Abstract:
In recent years, there has been much interest in automatic character recognition. Between handwritten and printed forms, Handwritten Character Recognition (HCR) is more challenging. A handwritten character written by different persons is not identical but varies in both size and shape. Numerous variations in writing styles of individual character make the recognition task difficult. The similarities in distinct character shapes, the overlaps, and the inter-connections of the neighboring characters further complicate the problem. Recently, the Convolutional Neural Network (CNN) has been shown noticeable success in the area of image-based recognition, video analytics, and natural language processing due to their unique characteristics of feature extraction and classification. This is mainly due to the fact that the design of a CNN is motivated by the close imitation of visual mechanism as compared to the conventional neural network. The convolution layer in a CNN performs the similar filtering function that is seen in the cells of visual cortex. As a result of replication of weight configuration of one layer to the local neighboring receptive field in the previous layer through the convolution operation, the features extracted by the CNN possess the invariance properties of scale, rotation, translation and other distortions of a pattern. A recently reported HCR technique that considers the Bangla characters uses shallow CNN by considering only two-level convolution layers and a fixed kernel size experimented on a small-size private dataset. In this thesis, a Deep CNN with three convolutional layers with different kernel sizes in different convolutional layers is used on a large dataset made of combining two datasets. Experimental result shows an accuracy in recognition that is 7% higher than that of previous work.
