siyiding1216 / dl-ocr Goto Github PK

Run main.m to start the driver licesne recognition or create an atlas.

Require:

1. Matlab image processing toolbox.
2. Tesseract binary.

Step 1. Create Atlas (template). An atlas needs to be created for each type / category of driver licenses.

An atlas contains the following data, template image: a driver license picture taken in a good lighting condition ROIs: a list of rectanglar regions, each for a specific character list to be recognized. For example, the rectangle surrounding the last name (LN) in california license. foreground image: The parts in template image that are the unchanged parts across different licenses. The reason to have this is that we want to exclude any parts that are variant to each different person, like photo, all to be recognized characters.

Step 2: Landmarks detection and correspondence finding. When a test image is provided, landmarks at image scale space are extracted from both the template and test image. The 64 feature vectors are also computed for these landmarks. The Eucledean distance between these feature vectors are used to establish point correspondence. We then took x number of most similar point pairs.

Step 3: Landmark registration. Once the point correspondence are established, a point registration is done to compute a transformation that can warp the test image to the reference image. Affine transformation is used here since the photo could be taken potentially at different angles. The transformation then warp the test image to reference space where we can apply directly the ROI coordinates to take out the parts that need to be recognized.

Step 4. Recognize image patches using tesseract. tesseract is called externally to recognize each image patch. Image patch are further preprocessed with histogram equalization and known character pattern is used as prior for tesseract to work more accurately.