download training data from : http://www.ee.surrey.ac.uk/CVSSP/demos/chars74k/

In transfer_dataset.py

1.set number of validation images _NUM_VALIDATION

2.set image size _IMAGE_SIZE

3.set dataset_dir to your dataset directory. The structure should be like /path/class1/xxx.png /path/class2/xxx.png or else you have to change some code.

4.it will generate xxx.tfrecord. you can change the name pattern in function _get_dataset_filename

In train.py

1.set FLAGS, usually set depth multiplier, checkpoint_dir, num_classes, and image_size. training_set_size. FLAGS.dataset_dir is where you store *.tfrecord

2.run train.py

3.notice that though the network may seem to converge after 2k steps, however it should run at least 10,000 steps to get a reasonable result.

4.in preprocessing_factory, function distorted_bounding_box_crop, change area_range and min_objected_covered to strengthen the ability of size invariance.

In eval.py

1.set the same thing as training.

2.run eval.py (in some operating systems you may have to point out the checkpoint you use by setting model_dir as like model.ckpt-12345 or it will report bug.

In test.py

1.set FLAGS.test_pic as the test_image, when you run the image, it will show you the image and print the character it predict. Click any key to recognize the next one.

in split_char.py

if you have a large image with a lot of characters and you want to sperate them to individual images.

1.run the main function, change the input image in main function, and press the corresponding key. If you see A just press "a" and it will save image to directory "A".