DeepConvFarsiOCR

A Deep Convolutional Approach Toward Farsi Character Recognition. Used for both machine printed and handwritten datasets.

Papers

• From machine generated to handwritten character recognition; a deep learning approach

Usage

Step1: Data

• Download all images from this link.
• Extract them to the home directory of this repository. After this, you should see PDB-Train and PDB-Test folders in your home directory.
• ensure that run() function in convert.images.lua is not commented.
• execute convert.images.lua with the following parameters:

th convert.images.lua --src PDB-Test --dest PDB-Test --bin PDB_Test.bin

and

th convert.images.lua --src PDB-Train --dest PDB-Train --bin PDB_Train.bin

• These commands will
  • Convert all bpm images in --src directory into PNG and store them in --dest
  • extract labels from file names and store them with them images in --bin binary file.

Note that all files paths given to --src and --dest should be RELATIVE and should NOT include any / in them.

Step2: Data Sourcing

This step converts the raw data stored in binary files to dp:DataSource. This operation is done using data.source.lua and it is called internally by cnn.v2.lua. Hence, it cannot accepts parameters and should be manually adjusted. The following parameters are important:

local validRatio = .5
local train_bin = './PDB_Train.bin'
local test_bin = './PDB_Test.bin'

• train_bin and test_bin should be qual to --dest parameters in the previous section.
• validRatio indicates what portion of the test dataset should be used for cross-validation.

Step3: Train

• run th cnn.v2.lua --progress. See the source file for more commands and options.

Always use the --id parameter. This name will be used to store the model and logs.