# Deep Learning Model
The goal of SIGNARA is to develop software that is capable of real-time translation of Arabic sign language into text. Due to resources and time constraints, the scope has been limited to implementing 9 Words and the whole Arabic alphabet.
We achieved the words sector using [Mediapipe](https://mediapipe.dev/) - a framework for building multimodal, cross-platform, applied ML pipelines. While the alphabets sector Uses transfer learning.
<p align="center">
<img src="img\Mediapipe.png" alt="Meidapipe"/>
</p>
Now we will discuss technical details of the 2 sectors:
## Words
Challenges we faced while doing this part for having a dynamic data that can be represented as series of motion we had many questions during the research process
### Dataset
After doing online research we found that there is no public dataset for the Arabic Sign Language words which contain continuous motion. So we started collecting our dataset by capturing video that takes 30 and 60 Frames of the word motion. We Captured 120 videos of 2 different data collectors to have a variety
### Model
Deep learning methods such as recurrent neural networks like as LSTMs and variations that make use of one-dimensional convolutional neural networks or CNNs have been shown to provide state-of-the-art results on challenging activity recognition tasks with little or no data feature engineering, instead using feature learning on raw data.
We applied our knowledge and tried many deep learning algorithms in order to get insghts of the performance by comparing different approaches.
- CONV1D
- LSTM
- CONV1DLSTM
- LSTM With CTC Loss
- Transformers
## Characters
<p align="center">
<img src="img\char.png" alt="Characters"/>
</p>
We found dataset consists of 54,049 images of [ArSL alphabets](https://data.mendeley.com/datasets/y7pckrw6z2/1) performed by more than 40 people for 32 standard Arabic signs and alphabets. The number of images per class differs from one class to another. The dataset gathered are of size 64 * 64 Pixels of grayscale.
Deep convolutional neural network models may take days or even weeks to train on very large datasets.
A way to short-cut this process is to re-use the model weights from pre-trained models that were developed for standard computer vision benchmark datasets, such as the ImageNet image recognition tasks. Top performing models can be downloaded and used directly, or integrated into a new model for your own computer vision problems. This way is called Transfer Learning
In our Approach we used VGG16 Model ,The model achieves 92.7% top-5 test accuracy in ImageNet, which is a dataset of over 14 million images belonging to 1000 classes.
没有合适的资源?快使用搜索试试~ 我知道了~
资源推荐
资源详情
资源评论
收起资源包目录
基于深度学习的聋哑人实时手语翻译器.zip (95个子文件)
基于深度学习的聋哑人实时手语翻译器
Demo Video.mp4 8.74MB
img
Logo.png 11KB
signara.jpg 870KB
Sign To Text
helper.py 7KB
Notebooks
download data.txt 82B
Sign Lang Translator Word Based.ipynb 38KB
Sign Lang Translator Character Based.ipynb 11KB
Logs
LSTM
train
events.out.tfevents.1634268063.PETER-DESKTOP.30040.2.v2 1.47MB
validation
events.out.tfevents.1634268069.PETER-DESKTOP.30040.3.v2 52KB
vgg16_transfer_learning
train
events.out.tfevents.1634839375.PETER-DESKTOP.20520.0.v2 11KB
events.out.tfevents.1634844986.PETER-DESKTOP.17376.0.v2 284KB
validation
events.out.tfevents.1634845014.PETER-DESKTOP.17376.1.v2 2KB
Conv1d
train
events.out.tfevents.1634318138.PETER-DESKTOP.36432.0.v2 1.5MB
validation
events.out.tfevents.1634318141.PETER-DESKTOP.36432.1.v2 79KB
vgg16_finetuning
train
events.out.tfevents.1634845125.PETER-DESKTOP.17376.2.v2 355KB
validation
events.out.tfevents.1634845136.PETER-DESKTOP.17376.3.v2 3KB
Conv1d_LSTM_custom
train
events.out.tfevents.1634781709.PETER-DESKTOP.24684.6.v2 248KB
validation
events.out.tfevents.1634781746.PETER-DESKTOP.24684.7.v2 3KB
transformer_encoder
train
events.out.tfevents.1634271989.PETER-DESKTOP.30040.4.v2 40B
events.out.tfevents.1634271997.PETER-DESKTOP.30040.5.v2 28.63MB
validation
events.out.tfevents.1634272011.PETER-DESKTOP.30040.6.v2 181KB
Conv1d_LSTM_tf
train
events.out.tfevents.1634789339.PETER-DESKTOP.24684.14.v2 769KB
validation
events.out.tfevents.1634789860.PETER-DESKTOP.24684.15.v2 1KB
LSTM2
train
events.out.tfevents.1634838285.PETER-DESKTOP.28980.3.v2 732KB
events.out.tfevents.1634838271.PETER-DESKTOP.28980.2.v2 394KB
validation
events.out.tfevents.1634838478.PETER-DESKTOP.28980.4.v2 9KB
LSTM3
train
events.out.tfevents.1634849165.PETER-DESKTOP.23756.2.v2 2.47MB
events.out.tfevents.1634846502.PETER-DESKTOP.23756.0.v2 529KB
validation
events.out.tfevents.1634849172.PETER-DESKTOP.23756.3.v2 61KB
events.out.tfevents.1634846696.PETER-DESKTOP.23756.1.v2 4KB
CTC
train
plugins
profile
2021_10_21_16_31_30
PETER-DESKTOP.input_pipeline.pb 2KB
PETER-DESKTOP.xplane.pb 100KB
PETER-DESKTOP.memory_profile.json.gz 19KB
PETER-DESKTOP.kernel_stats.pb 0B
PETER-DESKTOP.trace.json.gz 4KB
PETER-DESKTOP.overview_page.pb 4KB
PETER-DESKTOP.tensorflow_stats.pb 5KB
events.out.tfevents.1634833890.PETER-DESKTOP.profile-empty 40B
events.out.tfevents.1634833879.PETER-DESKTOP.28980.0.v2 911KB
validation
events.out.tfevents.1634833898.PETER-DESKTOP.28980.1.v2 55KB
Archs images
model Transformer.png 240KB
LSTM model.png 33KB
VGGmodel.png 135KB
model_conv1d.png 40KB
model CTC.png 29KB
model_convlstm.png 29KB
Models
LSTM3.h5 2.27MB
model_transformer_encoder.h5 16.3MB
model_conv1d.h5 1.33MB
VGG_transfer_learning.h5 80.73MB
main.py 6KB
models.py 1KB
utils
record vids.ipynb 2KB
Create Videos from Frames and save it in Dataset Folder.ipynb 3KB
Extract Videos and Keypoints.ipynb 14KB
Chars Dataset Gathering.ipynb 6KB
img
Mediapipe.png 31KB
char.png 13KB
const.py 2KB
README.md 3KB
fonts
Sahel.ttf 72KB
SIGNARA_Presentation.pptx 24.57MB
Final Demo.mp4 19.29MB
Text To Sign
input.txt 11B
ner.py 669B
Notebooks
ANERCorp.xlsx 2.11MB
ner.py 669B
ner_model.sav 5.84MB
namedEntity.ipynb 16KB
main.py 3KB
ModifiedVincent2.blend 18.87MB
logo.png 12KB
utils
main.ipynb 4KB
Calculate Transition - accross multiple Moves.ipynb 3KB
Moves
base_coord.bvh 562KB
live.bvh 255KB
egypt.bvh 365KB
hru_motion_coord.bvh 310KB
base_coord_main.bvh 83KB
hello_motion_coord.bvh 213KB
i.bvh 209KB
img
Animation.jpg 213KB
render.py 1KB
multithreading 3KB
maps.py 401B
anime.py 1KB
models
ner_model.sav 5.84MB
bvh.py 8KB
spell_checker.py 162B
How it Works.mp4 3.29MB
test.py 2KB
outputs
output.mp4 1.55MB
README.md 795B
ModifiedVincent2.blend1 18.87MB
README.md 2KB
共 95 条
- 1
资源评论
小码蚁.
- 粉丝: 2672
- 资源: 4542
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
最新资源
- 新建 Microsoft Word 文档 (2).docx
- 1652189183080063_42AD52629C553C7BC65694A425954A3F_1.apk
- 王都创世物语魔改版.apk
- Linear_regrassion_with_gradien_decent_example.ipynb
- 1719163625521533_1e90151d86c4cc2fadb97a0a82d0fc0e
- 虚拟现实课件-第五章 3物理建模
- 虚拟现实课件-第五章 2运动建模
- A1-ERP(资源协同)管理平台需求说明书.pdf
- 基于springboot+sureness的面向REST API资源无状态认证权限管理系统(源码+数据库+说明文档)
- programming LSQUIC
资源上传下载、课程学习等过程中有任何疑问或建议,欢迎提出宝贵意见哦~我们会及时处理!
点击此处反馈
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功