![tamil letters tamil letters](https://oppidanlibrary.com/wp-content/uploads/2017/04/Tamil-Letters-Concept.gif)
After each batch the model was validated over validation dataset. The CNN was trained in batches of 128 samples from traning dataset. Output layer of 155 neurons(corresponding each tamil alphabet) - with activation function softmax. It was follwed by a hidden layer of 500 neurons - with activation function tanh, and ModelĬNN :- The were two sets of, Conv + ReLu + MaxPooling neural layers. The model are their accuracy of prediction after traning are as follows. We trained totally four model of neural networks, for recognizing the tamil charaset. Those samples had to be discarded, based on the structural similarity with a standard tamil font set. Some of the samples were extremely damaged, using them for traning will actually worsen the learned model. The samples, were resized to 30x30 and conveted to. The samples were centered(based on contour detection) and padded with 10 pixels on all the sides. The sample images were preprocessed and the preprocessed images were of size 30x30. The whole dataset was divided into three sets, training, validation, test datasets. The dataset contained 77.7k samples collected from 170 users, for 155 alphabets(approx. The tamil alphabets dataset was downloaded from Isolated Handwritten Tamil Alphabet Dataset. Currently starting out as a simple OCR project that recognizes tamil characters using a deep Convolutional Neural Network model similar to LeNet5. Research on Deep Models for recognizing tamil characters and translation.