Abstract
Automated detection of diseases in the poultry farming industry is seriously challenged in resource-limited farming environments where computational resources and technical expertise are scarce. This work fills this gap, via systematic evaluation of lightweight transfer learning architectures for practicalpro deployment. Two state-of-the-art pre-trained Convolutional Neural Network (CNN) models, MobileNetV2 and MobileNetV3Small, were tested along with three traditional Machine Learning Models (Support Vector Machine (SVM), Logistic Regression (LR) and K-Nearest Neighbours (KNN)) by using a balanced dataset containing 6436 images of faecal samples from three classes: Coccidiosis, Salmonella and Healthy. MobileNetV2-SVM showed better performance with 96.17% test accuracy (96% precision, recall, and F1-score), which was much better than other pipelines based on MobileNetV3Small (maximum 83.94% accuracy). The optimized pipeline achieves real-time inference at 61 milliseconds per image, enabling deployment on standard hardware. A publicly accessible web-based application was developed, allowing farmers and veterinary practitioners to perform smartphone-based disease classification without specialized expertise, democratizing AI-powered diagnostics for resource- limited agricultural settings. This research establishes a systematic benchmark for lightweight feature extraction architectures combined with traditional machine learning classifiers in poultry disease detection and demonstrates that practical, farmer-accessible AI diagnostics can achieve clinical-grade accuracy even in resource constrained environments.