Abstract:
Handwritten word recognition is an active research area due to numerous commercial applications in offline and online recognition systems. The diversity and complexity of...Show MoreMetadata
Abstract:
Handwritten word recognition is an active research area due to numerous commercial applications in offline and online recognition systems. The diversity and complexity of Persian handwritten words makes them more difficult to recognize. In current methods, discriminative features are manually extracted from images by humans so their performance depends on human creativity. This process is called shallow learning. In this study, deep Convolutional Neural Networks (CNNs), a widely used type of deep learning, is employed to automatically extract the discriminative features. Deep learning is able to discover complex structure (discriminative feature here) in large datasets. First in the proposed method, a preprocessing algorithm converts the images to equal size while maintaining handwritten words structure. Then, the images are given to two different architectures of CNNs, AlexNet and GoogLeNet with and without batch normalization. Finally, the proposed method is evaluated on “IRANSHAHR” dataset which includes 15383 images of 503 different city names of Iran. Experimental results show that GoogLeNet with preprocessed data and batch normalization achieves higher accuracy (99.13%) and outperforms the current methods.
Date of Conference: 25-27 October 2017
Date Added to IEEE Xplore: 26 March 2018
ISBN Information: