Identity card identification of pattern recognition

In the identification of various documents, identity card recognition is relatively simple, because the font and position fixed, and the color and background of a large difference.

Identification process:

                                          

(1) Input image (2) ID image coarse positioning and segmentation (3) fine positioning of various information and segmentation This step can be selected or not extracted on demand

( only part of the information is extracted, other similar )

(4) Identification of various information

preprocessing of the input image : The preprocessing of ID image refers to the grayscale and de-noising of the acquired ID image, in order to improve the quality of the image of the identity card image, while preserving and enhancing the information of the texture and color in the ID card, removing the noise that may affect the texture and color information of the identity card area. , it is convenient for locating the identity card image. The main methods: image grayscale, image grayscale stretching, spatial filtering median filtering.

ID Image Coarse positioning : So that the original image after the processing of various algorithms can clearly display the identity card image area, while the image of the non-ID area weakened, so as to accurately and effectively locate the identity card in the image position. The algorithms are: edge detection, mathematical morphology, location based on texture analysis, line detection and edge statistics, genetic algorithm, Hough change and Contour line method, Wavelet transform based method, neural network method, etc.

fine positioning of various information : Because the identity card information location fixed, you can directly specify the ROI region sub-image, or vertical projection and horizontal projection positioning.

identification : This step is also part of a larger computational volume. Methods there are template matching character recognition algorithm, statistical feature matching method (13-line method), neural network character recognition algorithm, svm+ feature extraction, and recent comparison of fire deep learning CNN. The recognition rate of template matching is low in testing, and it is better to use svm+ feature extraction.

Code can not be open, we all know ~