, ResNet50, DenseNet121, VGG16, InceptionV3, MobileNet, VGG19).

Main Results:

• MobileNet algorithm demonstrated the best overall performance, predicting 15 out of 22 parameters accurately.
• DenseNet121 achieved the highest classification accuracy (0.9894) for predicting purchasability, the critical freshness indicator.
• Various performance metrics including accuracy, precision, recall, F1-score, mean absolute error, and tolerance-based error were used for evaluation.

Conclusions:

• Deep learning-based image analysis is a promising and viable method for objective fish freshness evaluation.
• Transfer learning algorithms can be effectively applied to predict multiple fish quality parameters from visual data.
• The study highlights the potential of AI in enhancing food quality and safety monitoring systems.