Penggabungan Segmentasi Pixel-Level dan Superpixel-Level untuk Pelabelan Fesyen Item

Putra, Migel Aulia Mandiri (2023) Penggabungan Segmentasi Pixel-Level dan Superpixel-Level untuk Pelabelan Fesyen Item. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 05111940000194-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only until 3 October 2025. Download (4MB) | Request a copy

Abstract

Fesyen terdiri dari berbagai macam jenis dan bentuk. Untuk melakukan identifikasi terhadap semua jenis dan bentuk fesyen, dapat memanfaatkan segmentasi citra yang berfokus pada fesyen item. Segmentasi citra yang berfokus pada fesyen item mempunyai manfaat dalam beragam aplikasi, mulai dari rekomendasi gaya berpakaian (style recommendation) hingga human surveillance. Hingga kini, metode segmentasi pixel-level menjadi salah satu model berbasis Convolutional Neural Network yang banyak digunakan dalam segmentasi citra dengan obyek yang bukan fesyen. Metode segmentasi pixel-level menerapakan algoritma Mask R-CNN dalam melakukan deteksi obyek. Dengan mempertimbangkan bahwa citra fesyen rentan terhadap deformasi yang disebabkan oleh pose tubuh manusia, pemanfaatan metode segmentasi pixel-level secara langsung dalam segmentasi citra fesyen tidak dapat memberikan hasil yang optimal bila tanpa memperhatikan pose pada citra.
Dalam tugas akhir ini, diajukan sebuah metode mengidentifikasi fesyen item dengan menggabungkan dua metode segmentasi, yaitu metode segmentasi pixel-level dan segmentasi superpixel-level. Tujuan penggabungan dua metode segmentasi ialah untuk memperbaiki hasil dan performa yang dihasilkan hanya pada satu metode segmentasi saja yaitu metode segmentasi pixel-level. Selain menggunakan dua metode segmentasi, pada penelitian ini terdapat dua model yang digunakan, yaitu model yang berfokus pada fesyen item bernama Clothing Segmentation dan model yang berfokus pada bentuk tubuh manusia bernama Body Segmentation. Langkah yang akan diambil, kedua model akan dilatih menggunakan metode segmentasi pixel-level terlebih dahulu. Hasil yang didapatkan dari metode segmentasi pixel-level, akan digabungkan dan diproses dengan menggunakan metode segmentasi superpixel-level.
Eksperimen terhadap benchmark dataset menunjukkan performa metode usulan lebih unggul dibandingkan metode segmentasi pixel-level, dengan selisih rata-rata pada akurasi sebesar 1,386% dan intesection over union sebesar 4,197%. Hal ini menunjukkan bahwa metode usulan dapat melakukan identifikasi dan pelabelan terhadap fesyen item secara lebih baik, dibandingkan dengan pelabelan fesyen item menggunakan metode segmentasi pixel-level.
=============================================================================================================================
Fashion consists of various types and shapes. To provide assistance to all types and forms of fashion, you can utilize image segmentation that focuses on fashion items. Image segmentation that focuses on fashion items has benefits in a variety of applications, from style recommendations to human monitoring. Until now, the pixel-level segmentation method has become one of the models based on the Convolutional Neural Network which is widely used in image segmentation with non-fashionable objects. The pixel-level segmentation method applies the Mask R-CNN algorithm in object recognition. Taking into account that fashion images are susceptible to deformation caused by human body poses, the use of the pixel-level segmentation method directly in fashion image segmentation cannot provide optimal results without paying attention to the pose in the image.
This final project proposes a method for identifying fashion items by combining two segmentation methods, namely the pixel-level segmentation method and the superpixel-level segmentation method. The purpose of combining the two segmentation methods is to improve the results and performance produced by only one segmentation method, namely the pixel-level segmentation method. In addition to using two segmentation methods, in this study there are two models used, namely a model that focuses on fashion items called Clothing Segmentation and a model that focuses on the shape of the human body called Body Segmentation. Steps to be taken, both models will use the pixel-level segmentation method first. The results obtained from the pixel-level segmentation method will be combined and processed using the superpixel- level segmentation method.
Experiments on the benchmark dataset show that the performance of the firing method is superior to the pixel-level segmentation method, with an average difference in accuracy of 1,386% and intersection over union of 4,197%. This shows that the visiting method can make better withdrawals and labeling of fashion items, compared to labeling fashion items using the pixel-level segmentation method.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Body Segmentation, Clothing Segmentation, Fashion Items, Mask R-CNN, Pixel-level Segmentation, Superpixel-level Segmentation.
Subjects:	Q Science > Q Science (General) > Q325.5 Machine learning.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Migel Aulia Mandiri Putra
Date Deposited:	02 Aug 2023 02:46
Last Modified:	02 Aug 2023 02:46
URI:	http://repository.its.ac.id/id/eprint/100402

Actions (login required)

View Item