A Modified Hierarchical Vision Transformer Model for Poultry Disease Detection

Abstract

Poultry production faces challenges from diseases like newcastle, salmonella, and coccidiosis, which are critical to global food security, resulting in economic losses and public health concerns. Current detection technologies, such as human inspections and PCR-based procedures, are time-consuming and costly, limiting scalability. Convolutional neural networks (CNNs) like ResNet50 and VGG16 have shown promise for automating disease identification, but they struggle with generalization and collecting fine-grained local and global information. In this study, we propose a deep learning solution based on a hierarchical vision transformer (HViT) model to detect poultry diseases from fecal images. We compare the performance of our HViT model with traditional CNNs (ResNet50, VGG16), lightweight architectures (MobileNetV3_Large_100, XceptionNet), and standard vision transformers (ViT) (ViT-B/16). The experimental results demonstrate that our HViT model outperforms other models, achieving an average validation accuracy of 90.90% with a validation loss of 0.2647. The HViT's ability to balance local and global feature recognition highlights its potential as a scalable solution for real-time poultry disease detection. These findings underscore the significance of hierarchical attention in addressing complex image analysis tasks, with implications for broader applications in agriculture and medical imaging.