DEVELOPMENT OF SHERPA MODEL BASED ON DEEP CONVOLUTIONAL NEURAL NETWORK (CNN) AS THE CORE SYSTEMS OF LOVI APPLICATION
The high number of visual impairment or difabel netra (DN) in Indonesia triggers public attention to provide maximum service for their rights as fellow human beings. Physical limitations make it difficult for DNs to carry out activities inside and outside the room, and often experience disorienta...
Saved in:
| Main Author: | |
|---|---|
| Format: | Final Project |
| Language: | Indonesia |
| Online Access: | https://digilib.itb.ac.id/gdl/view/68748 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Institut Teknologi Bandung |
| Language: | Indonesia |
| Summary: | The high number of visual impairment or difabel netra (DN) in Indonesia triggers
public attention to provide maximum service for their rights as fellow human
beings. Physical limitations make it difficult for DNs to carry out activities inside
and outside the room, and often experience disorientation which can endanger
DNs. In this modern era, the use of deep learning technology, especially in terms
of image classification, will help DN overcome these problems. Therefore, we
developed the LoVi application that can help DNs, especially low vision (LoVi) in
their daily activities. The LoVi application has three modes, which are outdoor
mode, indoor mode, and currency mode with core systems in the form of Sherpa
models (TrotoarNet, IndoorNet, and CurrencyNet). Based on the test, from the final
architecture of the Sherpa model, the test accuracy reached 95.5% on IndoorNet,
96.79% on TrotoarNet, and 93.35% on CurrencyNet, with FPS reaching 13 img/s,
40 img/s, and 26 img/s, respectively. Also, the final model size is much smaller than
the Sherpa baseline model, which is 81MB. Thus, the built Sherpa model already
meets the specified system specifications. |
|---|