Hard Voting Ensemble of CNNs for Accurate Detection of Poultry Diseases using Faecal Images

Nigeria's agricultural GDP greatly benefits from the poultry industry, which is also necessary for the nation's protein needs. Poultry farming is a critical sector in Nigeria, contributing 58% of the country’s total livestock production. However, poultry diseases such as Salmonellosis, Coccidiosis, and Newcastle significantly impact productivity. Prior studies have primarily focused on individual model performances without addressing the trade-offs between computational efficiency and classification accuracy, nor have they exploited ensemble strategies to enhance diagnostic reliability. To address these limitations, this study evaluates the performance of five pre-trained convolutional neural network models—VGG-16, Inception, MobileNet, ResNet, and EfficientNet—and their ensemble via Hard Voting for the early detection of poultry diseases using faecal imagery. Leveraging transfer learning and selective layer freezing, experiments on two distinct datasets revealed that models such as MobileNet and EfficientNet maintained high accuracy levels—up to 99% accuracy with 98% precision and recall—when trained without layer freezing or with only one-fifth of layers frozen, thereby significantly reducing training time. In contrast, models like ResNet and VGG-16 exhibited notable performance drops when a larger fraction of layers was frozen. The Hard Voting ensemble approach consistently achieved an accuracy of approximately 98% on Dataset 1, demonstrating improved robustness over individual models. But performance on Dataset 2 wasn’t encouraging because of the low quality of the dataset. While the proposed method shows promise in improving early poultry disease detection, limitations such as reduced performance on more heavily frozen models and the need for broader validation on diverse datasets highlight avenues for future work, including real-time deployment and the exploration of adaptive freezing techniques