Benchmarking CNN and YOLO Models for Automated Classification of Fish Freshness

Abstract

Fish freshness assessment is essential for ensuring food quality and consumer safety; however, conventional visual inspection remains subjective and inconsistent. Although deep learning has shown promising performance in image classification, standardized benchmarking of Convolutional Neural Networks (CNN) and YOLOv8 Classification under identical experimental settings for fine-grained fish freshness classification remains limited. This study compares both models using the same dataset, preprocessing pipeline, augmentation strategy, and training configuration to evaluate predictive performance and computational efficiency. The dataset comprised digital images of tongkol and slungsung fish categorized into four classes: Fresh Tongkol, Rotten Tongkol, Fresh Slungsung, and Rotten Slungsung. Model performance was evaluated using accuracy, precision, recall, F1-score, training time, and inference speed. CNN achieved superior predictive performance with 99.25% accuracy, 99.13% precision, 99.13% recall, and 99.13% F1-score, whereas YOLOv8 Classification achieved 89.88% accuracy, 89.96% precision, 89.88% recall, and 89.89% F1-score. Conversely, YOLOv8 required only 15 minutes for training and 9 ms per image for inference, compared with 23 minutes 20 seconds and 18 ms for CNN. These findings establish a robust benchmark for selecting deep learning architectures by balancing predictive accuracy and computational efficiency in automated fish freshness inspection systems.