K. Manjusha

Hindi Character Segmentation in Document Images using Level set Methods and Non-linear Diffusion

is the national language of India, spoken by more than 500 million people and is the second most popular spoken language in the world, after Chinese. Digital document imaging is gaining popularity for application to serve at libraries, government offices, banks etc. In this paper, we intend to provide a study on character binarization and segmentation of Hindi document images, which are the essential pre-processing steps in several applications like digitization of historically relevant books. In the case of historical documents, the document image may have stains, may not be readable, the background could be non-uniform and may be faded because of aging. In those cases the task of binarization and segmentation becomes challenging, and it affects the overall accuracy of the system. So these processes should be carried out accurately and efficiently. Here we experiment level set method in combination with diffusion techniques for improving the accuracy of segmentation in document process task. KeywordsSet Method, Binarization, Segmentation, Convex Optimization.

Novel SVD Based Character Recognition Approach for Malayalam Language Script

The research on character recognition for Malayalam script dates back to 1990’s. Compared to other Indian languages the research and developments on OCR reported for Malayalam script is very less. The character level and word level accuracy of the existing OCR tools for Indian languages can be improved by implementing robust character recognition and post-processing algorithms. In this paper, we are proposing a character recognition procedure based on Singular Value Decomposition (SVD) and k- Nearest Neighbor classifier (k-NN). The proposed character recognition scheme tested with the dataset created from Malayalam literature books and it could classify 94% of character images accurately.

Convolutional Neural Networks for the Recognition of Malayalam Characters

Optical Character Recognition (OCR) has an important role in information retrieval which converts scanned documents into machine editable and searchable text formats. This work is focussing on the recognition part of OCR. LeNet-5, a Convolutional Neural Network (CNN) trained with gradient based learning and backpropagation algorithm is used for classification of Malayalam character images. Result obtained for multi-class classifier shows that CNN performance is dropping down when the number of classes exceeds range of 40. Accuracy is improved by grouping misclassified characters together. Without grouping, CNN is giving an average accuracy of 75% and after grouping the performance is improved upto 92%. Inner level classification is done using multi-class SVM which is giving an average accuracy in the range of 99-100%.

Scattering representation in Malayalam character recognition

Feature extraction is the process of mapping input signal to informative representation that can easily be handled by the classifier systems to build decision boundary in between the participating pattern classes. Scattering representation build invariant signal representation by applying a cascade of wavelet decompositions and complex modulus, followed by low-pass filtering. The objective of this paper is to analyze the performance of scattering representation over Malayalam character recognition process. Malayalam character recognizers built from image pixel features and the features extracted from scattering network are tested over real world document images. Soft-max Regression classifier is utilized for building the classification models. Scattering representation based recognition system could achieve a 2% increase in recognition accuracy compared to image pixel value based features.

Integrating scattering feature maps with convolutional neural networks for Malayalam handwritten character recognition

Convolutional neural network (CNN)-based deep learning architectures are the state-of-the-art in image-based pattern recognition applications. The receptive filter fields in convolutional layers are learned from training data patterns automatically during classifier learning. There are number of well-defined, well-studied and proven filters in the literature that can extract informative content from the input patterns. This paper focuses on utilizing scattering transform-based wavelet filters as the first-layer convolutional filters in CNN architecture. The scattering networks are generated by a series of scattering transform operations. The scattering coefficients generated in first few layers are effective in capturing the dominant energy contained in the input data patterns. The present work aims at replacing the first-layer convolutional feature maps in CNN architecture with scattering feature maps. This architecture is equivalent to utilizing scattering wavelet filters as the first-layer receptive fields in CNN architecture. The proposed hybrid CNN architecture experiments the Malayalam handwritten character recognition which is one of the challenging multi-class classification problems. The initial studies confirm that the proposed hybrid CNN architecture based on scattering feature maps could perform better than the equivalent self-learning architecture of CNN on handwritten character recognition problems.