Biometric Identification of Cattle via Muzzle Print Patterns and Deep Learning in a Few-Shot Learning Context

Abstract

<p>Over the past decade, increasing global demand for livestock products has encouraged farmers to raise more stock. Tracking and monitoring the welfare of large numbers of livestock is, however, not an easy task. Poor animal welfare impacts greatly on consumer preference and negatively impacts productivity. Accurate recording of herd production depends on the identification of the animals. Individual animal identification allows producers to record and manage important animal information. A successful identification system should be able to identify an animal accurately and quickly. A wide range of traditional approaches including ear tagging, ear tattooing, hot ironing, freeze branding, ID collars, microchipping and visual markers such as paint have been used to track and identify individual animals within herds, but most of these approaches are invasive with the potential to cause pain and morbidity to stock. The current cattle identification approaches have significant limitations, however, recent innovations in human biometrics research such as face recognition systems offer promising alternative options. Biometric identification offers less invasive monitoring with corresponding benefits to livestock welfare.</p> <p>In this thesis, a general framework is proposed that is focused on using cattle muzzle print patterns as a biometric. Muzzle print patterns were selected based on prior research indicating that they are a highly distinctive biometric which is similar to a fingerprint. A full biometric identification system is developed spanning the automated detection and extraction of the muzzle region through to convolutional neural network (CNN) based classification of individual identity. Application specific challenges are also addressed, notably: (i) obtaining the large and representative cattle biometric identification dataset, (ii) developing and evaluating few-shot learning and metric matching approaches to minimise the number of training images required per individual and (iii) exploring a model updating approach to ensure that the biometric recognition model is robust as herd composition changes due to addition and removal of stock. Three different CNN network architectures (ResNet-50, VGG16 and MobileNetv2) are assessed along with two model training strategies (transfer learning and fine-tuning) and three metric matching assessments (Euclidean, Cosine, BrayCurtis). Model performance was evaluated with a particular emphasis on both performance (accuracy) and suitability for operational applications (smartphone or smart camera based biometric identification). This thesis focuses on three key studies and the collection of a novel dataset to support cattle biometric research.</p> <p>Dataset: There is a lack of publicly available datasets involving cattle muzzle biometrics. Furthermore, those studies conducted in the past were of limited scope and not reflective of many livestock commercial production settings. Specifically, both the diversity of breeds and the numbers of individual animals studied was limited. This thesis collects and publicly releases a large cattle muzzle biometrics dataset obtained under a controlled protocol from a commercial scale facility. The size of the dataset permits both investigations into biometric recognition algorithms and ensures the evaluations are of a sufficiently sized herd (300 individuals) to be relevant in livestock production environments.</p> <p>Study 1: Automated Cattle Biometric Identification System</p> <p>This study is composed of two parts: (i) muzzle detection and extraction and (ii) CNNbased individual cattle identification using muzzle patterns. Both parts comprise an integrated workflow within an automated cattle biometric identification system.</p> <p>The primary purpose of the muzzle detector is to detect the muzzle region-of-interest from frontal face images of cattle. Once the muzzle region is detected its corresponding bounding box coordinates are used to extract the sub-region of the image corresponding to the cattle’s muzzle. A muzzle detector was developed and applied using a YOLOv3 model with three different network resolutions. Results indicated that the higher resolution (1024x1024) network provides better precision detection and achieved 99.13% accuracy on a test dataset.</p> <p>Post muzzle detection a CNN classifier is applied to the muzzle extracts to classify the individual identity of an animal. A ResNet-50 network architecture utilising a transfer learning with fine-tuning strategy was employed to achieve a high classification accuracy (99.1%) on a test data set. Achieving such high muzzle detection and classification accuracies on a large and diverse test set of animals signifies that an automated cattle biometric identification system is feasible and fast becoming a reality.</p> <p>Study 2: Deep Feature Extraction, Metric Distances and Few-shot Learning.</p> <p>Study 1 indicated that a highly accurate biometric identification algorithm was feasible for small-medium scale cattle herd sizes. A major barrier, however, to the implementation of biometric identification systems in livestock production settings is the intensive data requirements for accurate model development. Few-shot learning (≤ 5 images per individual animal) was investigated as an option to reduce the intensive data requirements. Three CNN network architectures were examined (ResNet-50, VGG16, MobileNetv2)</p> <p>Another major challenge when implementing CNN models is that model memory requirements and computational resources required for model training grow greatly as the number of classes increase. This issue poses a problem for the future practical implementation of livestock biometrics on herds of hundreds or potentially thousands of animals. Solutions to these barriers are explored using deep feature learning (via fine-tuned CNN models to extract image features and removal of the CNN classification layer) and distance metric matching (Euclidean, Cosine, and BrayCurtis distances) comparing the similarities between query image CNN features and database image CNN features. The results obtained demonstrate that this procedure can successfully train a CNN muzzle biometric recognition model with high accuracy for both single (ResNet50 1-shot Cosine distance 95.73% Accuracy) and few training images per individual (ResNet50 5-shot Bray-Curtis distance 98.57% Accuracy). The ResNet50 architecture was found to provide the highest overall accuracy (98.57%) but required considerable computational resources and is best suited for cloud server applications but the MobileNetv2 architecture was also found to produce highly competitive and accurate results (97.86%), which is most suitable for smartphone or smart camera (single-board computer options).</p> <p>Study 3: Model Updating and avoiding Catastrophic Forgetting.</p> <p>Another challenge to the implementation of a livestock biometric identification system is the fact that herd structure is dynamic; individual animals enter and leave the herd on a routine basis. The set of individuals used to develop the training model is unlikely to correspond to the herd composition at future dates, so the biometric model has a temporal constraint before it is outdated and unable to identify all of the individual animals in the herd. Excellent results were obtained (ResNet50 3-shot Cosine 95.98% accuracy) demonstrating that the deep feature learning and metric matching strategy is sufficiently robust to the addition of new individuals to the herd. Importantly, the proposed approach can maintain model efficiency and reduce training cost and dependency on previous data whilst also learning new classes.</p> <p>Overall, this study reveals that computer vision can prove highly useful in terms of non-invasive animal identification, with excellent potential to improve livestock welfare. The results are anticipated to contribute to building a practical automatic biometric-based cattle identification system. The deep feature extraction coupled with metric matching strategy was demonstrated to produce excellent results in a manner suitable for this livestock biometric identification application. Finally, this study provides recommendations for future studies, concerning other data types, regions of interest, and classification strategies.</p>