Klasifikasi Individu Sapi Berbasis ResNet50 Dan ArcFace

Thariq, Ivan Anendar (2026) Klasifikasi Individu Sapi Berbasis ResNet50 Dan ArcFace. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221013-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Identifikasi individu sapi penting dalam manajemen peternakan karena mendukung pemantauan kesehatan, pencatatan riwayat produksi, dan pengelolaan populasi ternak secara akurat. Metode konvensional seperti ear tag, tato, dan radio frequency identification (RFID) masih memiliki keterbatasan berupa risiko kehilangan, biaya implementasi, dan potensi stres pada hewan. Penelitian ini mengusulkan pendekatan berbasis citra untuk klasifikasi identitas individu sapi menggunakan model deep learning. Model yang diuji memanfaatkan ResNet50 dan IResNet50 sebagai backbone, serta Softmax dan ArcFace sebagai classification head. Dataset publik citra sapi diperiksa, divalidasi, dibagi secara stratified, dipraproses, dan diuji melalui lima skenario, yaitu dataset tidak seimbang, dataset terkontrol, offline augmentation pada kelas minoritas, studi ablasi kombinasi backbone dan classification head, serta analisis representasi embedding. Evaluasi dilakukan menggunakan top-1 accuracy, top-5 accuracy, macro precision, macro recall, macro F1-score, dan analisis threshold. Analisis embedding dilakukan melalui PCA, ray plot, within-cluster variance, dan silhouette coefficient untuk mengamati kekompakan fitur dalam kelas serta keterpisahan fitur antarkelas. Hasil menunjukkan bahwa IResNet50 + ArcFace memberikan performa terbaik berdasarkan macro F1-score dan top-1 accuracy. Pada dataset tidak seimbang, model ini mencapai macro F1-score 98,6789% dan top-1 accuracy 99,0627%. Analisis embedding juga menunjukkan struktur fitur yang lebih kompak dan lebih terpisah dibandingkan ResNet50 + Softmax. Secara keseluruhan, performa klasifikasi dipengaruhi oleh arsitektur model, fungsi classification head, kualitas distribusi data, dan representasi fitur yang dibentuk.
==============================================================================================================================
Individual cattle identification is important for livestock management because it supports health monitoring, production record tracking, and accurate herd population management. Conventional methods such as ear tags, tattoos, and radio frequency identification (RFID) still have limitations, including loss, implementation cost, and potential stress to animals. This study proposes an image-based approach for individual cattle identity classification using deep learning models. The evaluated models employ ResNet50 and IResNet50 as backbones, with Softmax and ArcFace as classification heads. A public cattle image dataset was inspected, validated, stratified, preprocessed, and evaluated under five scenarios: an imbalanced dataset, a controlled dataset, offline augmentation for minority classes, an ablation study of backbone and classification head combinations, and embedding representation analysis. Model performance was evaluated using top-1 accuracy, top-5 accuracy, macro precision, macro recall, macro F1-score, and threshold analysis. Embedding analysis was conducted using PCA, ray plots, within-cluster variance, and silhouette coefficient to observe intra-class compactness and inter-class separability. The results show that IResNet50 + ArcFace achieved the best performance based on macro F1-score and top-1 accuracy. On the imbalanced dataset, this model obtained a macro F1-score of 98.6789% and a top-1 accuracy of 99.0627%. Embedding analysis also showed a more compact and better-separated feature structure than ResNet50 + Softmax. Overall, performance was influenced by model architecture, classification head, data distribution, and learned feature representation.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Individual Cattle Classification, ResNet50, ArcFace, Deep learning, Klasifikasi Individu Sapi, ResNet50, ArcFace, Deep learning.
Subjects:	Q Science Q Science > QA Mathematics > QA336 Artificial Intelligence
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Thariq Ivan Anendar
Date Deposited:	23 Jun 2026 06:24
Last Modified:	23 Jun 2026 06:24
URI:	http://repository.its.ac.id/id/eprint/133981

Actions (login required)

View Item