Fish Freshness Detection Based on Eyes and Gills Using YOLOv8 Model

Abstract

This research presents an experimental fish freshness detection system utilizing the YOLOv8 model for automatic
assessment of fish quality based on visual indicators, specifically the clarity of eyes and the condition of gills. Fish
freshness assessment is critical in the seafood industry, especially for export markets where international safety
standards apply. Traditional methods such as organoleptic tests are often subjective and inefficient, highlighting
the need for scalable, automated solutions. This study leverages computer vision and transfer learning, using
YOLOv8 pretrained on COCO weights, and analyzes a dataset of 3,916 high-resolution images expanded to 4,600
through strategic data augmentation and perfect class balancing. Four freshness classes—fresh eyes, non-fresh
eyes, fresh gills, and non-fresh gills—are evenly represented. The dataset was resized to 640×640 pixels and split
into 3,200 training, 800 validation, and 600 testing images. The YOLOv8 model achieved 99.83% detection
accuracy and a [email protected] of 0.995 after 20 training epochs under controlled conditions. Confusion matrix
analysis revealed near-perfect classification across all categories. A comparative experiment using EfficientDet
demonstrated YOLOv8’s superior performance with a 24.43% accuracy improvement. These results validate the
effectiveness and robustness of the approach for real-world deployment. The study provides a scalable foundation
for implementing computer vision-based seafood quality control systems, with strong potential to enhance food
safety, operational efficiency, and consumer trust in the seafood supply chain.