Egg fertility assessment system using deep learning technology

In the poultry industry, accurate distinction between fertilized, unfertilized, and hatching failure eggs is a key factor for improving hatching efficiency and reducing resource waste. This study comprehensively evaluated 12 CNN architectures for non-destructive egg fertility classification based on candling-style images captured on days 5-6 of incubation, under four experimental scenarios: Case A (image-level random split, three classes), Case B (egg-level separation, three classes), Case C (image-level random split, two classes), and Case D (egg-level separation, two classes). In Cases A and B, overall accuracy ranged from 93.65%-98.65% and 90.82%-95.79%, respectively. Per-class recall for fertilized eggs exceeded 98% across all models, while hatching failure egg recall was highly variable (5.88%-76.47% in Case A; 0%-27.78% in Case B), reflecting severe class imbalance and visual ambiguity. In Cases C and D (fertilized vs. unfertilized only), overall accuracy improved to 97.02%-99.80% and 95.84%-98.81%, respectively, with unfertilized egg recall ranging from 74.47%-97.87% (Case C) and 58.33%-87.50% (Case D). Egg-level evaluation (Cases B and D) consistently yielded lower accuracy than image-level splitting (Cases A and C), confirming that image-level splitting overestimates real-world generalization performance. CNN selection should be guided by the balance between computational resources and required accuracy: lightweight models (SqueezeNet: 4.7 MB; ShuffleNet: 5.5 MB) suit resource-constrained deployments, while larger models (Inception-v3, VGG-16, VGG-19) offer superior accuracy for high-performance applications. Classification of hatching failure eggs remains challenging due to their visual similarity to both fertilized and unfertilized eggs depending on the timing of developmental arrest, and represents an important direction for future research.