Putra, Sindunuraga Rikarno (2015) Implementasi Convolutional Neural Network Untuk Klasifikasi Obyek Pada Citra. Undergraduate thesis, Institut Technology Sepuluh Nopember.
Preview |
Text
5111100076-Undergraduate Thesis.pdf - Published Version Download (3MB) | Preview |
Abstract
Deep Learning adalah sebuah bidang keilmuan baru dalam bidang Machine Learning yang akhir-akhir ini berkembang karena perkembangan teknologi GPU acceleration. Deep learning memiliki kemampuan yang sangat baik dalam visi komputer. Salah satunya adalah pada kasus klasifikasi obyek pada citra yang telah lama menjadi problem yang sulit diselesaikan.
Tugas akhir ini mengimplementasikan salah satu jenis model Deep Learning yang memiliki kemampuan yang baik dalam klasifikasi data dengan struktur dua dimensi seperti citra, yaitu Convolutional Neural Network (CNN). Digunakan dataset CIFAR-10 yang telah lama menjadi benchmark dalam kasus klasifikasi citra. Model CNN akan dikembangkan menggunakan library Theano yang memiliki kemampuan baik dalam memanfaatkan GPU acceleration.
Dalam penyusunan model, dilakukan optimasi hyperparameter jaringan dan analisa penggunaan memori untuk mendapatkan intuisi yang lebih baik terhadap perilaku model CNN. Tugas akhir ini membandingkan tiga arsitektur CNN, yaitu DeepCNet, NagadomiNet, dan NetworkInNetwork, dengan kedalaman convolution layer maksimal lima layer pada setiap arsitektur.
viii
Dari hasil uji coba, didapatkan nilai error klasifikasi terkecil yaitu 17.69% dengan menggunakan arsitektur NagadomiNet yang terdiri dari convolution layer dengan ukuran kernel 3x3, penggunaan Global Average Pooling sebelum Softmax Layer, serta implementasi inverted dropout dengan nilai drop rate incremental sebesar 0.1, 0.2, 0.3, 0.4, dan 0.5 pada masing-masing convolution layer.
======================================================================================================
Deep Learning is a new branch of knowledge in the field of Machine Learning that in the past few years has developed due to the improvement in GPU Acceleration technologies. Deep Learning has great capabilities in the field of computer vision. One of its capabilities is in the case of object classification in images, a problem that has been unsolved for a long time.
This final project will implement a form of Deep Learning that is designed for the processing of two dimensional structured data such as images, which is called The Convolutional Neural Network (CNN). The CIFAR-10 dataset is used because it has been a classic benchmark for the case of image classification. The CNN model is implemented using the pyhton Theano Library because it is designed for use with GPU Acceleration.
In the creation of the model, hyperparameter tuning and memory usage analysis will also be done to get a better intuition on the characteristics of CNN. In this final project, three CNN model architectures is compared, which respectively is DeepCNet, NagadomiNet, and Network in Network, each with a maximum convoutional layer depth of 5.
x
From the experiment, the lowest classification error gotten is 17.69% from a model that uses the NagadomiNet architecture which consists of convolutional layers with a kernel of size 3x3, the usage of Global Average Pooling before the Softmax Layer, and also an implementation of inverted dropout using an incremental drop rate with a value of 0.1, 0.2, 0.3, 0.4, and 0.5 on each convolutional layers.
| Item Type: | Thesis (Undergraduate) |
|---|---|
| Additional Information: | RSIf 006.3 Put i |
| Uncontrolled Keywords: | Convolutional Neural Network, Deep Learning, GPGPU, Klasifikasi Citra, CIFAR |
| Subjects: | Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques |
| Divisions: | Faculty of Information Technology > Informatics Engineering > 55201-(S1) Undergraduate Thesis |
| Depositing User: | Mr. Tondo Indra Nyata |
| Date Deposited: | 21 Oct 2019 05:13 |
| Last Modified: | 21 Oct 2019 05:13 |
| URI: | http://repository.its.ac.id/id/eprint/71292 |
Actions (login required)
 |
View Item |