Synergizing meat science and interpretable AI: Quantifying crispness gradients for quality authentication of Tilapia fillet processing

Abstract

Crispy tilapia has become a popular aquatic product due to its unique texture and high market demand. However, fillets at different stages of crispiness vary significantly in nutritional value and taste, directly affecting product quality and consumer experience. Therefore, rapid and accurate identification of the crispiness of tilapia fillets is crucial for farmers, traders and consumers. Hyperspectral imaging (HSI) technology has emerged as a powerful tool in food quality testing, offering rich spectral information that can be leveraged for detailed analysis. In this study, we propose a method combining HSI and dual-branch convolutional neural network (DB-CNN) for classifying tilapia fillets at different stages of crispness. By separately processing VNIR and SWIR data and fusing them in the feature space, the DB-CNN achieved 95.74 % classification accuracy, outperforming traditional fusion methods. Grad-CAM++ visualization validated the model's recognition of key spectral features. This approach offers an effective solution for authenticity identification and quality control of crispy tilapia fillets and showcases its potential for broader applications in food and aquatic product classification.