Skip to content

American Sign Language Alphabet recognition with Deep Learning's CNN architecture

Notifications You must be signed in to change notification settings

MelihGulum/ASL-Recognition-CNN-OpenCV

Repository files navigation

ASL Recognition CNN OpenCV

In this project, I wanted to raise awareness for the hearing impaired. Do not forget every learning starts with the alphabet. You can make predictions about the American Sign Language Alphabet to the model by camera. So let's dig out the project some more.

I cover this project on the four step.

  • About Dataset
  • Deep Learning - CNN
  • OpenCV
  • How to run

1. Dataset

I use two different dataset for this project. The first one is, asl-alphabet, used for training and testing parts. It consist of 29 classes and each class has 3000 image. For Training 78300 image and testing 8700 image used. Second dataset is, asl-alphabet-test, is used for evaluating. These two dataset is too diffrent from eachother. Later we see some consequence of this.

2. Deep Learning - CNN

First things first we should some pre-process before training. One hot encoding and RGB normalization. These links are usefull if you don't know about them. Tradionally CNN is for mostly used for Image Processing and i kept this tradition. The following two figure are Confusion Matrix of respectively test and evaluation. Before, i mention about the difference about two dataset. That is the consequence of difference. Test images are so similar with the training set so accuracy and values of Confusion matrix is so high. On the other hand evaluation set is so much diffrent than training that is the reasen why evaluation scores too low.

Finally here some different training result:

OPTIMIZER EPOCH TEST ACCURACY EVALUATION ACCURACY
MODEL-1 rmsprop 5 %99.609 %40.46
MODEL-2 rmsprop 10 %99.989 %52.529
MODEL-3 adam 5 %99.494 %38.506
MODEL-4 adam 10 %99.805 %38.62

NOTE: Model-2 has been selected because our benchmark is evaluation accuracies.

Lack of hardware such as GPU and having a large datasets, ı use Google Colaboratory and suggest you to use that kind of cloud systems

3. OpenCV

Now, it is time for the real-time predictions. But there was a problem. Hand detection is not simple as face detection nor common. To overcome this difficulty, I first detected the hand from the camera (hand_detection.py). Then CNN is implemented to the OpenCV. But predictions were way ridiculous. Here is the reason of it 👇 :

  • I thought that if i make the sign to the window (which is opened by Opencv), model can easily predict. But it didn't turn out the way I thought. Because model took whole the window as an input. Including myself, door etc.
  • Then the idea of drawing a square came to my mind. Unfortunately that's a another fail. Model took the whole window again.
  • Finally i managed the crop the square. That's why prediction only works in the square.

  • I made it this square brackets to make basic words and abbrevations. If there are 10 consecutive identical signs, it appends it to a list.

4. How to run

  1. Fork this repository.
$ git clone https://github.com/MelihGulum/ASL-Recognition-CNN-OpenCV.git
  1. Load the dependencies of the project

NOTE: This dependencies not including the Deep Learning part. Colab meet all dependencies (such as tensorflow).

pip install -r requirements.txt
  1. Now you can run ASL_Recognition.py. But if you want you can run .ipynb and you can build your own model. It is up to you.

About

American Sign Language Alphabet recognition with Deep Learning's CNN architecture

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published