Reducing the Individual Labelling Effort of Holstein-Friesian Cattle with Deep Learning

Abstract

This thesis addresses the challenge of minimising or eliminating the need for manual labelling in the identification of individual Holstein-Friesian cattle within real-world farm settings. By capitalising on the unique coat patterns of the cattle, we introduce a non-invasive identification method that leverages visual information and integrates deep learning techniques to reduce labelling efforts. This research comprises three pivotal aspects:

Firstly, the thesis proposed a detector capable of identifying both the heading orientation and torso regions of cattle, providing directional-normalised data to alleviate the manual labelling burden and support subsequent identification tasks.

Secondly, a data-driven, self-guided system for cattle identification using video footage is developed. This approach leverages unlabelled tracklet data from the cattle detector, combined with self-supervised learning, and requires only a minimal amount of human labelling. It significantly reduces the labelling workload and demonstrates the efficient use of large volumes of unlabelled data for recognition tasks. Experimental results, featuring 155 individuals, demonstrate an accuracy of 92.4% (Adjusted Rand Index = 0.93) with only 10 minutes of labelling effort.

Lastly, an identifier is introduced to reduce the labelling burden for new farms, with a focus on video footage utilisation.
By exploiting domain adaptation and self-supervised learning with tracklet information, a labelled source group can facilitate clustering for the individual identification of targeted new individuals. The proposed method demonstrates the ability to reduce the cattle labelling burden while maintaining accuracy across various farms or open-set scenarios. Experimental results, featuring 59 individuals, indicate an accuracy of 94.1% (Adjusted Rand Index = 0.95).

This thesis underscores the application of deep learning in developing efficient visual identification systems for livestock, resulting in a significant reduction of the manual labelling workload. It suggests promising implications for real-world applications in precision agriculture and the enhancement of animal welfare.

Date of Award	10 Dec 2024
Original language	English
Awarding Institution	University of Bristol
Supervisor	Neill W Campbell (Supervisor) & Tilo Burghardt (Supervisor)