+86 138 1226 9927


Home  >  News

Recognition Accuracy Up To 99.8% For License Plates

May. 17, 2019

Here is ANPR Parking System Supplier talking about Recognition Accuracy Up To 99.8% For License Plates.

If you have any idea about ANPR Parking System, welcome to contact us and discuss.

The big-name test uses a ratio of 90:36 license plate to train the AdaBoost detection based on the haar feature.

Prepare the sample:

Positive sample: Sample processing and selection are very skillful. My standard is to frame the entire license plate and leave a border, which preserves the original character characteristics of the license plate, the character group features and the frame features of the license plate. In the two-way license plate, I only take the bottom line. And it is best not to pre-process the sample, and what graphics are used by the input source. The specific method of drawing can refer to the other three aspects of the license plate recognition technology of my blog. The positive and negative samples of the character detection are obtained (using the mouse frame to draw the picture).

Negative samples: Negative sample selection is also very tricky. Try to capture the background image of the license plate usage environment, and need to take a part of the license plate characters but not a positive sample of the negative sample around the license plate.

Anpr Parking System Supplier

Third, the single license plate character segmentation is based on haar+cascade plus logic screening.

1. One of the most difficult modules in image recognition: target detection and segmentation

The most difficult part of the identification application is the segmentation, the image segmentation is good, and the backend is easier to identify.

* It is not necessary to do image pre-processing before detection: it is recommended to be simple according to the actual situation. Commonly used such as cvNorm, but only on the backup image, the original image is not moved as much as possible, and the original image is left for identification.

* Train a classifier for target detection. Take haar+adaboost as an example. Refer to the positive and negative samples of character detection for details (using the mouse to draw a picture) and prepare samples.

2. Often the classifier can only get the above preliminary effect. At this time, some adjustments need to be made according to the characteristic rules inherent in the actual image of the project.

Fourth, the recognition supports blstm+ctc full image recognition, single character segmentation recognition, and FCN full convolution recognition.

Previous: None

Next: Valet Parking