The goal is to identify the vowel and the consonant of each character image using Convolutional Neural Networks and its variations. The model was developed in such a way that given an image of hindi alphabet it will classify both the vowel and consonant present in it.
Each image in dataset is of 64x64 pixels and contains the character to be classified superimposed over a random scenic background. Training set comprises of 10,000 images.10 vowels and 10 consonants have been selected for the problem and dataset of 10,000 images is made using all possible permutations of vowel-consonant pairs
In all the standard CNN models the convolution part was pre-trained and the fully connected part was specifically designed to cater to the problem to classify both the vowel and consonant.On the feature maps generated from the various pre-trained CNN models two different fully connected layers were created for vowels and consonants respectively. The models were trained for 16 epochs using ADAM optimizer
After trying various variations of vanilla CNN,ResNet50 and VGG16 it was found that ResNet50 gave comparatively better result though with considerable overfitting happening
Accuracy Metrics
Model | Training Accuracy | Testing Accuracy |
---|---|---|
Vanilla CNN | 19.20 | 17.20 |
VGG16 | 40.53 | 37.70 |
ResNet50 | 64.29 | 58.40 |