Segmentation-free Offline Handwritten Malayalam Word Recognition using Transfer Learning Based Deep Neural Network

Offline Handwriting Recognition is a challenging open research problem under the domain of document analysis and recognition. In this paper we propose a method for offline handwritten Malayalam word recognition from degraded scanned images. Here the major complexities arise from variations in writing styles, touching of characters and degraded noisy scanned images. To avoid the challenges while decomposing the words into pseudo-characters, we treat the word as an individual entity for the recognition. We created a dataset by considering the repeated Malayalam words from the degraded scanned documents of Police First Information statemenst and increased the volume of samples by collecting handwritten words of the same lexicon. In this lexicon based work, we extract the holistic features of the word image using convolutional neural network based transfer learning method. Six different transfer learning architectures, namely VGG16, ResNet50, Xception, MobileNetV2, InceptionV3 and DenseN et121 are experimented for feature extraction and VGG16 outperforms the others. Then for improved performance, an ensemble learning technique, Extremely Randomized Trees Classifier is used for feature selection from the extracted features. Finally, an 8-layer Deep Neural Network is used for the recognition and achieved 97.98% accuracy.