Evaluating synthetic training data for character recognition in natural images
Abstract: This thesis is centered around character recognition in natural images. More specifically, evaluating the use of synthetic font images for training a Convolutional Neural Network (CNN), compared to natural training data. Training a CNN to recognize characters in natural images often demands a large amount of labeled data. One alternative is to instead generate synthetic data by using digital fonts. A total of 41,664 font images were generated, which in combination with already existing data yielded around 99,000 images. Using this synthetic dataset, the CNN was trained by incrementally increasing synthetic training data and tested on natural images. At the same time, different preprocessing methods were applied to the synthetic data in order to observe the effect on accuracy. Results show that even when using the best performing pre-processing method and having access to 99,000 synthetic training images, a smaller set of natural training data yielded better results. However, results also show that synthetic data can perform better than natural data, provided that a good preprocessing method is used and if the supply of natural images is limited.
AT THIS PAGE YOU CAN DOWNLOAD THE WHOLE ESSAY. (follow the link to the next page)