Optimisasi Dynamic Routing pada Dual Attention Capsule Network untuk Pedestrian Attribute Recognition

Salsabila, Nada (2026) Optimisasi Dynamic Routing pada Dual Attention Capsule Network untuk Pedestrian Attribute Recognition. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6003241038-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (2MB) | Request a copy

Abstract

Pedestrian Attribute Recognition (PAR) merupakan salah satu permasalahan penting dalam bidang computer vision yang memiliki berbagai aplikasi, seperti sistem pengawasan, keamanan publik, dan smart city. Berbeda dengan klasifikasi citra tunggal, PAR termasuk ke dalam tugas multi-label fine-grained recognition karena setiap pedestrian dapat memiliki lebih dari satu atribut secara bersamaan. Penelitian sebelumnya banyak mengandalkan Convolutional Neural Network (CNN) yang mampu mencapai akurasi tinggi, namun CNN memiliki keterbatasan dalam mempertahankan informasi spasial, sehingga kurang robust terhadap perubahan posisi, occlusion, dan atribut berukuran kecil. Capsule Network (CapsNet) diperkenalkan sebagai alternatif karena mampu merepresentasikan relasi spasial antar fitur melalui mekanisme dynamic routing. Pengembangan lebih lanjut dilakukan dengan mengintegrasikan mekanisme dual Attention sehingga membentuk arsitektur Dual Attention Capsule Network (DA-CapsNet), yang berpotensi meningkatkan kemampuan pengenalan atribut pedestrian. Namun demikian, mekanisme dynamic routing pada CapsNet cenderung kurang stabil dan dapat memengaruhi proses konvergensi model. Oleh karena itu, penelitian ini mengusulkan penerapan optimisasi pada mekanisme dynamic routing dalam arsitektur DA-CapsNet untuk meningkatkan kestabilan pembelajaran sekaligus memperbaiki kinerja model. Evaluasi dilakukan menggunakan subset dari dataset PA-100K yang terdiri dari 50.000 citra pedestrian dengan anotasi multi-atribut. Sebanyak 7 atribut dipilih untuk multi label pada gambar tersebut. Kinerja model dievaluasi menggunakan metrik Area Under the Curve (AUC), baik secara per atribut maupun rata-rata. Hasil eksperimen menunjukkan bahwa DA-CapsNet dengan dynamic routing yang dioptimisasi menghasilkan kinerja yang lebih baik dan lebih stabil dibandingkan CapsNet standar serta DA-CapsNet tanpa optimisasi. Temuan ini menunjukkan bahwa kombinasi mekanisme dual Attention dan optimisasi dynamic routing dapat meningkatkan kemampuan diskriminasi model dalam pengenalan atribut pedestrian.
==================================================================================================================================
Pedestrian Attribute Recognition (PAR) is an important problem in the field of computer vision with various applications, such as surveillance systems, public security, and smart cities. Unlike single-image classification, PAR is a multi-label fine-grained recognition task, as each pedestrian may possess multiple at-tributes simultaneously. Previous studies have predominantly relied on Convolu-tional Neural Networks (CNNs), which have demonstrated high accuracy; how-ever, CNNs have limitations in preserving spatial information, making them less robust to changes in pose, occlusion, and small-scale attributes. Capsule Net-works (CapsNet) were introduced as an alternative approach due to their ability to represent spatial relationships among features through the dynamic routing mechanism. Further development has been achieved by integrating a dual Atten-tion mechanism, resulting in the Dual Attention Capsule Network (DA-CapsNet), which has the potential to enhance pedestrian attribute recognition performance. Nevertheless, the dynamic routing mechanism in CapsNet tends to be unstable and may affect the model’s convergence process. Therefore, this study proposes an optimization of the dynamic routing mechanism within the DA-CapsNet ar-chitecture to improve learning stability while enhancing model performance. The evaluation was conducted using a subset of the PA-100K dataset consisting of 50,000 pedestrian images with multi-attribute annotations. A total of seven at-tributes were selected for multi-label classification. Model performance was evaluated using the Area Under the Curve (AUC) metric, both on a per-attribute basis and as an average measure. Experimental results demonstrate that DA-CapsNet with optimized dynamic routing achieves better and more stable per-formance compared to the standard CapsNet and DA-CapsNet without optimiza-tion. These findings indicate that the combination of a dual Attention mechanism and dynamic routing optimization can improve the model’s discriminative capa-bility in pedestrian attribute recognition.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Deep learning, Attribute Recognition, pedestrian dataset, Pedestrian dataset, Capsule Network, Attention mechanism, Dynamic routing optimization, Deep learning, Attribute Recognition
Subjects:	Q Science > QA Mathematics > QA76.76.P37 Software patterns. Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) Q Science > QA Mathematics > QA9.58 Algorithms
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Statistics > 49101-(S2) Master Thesis
Depositing User:	Nada Salsabila
Date Deposited:	28 Jan 2026 08:03
Last Modified:	28 Jan 2026 08:03
URI:	http://repository.its.ac.id/id/eprint/130870

Actions (login required)

View Item