Analisis Klasifikasi Few-Shot Learning Untuk Identifikasi Produk Gambar Di E-Commerce Menggunakan Siamese Convolutional Neural Network Dan Transfer Learning

Choiruddin, Hanif (2025) Analisis Klasifikasi Few-Shot Learning Untuk Identifikasi Produk Gambar Di E-Commerce Menggunakan Siamese Convolutional Neural Network Dan Transfer Learning. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5003211063-Undergraduate_Thesis.pdf Download (4MB)

Abstract

Pertumbuhan pesat e-commerce mendorong kebutuhan akan sistem pencarian produk berbasis gambar yang efisien, terutama untuk mengidentifikasi produk baru dengan jumlah data yang terbatas. Penelitian ini mengusulkan pendekatan Few-Shot Learning menggunakan arsitektur Siamese Network untuk tugas klasifikasi kemiripan gambar produk pada platform e-commerce.Penelitian ini mengkaji pendekatan Few-Shot Learning berbasis arsitektur Siamese Network dalam mengidentifikasi kemiripan citra produk e-commerce dengan keterbatasan data. Dua arsitektur backbone dibandingkan, yakni VGG-16 pretrained dan CNN, pada skenario 5-Way K-Shot (K = 1 hingga 5). Data diperoleh dari kumpulan gambar produk Shopee di platform Kaggle, yang kemudian dipasangkan menjadi support-query untuk membentuk episode pelatihan. Proses preprocessing mencakup resize ke 256×256 piksel dan normalisasi sesuai standar VGG-16. Evaluasi dilakukan pada data validasi dan uji menggunakan metrik akurasi dan presisi antar kelas. Hasil eksperimen menunjukkan bahwa model dengan CNN memperoleh akurasi validasi tertinggi sebesar 54,17% pada konfigurasi 4-Shot, serta akurasi uji tertinggi 49,90% pada konfigurasi yang sama. Sementara itu, VGG-16 menunjukkan performa yang lebih stabil dan efisien secara komputasi, terutama pada skenario 1-Shot dan 5-Shot. CNN unggul pada jumlah support menengah, namun menunjukkan fluktuasi yang lebih besar. Berdasarkan analisis menyeluruh terhadap akurasi, presisi antar kelas, serta efisiensi pelatihan, dapat disimpulkan bahwa pemilihan backbone sangat dipengaruhi oleh jumlah data dan kebutuhan sistem, CNN cocok untuk kondisi data terbatas, sedangkan VGG-16 lebih unggul untuk prediksi yang konsisten dan seimbang.
======================================================================================================================================
The rapid growth of e-commerce has driven the need for efficient image-based product search systems, particularly for identifying new products with limited data availability. This study proposes a Few-Shot Learning approach using a Siamese Network architecture to classify product image similarity on e-commerce platforms. Two backbone architectures are compared: a pretrained VGG-16 and a CNN, evaluated under a 5-Way K-Shot setting (K = 1 to 5). The dataset is derived from Shopee product images available on the Kaggle platform, which are paired into support-query episodes for training. Preprocessing includes resizing images to 256×256 pixels and normalizing them according to the VGG-16 standard. Evaluation is performed on both validation and test data using accuracy and per-class precision as performance metrics. Experimental results show that the CNN model achieved the highest validation accuracy of 54.17% and the highest test accuracy of 49.90%, both under the 4-Shot configuration. In contrast, VGG-16 demonstrated more stable and computationally efficient performance, especially in the 1-Shot and 5-Shot scenarios. While the CNN excelled in mid-Shot settings, it exhibited greater performance fluctuations. Overall, a comprehensive analysis of accuracy, class-wise precision, and training efficiency suggests that the choice of backbone is highly dependent on data volume and system requirements. The CNN is better suited for limited-data scenarios, whereas VGG-16 offers more consistent and balanced predictions.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	E-Commerce, CNN, Few-Shot Learning, Siamese Network, VGG-16, Identifikasi Produk. E-Commerce, CNN, Few-Shot Learning, Siamese Network, VGG-16, Product Identification.
Subjects:	B Philosophy. Psychology. Religion > BF Psychology > BF318 Learning, Psychology of (Deep learning) Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines. Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.754 Software architecture. Computer software Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) T Technology > TE Highway engineering. Roads and pavements > TE228.37 Vehicular ad hoc networks (Computer networks)
Divisions:	Faculty of Science and Data Analytics (SCIENTICS)
Depositing User:	Hanif Choiruddin
Date Deposited:	04 Aug 2025 08:35
Last Modified:	04 Aug 2025 08:35
URI:	http://repository.its.ac.id/id/eprint/127121

Actions (login required)

View Item