<![CDATA[Manually assessing the freshness of tuna (Euthynnus affinis) within traditional auction markets is frequently subjective, inconsistent, and prone to human error. To address these limitations, this study implemented a robust Deep Learning model utilizing the MobileNetV2 architecture to classify fish freshness levels based on ocular imagery into three distinct categories: Fresh, Less Fresh, and Not Fresh. Although the MobileNetV2 architecture is inherently reliable for image classification, conventional manual hyperparameter tuning often leads to suboptimal model configurations and severe overfitting challenges. As a strategic solution, this study implemented the advanced Tree-structured Parzen Estimator (TPE) algorithm through the Optuna framework to automatically optimize critical hyperparameters, including learning rate, dropout rate, optimizer type, and the fine-tune depth of the convolutional layers. Empirical testing results rigorously proved that the strategic integration of Optuna TPE with the MobileNetV2 model successfully provided a significant performance boost. The optimized model achieved an accuracy of 93.51%, representing a substantial increase of 32.04% compared to the 61.47% accuracy observed in the manual baseline model. Correspondingly impressive improvements were recorded in key performance metrics, specifically achieving a Precision of 93.64%, a Recall of 93.37%, and an F1-Score of 93.49%. In conclusion, this research demonstrates that automated hyperparameter optimization using the Optuna TPE framework significantly enhances the reliability, efficiency, and generalization capability of Deep Learning models, establishing a highly accurate and practical solution for the automated detection of tuna freshness in real-world fishery environments.]]>
Copyrights © 2026
