Scribd
Upload
55 views

ALGORITHM

The document describes an algorithm that uses a convolutional neural network (CNN) to recognize stationary signs from an image dataset. The CNN is implemented using Keras and TensorFlow in Python. The algorithm is trained on multiple iterations called epochs, with accuracy improving the more epochs that are run, up to a maximum of 95% accuracy for the training data and 85% for validation data. Graphs show how accuracy increases and loss decreases with more epochs. The trained model generates confidence intervals to identify test images.

Uploaded by

Srushti Pawar
Copyright
© © All Rights Reserved
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
55 views

ALGORITHM

The document describes an algorithm that uses a convolutional neural network (CNN) to recognize stationary signs from an image dataset. The CNN is implemented using Keras and TensorFlow in Python. The algorithm is trained on multiple iterations called epochs, with accuracy improving the more epochs that are run, up to a maximum of 95% accuracy for the training data and 85% for validation data. Graphs show how accuracy increases and loss decreases with more epochs. The trained model generates confidence intervals to identify test images.

Uploaded by

Srushti Pawar
Copyright
© © All Rights Reserved
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
You are on page 1/ 6

ALGORITHM

• The implementation for this system would be done in Python. This is because
Python is a high-level, interpreted language that finds application in systems
relating to artificial intelligence.
• A dataset was compiled of stationary signs from a pre-existing, open source
GitHub repository . This dataset comprises of ISL signs as signed by different
users.
• The Convolutional Neural Network (CNN) algorithm makes use of Keras Deep
Learning library.
• Keras offers pre-trained networks that can aid with the identification and
classification of object categories with high accuracy. In order to customize
these classifications, Keras is run on top of TensorFlow.
• For the model to be able to attain a high accuracy in gesture recognition,
it is necessary for it to run multiple iterations on the training dataset.
• These iterations can be defined by determining the number of
Epoch cycles to be run.

Fig.2: Model Architecture of the Convolutional Neural Network


• The results in this case, would be the text output for the sign input.
This output is then processed by running it through a text to speech
translator.
• Once run through this translator the outputs can either be played or
stored and kept in audio file format.
IMPLEMENTATION
• Multiple instances of the input dataset were generated and
processed through the algorithm for varying number of Epoch cycles
to verify the accuracy of the model generated.
• For a given input dataset comprising of a set of gestures, the model
was trained for 25, 50, 75 and 100 Epoch cycles.
• It was found that higher the number of Epoch cycle, greater the
stability of the accuracy.
• The model accuracy grew up to the range of 85-95% for the training
dataset and in the range of 75-85% for the validation dataset.
• As the accuracies increase with every cycle, the loss decreases. The graphical
representation of the loss function and accuracies are provided in the following
images.

Fig.: Loss Function for a dataset of 15 signs and 100 Epoch cycles Fig: Accuracy for a dataset of 15 signs and 100 Epoch cycles
• The model generated a confidence interval for a test image by generating a
confidence interval for every sign it was trained on and then opting for the
highest interval.

Fig: Correctly identified test signs along with their confidence intervals as identified by the trained model, developed using a training dataset of
15 signs and 75 Epoch cycles

You might also like