Handwriting Text-line Detection and Recognition in Answer Sheet Composition with Few Labeled Data

Automatic location and recognition are the new trends in today's education industry. However, the dataset based on the scene of English text-line written by school students. Works always focus on text detector' improvement other than a faster detection training period. This article introduces a data synthesis method for synthesizing a dataset that can be used to locate composition text-lines from scanned answer sheet images with few labeled data. The synthetical handwritten text-line dataset includes 5k composite images and more than 2.5k coordinate annotations. This method can also make CTPN text-line detecting network trained from scratch. Besides, Handwriting Recognition (HWR) is a key to revising large-batch English composition on answer sheet. However, handwriting feature is very different from the scene text feature, which challenges the traditional recognition. Hence, this article introduces MLC-CRNN, a refined handwritten text-line recognizer based on CRNN, which can improve recognition accuracy. The proposed method focuses on depth of network and MLC module respectively, and shows can both contribute to handwritten text-line recognition.