/ Images

Visual Attention based OCR

Visual Attention based OCR


The model first runs a sliding CNN on the image (images are resized to height 32 while preserving aspect ratio). Then an LSTM is stacked on top of the CNN. Finally, an attention model is used as a decoder for producing the final outputs.


Most of our code is written based on Tensorflow, but we also use Keras for the convolution part of our model. Besides, we use python package distance to calculate edit distance for evaluation. (However, that is not mandatory, if distance is not installed, we will do exact match).

Tensorflow: Installation Instructions (tested on 0.12.1)

Distance (Optional):

wget http://www.cs.cmu.edu/~yuntiand/Distance-0.1.3.tar.gz
tar zxf Distance-0.1.3.tar.gz
cd distance; sudo python setup.py install


Note: We assume that the working directory is Attention-OCR.


Data Preparation

We need a file (specified by parameter data-path) containing the path of images and the corresponding characters, e.g.:

path/to/image1 abc
path/to/image2 def

And we also need to specify a data-base-dir parameter such that we read the images from path data-base-dir/path/to/image. If data-path contains absolute path of images, then data-base-dir needs to be set to /.

A Toy Example

For a toy example, we have prepared a training dataset of the specified format, which is a subset of Synth 90k

wget http://www.cs.cmu.edu/~yuntiand/sample.tgz
tar zxf sample.tgz
python src/launcher.py --phase=train --data-path=sample/sample.txt --data-base-dir=sample --log-path=log.txt --no-load-model

After a while, you will see something like the following output in log.txt:

2016-06-08 20:47:22,335 root  INFO     Created model with fresh parameters.
2016-06-08 20:47:52,852 root  INFO     current_step: 0
2016-06-08 20:48:01,253 root  INFO     step_time: 8.400597, step perplexity: 38.998714
2016-06-08 20:48:01,385 root  INFO     current_step: 1
2016-06-08 20:48:07,166 root  INFO     step_time: 5.781749, step perplexity: 38.998445
2016-06-08 20:48:07,337 root  INFO     current_step: 2
2016-06-08 20:48:12,322 root  INFO     step_time: 4.984972, step perplexity: 39.006730
2016-06-08 20:48:12,347 root  INFO     current_step: 3
2016-06-08 20:48:16,821 root  INFO     step_time: 4.473902, step perplexity: 39.000267
2016-06-08 20:48:16,859 root  INFO     current_step: 4
2016-06-08 20:48:21,452 root  INFO     step_time: 4.593249, step perplexity: 39.009864
2016-06-08 20:48:21,530 root  INFO     current_step: 5
2016-06-08 20:48:25,878 root  INFO     step_time: 4.348195, step perplexity: 38.987707
2016-06-08 20:48:26,016 root  INFO     current_step: 6
2016-06-08 20:48:30,851 root  INFO     step_time: 4.835423, step perplexity: 39.022887

Note that it takes quite a long time to reach convergence, since we are training the CNN and attention model simultaneously.

Test and visualize attention results

The test data format shall be the same as training data format. We have also prepared a test dataset of the specified format, which includes ICDAR03, ICDAR13, IIIT5k and SVT.

wget http://www.cs.cmu.edu/~yuntiand/evaluation_data.tgz
tar zxf evaluation_data.tgz

We also provide a trained model on Synth 90K:

wget http://www.cs.cmu.edu/~yuntiand/model.tgz
tar zxf model.tgz
python src/launcher.py --phase=test --visualize --data-path=evaluation_data/svt/test.txt --data-base-dir=evaluation_data/svt --log-path=log.txt --load-model --model-dir=model --output-dir=results

After a while, you will see something like the following output in log.txt:

2016-06-08 22:36:31,638 root  INFO     Reading model parameters from model/translate.ckpt-47200
2016-06-08 22:36:40,529 root  INFO     Compare word based on edit distance.
2016-06-08 22:36:41,652 root  INFO     step_time: 1.119277, step perplexity: 1.056626
2016-06-08 22:36:41,660 root  INFO     1.000000 out of 1 correct
2016-06-08 22:36:42,358 root  INFO     step_time: 0.696687, step perplexity: 2.003350
2016-06-08 22:36:42,363 root  INFO     1.666667 out of 2 correct
2016-06-08 22:36:42,831 root  INFO     step_time: 0.466550, step perplexity: 1.501963
2016-06-08 22:36:42,835 root  INFO     2.466667 out of 3 correct
2016-06-08 22:36:43,402 root  INFO     step_time: 0.562091, step perplexity: 1.269991
2016-06-08 22:36:43,418 root  INFO     3.366667 out of 4 correct
2016-06-08 22:36:43,897 root  INFO     step_time: 0.477545, step perplexity: 1.072437
2016-06-08 22:36:43,905 root  INFO     4.366667 out of 5 correct
2016-06-08 22:36:44,107 root  INFO     step_time: 0.195361, step perplexity: 2.071796
2016-06-08 22:36:44,127 root  INFO     5.144444 out of 6 correct

Example output images in results/correct (the output directory is set via parameter output-dir and the default is results): (Look closer to see it clearly.)

Format: Image index (predicted/ground truth) Image file

Image 0 (j/j): example image 0

Image 1 (u/u): example image 1

Image 2 (n/n): example image 2

Image 3 (g/g): example image 3

Image 4 (l/l): example image 4

Image 5 (e/e): example image 5


  • Control

    • phase: Determine whether to train or test.
    • visualize: Valid if phase is set to test. Output the attention maps on the original image.
    • load-model: Load model from model-dir or not.
  • Input and output

    • data-base-dir: The base directory of the image path in data-path. If the image path in data-path is absolute path, set it to /.
    • data-path: The path containing data file names and labels. Format per line: image_path characters.
    • model-dir: The directory for saving and loading model parameters (structure is not stored).
    • log-path: The path to put log.
    • output-dir: The path to put visualization results if visualize is set to True.
    • steps-per-checkpoint: Checkpointing (print perplexity, save model) per how many steps
  • Optimization

    • num-epoch: The number of whole data passes.
    • batch-size: Batch size. Only valid if phase is set to train.
    • initial-learning-rate: Initial learning rate, note the we use AdaDelta, so the initial value doe not matter much.
  • Network

    • target-embedding-size: Embedding dimension for each target.
    • attn-use-lstm: Whether or not use LSTM attention decoder cell.
    • attn-num-hidden: Number of hidden units in attention decoder cell.
    • attn-num-layers: Number of layers in attention decoder cell. (Encoder number of hidden units will be attn-num-hidden*attn-num-layers).
    • target-vocab-size: Target vocabulary size. Default is = 26+10+3 # 0: PADDING, 1: GO, 2: EOS, >2: 0-9, a-z