wenhuach/symmetry_text_line_detection

###################################################################

#                                                                                                                                  #

#                   Symmetry-based text detection                                                             #

#                                                                                                                                 #

#                                                                                                                                 #

###################################################################

1. Introduction.

The source code of symmetry-based text detection algorithm. We adopt approximate calculation technology and parallel technology to speed up the proposed algorithm.

For this reason, the probability output has a slightly difference but very close result on ICDAR dataset, compared to the CVPR'15 version.

###################################################################

2. Installation.

a) This code is written for the Windows x64 and Visual Studio 2012, and you may needs OpenCV 2.4.10 and VL-feat. 

b) You should check the config.txt for parameter setting.

   The config.txt is consists of several lines:

   line #1: Dataset Path

   line #2: Mode flag(0 or 1), 0 is testing mode and 1 is training mode

   line #3: The normalized height of input image

   line #4: number of sliding window scales for each octave

   line #5: the minimum sliding window scale, it would be 2^min_scale

   line #6: the maximum sliding window scale, it would be 2^min_scale

   line #7: Only for training stage. This parameter indicate an absolute distance. 

   Pixels whose distance greater than this distance are regard as negative samples.

   line #8: Only for training stage. This parameter indicate an relative distance(relative to the height of ground truth, Eg: if the height of ground truth is 10 and the paprameter is 0.2, the distance is 2). 

   Pixels whose distance less than this distance are regard as positive samples

   line #9: Only for training stage. The maximum negative samples in training stage.

   line #10: Only for training stage. The maximum positive samples in training stage.

   line #11: the bin number of lab channel

   line #12: the bin number of gradient channel

   line #13: the bin number of texture channel. In current version, this parameter is fixed by 58

   line #14: the angle threshold for symmetry line linkage

   line #15: the core number for parallel

   line #16: the shrink step, it should be 2^n

   and We support an example in the package. Note that the setting of config.txt(line #11 ~ line#13) should be keep consistent during training and testing.

c) Prepare the dataset and create folders.

   The directory structure is described as bellow:

   ->Dataset\

->Annotation\ 

->train\

->Feature

->train\

->Images

->train\

->test\

->Model

->Result

    And for annotation files, each of them should contains several lines, and each line should contain 4 numbers: x, y, width, height.

    For more details, please refer to the sample in the package.

###################################################################

3. Getting Started.

 - Make sure to carefully follow the installation instructions above.

 - Set config.txt.

 - Run TextlineDetection.exe config.txt in command line.

###################################################################