<![CDATA[Reliable assessment of fish freshness supports product quality, food safety, and consumer acceptance throughout fisheries supply chains. Sensory inspection remains common, but reliance on individual judgment can produce inconsistent assessments and limit monitoring efficiency. Here, a multimodal framework combines Computer Vision, IoT sensing, and Deep Learning to monitor Nile tilapia freshness in realtime. Its visual branch uses YOLO26 to locate fish eyes and EfficientNet B3 with CBAM to classify the extracted regions. A BiLSTM network learns temporal patterns from MQ136 and MQ137 measurements of H₂S and NH₃ associated with spoilage. An MLP then combines the class probabilities from both branches to assign the final freshness category. In the detection experiment, YOLO26 recorded 0.930 precision and 0.953 mAP@0.5. Visual classification with EfficientNet B3 plus CBAM reached 92% accuracy, whereas sensor classification with BiLSTM reached 93%, with AUC values as high as 0.99. Combining the modalities yielded 97% accuracy and a macro average F1 score of 0.97, exceeding the performance of either branch alone. Grad CAM maps located the main visual responses within eye regions relevant to freshness deterioration. These results support the framework's interpretability and its potential use for intelligent fish freshness assessment under near realtime monitoring conditions.]]>
Copyrights © 2026
