Random Polygon Cover for Oracle Bone Character Recognition

Abstract

Deep Convolutional neural networks are widely used in computer vision research because of their good feature extraction ability, which can often show good performance in related tasks. Performance of deep convolution network models is not only related to its own architecture design, but also closely related to training data. When training model with large dataset and the images are clear and noise in images is less, it can get good result. But in the case of small dataset and low image quality, it is easy to appear that the model can fit the data in the training process and perform badly in testing, that is, overfitting problem. Our work proposes random polygon cover algorithm to simulate the possible damage object and partial content loss in training dataset, which is also a data augmentation technique. We'll use experiments to prove the effectiveness of this approach, while trying to reveal how data augmentation works and how our method differs from dropout.