Sulthan Saalim Rabbani Atmadja, Aris Haris Rismayana

Klasifikasi Brand Sepatu Nike Berbasis Citra dengan Algoritma Convolutional Neural Network (CNN)

Introduction

Klasifikasi brand sepatu nike berbasis citra dengan algoritma convolutional neural network (cnn). Deteksi sepatu Nike asli vs. palsu berbasis citra menggunakan CNN MobileNetV2. Capai akurasi 96% untuk klasifikasi produk. Solusi autentikasi otomatis via deep learning.

Abstract

Peredaran sepatu Nike tiruan (KW) yang semakin luas menimbulkan kesulitan dalam membedakan produk asli dan palsu karena kemiripan visual yang tinggi. Penelitian ini bertujuan membangun sistem klasifikasi sepatu Nike berbasis citra digital menggunakan pendekatan kuantitatif dengan metode eksperimen. Model yang digunakan adalah Convolutional Neural Network (CNN) berbasis arsitektur MobileNetV2, dengan pendekatan transfer learning. Dataset terdiri atas 200 gambar sepatu Nike yang dibagi menjadi dua kelas (original dan KW), dan telah melalui proses augmentasi serta pelabelan manual. Model dilatih menggunakan TensorFlow dan Keras, kemudian dievaluasi berdasarkan akurasi, precision, recall, dan confusion matrix. Hasil evaluasi menunjukkan bahwa MobileNetV2 mencapai akurasi 96% pada data uji, dengan recall 0.94 untuk kelas KW dan 0.99 untuk kelas original. Dibandingkan dengan VGG16 sebagai model pembanding, MobileNetV2 terbukti lebih ringan, cepat, dan seimbang dalam performa klasifikasi. Penelitian ini menunjukkan bahwa pendekatan deep learning dengan MobileNetV2 efektif untuk mendeteksi sepatu palsu, serta berpotensi diterapkan dalam aplikasi mobile sebagai solusi autentikasi produk secara otomatis.

Review

This paper presents a timely and relevant study on classifying Nike shoes as original or counterfeit using image-based deep learning, an increasingly important task given the proliferation of high-quality replicas. The authors effectively leverage a Convolutional Neural Network (CNN) with the MobileNetV2 architecture, employing transfer learning, which is a sound methodological choice for image classification tasks, especially when dealing with potentially limited datasets. The reported accuracy of 96% on the test data, coupled with strong recall scores for both classes (0.99 for original and 0.94 for KW), demonstrates the significant potential of this approach. Furthermore, the explicit comparison with VGG16 highlights MobileNetV2's advantages in terms of efficiency and balanced performance, making a strong case for its suitability for practical deployment, particularly in mobile applications. While the results are promising, there are several areas where the research could be strengthened. The dataset size of 200 images, even with augmentation, is relatively small for deep learning models, especially for a complex task like distinguishing subtle differences between original and counterfeit products. A larger and more diverse dataset, encompassing a wider variety of Nike models, angles, lighting conditions, and crucially, a broader spectrum of counterfeit qualities and sources, would significantly enhance the model's generalizability and robustness. The abstract also lacks specific details on the augmentation techniques used, which would be valuable for reproducibility. Additionally, while the evaluation metrics are comprehensive, a deeper discussion on the implications of the slightly lower recall for the 'KW' class, and how this might impact real-world application where false negatives (misclassifying a fake as original) could be particularly damaging, would be beneficial. Despite these points, the research provides a valuable contribution to the field of automated product authentication. The demonstrated effectiveness of MobileNetV2 for this specific classification task lays a solid foundation for future work. Moving forward, the authors could explore expanding the dataset significantly, potentially through web scraping or collaborations with brands, and investigate the model's performance on unseen counterfeit types. Integrating explainable AI (XAI) techniques could also offer insights into the features the CNN uses to make its distinctions, further building trust and understanding. The envisioned application in mobile environments is particularly exciting, and future iterations could delve into optimizing inference speed and robustly handling real-world image capture variations to truly bring this solution to practical implementation.

Full Text

Comments

You need to be logged in to post a comment.