A tensorflow implementaion for text detection and text recognition.
DeepLetters consists of two networks.
- SSD(Single Shot MultiBox Detector) is used to detect text regions in images.
- Convolutional Recurrent Neural Network recognizes the textual content of the identified text regions.
First you need to clone the tensorflow/models repository and download the pre-trained SSD model.
$ git clone https://github.com/tensorflow/models.git
$ cd models/research/object_detection
$ wget http://download.tensorflow.org/models/object_detection/ssd_inception_v2_coco_2018_01_28.tar.gz
$ tar zxvf ssd_inception_v2_coco_2018_01_28.tar.gz
After cloning the repository, you should append the tensorflow/models/research/ and slim directories to PYTHONPATH.
# From tensorflow/models/research/
$ export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/slim
Then clone the DeepLetters repository and create symbolic link to tensorflow/models directory.
$ git clone https://github.com/satojkovic/DeepLetters
$ cd DeepLetters
$ ln -s <OBJECT_DETECTION_API_DIR>/ssd_inception_v2_coco_2018_01_28 ssd_inception_v2_coco_2018_01_28
Before training text detector, you need to convert cocotext.v2.json and synthtext gt.mat into tfrecord format.
# COCO dataset
# Train
$ python gen_coco_tfrecord.py --train_or_val train --cocotext_json <DATA_ROOT_DIR_PATH>/cocotext/cocotext.v2.json --coco_imgdir <DATA_ROOT_DIR_PATH>/COCO/images --output_path coco_train.tfrecord
# Validation
$ python gen_coco_tfrecord.py --train_or_val val --cocotext_json <DATA_ROOT_DIR_PATH>/cocotext/cocotext.v2.json --coco_imgdir <DATA_ROOT_DIR_PATH>/COCO/images --output_path coco_val.tfrecord
# SynthText
# Train and Test
$ python gen_synthtext_tfrecord.py --gt_mat_path <DATA_ROOT_DIR_PATH>/SynthText/gt.mat
<DATA_ROOT_DIR_PATH> is the path to root directory of COCO and SynthText datasets.
Now, to start a new training job, type the following command.
$ python <OBJECT_DETECTION_API_DIR>/legacy/train.py --logtostderr --pipeline_config_path=ssd_inception_v2.config --train_dir=training
$ python deep_letters.py --input <input image or video> --detection_model_path <detection_model_pb> --detection_th <th> --recognition_model_path <recognition_model.pth>Download detection model file (pb file) from GoogleDrive.
Clone crnn.pytorch repository and place it on same level with DeepLetters repository. (DeepLetters use crnn.pytorch to recognize texts in detection results)