A Multilevel CNN Architecture for Character Recognition from Palm Leaf Images

Deep Learning networks have proven its significance in almost all the areas related to object recognition. With the advent of deep learning concepts, there is a drastic improvement for object recognition problems in various machine learning domain. Convolutional Neural Networks (ConvNets or CNNs) are a special category of Neural Networks that have proven very effective in areas such as image recognition and classification. ConvNets have been successful in most of the pattern recognition problems. Due to the high impact produced by deep learning networks in machine learning applications, different variations of CNN have been developing in a competing manner overriding the performance of the just previous version. In this work, a modified version of CNN is proposed constituting multilevel layers of CNN. Here the input image entrenched in the form of a pyramid is given as input to the system. The performance of the proposed work is tested with degraded grantha characters extracted from palm leaf documents and characters from MNIST dataset. Grantha characters from palm leaves documents are chosen to test the performance of the system with very small sized and highly degraded dataset.