Indonesian J our nal of Electrical Engineering and Computer Science V ol. 32, No. 1, April 2025, pp. 555 568 ISSN: 2502-4752, DOI: 10.11591/ijeecs.v32.i1.pp555-568 555 HorseNet: a no v el deep lear ning appr oach f or horse health classication Nesrine Atitallah 1 , Ahmed Abdel-W ahab 1 , Anas A. Hadi 1 , Hussein Abdel-J aber 1 , Ali W agdy Mohamed 2 , Mohamed Elsersy 3 , Y usuf Mansour 1 1 F aculty of Computer Studies, Arab Open Uni v ersity , Riyadh, Saudi Arabia 2 Department of Operations Research, F aculty of Graduate Studies for Statistical Research, Cairo Uni v ersity , Giza, Egypt 3 Department of Computer Information Systems, Higher Colle ges of T echnology , Al Ain, Ab u Dhabi, United Arab Emirates Article Inf o Article history: Recei v ed Jan 28, 2024 Re vised Oct 15, 2024 Accepted Oct 30, 2024 K eyw ords: Con v olutional neural netw orks Deep learning Horse wellness classication Inception T ransfer learning V GG16 ABSTRA CT In equestrian sports and v eterinary medicine, horse welf are is paramount. Horse tiredness, lameness, col ic, and anemia can be identied and classied using deep learning (DL) models. These technologies analyze horse images and videos to help v ets and researchers nd symptoms and trends that are hard to see . Early detection and better treatment of certain disorders can impro v e horses’ health. DL models can also impro v e with ne w data, impro ving diagnosis accurac y and ef cienc y . This study comprehensi v ely e v aluates three con v olutional neural net- w ork (CNN) models to distinguish normal and abnormal horses using the gen- erated horse dataset. F or this study , a unique dataset of horse breeds and their normal and abnormal states w as collected. The dataset includes mobility pat- terns from this study’ s initial data collection. DL models lik e CNNs and trans- fer learning (TL) models (visual geometry group (V GG)16, InceptionV3) were emplo yed for cate gorization. T he InceptionV3 model outperformed CNN and V GG16 with o v er 97% accurac y . Its depth a nd multi-le v el structure allo w the InceptionV3 model to recognize characteristics in images of v aried scales and comple xities, e xplaining its e xcellent performance. This is an open access article under the CC BY -SA license . Corresponding A uthor: Ahmed Abdel-W ahab F aculty of Computer Studies, Arab Open Uni v ersity Riyadh, Saudi Arabia Email: a.rakha@arabou.edu.sa 1. INTR ODUCTION The surv eillance of animal health, specically i n horses, is of ut most importance in guaranteeing the ir welf are, producti vity , and economic signicance in both professional and leisure conte xts [1]. Con v entional approaches to monitor horse health often include labor -intensi v e, subje cti v e, and time-consuming manual ob- serv ation and assessment conducted by v eterinarians or care gi v ers [2]. There are numerous inherent limitations associated with traditional techniques for monitoring the health of horses. Initially , it is important to note that the process of manually observing and e v aluating the health of horses by v eterinarians or caretak ers is often characterized by a signicant amount of labor and time e xpenditure [3]. Ho we v er , shifts in the e xpertise and background of those caring for or treating the horse can af fect the reliability of these e v aluations. This can lead to inconsistencies when judging the animal’ s o v erall well-being. This may lead to inconsistencies when judg- ing the animal’ s o v erall well-being. Furthermore, it is important to note that these methodologies may e xhibit limitations in identifying nuanced alterations in indi viduals’ health statuses, including the rst indications of J ournal homepage: http://ijeecs.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
556 ISSN: 2502-4752 ailments or injuries. If left undetected and untreated in a timely manner , these illnesses ha v e the potential to deteriorate progressi v ely . These issues ha v e the potential to undermine the ef cac y of con v entional procedures, hence emphasizing the need for ne w methodologies that may ef fecti v ely tackle these limits and enhance the precision and ef cienc y of horse health monitoring [4]. The use of cutting-edge technology , such as articial intelligence (AI), presents no v el prospects for the automation and enhancement of animal health monitoring, hence enabling a more complete and non-intrusi v e approach to e v aluate animal well-being. Through the adop- tion of no v el approaches, it is possible to augment the well-being, ef cac y , and economic w orth of horses, guaranteeing their sustained prosperity in both leisure and occupational conte xts [5]. These technologies ha v e the capability to automate and impro v e the precision of animal health mon- itoring, therefore of fering a more complete and non-intrusi v e approach to e v aluating animal well-being. An illustration of the use of wearable sensor s is their utilization in the continuous monitoring of essential ph ysio- logical indicators, including heart rate, breathing rate, and temperature. These sensors posses s the capabi lity to communicate data in real-time to a system based on cloud computing [6]. This system f acilitates the analysis of the data via the use of machine learning (ML) algorithms, thereby enabling the identication of the rst indications of disease or damage. In a similar v ein, computer vision algorithms may be used to e xamine video recordings of horses to identify alterations in their g ait or beha vior , which may serv e as potential indicators of underlying health conditions [7]. These adv anced technologies can impro v e horse health monitoring accurac y , ef cienc y , and ef cac y . T echnology can increase ani mal well being and benet horse o wners, care gi v ers, and v ets. The y can pro vide timely and accurate horse health information to help them mak e informed horse care and treatment decisions [8]. Deep learning (DL) has the potential to profoundly transform the eld of animal health monitoring. W ith e xtensi v e datasets, DL algorithms ha v e the capacity to acquire kno wledge about patterns and then pro- vide precis e forecasts pertaining to the well-being of animals [9]. This technol o gi cal adv ancement has se v eral adv antages in comparison to con v entional approaches for monitoring animal well-being, including enhanced precision, automation, and timely identication of health concerns. The use of DL techniques enables the continuous and non-in v asi v e monitoring of animal health, hence f acilitating a more thorough e v aluation of an animal’ s o v erall health condition. Furthermore, DL has the capability to automate the process of monitor - ing, therefore mitig ating the labor -intensi v e charact eristics associated with con v entional monitoring techniques [10]. The timely identication of health concerns is of utmost importance in mitig ating the progression of se v ere medical ailments. In this re g ard, the use of DL algorithms may play a pi v otal role in promptly detecting small alterations in an animal’ s health, hence f acilitating timely interv ention and treatment. These adv antages illustrate the capacity of DL to signicantly transform the eld of animal health monitoring and enhance the welf are of animals [11]. T ransfer learning (TL) is a method of reusing a model that has already been trained on one task to solv e a ne w , related task. This can sa v e a lot of time and ef fort, and often leads to better results than training a ne w model from scratch, especially when there is limited data a v ailable. In f act, se v eral w orks used TL to classify images. Noor et al. [12] proposed a dataset of sheep f acial images and a frame w ork that le v erages TL and ne-tuning to automatically dif ferentiate between images of sheep f aces sho wing pain and those that appear normal. In addition, T ammina [ 1 3] emplo yed the V GG-16 model, which is a pre-trained deep CNN, to classify images. This study presents an inno v ati v e methodology for the identication of horse health status via the use of DL methodologies. This study introduces se v eral k e y inno v ations in horse health monitoring: i) HorseSet 1.0: a comprehensi v e dataset of horse wellness images, capturing di v erse situations, weather conditions, lighting, and angles. ii) Expert-v alidated data: annotated and re vie wed by v eterinary specialists with o v er a decade of e xperience, ensuring reliability and accurac y . iii) HorseNet: an ef cient DL approach for rapid and accurate detection and classication of horse wellness. i v) Performance e v aluation: implementation and assessment of multiple DL algori thms (V GG16, Incep- tionV3, proposed CNN) using the no v el dataset. This w ork is structured in the follo wing manner: section 2 pro vided a thorough e xamination of the current body of research pertaining to the use of DL algorithms in horse monitoring. Section 3 of the document pro vides a comprehensi v e e xposition of the proposed model, including a thorough depiction of the dataset and the DL methodologies used within it. Section 4 presents the results of the study and ho w the y were found. It includes an y statistical analyses or pictures of the data, lik e the F1-score, accurac y , confusion matrix, loss, precision, recall, and loss. Section 5 encompasses an analysis and interpretation of the obtained results. This Indonesian J Elec Eng & Comp Sci, V ol. 32, No. 1, April 2025: 555–568 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 557 includes a comprehensi v e discussion on the implications of the ndings for the proposed model as well as a thorough performance study of the model’ s outcomes. Additionally , this part pro vides a concise comparison of the DL technologies included in the proposed model. The last part presents a concise o v ervie w of the primary disco v eries and ramications of the in v estig ation. Furthermore, this part pro vides an analysis of potential a v enues for further study . 2. LITERA TURE REVIEW AND THEORETICAL FRAMEW ORK 2.1. Literatur e r e view The application of ML and DL in animal health monitoring has g ained momentum, with studies focus- ing on v arious species, including horses. F or instance, P allathadka et al. [14] demonstrated the ef fecti v eness of DL in identifying lameness in horses through video analysis, achie ving a 91% accurac y , outperforming tradi- tional methods. DL has also been applied to detect colic in horses using data from wearable sensors, sho wing 90% precision [15], and has pro vided insights into horse beha vior , tness, and athletic abilities [16]. In pig health monitoring, Ka vlak et al. [17] emplo yed ML (e xtreme gradient boosting (XG-boost)) to predict health issues by analyzing feeding beha viors. Although the approach sho wed good sensiti vity and specicity , it f aced challenges due to unbalanced data and lo w symptom pre v alence, emphasizing the need for better data quality and swine management strate gies [17]. Similarly , TL, which a dapts pre-trained models for ne w ta sks, has been e xplored for wildlife monitoring by Nguyen et al. [18]. Their DL-based system sho wed promise in identifying animals in images, e v en with imbalanced datasets. The study suggests further enhancements through augmented datasets and adv anced CNN models. Thermal image processing combined with ML has been used to detect a vian diseases. Sade ghi et al. [19] achie v ed high accurac y using support v ector machines (SVM) and articial neural netw orks (ANN) to classify diseases lik e ne wcastle disease and a vian inuenza in broilers. After optimization, the SVM with Dempster–Shafer e vidence theory outperformed ANN, with o v er 97% accurac y in disease classication [19]. Another study by Quaderi et al. [20] applied DL to detect beehi v e sounds on their o wn datasets and using v arious methods to reduce features. Sequenti al neural netw orks with AdaMax and sigmoid acti v ation functions performed well, outperforming other methods. Random forests were also ef fecti v e. When combining dif ferent types of data, sequential neural netw orks ag ain sho wed the best results. Recurrent neural netw orks were particularly good at distinguishing bee sounds from noise [20]. DL has also been inte grated with internet of things (IoT) technology for animal care and surv eillance. P atil and Ansari [21] de v eloped an intelligent system using CNNs and recurrent neural netw orks (RNNs) to monitor animal health, recognize acti vities, and detect en vironmental anomalies. The study highlights t h e potential of DL to impro v e animal welf are through proacti v e care. Sreede vi and Anitha [22] focused on using DL for wildlife detection vi a images and videos, sho wing that CNN-based approaches are ef fecti v e for identifying dif ferent wild animal species, with practical implications for wildlife conserv ation. The welf are of animals has become a critical concern, leading to the de v elopment of wearable health monitoring de vices using IoT . These de vices collect vital signs such as body temperature, heart rate, and respiration rate, which can be transmitted to v eterinary professionals for timely interv ention. The cattle industry , in particular , could benet from such systems, which allo w continuous mon- itoring of indi vidual animals’ welf are. A study described a telemonitoring system prototype using wearable technology to enhance decision-making and impro v e horse welf are. Digital tools ha v e the potential to enhance equine health monitoring by impro ving the accurac y and ef cienc y of health assessments [23]. While there ha v e been signicant adv ances in using DL for equine heal th monitoring, there are still areas that need impro v ement. Research has primarily focused on detecting lameness and colic in horses, with limited studies e xploring respi- ratory and metabolic issues. DL requires lar ge, high-quality datasets, making data acquisition and standardiza- tion challenging. Additionally , user -friendly tools are needed to inte grate DL into health monitoring practices. Despite these challenges, DL of fers promising opportunities to enhance horse welf are, performance, and eco- nomic v alue. Future research should address these g aps and de v elop practical tools for equine health monitoring [24]. T o g ain a comprehensi v e understanding of the e xisting approaches for horse health classication, we con- ducted a detailed o v ervie w of the v arious methodologies emplo yed in prior studies as illustrated in T able 1. This re vie w pro vides an o v ervie w of v arious methodologies used in horse health cl assication, highlighting dataset characteristics, classication algorithms, and performance metrics. The analysis serv es as a v alu- able resource for researchers and identies g aps and opportunities for further impro v ement in horse health monitoring. Hor seNet: a no vel deep learning appr oac h for hor se health classication (Nesrine Atitallah) Evaluation Warning : The document was created with Spire.PDF for Python.
558 ISSN: 2502-4752 T able 1. Comparison of related w orks Reference Y ear Domain Algorithm Problem to Performance No. Used be solv ed [14] 2023 Identication of DL Lameness identication Accurac y of 91% horse lameness of horse [15] 2023 Analysis from wearable DL Detect of Precision 90% sensors to detect colic the CO VID-19 virus issue in horse from wearable sensors [16] 2023 The heal th and CNN Monitoring the horses DeepLabCut DLC tness of horses using g ait analysis V er2.2 tool used [17] 2023 Enhanced da ta XGBoost Data quality Animal diseases quality enhancement management [18] 2017 Automated wi ldlife CNN Detection and identication W ildlife spotter monitoring TL of the classied species image dataset, species accurac y of 96.6% [19] 2023 Classication of SVM with Ef cac y of thermograph y and Accurac y of o v er ne wcastle disease Dempster–Shafer ML techniques in the 97% e vidence theory classication of ne wcastle disease and a vian inuenza among broilers [20] 2022 Beehi v e sound CNN, RNN Classify bee sounds from AdaMax optimizer , analysis the non-beehi v e noises accurac y of 85% [21] 2020 Smart surv eillance CNN, RNN Monitor stray dog Accurac y of 85-90 % of stray animals dog animals in a particular area [22] 2022 W ild animal CNN, rectied W ildlife animal detection Accurac y of 87.8% classication linear unit (ReLU) IW ildCam dataset [24] 2023 Inte grat ion of DL Y OLO v7 A h ybrid technique of Accurac y of 96.2% monitoring for horse automated muzzle feature health surv eillance e xtraction empo wered by Y olo and SIFT , and feature matching using FLANN 2.2. T ransfer lear ning con v olutional neural netw orks TL in v olv es le v eraging a model pre-trained on one problem to solv e a dif ferent problem. This ap- proach is g aining traction in the realm of deep neural netw orks, gi v en their substantial data and computational demands. It allo ws for a more ef cient training process by rening a pre viously trained DL model for a similar task. Essentially , it harnesses the insights a model has g arnered from a data-rich task and applies it to a ne w task with limited data. T o circumv ent the challenges of training duration and data v olume, we emplo yed TL on tw o distinct pre-trained CNN models: V GG16, pioneered by Simon yan and Zisserman [25], and InceptionV3, crafted by Google’ s research team. 2.2.1. V GG16 The V GG16 model, de v eloped by the visual geometry group at Oxford Uni v ersity , is a reno wned neu- ral netw ork architecture for image classication. Its simple yet ef cient design comprises 16 layers (13 con- v olutional and 3 fully connected), enabling the e xtraction of hierarchical features from images. The model’ s depth allo ws it to learn comple x i mage representations, making it v ersatile for v arious visual recognition tasks. V GG16’ s straightforw ard structure and high performance ha v e made it inuential in adv ancing DL and com- puter vision, cementing its popularity among researchers and practitioners in the eld [26]. 2.2.2. InceptionV3 InceptionV3, de v eloped by Google, is a sophisticated CNN architecture designed for lar ge-scale image recognition and classication. Its inno v ati v e desi gn features inception modules that capture information at multiple scales, striking an optimal balance between depth and computational ef cienc y . This approach enables high accurac y in image classication tasks while maintaining scalability for lar ge datasets. InceptionV3’ s v ersatility has l ed to its widespread adoption across v arious domains, consistently ac hie ving top-tier results in benchmark competitions. Its combination of performance and ef cienc y mak es it a popular choice for comple x image recognition tasks, particularly when processing e xtensi v e datasets [27]. Due to its impressi v e performance and ef cienc y , t he InceptionV3 model has become a reference architecture in the eld of DL and image classication. It has serv ed as a foundation for subsequent adv ancements in CNN architecture and Indonesian J Elec Eng & Comp Sci, V ol. 32, No. 1, April 2025: 555–568 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 559 has inuenced the de v elopment of ne wer models. Researchers and practitioners alik e continue to le v erage the capabilities of InceptionV3 to tackle comple x image recognition challenges and push the boundaries of computer vision applications. 3. METHODS AND MA TERIALS 3.1. Dataset In order to ef fecti v ely classify and detect the well-being of horses using camera and video foota ge, we utilized adv anced computer vision techniques and DL algorithms. This inno v ati v e approach a llo wed us to de v elop a system capable of analyzing visual data and e xtracting meaningful insights related to t he health and condition of horses. In the upcoming section, we will introduce and elaborate on our meticulously curated dataset, kno wn as HorseNet. This dataset plays a crucial role in our research as it serv es as the foundation for training and e v aluating the performance of our models. The subsequent discussion will encompass a detailed description of the dataset, including its composition, size, and the specic attrib utes and features captured within. By shedding light on the intricacies of HorseNet, we aim to pro vide a comprehensi v e understanding of the underlying data that po wers our horse wellness classication and detection system. 3.1.1. Study ar ea The research study w as conducted at the esteemed horse club located in Ahsa, a re gion kno wn for i ts rich equestrian culture. During the initial stages of the research, the authors encountered a challenge in nding a suitable dataset that adequately represented the di v erse range of horses and conditions rele v ant to their study . T o o v ercome this hurdle, the authors took the initiati v e to g ather and curate the necessary data themselv es. The y meticulously collected a comprehensi v e set of images featuring horses, ensuring that it encompassed a wide spectrum of breeds, ages, and ph ysical attrib utes. These images wer e specically sourced from the horse club situated in the capti v ating Ahsa re gion of Saudi Arabia, which boast s a notable reputation for its dedication to the equestrian arts. By acti v ely eng aging with the horse club and obtaining their cooperation, the authors were able to acquire a di v erse and representati v e dataset that formed the foundation of their research. 3.1.2. Data collection Our research endea v ors led us to amass a v ast and comprehensi v e collection of horse images and videos, encompassing a di v erse range of breeds, ages, and v arying states of health conditions. T o accomplish this, we es tablished a fruitful partnership with kno wledgeable members of the esteemed horse club . W orking in collaboration, we utilized smartphones equipped with high-quality cameras to capture the visual data. The con v enience and portability of these de vices allo wed us to ef ciently document the horses in thei r natural en vi- ronments, ensuring the authenticity and representati v es of the collected media. Figure 1 illustrates the process, sho wcasing the use of smartphones in capturing the images and videos that constitute our e xtensi v e dataset. Through this diligent and collaborati v e ef fort, we were able to create a rich and di v erse resource that forms the backbone of our research on horse wellness. Figure 1. Horse wellness dataset samples Hor seNet: a no vel deep learning appr oac h for hor se health classication (Nesrine Atitallah) Evaluation Warning : The document was created with Spire.PDF for Python.
560 ISSN: 2502-4752 The pictures were stored as PNG les in multiple resolutions. Notably , the primary focus of the dataset is on images depicting horses in di v erse settings. These photos were g athered from January to April 2023 and were inuenced by dif ferent weather and illumination conditions. The y were also shot from a range of perspecti v es. Out of the entire collection, onl y 1,218 pictures met the criteria for the research. The rest, which were either out of focus or didn’ t portray the necessary scenarios in v olving horses, were e xcluded. 3.1.3. Horse wellness classications DL techniques, specically CNNs, ha v e sho wn grea t potential in v arious image recognition tasks. While s ome aspects of f atigue, lameness, colic, and anemia in horses can potentially be detected usi n g images and DL. In the follo wing, there are some aspects that might be detected using images and DL: i) Posture and g ait analysis: DL algorithms can analyze images or videos of a horse’ s posture and g ait to detect abnormalities, such as limping, sti f fness, or une v en weight distrib ution. This analysis can help classify the se v erity and type of lameness. ii) Ph ysical appear ance: f atigue, anemia, and colic can cause changes in a horse’ s ph ysical appearance that may be detectable using DL. F or e xample, an anemic horse may e xhibit pale mucous membranes (gums), and a horse with colic may sho w signs of discomfort or abdominal distension. DL models can be trained to recognize these features from images and identify horses that may need further e xamination. iii) Beha vioral patterns: horses suf fering from f atigue, lameness, colic, or anemia may e xhibit abnormal be- ha viors such as restlessness, rolling, or frequent changes in position. By analyzing a series of images or videos, DL models may be able to detect these beha vior patterns and help in identifying af fected horses. T able 2 presents a comprehensi v e cate gorization of horse wellness, outlining six distinct classes of horse be- ha vior and health. It ranges from the ”Normal Horse” cate gory , which comprises the majority with 178 horses, to more specic beha viors lik e ”Rolling Horse” and ”Stretching Horse”. T able 2. Horse wellness cate gories W ellness class Class name Number Normal horse 0 478 Rolling horse 1 259 Stretching horse 2 21 P a wing 3 358 L ying do wn 4 38 Biting at sides 5 62 Kicking belly 6 2 3.1.4. Data distrib ution In T able 3, we present a comprehensi v e o v ervie w of the horse wellness datasets used in our study . Upon analyzing these datasets, we observ ed an inherent imbalance in the distrib ution of instances across the v arious horse wellness cate gories. This imbalance posed a challenge as it could potentially bias the performance of our DL algorithms during training. T o address this issue and promote f airness in our model training, we em- plo yed image enhancement methods. These techniques allo wed us to manipulate and augment the dataset, ensuring a more balanced distr ib ution of instances across the dif ferent horse wellness cate gories. By equal- izing the representation of each cate gory , we aimed to enhance the algorithm’ s ability to learn and generalize patterns associated with the full spectrum of horse wellness. Through the application of these image enhance- ment methods, we aimed to mitig ate an y potential biases that could arise due to the imbalanced nature of the original dataset, ultimately fostering more accurate and reliable results. T able 3. Distrib ution in our dataset No. Class name Number of samples 1 Normal horse 478 2 Abnormal horse 740 T otal 1,218 Indonesian J Elec Eng & Comp Sci, V ol. 32, No. 1, April 2025: 555–568 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 561 3.2. Methods The proposed detection and classication frame w ork entails a well-dened w orko w comprising v e k e y steps: g athering data, rening, and augmenting this data, emplo ying v arious CNN models, and conducting e xperiments and assessments as depicted in Figure 2. These steps collecti v ely contrib ute to the accurac y and ef fecti v eness of the frame w ork. Figure 2. The e xperimentation w orko w 3.2.1. Data pr e-pr ocessing A critical phase in an y computer vision (CV) system in v olv es prepossessing the images. Initi ally , e v ery image w as resized to dimensions of 300×300 to ensure square-shaped images and consistenc y throughout the dataset. Subsequently , these images were altered to match the input dimensions of v arious models. F or the custom designed HorseNet, the images were adjusted to 256×256 pix els. F or V GG16 and traditional CNN models, the y were changed to 224×224 pix els. Meanwhile, the input for InceptionV3 w as resized to 71×71 pix els. 3.2.2. Data augmentation DL models typicall y need a v ast am o unt of data for opti mal performance. When there’ s limited t rain- ing data a v ailable, image augmentation is often used to bolster the rob ustness o f image classiers. Image augmentation articially produces training samples using techniques lik e rotation, noise addition, shifting, mir - roring, and blurring. From the initial dataset of 1,218 images, after augmentation, it w as split: 90% w as used for training and v alidation, and 10% for testing, as illustrated in T able 4. In this study , we emplo yed four augmentation methods: rotation, ipping, zoom, and brightness adjustments. Rotating images is a common technique, e xpanding the dataset by producing v ariants of the original images rotating an ywhere from 0 to 360 de grees. Flipping, on the other hand, can be seen as a subset of rotation, creating mirrored v ersions of the original. T able 4. Distrib ution of images dataset after data augmentation Normal Abnormal T otal T raining data 6,920 9,740 16,658 V alidation data 2,488 2,170 4,658 T esting data 1,485 1,476 2,961 T otal 10,893 13,386 24,277 Hor seNet: a no vel deep learning appr oac h for hor se health classication (Nesrine Atitallah) Evaluation Warning : The document was created with Spire.PDF for Python.
562 ISSN: 2502-4752 3.3. Model design 3.3.1. Pr oposed CNN This subsection tackled the architecture of the proposed traditional CNN designed for the clas sica- tion of horse wellness. The model is composed of six main components: 10 con v olutional layers paired with 4 pooling layers, tw o fully connected layers, and a ReLU acti v ation layer . Dif ferent lters were utilized in each con v olution layer to e xtract v aried features [28]. T o counteract o v er tting, dropout w as inte grated as a re gularization technique within both the max-pooling and the fully connected layers. The input images had dimensions of 32×32×3. By choosing a batch size of 64 and a learning rate of 0.0001, the training speed w as optimized. T w o deep layers were established: the inaugural layer took an input channel of one, featuring a 3×3 k ernel, a stride of one, and a padding v alue of tw o. Post-con v olution, the image dimensions shrank; ho w- e v er , padding w as adjusted to zero to retain the original size. The ReLU function w as f a v ored as the model’ s acti v ation due to its resilience ag ainst saturation and its gradient performance relati v e to other acti v ation. The max-pooling layer emplo yed both a k ernel and stride of tw o. F or the subsequent con v olutional layer , the input dimension w as 32, output w as 64, using a 5x5 k ernel, a stride of one, and padding of tw o. The ReLU and max pooling remained consistent across both layers. A dropout layer w as inte grated to alle viate and mitig ate o v er tting. Conclusi v ely , tw o fully connected layers ensured the interlinking of all neurons. The CNN’ s blueprint can be visualized in Figure 3. Figure 3. Architecture of the proposed CNN 3.3.2. Pr oposed appr oach In this research, we utilized three distinct models: a custom-b uilt CNN, V GG-16, and InceptionV3. W e trained our models using augmented data from the primary dataset, as detailed in section 3.1.. Post- augmentation, the e ntire dataset, consisting of 24,277 images, w as split into 16,658 training, 4,658 v alidation, and 2,961 test images, as illustrated in T able 4. W e adopted TL method to address the challenges of e xtended training times and limited data a v ailabil ity . Moreo v er , the top layers, along with the gully connected (FC) layers added to the tail-end of the pre-e xisting models, were set to a static state, and subsequently retrained on our specic dataset to achie v e the tar geted outcomes. In our methodology , we tha wed the concluding layers of these established models, then retrained them using our dataset, while k eeping the preliminary l ayers frozen. This process is visualized in Figure 4. 4. RESUL TS AND DISCUSSION In this section, we will pro vide a comprehensi v e summary of the results obtained from the e xperi- ments conducted o n the dataset pertaining to horse well-being. These e xperiments were conducted with the aim of utilizing adv anced CV techniques and DL algorithms to g ain insights into the health and condition of horses. The collected dataset, comprising a di v erse collection of horse images and videos, formed the basis for the e xperiments. Using this dataset, we trained and e v aluated se v eral CNN models specically designed for horse well-being analysis. These models were selected based on their established performance in image classication tasks and their suitability for the domain of horse well-being. Through rigorous e xperimentation and analysis, we obtained v aluable ndings re g arding the classication and detection of horse well-being using Dl techniques. The results shed light on the ef cac y of the emplo yed CNN models in accurately assessing v ar - ious aspects of horse well-being, such as o v erall health, body condition, and an y potential signs of discomfort or distress. Furthermore, in order to pro vide a comprehensi v e analysis, we compared the performance of the dif ferent CNN models used in the e xperiments. This comparati v e analysis allo wed us to identify the strengths and weaknesses of each model, enabling us to mak e informed decisions re g arding their suitability for specic Indonesian J Elec Eng & Comp Sci, V ol. 32, No. 1, April 2025: 555–568 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 563 horse well-being assessment tasks. By presenting these summarized results and conducting a thorough analy- sis, we aim to contrib ute to the e xisting body of kno wledge utilizing CV and DL for horse well-being analysis. The ndings obtained from these e xperiments ha v e the potential to adv ance the understanding of horse health assessment, potentially leading to impro v ed care and well-being for these magnicent animals. Figure 4. Proposed approach for horse wellness detection and binary-classication 4.1. Experimental setup In our research, we emplo yed the Jupiter notebook to script the entire w orko w using Python 3.8. T o b uild the neural netw orks, we incorporated both the K eras library and T ensorFlo w as the back end. Furthermore, OpenCV f acilitated data loading and pre-processing, while S ci-Kit Learn w as instrumental in generating classi- cation sum maries. F or e xpedited computational performance, we utilized the Nvidia GeF orce MX 250 GPU, supplemented by CUD A and cuDNN libraries. cuDNN is a GPU-enhanced library tailored to boost v arious DL frame w orks. The system underpinning our w ork had these specications: a 64-bit processor , an Intel Core i7-8565U CPU clocking at 1.80 GHz, and 16 GB of RAM. It operated on a W indo ws 10 platform equipped with NVIDIA GeF orce MX. 4.2. Experimental r esults The main objecti v e of the suggested approach is to accurately identify and cate gorize the health of horses. T o achie v e this, we utilized three distinct CNN models. T able 5 presents the performance metrics of the CNN models post ne-tuning and t raining on the horse wellness dataset. Among the models, InceptionV3 stands out with commendable accurac y , precision, and recall rate, each at 97%. This is trailed by V GG16 and subsequently the t raditional CNN. Ov erall, the models grounded in TL sho wcased rob ust performance, boast- ing an accurac y surpassing 95%. InceptionV3 emer ges as the top performer with all metrics at or near 97%. This indicates not only high accurac y b ut also a balanced capability in precision and recall, suggesting a fe w f alse positi v es and f alse ne g ati v es. This balance is crucial in medical or wellness conte xts where both types of errors carry signicant consequences. V GG16 follo ws closely , demonstrating that while slightly less ef fecti v e than InceptionV3, it still pro vides a highly reliable met hod for classifying horse health, with metrics around 95%. Proposed CNN, while trailing behind the other tw o models, still sho ws respectable performance metrics (90%). This indicates a viable option when computational resources are limited or for preliminary e xplorations. Hor seNet: a no vel deep learning appr oac h for hor se health classication (Nesrine Atitallah) Evaluation Warning : The document was created with Spire.PDF for Python.
564 ISSN: 2502-4752 T able 5. Performance metrics of the deplo yed TL-based models and the proposed CNN Model Accurac y Precision Recall F1 score Proposed CNN 90.20% 90% 90% 90% V GG16 96.18% 96% 96% 96% InceptionV3 96.92% 97% 97% 97% Figure 5 sho ws the plots of confusion matrices for each horse wellness class produced by the pro- posed CNN, V GG16, and InceptionV3 models. Figure 5(a) represents the model with the lo west performance, demonstrating greater dif culty in distinguishing between normal and abnormal conditions compared to the others. Figure 5(b) sho ws mark ed impro v ement in both sensiti vity and specicity , with a balanced decrease in both f alse positi v es and f alse ne g ati v es, suggesting better o v erall performance. Figure 5(c) stands out as the best performer , with the highest true positi v e and true ne g ati v e rates, coupled with the lo west f alse positi v e and f alse ne g ati v e rates, indicating a highly accurate model. The progression observ ed is consistent with the general e xpectation that more adv anced models with more parameters and sophisticated ar chitectures, often utilizing TL, will generally outperform simpler models, particularly on comple x tasks lik e image classication. (a) (b) (c) Figure 5. The confusion matrix of: (a) the proposed CNN model, (b) the pre-trained V GG16, and (c) the pre-trained InceptionV3 approach The analysis of curv es depicted in Figures 6, which includes the smoothed training curv es and v alida- tion loss and accurac y curv es of the proposed CNN model in Figure 6(a), the ore-trained V GG16 approach in Figure 6(b), and the pre-trained InceptionV3 approach in Figure 6(c). This Figure suggests that the pre-trained InceptionV3 model is the best performer in terms of learning ef fecti v ely and general izing from the training data to the v alidation data. It also manages to achie v e high accurac y while a v oiding o v er -tting, making it the most suitable model for deplo yment in real-w orld scenarios where model rob ustness and reliability are crucial. Indonesian J Elec Eng & Comp Sci, V ol. 32, No. 1, April 2025: 555–568 Evaluation Warning : The document was created with Spire.PDF for Python.