Deep Learning Based Marathi Sentence Recognition using Devnagari Character Identification

There are multiple algorithms available to recognize Marathi Devnagari characters. Most of these methods are limited because of the large variety of character variations due to Kana, Matra, Ukar, Velanti, and Anusvar, which are specific to the Marathi grammar called Barakhadi. There is a need to have a dictionary-based word formulation to achieve full Marathi sentence recognition. In the proposed work, a Marathi sentence is recognized using a combination of full 454 variation detection of Devnagari characters and nearest dictionary word mapping using the k-nearest neighbour (KNN) model to achieve full sentence recognition. This is the first time full 454 (Vyanjan variation as per Barakhadi) character recognition instead of the traditional 58 characters (Vyanjans) has been attempted which leads to sentence recognition. The proposed method could achieve a sentence recognition accuracy of 86.84%, a 454 character classification accuracy was 89.52%, and the execution speed of the proposed system was 1.464 secs per word. For the training of the character recognition network, a separate dataset was created for all Vyanjan variations as per Barakhadi. This novel contribution of the proposed system will surely inspire researchers to explore Devnagari sentence recognition.