The license plate recognition technology refers to the technology of automatically reading the license plate number, license plate type, license plate color and other information after the vehicle image or video sequence captured by the camera is processed by algorithms such as machine vision, image processing and pattern recognition. It is an artificial intelligence technology. Important branch. Its hardware foundation includes triggering equipment, camera equipment, lighting equipment, image acquisition equipment, and processor for identifying license plate numbers. Its software core includes algorithms such as license plate positioning, character segmentation, and character recognition.
At present, it has been widely used in various occasions of intelligent transportation systems, such as highway tolls, parking management, weighing systems, traffic guidance, traffic law enforcement, highway inspection, vehicle scheduling, vehicle detection, etc. It is of practical significance to maintain traffic safety and urban security, prevent traffic jams, and realize fully automated traffic management.
License plate recognition process
1 image acquisition
According to different vehicle detection methods, image acquisition is generally divided into two types. One is image acquisition in static mode. The vehicle triggers a ground sensing coil, infrared or radar and other devices to give the camera a trigger signal. The camera receives the trigger signal An image will be captured afterwards. The advantage of this method is that it has a high trigger rate and stable performance. The disadvantage is that it needs to cut the ground and lay the coil, which requires a large amount of construction. The other is image acquisition in video mode. The video stream image will be recorded in real time. The advantage of this method is that it is convenient to construct, there is no need to cut the ground to lay the coil, and there is no need to install components such as car detectors, but its disadvantages are also very significant. The rate and recognition rate are lower than the peripheral trigger.
2 pretreatment
Since image quality is easily affected by factors such as lighting, weather, camera position, etc., it is necessary to do some preprocessing on the camera and image before recognizing the license plate to ensure the clearest image of the license plate. Generally, conclusions will be drawn based on the analysis of the on-site environment and the images that have been taken, and the camera ’s automatic exposure processing, automatic white balance processing, automatic backlight processing, and automatic burst processing will be implemented, and the image will be filtered for noise, contrast enhancement, Image scaling and other processing.
Denoising methods include mean filtering, median filtering, and Gaussian filtering; contrast enhancement methods include contrast linear stretching, histogram equalization, and homomorphic filters; the main methods of image scaling are nearest neighbor interpolation and bilinear interpolation Method and cubic convolution interpolation.
3 License plate positioning
Accurately detecting the license plate area from the entire image is an important step in the license plate recognition process. If the positioning fails or the positioning is incomplete, it will directly lead to the final recognition failure. Due to the complex image background and the unclear license plate positioning must be considered, it is easy to treat noise such as fences and billboards as license plates, so how to eliminate these fake license plates is also a difficult point for license plate positioning. In order to improve the accuracy of positioning and improve the recognition speed, the general license plate recognition system will design an external interface, allowing users to set different recognition areas according to the site environment.
4 License plate correction
Due to the influence of factors such as shooting angle, lens, etc., the license plate in the image is deformed such as horizontal tilt, vertical tilt, or trapezoidal distortion, which brings difficulties to the subsequent recognition process. If the license plate correction process is performed after the license plate is located, it is beneficial to remove noise such as the license plate frame and is more conducive to character recognition.
At present, the commonly used correction methods are: Hough transform method, which calculates the inclination angle by detecting the straight lines of the license plate up and down, left and right borders; rotary projection method, by vertically projecting the image on the horizontal axis at different angles, the projection value is the sum of 0 points The maximum angle is the vertical tilt angle, and the calculation method for the horizontal angle is similar to it. The principal component analysis method is based on the characteristic that the color at the boundary of the license plate background and the character has a fixed collocation. Is the horizontal tilt angle of the license plate; the minimum variance method, which derives the closed expression of the vertical tilt angle according to the coordinate variance of the projected point of the character in the vertical direction to determine the vertical tilt angle; perspective transformation, using the four vertices of the detected license plate to pass After the correlation matrix is transformed, the license plate distortion correction is realized.
5 character split
After locating the license plate area, since you do not know how many characters there are in the license plate, the positional relationship between the characters, and the width and height of each character, so in order to ensure that the license plate type matching and character recognition are correct, character segmentation is essential Step.
The main idea of character segmentation is that, based on the binarization result of the license plate or the edge extraction result, using the structural characteristics of the characters, the similarity between the characters, the spacing between the characters and other information, on the one hand, the individual characters are extracted separately, including adhesion and Handle special cases such as broken characters; on the other hand, classify characters with similar width and height into one category to remove the license plate border and some small noises. The commonly used algorithms are: connected domain analysis, projection analysis, character clustering and template matching.
6 character recognition
The grayscale image of the segmented characters is normalized, feature extraction, and then machine learning or matching with the character database template, and finally the result with the highest matching degree is selected as the recognition result.
The currently popular character recognition algorithms are: template matching method, artificial neural network method, support vector machine method and Adaboost classification method. The advantages of the template matching method are fast recognition and simple method, but the disadvantage is that there are some difficulties in dealing with fractures and fouling; the artificial neural network method has strong learning ability, adaptability, and strong classification ability but is time-consuming; support vector The machine method has better recognition ability for unseen test samples and requires fewer training samples; Adaboost classification method can focus on more important training data, with fast recognition speed and high real-time performance. China's license plate is composed of three characters: Chinese characters, English letters and Arabic numerals, and has a unified style, which is also convenient for the recognition process.
7 Output of license plate recognition results
Output the license plate recognition results in text format, including license plate number, license plate color, license plate type, etc.
For more information, please search for intelligent transportation