Place Recognition project: Signature Detection and Recognition

People: Sobhan Naderi Palizi, Alireza Tavakoli Targhi, Omid Aghazadeh, Jan-Olof Ekhulnd

User takes an image from an urban place using his/her mobile phone and sends it to a server. Some target objects within the image is detected (the street plate) such that location of the person can be identified by recognizing some uniquely discriminant sign within the detected area of interest (the street name).

The MOBVIS project aimed at design and implementation of mobile vision technology in urban scenarios. Many ideas bloomed in this regard (e.g. place recognition using landmarks). Among those theories, our idea was to extract information from street plates in images taken by a cell phone. We took three steps to approach this problem:

1. Detect the street plates in the image (if any).

2. Segment the text region (the street name) out of the detected plate.

3. Recognize the street name by finding the best match of the segmented text with a database of street names.

For detection, we used a version of Boosting method which is invariant to scale and affine transforms to a great extent. Text segmentation is done by a texture transform and multi-threshold tree-based Connected Component technique. Dynamic Time Warping integrated with projected features solves our image matching (text recognition) problem.

Database:

We also compiled a new database to evaluate performance of our method. This database consists of two sets of images. The first set contains 86 images taken by an ordinary cell phone camera with 1280 * 960 pixels of resolution. In the second set, there are 120 images taken by a 2448 * 3264 pixels camera. Each image contains at least (in most of the cases exactly) one street plate inside. Each target plate ideally appears 9 times in the database since we considered 3 different scales and 3 different orientations.

Streets included in the database are marked as red in the map of Graz city. The long tilt street is Herrengasse.

The database contains a wide variety of outdoor images each of which contains a street plate inside and comes with 3 different viewpoints f-45, 0, 45g and 3 scales f large, medium, small}. So, for each street plate we have 9 images in the database. Range of viewpoint and scale variation is represented by these four pictures.