International Core Journal of Engineering 2020-26 | Page 40

2019 International Conference on Artificial Intelligence and Advanced Manufacturing (AIAM) An English Handwriting Evaluation Algorithm Based on CNNs Yingguo Gao Runze Yu Xiaohui Duan Center for Wireless Communication and Signal Processing EECS, Peking University Beijing, China e-mail: [email protected] Center for Wireless Communication and Signal Processing EECS, Peking University Beijing, China e-mail: [email protected] Center for Wireless Communication and Signal Processing EECS, Peking University Beijing, China e-mail: [email protected] deformation are poor, as well as the conformity for similar characters. The methods based on structural features, such as skeleton, outline, shape features, etc., are robust to deformation, but have a very high complexity. So, a more general approach which automatically learns effective visual representations for handwritten document images is desired. Abstract—English handwriting evaluation is an essential part in elemental English teaching. An automatic evaluation algorithm for English handwriting quality is proposed in this paper. Generally, conventional document image processing approaches rely on hand-crafted features for capturing statistical or structural information. In contrast, we take advantage of Convolutional Neural Networks (CNNs) for extracting features from raw image pixels. The performance of this algorithm is more effective than traditional machine learning methods and the accuracy is greater than 94% in our experiment. Based on this algorithm, an intelligent English handwriting marking system is designed and it is already online. Recent advances in deep learning have made it possible to extract high-level features from raw sensory data, leading to breakthroughs in computer vision. Yann LeCun et al. [7] first proposed to use Convolutional Neural Networks for handwritten digit recognition. Le Kang et al. [8] presents a general approach for document classification using Convolutional Neural Networks. Xu-Yao Zhang et al. [9] set new benchmarks for both online and offline handwritten Chinese character recognition applying deep learning. Aiquan Yuan et al. [10] proposed a method for offline handwritten English character recognition based on Convolutional Neural Networks. Keywords—English handwriting quality, automatic evaluation, Convolutional Neural Networks, marking system I. I NTRODUCTION English handwriting quality evaluation has been a hot topic in elemental English teaching. One of the important method for assessment is English calligraphy competition. Current evaluation methods almost depend on pure manual work, which have some significant disadvantages, such as different evaluation standards from person to person, relatively low efficiency, and high labor costs. Therefore, the development of automatic evaluation methods has recently attracted much attention due to the high accuracy and efficiency. In this paper, an automatic evaluation algorithm is presented for English handwriting quality using Convolutional Neural Networks (CNNs). Experiments on real-world data set show that our approach is more effective than conventional machine learning methods. Moreover, an intelligent English handwriting marking system based on our algorithm is designed and put into use. The reminder of this paper is outlined as follows. Section II describes the system model, while Section III describes CNNs for English handwriting evaluation. In section IV, we present our experimental results after applying our algorithm on our data set. Conclusions are drawn in Section V. Document image processing has been widely used in the field of character recognition, text detection, and document image classification, but rarely seen in handwriting-quality evaluations. In the current literature, Chongbiao Zhuang [1] and Yi Peng [2] have studied the quality evaluation algorithm of handwritten Chinese characters. The differences between the two above algorithms mainly lie in the feature extraction methods. In the paper [1], the feature extraction algorithm is based on Gabor transform combining with the rule-based feature extraction method. In contrast, the algorithm in [2] is based on direction rules and connected component detections. II. S YSTEM M ODEL We design an intelligent English handwriting marking system based on CNNs. Our system consists of four parts: one is data source interface. Data set is divided into two categories, award-winning and non-award-winning. There are about 24000 samples. The second part is data preparation. We detect, cut, and calibrate the English writing area for standard data and label input. A Classification model based on CNNs is described in the third part. We first design the network architecture according to our application scenario. After training and refining, we got the final model. The last part is an application platform. We integrated the algorithms and setup a web server. The whole system model is depicted in Fig. 1. Feature extraction is an important part of handwritten document image processing. Previous approaches for feature extraction most rely on hand-crafted features for capturing statistical or structural information. The methods based on statistical features, such as template matching, zoning [3], moments [4, 5], n-tuples [4, 6] etc., have a low complexity and a simple training. However, their abilities to resist 978-1-7281-4691-1/19/$31.00 ©2019 IEEE DOI 10.1109/AIAM48774.2019.00010 18