Tuna Loin Quality Grading Using Image Processing and EfficientNetV2
Narendra Agastya, Ledya Novamizanti, Gelar Budiman · 2025
The global tuna industry struggles with inconsistent quality grading due to the subjective and labor-intensive nature of manual methods. This study develops a tuna loin quality grading computer vision model based on EfficientNetV2 using image preprocessing techniques. Tuna loins are classified into three quality grades (A, B, and C) based on color and texture features. To address lighting inconsistencies of the environment and enhance texture detail of the tuna loins, the study evaluates pre-processing methods including Shades of Gray (SOG), Self-Adaptive Illumination Correction (SAIC), and Contrast Limited Adaptive Histogram Equalization (CLAHE). Various optimizers (Adam, AdamW, and SGD) and learning rate schedulers (Cosine Annealing and Cyclic) are also tested. The final model—combining SAIC and CLAHE pre-processing with EfficientNetV2M, the Adam optimizer, and Cyclic Learning Rate Scheduling–achieves $96.9 \%$ validation accuracy and 96.0% test accuracy. This demonstrates the effectiveness of the proposed approach in improving grading reliability for real-world applications.