Integrating Wavelet Coefficients and CNN for Recognizing Handwritten Characters
Madhuri Yadav, Ravindra Kr. Purwar · 2018
Convolutional Neural Network (CNN) based image recognition has shown significant progress in recent years. It has achieved state-of-art results in field of pattern recognition. The proposed work uses convolutional network for Hindi handwritten character recognition. In traditional CNN, the raw images are fed as input to the network, which along with relevant information also contain redundant data. This redundant data makes feature extraction complex and increases the training time of the network. To improve the feature learning capacity of CNN, this work integrates wavelets and convolutional network for recognizing Hindi characters. The wavelet coefficients in three directions i.e. horizontal, vertical and diagonal are extracted from raw images and fed as input to the network. These coefficients are trained independently on different networks and the extracted features are merged at dense layer. The proposed framework outperforms the traditional CNN as demonstrated by experiments. It achieves significant results in much lesser time as compared to traditional CNN.