Muhammad, Brillianta Zayyan (2025) Implementation of instance segmentation using yolov8 for Indonesian sign language alphabet letters detection (bisindo) / Brillianta Zayyan Muhammad</p>. Diploma thesis, Universitas Negeri Malang.
Full text not available from this repository.Abstract
p Sign language is a non-verbal communication system that uses hand movements lip movements and facial expressions to convey messages. This communication system has various variations across countries and regions. In Indonesia there are 2 variations of communication systems used one of which is BISINDO. This communication system is widely used by deaf people in Indonesia because of its more natural and easy-to-understand characteristics. However there are still some challenges in implementing BISINDO one of which is the communication gap between deaf people and people without hearing loss. To handle these challenges there is one potential that can be used namely by implementing a sign language detection system. However to develop an accurate and efficient sign language detection system an optimal model is needed. YOLOv8 emerged as one of the relevant models used to detect sign language. It has various model variants developed to meet specific system needs. In addition YOLOv8 is known as a model with its accuracy and speed in detecting objects. YOLOv8 also has more complex features besides object detection namely by applying instance segmentation through the YOLOv8-seg model. The model has the ability to detect and segment objects with more precision down to the pixel level. With this ability the system will be easier to detect sign language especially in alphabet letters that have hand shapes with various variations. This research implements two variants of the YOLOv8-seg model namely YOLOv8n-seg and YOLOv8s-seg to detect alphabet letters in BISINDO. Both variants of the model are trained using four data sharing scenarios that have been annotated to data augmentation. The results of the training process are then evaluated with various approaches ranging from the evaluation of accuracy and speed metrics learning curve analysis and confusion matrix. The results show that the two model variants have a relatively small difference in accuracy metrics where YOLOv8s-seg is superior on average in all scenarios. Meanwhile the difference is apparent in the speed metric where YOLOv8n-seg shows lower inference time than YOLOv8s-seg. However the data sharing scenario does not really affect the inference time generated. Furthermore the learning curve graph shows that the YOLOv8s-seg model is superior to YOLOv8n-seg in terms of stability convergence and gap in the training and validation loss curves. Then based on the confusion matrix both model variants have difficulty in detecting similar letters (M and N) in scenarios with less training data. /p
| Item Type: | Thesis (Diploma) |
|---|---|
| Divisions: | Fakultas Teknik (FT) > Departemen Teknik Elektro (TE) > S1 Teknik Informatika |
| Depositing User: | library UM |
| Date Deposited: | 25 Jun 2025 04:29 |
| Last Modified: | 09 Sep 2025 03:00 |
| URI: | http://repository.um.ac.id/id/eprint/426185 |
Actions (login required)
 |
View Item |