Evaluating CNN Models for Chicken Disease Identification Using Feces Image

Abstract

Poultry farming faces significant challenges due to infectious diseases such as Newcastle Disease, coccidiosis, and salmonellosis, which lead to high mortality rates and economic losses. The purpose of this study is to evaluate and compare the performance of multiple Convolutional Neural Network (CNN) architectures for automatic chicken disease classification using feces images. The methodology employs a transfer learning approach using five pretrained CNN models: VGG16, ResNet50, InceptionV3, MobileNetV2, and EfficientNetB0. The dataset consists of 8,067 images from four classes: Healthy, Newcastle Disease, Coccidiosis, and Salmonella. Image preprocessing includes resizing, normalization, and data augmentation, while model performance is evaluated using accuracy, macro precision, macro recall, and macro F1-score. The results show that all models achieve strong performance with accuracies above 96%. VGG16 and ResNet50 obtain the highest accuracy of 97.52%, while EfficientNetB0 achieves 97.19% with faster inference speed. The conclusions indicate that ResNet50 provides the most optimal balance between accuracy and computational efficiency. The implications of this study suggest that CNN-based transfer learning can be effectively applied as an early disease detection tool in poultry farming, especially for real-time and resource-limited applications.