Adam dan SGD pada Faster RCNN ResNet dan MobileNet untuk Deteksi Gestur Tangan Bahasa Isyarat

Sinurat, Ryan Christofer (2024) Adam dan SGD pada Faster RCNN ResNet dan MobileNet untuk Deteksi Gestur Tangan Bahasa Isyarat. Undergraduate thesis, UPN Veteran Jawa Timur.

	Text (Cover) 20081010210.-cover.pdf Download (1MB)
	Text (Bab 1) 20081010210.-bab1.pdf Download (223kB)
	Text (Bab 2) 20081010210.-bab2.pdf Restricted to Repository staff only until 18 December 2026. Download (571kB)
	Text (Bab 3) 20081010210.-bab3.pdf Restricted to Repository staff only until 18 December 2026. Download (1MB)
	Text (Bab 4) 20081010210.-bab4.pdf Restricted to Repository staff only until 18 December 2026. Download (3MB)
	Text (Bab 5) 20081010210.-bab5.pdf Download (210kB)
	Text (Daftar Pustaka) 20081010210.-daftarpustaka.pdf Download (199kB)

Abstract

A person who is deaf or hard of hearing is a non-verbal communicator. In 2021, 1.3 billion or 16% of the global population will be. This large number cannot be matched by the number of people who understand sign language. With the amount of difference, it is undoubtedly challenging to communicate with people with disabilities. One of the non-verbal communication methods of sign language is the use of American Sign Language. Recognizing American Sign Language alphabets is very important in supporting people who are deaf or hard of hearing and speech impaired in translating their signs or the intentions they want to convey. Faster Regional Convolutional Neural Network (Faster R-CNN) is a deep learning and computer vision model designed to detect and recognize objects in images or videos. This research uses ResNet-50 and MobileNet backbone models with Adam and SGD optimization. In addition, the non-maximum suppression (NMS) technique is applied, and it is expected to be detected accurately. Based on this explanation, the author conducted American Sign Language object detection by comparing the two backbones and optimizations. The result of this research is the best model obtained: MobileNet with Adam Optimization. The mean Average Precision (mAP) accuracy on the test data is 88.97%. The consistency of this model performance is reinforced by the average f1-score value of 88.54% on the test data. This shows a better balance between precision and recall in these classes.

Item Type:

Thesis (Undergraduate)

Contributors: