American Sign Language Detection/abhisheks008/DL-Simplified#312
to predict correct sign language labels corresponding to their corresponding sign images
Full name : Aditya Narayan Jha
GitHub Profile Link : https://github.com/aditya0929
Email ID : adityajha8906@gmail.com
Participant ID (if applicable): (certificate_id) - 7e43ecd9-76ff-4713-917e-f884b9fedf02
Approach for this Project : Image segmentation
is a crucial task in computer vision that involves identifying and classifying different regions or objects within an image. In this project, I will explore three different approaches for image segmentation using deep learning models: VGG16, MobileNet . InceptionV3 and InceptionResNetV2. the sign langauge dataset images had black background behind of hand signs and it the resolution was 680x480 pixels
InceptionResNetV2
When implementing the InceptionResNetV2 model in code, we leverage its powerful architecture to enhance our image classification tasks. By loading the pre-trained InceptionResNetV2 model with weights from the ImageNet dataset, we benefit from its extensive knowledge. To fine-tune the model for our specific dataset, we freeze the layers to preserve the learned representations. The combination of residual blocks and Inception modules allows the network to effectively capture multi-scale features and learn intricate representations. The skip connections in the residual blocks address the vanishing gradient problem and facilitate smooth gradient flow during training. With its exceptional performance on image classification, InceptionResNetV2 serves as an excellent choice for deep learning practitioners seeking accurate and efficient models.
MobileNet
By utilizing transfer learning with the MobileNet model, we can leverage pre-trained weights and significantly reduce the training time required for our image classification task. This approach is particularly useful when working with limited training data, as we can benefit from the rich representations learned by the base model on a large-scale dataset like ImageNet.
InceptionV3
To implement InceptionV3, we start by loading the pre-trained model, which comes with weights learned from the ImageNet dataset. We freeze the layers of the pre-trained model to prevent them from being updated during training, preserving their valuable representations. Next, we add custom layers on top of the pre-trained model, including BatchNormalization, Dense, Dropout, and a final Dense layer with softmax activation for classification. These additional layers enable us to adapt the model to our specific dataset. Finally, we compile the model by specifying an optimizer, a suitable loss function, and metrics for evaluation. This compilation step prepares the model for fine-tuning and training on our dataset.
vgg16
I will utilize the VGG16 (Visual Geometry Group) architecture, which have deeper and complex structures. These models are renowned for their exceptional performance on various image recognition tasks. By leveraging the pre-trained weights of VGG, I can benefit from the learned features and fine-tune the network for image segmentation on the Lemon Quality Dataset.
Accuracy Comparison
Model |
Accuracy |
|---|---|
| VGG16 | 97% |
| InceptionV3 | 88% |
| InceptionResNetV2 | 88% |
| MobileNet | 97% |
Since the models' decent levels of accuracy(88% and above) means that most of their pictures will be almost havinG similar predicted labels with a small room for mistake, the anticipated labels for the sign image labels are as are visualised as follows.
Throughout the project,
I will preprocess the dataset by resizing the images and splitting it into training,validation and testing sets. For training, I will employ a loss function suitable for image segmentation, such as cross-entropy loss, and optimize the models using technique like Adam optimization
After training the models,
I will evaluate their performance using appropriate metrics. Additionally, I will visualize the segmentation results to gain insights into how well the models can accurately identify and classify different regions within the vegetable images.
after evaluation, MobileNet or VGG16 model looks to be the best fit model in this case of American Sign Language Classification .
Future Scope
ON DEPLOYMENT THESE MODEL WILL HAVE A PURPOSE OF DEVELOPING SENSE AND INTERPRETATION OF THE SIGN LANGUAGES