Multi-stage hybrid YOLO-driven and MobileNetV2-CNN variants for robust fish freshness classification

Abstract

This study presents a multi-stage hybrid representation learning framework for robust fish freshness classification to address the critical challenge of reliable quality assessment in unpreserved food supply chains. The proposed pipeline operates in two stages: stage-1 employs you only look once (YOLO)v8n as feature-gated detector to validate inputs and eliminate non-fish images, while stage-2 leverages transfer-learned MobileNetV2 variants enhanced with convolutional neural network (CNN) layers, fine-tuning, adaptive learning rate schedulers, and expanded fully connected layers for hierarchical classification into three freshness classes i.e., highly fresh, fresh, and not fresh. The framework has been trained and evaluated on two curated datasets comprising 4,500 fish and non-fish images and 11,111 freshness-labeled including augmented dataset to increase diversity. The experimental results showed significant improvements. The best performance model transfer learning (TL)-MobileNetV2 + CNN + fine-tuning achieved 98.07% training accuracy and 67.72% validation accuracy on 80:20 split, and training accuracy of 97.57%, validation accuracy of 97.21% through 10-fold cross-validation. The comparative benchmarking confirmed that dual-stage design outperformed baseline MobileNetV2 and YOLOv5s models across precision, recall, and F1-score. The findings highlighted significant value of integrating detection-driven validation with transfer-learning classification, and propose new benchmark for intelligent freshness monitoring. For future work, this study aims to explore attention-based models, data integration, and species diversity.