A Systematic Review of Machine Learning Techniques for Cattle Identiﬁcation Datasets Methods and Future Directions Md Ekramul Hossainae Muhammad Ashad Kabirabe Lihong Zhengae Dave L. Swainbce Shawn McGrathbde

2025-04-30 0 0 1.19MB 34 页 10玖币

侵权投诉

A Systematic Review of Machine Learning Techniques for Cattle Identiﬁcation:

Datasets, Methods and Future Directions

Md Ekramul Hossaina,e, Muhammad Ashad Kabira,b,e,∗

, Lihong Zhenga,e, Dave L. Swainb,c,e, Shawn McGrathb,d,e,

Jonathan Medwayb,e

aSchool of Computing, Mathematics and Engineering, Charles Sturt University, Bathurst, NSW 2795, Australia

bGulbali Institute for Agriculture, Water and Environment, Charles Sturt University, Wagga Wagga, NSW, 2678, Australia

cTerraCipher Pty. Ltd., Alton Downs, QLD 4702, Australia

dFred Morley Centre, School of Animal and Veterinary Sciences, Charles Sturt University, Wagga Wagga, NSW 2678, Australia

eFood Agility CRC Ltd, Sydney, NSW 2000, Australia

Abstract

Increased biosecurity and food safety requirements may increase demand for eﬃcient traceability and identiﬁca-

tion systems of livestock in the supply chain. The advanced technologies of machine learning and computer vision

have been applied in precision livestock management, including critical disease detection, vaccination, production

management, tracking, and health monitoring. This paper oﬀers a systematic literature review (SLR) of vision-based

cattle identiﬁcation. More speciﬁcally, this SLR is to identify and analyse the research related to cattle identiﬁcation

using Machine Learning (ML) and Deep Learning (DL). This study retrieved 731 studies from four online scholarly

databases. Fifty-ﬁve articles were subsequently selected and investigated in depth. For the two main applications of

cattle detection and cattle identiﬁcation, all the ML based papers only solve cattle identiﬁcation problems. However,

both detection and identiﬁcation problems were studied in the DL based papers. Based on our survey report, the most

used ML models for cattle identiﬁcation were support vector machine (SVM), k-nearest neighbour (KNN), and arti-

ﬁcial neural network (ANN). Convolutional neural network (CNN), residual network (ResNet), Inception, You Only

Look Once (YOLO), and Faster R-CNN were popular DL models in the selected papers. Among these papers, the

most distinguishing features were the muzzle prints and coat patterns of cattle. Local binary pattern (LBP), speeded

up robust features (SURF), scale-invariant feature transform (SIFT), and Inception or CNN were identiﬁed as the

most used feature extraction methods. This paper details important factors to consider when choosing a technique or

method. We also identiﬁed major challenges in cattle identiﬁcation. There are few publicly available datasets, and

the quality of those datasets are aﬀected by the wild environment and movement while collecting data. The process-

ing time is a critical factor for a real-time cattle identiﬁcation system. Finally, a recommendation is given that more

publicly available benchmark datasets will improve research progress in the future.

Keywords: Cattle identiﬁcation, cattle detection, machine learning, deep learning, cattle farming.

∗Corresponding author: School of Computing, Mathematics and Engineering, Charles Sturt University, Panorama Ave, Bathurst, NSW 2795.

Ph.+61263386259, Email: akabir@csu.edu.au

Email addresses: mdhossain@csu.edu.au (Md Ekramul Hossain), akabir@csu.edu.au (Muhammad Ashad Kabir),

Published in Artiﬁcial Intelligence in Agriculture, vol 6, pp. 138-155, 2022, https: // doi. org/ 10. 1016/ j. aiia. 2022. 09. 002 October 18, 2022

arXiv:2210.09215v1 [cs.CV] 13 Oct 2022

1. Introduction

The demand for eﬃcient traceability and identiﬁcation systems for livestock is growing due to biosecurity and

food safety requirements in the supply chain. The advanced technologies of machine learning and computer vision

have been applied in precision livestock management, including critical disease detection, vaccination, production

management, tracking, health monitoring, and animal well-being monitoring (Awad,2016). ‘Cattle identiﬁcation’

refers to ‘cattle detection’ and ‘cattle recognition’ (Mahmud et al.,2021). Cattle identiﬁcation systems start from

manual identiﬁcation to automatic identiﬁcation with the help of image processing. Traditional cattle identiﬁcation

systems such as ear tagging (Awad,2016), ear notching (Neary and Yager,2002), and electronic devices (Ruiz-Garcia

and Lunadei,2011) have been used for individual identiﬁcation in cattle farming. Disadvantages of these individual

identiﬁcation methods include the possibility of losses, duplication, electronic device malfunctions, and fraud of the

tag number (Rossing,1999;Roberts,2006). These are the issues and challenges for cattle identiﬁcation in livestock

farm management.

With the advent of computer-vision technology, cattle visual features have gained popularity for cattle identiﬁ-

cation (Kusakunniran and Chaiviroonjaroen,2018;Andrew et al.,2016,2017;de Lima Weber et al.,2020). Visual

feature based cattle identiﬁcation systems are used to detect and classify diﬀerent breeds or individuals based on a set

of unique features. In recent years, machine learning (ML) and deep learning (DL) approaches have been widely used

for automatic cattle identiﬁcation using visual features (Andrew et al.,2016;Tharwat et al.,2014b;Andrew et al.,

2019;Qiao et al.,2019;Li et al.,2021b). ML and DL are subﬁelds of artiﬁcial intelligence that can solve complex

problems for automatic decision-making. ML is mainly divided into two approaches, such as supervised learning

and unsupervised learning. The supervised ML approach is deﬁned by its use of labelled datasets, whereas the unsu-

pervised learning uses ML algorithms to analyse and cluster unlabeled datasets. An unsupervised ML approach can

detect hidden patterns in data without human supervision (Janiesch et al.,2021). DL approaches are useful in areas

with large and high-dimensional datasets. Thus, DL models are usually outperformed over traditional ML models in

the area of text, speech, image, video, and audio data processing (LeCun et al.,2015). There are two main steps in the

development of ML and DL models. In the ﬁrst step, a training dataset is used to train the model, and in the second,

the model is validated using a separate validation dataset. Thus, a trained model is created that is later used on the test

dataset to determine its performance based on the test dataset. The dataset used for ML models includes the features

and their corresponding outcomes or labels. The features are extracted from the input data using a feature extraction

method. DL algorithms can automatically extract high-level features from the dataset and learn from these features.

Although the implementation of the ML and DL models is straightforward, there are some challenges with selecting

algorithms, tuning parameters, and features for better prediction accuracy (Janiesch et al.,2021).

Several important review studies have been completed in livestock farm management. Some recent literature re-

lzheng@csu.edu.au (Lihong Zheng), dave.swain@terracipher.com (Dave L. Swain), shmcgrath@csu.edu.au (Shawn McGrath),

jmedway@csu.edu.au (Jonathan Medway)

views have addressed various research challenges in livestock farming, such as identiﬁcation, tracking, and health

monitoring, using tag-based, ML, and DL approaches. Recently, Awad (2016) and Kumar and Singh (2020) reviewed

the literature on using diﬀerent classical and visual biometrics methods for cattle identiﬁcation and tracking. Li et al.

(2021a) reviewed the deep learning-based approaches for classiﬁcation, object detection and segmentation, pose es-

timation, and tracking for diﬀerent kinds of animals such as cattle, pigs, sheep, and poultry. A systematic literature

review based on applying ML and DL approaches in precision livestock farming by Garcia et al. (2020) focused on

grazing and animal health. Qiao et al. (2021) summarised the ML and DL approaches in precision cattle farming

for cattle identiﬁcation, body condition score evaluation, and live weight estimation. They reviewed a small number

of articles (n=13) related to cattle identiﬁcation using ML and DL approaches. Mahmud et al. (2021) conducted a

systematic literature review showing the recent progress of DL applications for cattle identiﬁcation and health mon-

itoring. Their review included only a few articles related to cattle identiﬁcation. Moreover, these review articles

focused on the combination of diﬀerent types of challenges (e.g., tracking, pose estimation, weight estimation, identi-

ﬁcation, and detection) solved by tag-based, ML, and DL methods in precision livestock farming. Thus, they lack in

providing a comprehensive review on cattle identiﬁcation. Also, the existing review articles lack information on ML

and DL applications combined for cattle identiﬁcation as they cover partly either ML or DL for cattle identiﬁcation.

Moreover, the details of the cattle dataset for identiﬁcation are not discussed. In this context, an extensive systematic

literature review is needed, particularly for the challenge of cattle identiﬁcation addressed by ML and DL approaches.

Also, the details of the dataset used in the relevant articles need to be discussed, and the current trend of using ML

and DL techniques in cattle identiﬁcation and future research opportunities with challenges need to be identiﬁed.

This systematic literature review (SLR) aims to summarise and analyse the ML and DL applications used exten-

sively in cattle identiﬁcation. A total of 55 articles for cattle identiﬁcation and detection have been selected for this

SLR. The reviewed articles are ﬁrst summarised, and then the datasets used in the selected articles are discussed. We

then analyse the reviewed articles for trends in using ML and DL approaches for cattle identiﬁcation in recent years

before presenting the feature extraction methods and performance evaluation metrics extracted from the reviewed

articles. Finally, the challenges and future research directions in this ﬁeld are discussed.

2. Methodology

2.1. Review process

The review process of an SLR is divided into three phases – planning, conducting, and reporting the review

(Kitchenham and Charters,2007). In the ﬁrst phase, the research questions for the SLR are identiﬁed. Based on the

research questions, the electronic databases and search terms or keywords were determined. The search keywords

are used to create a search string that is applied to the diﬀerent electronic databases to extract the related articles for

the SLR. This study used the IEEE Xplore, Science Direct, Scopus, and Web of Science databases. These databases

were selected to cover a wide range of studies in our targeted sector as they index most of the journals from various

iii

publishers such as Springer, ACM, Inderscience, Elsevier, Sage, Taylor & Francis, IOS, Wiley, and so on. In the

second phase, the relevant research studies are identiﬁed by searching the databases. After that, the selection criteria

are determined for the quality assessment of the primary studies. The eligible studies are selected by applying the

selection criteria, and then the relevant data are extracted from the selected articles based on the research questions.

In the ﬁnal phase, the extracted data are analysed and used to address the research questions. Then, the results are

reported in the form of tables and ﬁgures followed by a brief discussion of research challenges and future research

opportunities.

2.2. Research questions

This SLR focuses on published research studies into cattle identiﬁcation using ML and DL approaches. The search

process identiﬁes potential primary studies that address the research questions. The answers to the research questions

are discussed based on the data extracted from the selected studies. This study deﬁned the following seven research

questions (RQs) for the SLR.

•RQ1: What ML models are used in cattle identiﬁcation?

•RQ2: What DL models are used in cattle identiﬁcation?

•RQ3: What datasets are used in cattle identiﬁcation?

•RQ4: What feature extraction methods are used in cattle identiﬁcation?

•RQ5: What performance evaluation metrics are used for ML and DL models in cattle identiﬁcation?

•RQ6: What are the best ML and DL models used in a speciﬁc cattle identiﬁcation problem?

•RQ7: What are the challenges in solving cattle identiﬁcation problems?

2.3. Search strategy

A search strategy is applied to keep the search results within the scope of the SLR. In this study, the initial

search was performed using a string with four keywords. The search string was (“cattle” AND “identiﬁcation”)

AND (“machine learning” OR “deep learning”). Some articles were extracted from the search results, and the title,

abstract, and author-speciﬁed keywords were read to ﬁnd the synonyms for the basic search keywords. For “cattle”,

synonyms considered were “cow” and “livestock”. For “identiﬁcation”, synonyms considered were “recognition” and

“detection”. The keywords “neural network”, “image processing” and “vision” were added with “machine learning”

and “deep learning” as similar terms. Thus, the general search string was (“cattle” OR “cow*” OR “livestock”)

AND (“identiﬁcation” OR “recognition” OR “detection”) AND (“machine learning” OR “deep learning” OR “neural

network” OR “image processing” OR “vision”). The search keywords were used for articles in four databases (August

2021). The search strings for the databases are shown in Table 1.

Table 1: Search strings for the selected databases.

Database name Search string

IEEE Xplore ((cattle OR cow* OR livestock) AND (identiﬁcation OR recognition OR de-

tection) AND (“deep learning” OR “machine learning” OR “neural network”

OR “image processing” OR vision)) (anywhere).

Science Direct (cattle OR cow) AND (identiﬁcation OR recognition OR detection) AND

(“deep learning” OR “machine learning” OR “neural network” OR “image pro-

cessing”). It was used to search in the title, abstract and keywords.

Scopus TITLE-ABS-KEY ((“cattle identiﬁcation” OR “cow* identiﬁcation” OR “live-

stock identiﬁcation” OR “cattle recognition” OR “cow* recognition” OR “live-

stock recognition” OR “cattle detection” OR “cow* detection” OR “livestock

detection”) AND (“deep learning” OR “machine learning” OR “neural net-

work” OR “image processing” OR vision)). It was used to search in the title

(TITLE), abstract (ABS) and keywords (KEY).

Web of Science AB=((cattle OR cow* OR livestock) AND (identiﬁcation OR recognition OR

detection) AND (“deep learning” OR “machine learning” OR “neural net-

work”)) OR AK=((cattle OR cow* OR livestock) AND (identiﬁcation OR

recognition OR detection) AND “deep learning” OR“machine learning” OR

“neural network”)) OR TI=((cattle OR cow* OR livestock) AND (identiﬁca-

tion OR recognition OR detection) AND (“deep learning” OR “machine learn-

ing” OR “neural network”)). It was used to search in the title (TI), abstract

(AB) and author keywords (AK).

This study reduced some keywords from the search string for the Science Direct database as the maximum Boolean

connectors (AND/OR) for this database is eight. Since the Scopus database yielded many articles with the general

search string, the search results were reduced by putting two diﬀerent keywords together. In this SLR, we did not

limit the publication year during the search. After performing the above search strings, a total of 731 articles were

retrieved.

2.4. Study selection criteria

The selection criteria are used to identify the studies that can answer the research questions. In this study, inclusion

and exclusion criteria were deﬁned based on the research questions. The search results from all databases were

recorded on a spreadsheet for scrutiny using the inclusion and exclusion criteria. A study was selected for the SLR

when the inclusion criteria were true but the exclusion criteria were false. The exclusion criteria were: (i) publication

is not related to ML or DL for cattle identiﬁcation, (ii) publication is a survey or review paper, (iii) publication is

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

ASystematicReviewofMachineLearningTechniquesforCattleIdentication:Datasets,MethodsandFutureDirectionsMdEkramulHossaina,e,MuhammadAshadKabira,b,e,,LihongZhenga,e,DaveL.Swainb,c,e,ShawnMcGrathb,d,e,JonathanMedwayb,eaSchoolofComputing,MathematicsandEngineering,CharlesSturtUniversity,Bathurst,NSW2795,...

展开>> 收起<<

A Systematic Review of Machine Learning Techniques for Cattle Identiﬁcation Datasets Methods and Future Directions Md Ekramul Hossainae Muhammad Ashad Kabirabe Lihong Zhengae Dave L. Swainbce Shawn McGrathbde.pdf

共34页,预览5页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

A Systematic Review of Machine Learning Techniques for Cattle Identiﬁcation Datasets Methods and Future Directions Md Ekramul Hossainae Muhammad Ashad Kabirabe Lihong Zhengae Dave L. Swainbce Shawn McGrathbde

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: