Deep Learning-Based Assessment of European Seabass (Dicentrarchus labrax) Fillet Freshness and Quality Using Image...

Artificial intelligence and image-based deep learning offer promising non-destructive approaches for rapid fish quality assessment. This study evaluated the potential of deep learning models to classify freshness-related sensory attributes and predict color parameters of European seabass (Dicentrarchus labrax) fillets using image data acquired under controlled illumination. Seven sensory quality parameters and six color attributes were assessed. A structured image dataset was obtained at a fixed camera-to-sample distance of 45 cm, and 65 independently trained models, comprising 35 classification and 30 regression models, were developed using DenseNet121, InceptionV3, MobileNet, VGG16, and VGG19 architectures. Model training and evaluation were performed using group-based 10-fold cross-validation to ensure that images from the same fillet were not shared between the training and test sets. Classification performance depended strongly on both architecture and target parameter. DenseNet121 achieved the highest accuracy for purchasability (92.60%), water release (90.20%), odor (75.60%), color (77.40%), and disintegration (87.70%), whereas MobileNet performed best for texture (81.20%) and total score (57.70%). Purchasability was the most predictable sensory parameter across architectures, with a mean accuracy of 86.84%, whereas total score was the least predictable at 45.64%. In the regression tasks, DenseNet121 achieved the lowest MAE for a* (0.680), L* (2.467), hue (0.961), and whiteness (2.357), while MobileNet performed best for b* (1.200) and chroma (1.024). Grad-CAM visualizations indicated that model activations were predominantly associated with the fillet region, although more difficult classes showed more fragmented and occasionally edge- or background-associated activation patterns. Overall, DenseNet121 provided the strongest and most consistent performance, ranking first for nine of the 13 prediction targets. These findings demonstrate the potential of RGB imaging and deep learning for rapid, objective, and non-destructive assessment of selected freshness and quality attributes of European seabass fillets.