• People who are hearing impaired communication with the help of gesture-energetic movements of the hand accompanied by a living facial expression
• This language is also used at places where listening is not possible for ex... under water drivers, asutronuts….

• This dataset contain total 4972 jpg images in 24 classes A to y. as well as all classes data is imbalanced so we perform oversampling to balance the classes.

• Make subset of training, validation and testing as well as perform oversampling to balance the all classes
• Prepare training, validation and testing set
• Get all classes labels
• Convert [Train, validation, Test] set into (X_train,y_train)(X_valid,y_valid),(X_test,y_test) and convert categorical lebel into binary metrix
• Visualise training images
• Used Resnet50, VGG16 & EfficientnetB3 pre-trained model
• Model Compilation
• Model Training
• Model Evaluation
• Model Saving
• Prediction on test data

• Make a subset of data into three parts train, test, and validation and balance the data of each class with the help of splifolder library.

• Found 3586 images belonging to 24 classes.

• Found 1390 images belonging to 24 classes.

• Found 1365 images belonging to 24 classes.

• After perform oversampling 3586 samples for training, 1390 for validation and 1365 samples goes for testing

{'A': 0,'B': 1,'C': 2,'D': 3,'E': 4,'F': 5,'G': 6,'H': 7,'I': 8, 'K': 9,'L': 10,'M': 11,'N': 12,'O': 13,'P': 14,'Q': 15,'R': 16, 'S': 17,'T': 18,'U': 19,'V': 20,'W': 21,'X': 22,'Y': 23}

• Total 24 classes is present in this data we need to seprate the labels as well as encode the categorical integer into binary matrix

1. RESNET50:

• In resnet50 model we train the existing weights to get a better Result.
• Use dropout to supress overfitting problem as well as use batch Normalization
• At last output layer use softmax activation function because the Task is multiclass classification.

2. VGG16:

• In VGG16 we don’t train the existing weights
• Not use any dropout and batch normalization to check model Performance

3. EFFICIENTNET-B3

• In this model also we train initial weights to get better result.
• Use Dropout and batch normalization to supress overfittingProblem
• Add regularize at dense layer to reduce the losses

1. RESNET50:

• Here we use categorical cross-entropy loss as well as Adamax optimizer.
• Set learning rate and monitor validation loss and train Model on 300 epoch.
• validation loss is decreases after every epoch as well as validation accuracy is increases.

• 4/4 [==============================] - 1s 129ms/step - loss: 0.0067 - accuracy: 1.0000 Train loss & Accuracy: [0.00668429397046566, 1.0]
• 4/4 [==============================] - 1s 128ms/step - loss: 0.3905 - accuracy: 0.8828 Testing loss & Accuracy [0.3905099034309387, 0.8828125]

2. VGG16:

• Here we use categorical cross-entropy loss as well as Adamax optimizer.
• Set learning rate and monitor validation loss and train Model on 300 epoch.
• VGG16 model perform well on training data but in validation side accuracy is extreme less, as well as validation loss is also high

• 4/4 [==============================] - 1s 197ms/step - loss: 0.0771 - accuracy: 1.0000 VGG16 Training Loss & Accuracy [0.07713527977466583, 1.0]
• 4/4 [==============================] - 1s 198ms/step - loss: 1.7651 - accuracy: 0.4766 VGG16 Testing Loss & Accuracy [1.765055537223816, 0.4765625]

3. EFFICIENTNET-B3:

• Here we use categorical cross-entropy loss as well as Adamax optimizer.
• Set learning rate and monitor validation loss and train Model on 300 epoch.
• VGG16 model perform well on training data but loss is very high or validation side accuracy is around 75% , as well as validation loss is also high

• 4/4 [==============================] - 1s 149ms/step - loss: 3.4699 - accuracy: 1.0000 Training Loss & Accuracy [3.4699418544769287, 1.0]
• 4/4 [==============================] - 1s 147ms/step - loss: 4.0854 - accuracy: 0.9062 Testing Loss & Accuracy [4.085390090942383, 0.90625]

• Save model in training time with the help of model check point with .h5 extension

• Resnet50 model predict correctly on test image Not perform any testing on VGG16 because model not perform Well.
• Efficientnet B3 mode do some wrong prediction because of loss is very high.

• Resnet50 model perform well on training as well as testing data and both training and testing loss is less. as well as model predict correct result VGG16 model is not perform good on testing data
• EfficientnetB3 model training & Testing accuracy is good but training and testing loss is high, that's why model predict wrong output From above models we are select Resnet50 model

• Tensorflow
• Keras
• Matplotlib
• Glob
• Numpy
• Splifolder
• Resnet50
• VGG16
• Efficient-net B3

• Jupyter Notebook
• Pycharm