Handwritten Text Recognition (HTR) system implemented with TensorFlow (TF) and trained on the IAM off-line HTR dataset. This Neural Network (NN) model recognizes the text contained in the images of segmented words as shown in the illustration below. As these word-images are smaller than images of complete text-lines, the NN can be kept small and training on the CPU is feasible. More than 86% of the samples from the validation-set are correctly recognized. I will give some hints how to extend the model in case you need larger input-images or want better recognition accuracy.
Go to the model/
directory and unzip the file model.zip
(pre-trained on the IAM dataset).
Afterwards, go to the src/
directory and run python main.py
.
The input image and the expected output is shown below.
Tested with TF 1.3 on Ubuntu 16.04.
> python main.py
Init with stored values from ../model/snapshot-13
Recognized: "little"
The data-loader expects the IAM dataset (or any other dataset that is compatible with it) in the data/
directory.
Follow these instructions to get the dataset:
- Register for free at: http://www.fki.inf.unibe.ch/databases/iam-handwriting-database
- Download
words.tgz
- Download
words.txt
- Put
words.txt
into thedata/
directory - Create the directory
data/words/
- Put the content (directories
a01
,a02
, ...) ofwords.tgz
intodata/words/
- Go to
data/
and runpython checkDirs.py
for a rough check if everything is ok
If you want to train the model from scratch, delete the files contained in the model/
directory.
Otherwise, the parameters are loaded from the last model-snapshot before training begins.
Then, go to the src/
directory and execute python main.py train
.
After each epoch of training, validation is done on a validation set (the dataset is split into 95% of the samples used for training and 5% for validation as defined in the class DataLoader
).
Training on the CPU takes 6 hours on my system (VM, Ubuntu 16.04, 8GB of RAM and 4 cores running at 3.9GHz).
The expected output is shown below.
> python main.py train
Init with new values
Epoch: 1
Train NN
Batch: 1 / 500 Loss: 113.333
Batch: 2 / 500 Loss: 40.0665
Batch: 3 / 500 Loss: 24.2433
Batch: 4 / 500 Loss: 21.644
Batch: 5 / 500 Loss: 22.2018
Batch: 6 / 500 Loss: 18.6628
Batch: 7 / 500 Loss: 20.9978
...
Validate NN
Batch: 1 / 115
Ground truth -> Recognized
[OK] "," -> ","
[ERR] "Di" -> "D"
[OK] "," -> ","
[OK] """ -> """
[OK] "he" -> "he"
[OK] "told" -> "told"
[OK] "her" -> "her"
...
Correctly recognized words: 86.34782608695653 %
Either you convert your dataset to the IAM format (look at words.txt
and the corresponding directory structure) or you change the class DataLoader
according to your dataset format.
The model is a stripped-down version of the HTR system I implemented for my thesis. What remains is what I think is the bare minimum to recognize text with an acceptable accuracy. The implementation only depends on numpy, cv2 and tensorflow imports. It consists of 5 CNN layers, 2 RNN (LSTM) layers and the CTC loss and decoding layer. The illustration below gives an overview of the NN (green: operations, pink: data flowing through NN) and here follows a short description:
- The input image is a gray-value image and has a size of 128x32
- 5 CNN layers map the input image to a feature sequence of size 32x256
- 2 LSTM layers with 256 units propagate information through the sequence and map the sequence to a matrix of size 32x80. Each matrix-element represents a score for one of the 80 characters at one of the 32 time-steps
- The CTC layer either calculates the loss value given the matrix and the ground-truth text (when training), or it decodes the matrix to the final text with best path decoding (when inferring)
- Batch size is set to 50
Around 86% of the words from IAM are correctly recognized by the NN. If you need a better accuracy, here are some ideas on how to improve it:
- Data augmentation: increase dataset-size by applying random transformations to the input images. At the moment, only random distortions are performed
- Remove cursive writing style in the input images (see DeslantImg)
- Increase input size (if input of NN is large enough, complete text-lines can be used)
- Add more CNN layers
- Replace LSTM by multidimensional LSTM
- Decoder: either use vanilla beam search decoding (included with TF) or use word beam search decoding (see CTCWordBeamSearch) to constrain the output to dictionary words
- Text correction: if the recognized word is not contained in a dictionary, search for the most similar one