Identifying Chinese calligraphy with neural nets

Thursday, June 1 2017

I worked with Elliot Kerman to identify Chinese calligraphy using neural networks. We were specifically interested in how accurately we could identify the characters in Poetry on the Baotu Spring.

Calligraphy is an interesting task because the writing is an art form in itself and styles of calligraphy, like art, vary vastly over time.

We used techniques such as convolutional neural networks, generative adversarial networks, and transfer learning. We primarily used Tensorflow and TensorBoard.

Representation calligraphy in 3D space here. (Sadly I’m concerned I may get DDOS’d if I include the original gif)

T-SNE algorithm projecting the model’s representation of 10 characters of Chinese handwriting into 3D space. Each colour is one Chinese character written by many different authors.