M.S. AAI Capstone Chronicles 2024

A.S.LINGUIST

14

Results

CNN Model

As demonstrated by the learning curves in Figure 5, the CNN model was proved to be

effective in achieving its goal. Training and validation accuracy/loss progressively

increases/decreases with the number of epochs until reaching a plateau. This means that the

model does not overfit or underfit: it properly learned from the training data but it is also able to

generalize well on validation data (Bnomial, n.d.).

Figure 5

Representation of training (blue) and validation (red) accuracy (left panel) and loss (right panel)

over the total number of epochs for the CNN model

After training the CNN model, we verified its performance on the testing set, which

includes 8700 different sign language images. We observed the model capability to predict the

letters corresponding to the test images. Thus, we obtained that the macro/weighted averages of

the test precision, recall and f1-score are equal to 0.99 for all the classes and the model accuracy

is 0.99 as well. The confusion matrix calculated on the testing set is also given in Figure 6. The

194