Implementasi Fully Convolutional Network untuk Segmentasi Citra

Salomo, Gerald Parlindungan (2019) Implementasi Fully Convolutional Network untuk Segmentasi Citra. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 05111540000021-Undergraduate_Theses.pdf Restricted to Repository staff only Download (2MB) | Request a copy

Abstract

Machine learning adalah sebuah cabang dalam ilmu komputer yang meneliti bagaimana suatu mesin dapat menyelesaikan suatu masalah. Salah satu pengimplementasian Machine Learning adalah Deep Learning. Deep learning adalah salah satu metode implementasi dari Machine Learning yang bertujuan untuk meniru cara kerja otak manusia menggunakan Artificial Neural Network (ANN). Selain itu, Deep Learning cocok digunakan untuk memperhitungkan data yang tidak pasti, seperti bahasa, suara, gambar maupun gambar bergerak (video). Pada tugas akhir ini, penulis mengimplementasikan Deep Learning dengan arsitektur Fully Convolutional Network (FCN), untuk segmentasi citra. Penulis menggunakan pre-trained model VGG16 berdasarkan dataset ImageNet, dengan jumlah mencapai 1000 class objek. Penggunaan utama model ini adalah untuk melakukan klasifikasi objek. Maka dari itu, penulis melakukan modifikasi terhadap model tersebut agar bisa digunakan untuk melakukan segmentasi citra. Hal yang dilakukan adalah mengubah fully connected layer pada model VGG16 menjadi convolutional layer (disebut proses convolutionalization). Karena semua layer pada model merupakan convolutional layer, arsitektur yang dihasilkan disebut dengan Fully Convolutional Network. Hal unik dari arsitektur FCN adalah pada proses training, arsitektur mampu menerima berbagai ukuran citra input (non-fixed size), yang disebabkan oleh proses convolutionalization. Pada uji coba digunakan data testing dari PASCAL VOC 2012, yang memiliki 20 class objek (pesawat, manusia, sepeda, kapal, dan lainnya), dan tersedia citra asli beserta dengan ground truth. Langkah dari uji coba program yaitu dimulai dari input citra terhadap arsitektur FCN, akan menghasilkan citra semantic segmentation, citra semantic segmentation akan diubah menjadi citra object segmentation berdasarkan tipe uji coba, lalu dilakukan uji evaluasi pada hasil object segmentation terhadap citra ground truth. Penulis melakukan tiga tipe uji coba, yaitu satu class objek, dua class objek, dan tiga class objek. Metode evaluasi yang digunakan penulis, adalah Class Prediction Accuracy, Jaccard Coefficient atau Intersection over Union (IoU), dan Pixelwise Accuracy. Berdasarkan hasil uji coba, hasil evaluasi paling optimal adalah pada satu class objek sebesar 91% untuk Class Prediction Accuracy, 53% untuk Jaccard Coefficient, dan 66% untuk Pixelwise Accuracy.
===================================================================================================================================
Machine Learning is one of scientific study in computer science, that computer systems use in order to solve specific task effectively. One of the implementation of Machine Learning is Deep Learning. Deep Learning has a purpose of replicating how the human brain works, using Artificial Neural Network (ANN). Furthermore, Deep Learning is useful for computing inexact data, for example language/sentence, sound, still image or moving image (video) In this undergraduate thesis, the writer implements Deep Learning with Fully Convolutional Network (FCN) as the architecture, for image segmentation. The writer uses VGG16 pretrained model, based on ImageNet dataset with over 1000 classes. The main usage of the model is for object classification. Modifications were made to the model, so it can be used for image segmentation. What can be done is to transform the fully connected layer in VGG16 model to convolutional layer (the process is called convolutionalization). Because all the layer in the model consists of convolutional layer only, the architecture is called Fully Convolutional Network. The unique characteristic about FCN is when we train the model with image dataset, we can use arbitrary image size (non-fixed size) to train the model , effected by the convolutionalization. The testing process uses the testing datas from the PASCAL VOC 2012 dataset, which has 20 class object (aeroplane, human, bicycle, ship, and others), and it contains the original image as well as the ground truth image. The steps for testing the program starts with the input of an image to the FCN model, which will output a semantic segmentation image, then the semantic segmentation image will be transformed to object segmentation image based on the type of evaluation, then the application of evaluation methods to the object segmentation compared to the ground truth image. There are three types of evaluation in this program, those are one class object, two class object, and three class object. The evaluation methods applied in this program are Class Prediction Accuracy, Jaccard Coefficient or Intersection over Union (IoU), and Pixelwise Accuracy. Based on the evaluation results, the most optimal result is for one class object type which has 91% for Class Prediction Accuracy, 53% for Jaccard Coefficient, and 66% for Pixelwise Accuracy.

Item Type:	Thesis (Other)
Additional Information:	RSIf 006.42 Sal i-1 2019
Uncontrolled Keywords:	Fully Convolutional Network, Dataset ImageNet, Dataset PASCAL VOC 2012, Segmentasi Citra, VGG16, Class Prediction Accuracy, Jaccard Coefficient, Pixelwise Accuracy
Subjects:	T Technology > T Technology (General) > T57.5 Data Processing
Divisions:	Faculty of Information and Communication Technology > Informatics > 55201-(S1) Undergraduate Thesis
Depositing User:	Gerald Parlindungan Salomo
Date Deposited:	23 Jul 2026 03:35
Last Modified:	23 Jul 2026 03:35
URI:	http://repository.its.ac.id/id/eprint/65232

Actions (login required)

View Item