text_recognition

A ROS nodelet to read and locate texts in an image

Dependencies

object_detection_msgs

• https://github.com/yoshito-n-students/object_detection_msgs

Default deep neural networks

If you want to use default DNNs, run the following commands to download data.

roscd text_recognition/data
./download_data.sh

Size of downloaded files would be 2-3GB in total. May take minutes or hours. If a link in the script is dead, please post an issue (with a valid link if possible).

Nodelet: TextDetector

• former part of text recognition
• subscribe images and detect bounding boxes of word-level texts (text boxes)

Subscribed topics

image_raw (sensor_msgs/Image)

Published topics

image_out (sensor_msgs/Image)

• image in which text boxes are found
• never advertised and published unless ~republish_image is ture

texts_out (object_detection_msgs/Objects)

Parameters

~republish_image (bool, default: true)

• republish image if one text box at least is detected

~architecture_file (string, default: "$(find text_recognition)/data/textbox.prototxt")
~weights_file (string, default: "$(find text_recognition)/data/TextBoxes_icdar13.caffemodel")

• deep neural network to detect text boxes
• valid as arguments for cv::text::TextDetectorCNN::create()

~score_threshold (double, default: 0.4)
~nms_threshold (double, default: 0.5)
~eta (double, default: 1.0)
~top_k (int, default: 0)

• post-filtering to suppress overlapped detection
• valid as arguments for cv::dnn::NMSBoxes()

~image_transport (string, default: "raw")

Nodelet: WordRecognizer

• latter part of text recognition
• subscribe images and corresponding text boxes, and recognize words

Subscribed topics

image_raw (sensor_msgs/Image)

texts_in (object_detection_msgs/Objects)

• text boxes on subscribed images
• timestamp must match that of a subscribed image

Published topics

image_out (sensor_msgs/Image)

• image in which one word at least is recognized
• never advertised and published unless ~republish_image is ture

words_out (object_detection_msgs/Objects)

Parameters

~queue_size (int, default: 10)

• queue size of synchronized subscriber of images and text boxes

~republish_image (bool, default: false)

• republish image if one word at least is recognized

~min_probability (double, default: 0.5)

• minimum probability of published words

~architecture_file (string, default: "$(find text_recognition)/data/dictnet_vgg_deploy.prototxt")
~weights_file (string, default: "$(find text_recognition)/data/dictnet_vgg.caffemodel")
~words_file (string, default: "$(find text_recognition)/data/dictnet_vgg_labels.txt")

• deep neural network to recognize a word in a text box
• valid as arguments for cv::text::OCRHolisticWordRecognizer::create()

~image_transrpot (string, default: "raw")

Eamples

see launch/text.launch