Devanagari Character Recognition Using Image Processing And Deep Learning

This work allows optical character recognition of the Devanagari script, handwritten or printed, written in a paragraph or a line. Image processing techniques enable word and character segmentation, i.e., splitting words from paragraphs and separating characters from a word. The segmented characters are then recognized using a convolutional neural network.

We have also used boosting technique over the neural network for achieving more significant outcomes. We trained five neural networks with different number and type of layers, filters and pool size in the convolutional layer, dropout rate and the number of neurons in the dense layer. Then, while performing recognition, each neural network is first used for prediction separately, and then voting among them is performed. The one which occurred more frequently is chosen as the final prediction, resulting in better precision.

Sample images used for character recognition can be found in the Words directory of the repository.

To perform character recognition, accumulate the images in a directory and then provide the complete path to the directory in Main.py.

You can then either run the code directly on the visual studio using Devanagari-Character-Recognition.sln or can run Main.py.

• Split_Words.py is used for performing word segmentation, helpful in achieving character recognition of paragraphs
• Split_Characters.py is used for performing character segmentation, helpful in achieving character recognition of a word
• Predict_Characters.py is used for taking the prediction of a character using the boosting technique
• Split_Dataset.py is used for splitting the dataset into training and validation sets
• Model_1.py, Model_2.py, Model_3.py, Model_4.py, Model_5.py are used for training different convolution neural networks
• Evaluate_Model.py is used for evaluating and comparing the performance of different convolution neural networks

Gif 01. Word Segmentation

Gif 02. Character Segmentation

Table 01. Classification Report on Validation Data

Table 02. Classification Report on Sample Images

1. Word segmentation enables the character recognition of paragraphs, preserving the order of the words
2. Character segmentation enables the character recognition of a word
3. Image processing enables the segmentation of slanted words and characters, and svar having a matra
4. Data augmentation using image data generator class, rotated, shifted, sheared and zoomed
5. Convolution neural network helps in feature extraction
6. Boosting technique resulting in much more reliable efficiency

1. vyanjans having a matra cannot be segmented
2. Numerals and matra are not included in the dataset and hence can not be determined
3. Words with excess noise like this and highly slanted words might hinder efficiency

• Handwritten Devanagari Characters