HCCR

A Flux implementation for offline Handwritten Chinese Character Recognition

Requirements

• Flux
• Images
• BSON
• JLD2
• CuArrays

Get started

1. Clone the repository git clone https://github.com/afternone/HCCR.git
2. Download the trained_model.bson from here (access code: 47kh) and put it in the root directory HCCR

Usage

Command Line mode (PowerShell)

Open the PowerShell and change to the directory HCCR

• Single image

PS C:\HCCR> julia test.jl image/01.png
image/01.png    不

• Multiple images

PS C:\HCCR> julia test.jl image/01.png image/02.png
image/01.png    不
image/02.png    忘

• All images in a directory

PS C:\HCCR> julia test.jl (ls .\image\*.png).Fullname
C:\image\01.png       不
C:\image\02.png       忘
C:\image\03.png       初
C:\image\04.png       心
C:\image\05.png       牢
C:\image\06.png       记
C:\image\07.png       使
C:\image\08.png       命

REPL mode

julia> include("test.jl")

julia> recognition(["image/01.png"])
image/01.png    不

julia> recognition(joinpath.("image",readdir("image")))
image\01.png    不
image\02.png    忘
image\03.png    初
image\04.png    心
image\05.png    牢
image\06.png    记
image\07.png    使
image\08.png    命
julia>

Train your model

Download datasets train.jld2 and test.jld2 from here (access code: 47kh) and put them in HCCR. Then train your model as follows:

julia train.jl

The accuracy of the trained_model.bson is 95.22%.