Knowledge Distillation Network for Poultry Disease Classification from Chicken Dropping Images

Abstract

This paper develops an automated non-invasive system for disease detection in poultry farming by classifying chicken droppings. This study examines multiple deep learning architectures employing transfer learning to classify chicken droppings. Based on comparative analysis, a knowledge distillation network was proposed utilizing DenseNet as the teacher model due to its superior performance metrics. A lightweight student network was proposed and trained using knowledge distillation. This effectively transfers discriminative feature representations from the teacher model while significantly reducing computational complexity. Experimental evaluation on a publicly available poultry faecal image dataset demonstrated that the proposed knowledge distillation network achieved an accuracy of 98.40%, precision of 97.72%, specificity of 99.44%, recall of 98.21%, and F1-score of 97.96%.