Persian OCR with Cascaded Convolutional Neural Networks Supported by Language Model

Persian
¹
OCR is a difficult task because of some specific features of Persian writing style, like different styles of letters in different places of the word and similarity of letters to each other. Recognizing sub-words instead of individual letters can reduce these difficulties. In this manner sub-word segmentation is the critical task of pre-process step. In this paper, a cascaded Convolutional Neural Network is utilized to convert sub-word images into text. A large dictionary of Persian sub-word images with different font styles is used as training data and an Auto-Encoder enriches the features needed for constructing the cascade classifier structure. The initial classifier learns the overall structure of sub-word images that its training data is the result of applying k-means clustering on the huge sub-word image dataset. The later classifier finds the exact text equivalent of the sub-word image. A word segmentation method forms the words based on extracted sub-words. This method use contour distances as a measure for distinguishing words from sub-words. The initial OCR result is improved using Natural Language Processing techniques. Two fast search structures in word dictionaries with the help of a language model build the post-processing module and substitute the misspelled extracted words with the best alternative. Comparison results with Tesseract OCR engine shows the superiority of the algorithm.