Automated Fish Freshness Classification Using CNN: A Scalable Solution for Sri Lankan Coastal Regions

Abstract

Fish is a crucial component in sustaining human nutrition, serving as a major source of protein throughout history. Ensuring fish freshness is the key determinant of fish quality, consumer health and market value globally. In Sri Lanka, within traditional fish freshness assessment methods, manual sensory evaluations are often subjective and inaccurate due to humanized measurements. Chemical sensors-based evaluation methods are reliable but remain inaccessible to small-scale fishermen and local consumers, due to higher costs and resource limitations. To address this gap, this study employs Convolutional Neural Networks (CNNs) to automate fish freshness classification. Image samples were collected from Mannar coastal regions including Gulf and Island of Mannar and Vankalai Lagoon. The study comprised obtaining diverse dataset of 400 images (200 fresh and 200 non-fresh) per region of interest (ROI) including whole fish, fish eye and fish gill, total 1,200 images (600 fresh and 600 non-fresh). Each sample was manually categorized as fresh or non-fresh while capturing pictures. A singular dataset was created by combining all the images. Then data augmentation increased the dataset to 6500 images (3250 fresh and 3250 non-fresh). Then sequenced preprocessing techniques were applied including resizing, labelling and splitting dataset into training and testing sets in 7:3 ratio. A basic custom CNN model was developed with three convolutional layers with max-polling, dropout and dense layers. Adam optimizer was utilized for training with early stopping to prevent overfitting. The model achieved an impressive testing accuracy 94% along with excellent 94% precision, recall and F1-scores. The confusion matrix and precision-recall curve further validated the model’s effectiveness. By providing a scalable, cost-effective solution for fish freshness classification, outperforming traditional methods and holding significant potential to develop mobile applications. Future work will explore larger datasets and real-time implementation in fish markets.