Novel Deep Learning Framework For Bovine Iris Segmentation

TL;DR

This paper introduces a new deep learning framework for bovine iris segmentation that achieves high accuracy with minimal annotation, improving livestock biometric identification.

Contribution

It presents a novel deep learning approach using U-Net with VGG16 backbone that performs well even on corrupted images with limited labeled data.

Findings

Achieved 99.50% accuracy and 98.35% Dice score

Effective segmentation on corrupted images without extensive annotations

Advances livestock biometric identification methods

Abstract

Iris segmentation is the initial step to identify biometric of animals to establish a traceability system of livestock. In this study, we propose a novel deep learning framework for pixel-wise segmentation with minimum use of annotation labels using BovineAAEyes80 public dataset. In the experiment, U-Net with VGG16 backbone was selected as the best combination of encoder and decoder model, demonstrating a 99.50% accuracy and a 98.35% Dice coefficient score. Remarkably, the selected model accurately segmented corrupted images even without proper annotation data. This study contributes to the advancement of the iris segmentation and the development of a reliable DNNs training framework.