The algorithm for detecting the presence of text in the image

Question

With my new appointment, I am looking for a way to detect the presence of text in the image. An image is a map — for example, a google map. The task is to determine where the street / city mark is located.

I know that the opencv library has an algorithm that can detect functions (for example, human faces) - a classifier for a hara or a pig (a histogram of oriented gradients), but I heard that the process of learning such algorithms is quite complicated.

Do you know any algorithm, method or library that could do this (detect the presence of text in the image)?

Thanks, John

Answer 1

There is a standard problem in vision called text detection in images. this is a completely different difference from OCR. OCR agrees with what it says, while text detection is related to determining the presence of text in an image. The third Adi Shavit link is a method to solve this problem. You can find a well-quoted article from a google scientist in text detection .

Answer 2

There are several possible approaches.

• Use OCR. Searching for OCR in Stackoverflow will show many options. These include Tesseract and Ocropus .
• If your text uses a very specific fixed font, you can get away with a simple pattern matching .
• More generally, you can take a look at " Detecting text in natural scenes with stroke width conversion "

UPDATE January 2017
The OpenCV 3.2 contrib module now has a text detection module .
It also contains a sample on how to use it.

Answer 3

You need to configure this for a certain type of image on the map, or the problem will be very complicated (see the previous post about article links).

OCR is the way to go and you must use the existing library. However, OCR is mostly performed by text on a white background. To reduce the problem to a normal OCR problem, you should try to work with the color space on the map. The map text is likely to have a very specific color, and this may be enough to find these pixels. Then you can filter the detected pixels depending on the size of the connected areas.

If you literally want to find the location of text labels, you can do it, and pretty much just skip the OCR step. If the labels are not too close, you can find simple clustering algorithms to find their respective positions.