Poultry-CViTNet. A Hybrid ViT--CNN Network for Non-Invasive Poultry Health Monitoring Based on Fecal Images

Abstract

Poultry diseases remain a major challenge for sustainable poultry production because delayed diagnosis can increase mortality, treatment costs, and the risk of disease transmission across farms. Conventional diagnostic methods based on manual observation, veterinary inspection, and laboratory testing are often time-consuming, labor-intensive, and difficult to scale in low-resource farming environments. To address these limitations, this study proposes a hybrid Convolutional Neural Network (CNN) and Vision Transformer (ViT) framework for automated poultry health monitoring using fecal images. The proposed model combines the local texture extraction capability of CNNs with the global contextual modeling ability of ViTs, enabling effective discrimination between healthy and unhealthy fecal samples. The proposed CNN–ViT model was evaluated using accuracy, precision, recall, F1-score, and confusion matrix analysis, with additional ablation experiments. The experimental results show that the hybrid model achieved an accuracy of 97.47 ± 0.36%, precision of 97.50 ± 0.38%, recall of 97.47 ± 0.41%, and F1-score of 97.47 ± 0.39%, outperforming the CNN-only and ViT-only baseline models. The ablation study confirms that integrating local convolutional features with transformer-based global attention improves classification reliability. These results demonstrate the potential of the proposed framework as a non-invasive, low-cost, and scalable decision-support tool for early poultry disease screening in smart farming environments. Additionally, the model contained only 3.62 million parameters, required 2.60 GFLOPs, and processed 194.95 images per second. Furthermore, Grad-CAM-based interpretability was applied to visualize the model’s decision-making process.