Research on Data Enhanced Ancient Pictogram Recognition Method Based on Convolutional Neural Network

Abstract

As the carrier of national culture, words and pictograms record the unique culture and history of each nation, but the number of existing ancient pictogram is very small, and it is difficult to collect them, which makes it difficult for the academic research of ancient pictogram and the recognition by deep learning. In addition, due to the preservation environment and their own particularities, the traditional data enhancement methods will cause problems such as wrong data label, inability to simulate real scenes, etc. So, it can't effectively expand the large-scale data. To solve these problems, this paper proposes a set of data enhancement methods for small data sets and natural scenes. For the small data set enhancement method, firstly, we use artificial data enhancement to enhance original data, and then a limited random affine transform is used to limit the extent and extent of the enhancement. For natural scenes, we use the DCGAN to fuse the natural scene image and the ancient pictogram to simulate the natural environment. Finally, the paper designs a neural network model to recognize the ancient pictogram. It is proved that the data enhancement method can solve the problem of insufficient data, and finally achieve 99% accuracy.