Traditional Culture Encyclopedia - Traditional virtues - Research and Classification of License Plate Recognition Algorithm

Research and Classification of License Plate Recognition Algorithm

License plate recognition system should comprehensively use various means to extract the license plate area, accurately locate the car license plate, and finally complete the car license plate recognition. Therefore, the license plate recognition system should deal with various complex environments, such as traffic peak, illumination and reflection, license plate pollution and so on. Artificial neural network simulates the intelligence of human brain, and can carry out associative memory and reasoning when recognizing license plate, which can solve the problem of incomplete and unrecognizable characters.

Research on license plate recognition method

License plate recognition system mainly includes license plate location, character segmentation, character recognition and other working modules. At the same time, the system itself has good maintainability and expansibility, and can realize automatic vehicle detection without adding other special equipment.

Research on license plate location method

License plate location is to completely separate the license plate area from the vehicle image with complex background. It is to solve practical problems in image processing with various methods. At present, the most common positioning technologies are: edge detection, color segmentation, wavelet transform, genetic algorithm and artificial neural network technology.

License plate location method based on edge detection: Before locating the license plate, enhance and preprocess the automobile image by gray transformation and histogram equalization, then binarize it, and finally detect the edge of the image by edge detection operator. After the edge is detected, the area expands and corrodes irrelevant small objects. At this time, the image will present multiple connected judgment areas, and finally find out which one of all connected areas is the most likely license plate.

License plate location method based on color segmentation: it mainly consists of color segmentation, Japanese standard location and other modules. Before color segmentation, the original image needs to be converted from RGB color space to HSV space, and then the color analysis is carried out in HSV space. Specific segmentation operation: based on one of the four license plate background colors, the color component of each pixel in the image is compared with the comparison table 1, and the pixels beyond the definition range of reference color are directly set as the background color (white); Otherwise, the average of the three components of all pixels falling within the interval is counted as the color center of segmentation calculation, and then the color distance between the pixels and the color center of all intervals is calculated. If the distance is greater than the threshold, it is set as the background color and set as the color center.

Because of the complexity of the image background, the color filtered image may still contain multiple possible target areas, so it is necessary to further filter multiple target areas by using the body proportion characteristics of the license plate.

The license plate location method based on wavelet transform: firstly, the vehicle image is converted into an index image, and then the index image is wavelet transformed to get the wavelet coefficients of the image in different subbands. The feature extraction of license plate recognition is based on LH high frequency subband after wavelet transform of automobile image. According to the characteristics of large amplitude and high density of wavelet coefficients in license plate area in the image, the wavelet coefficients in non-license plate candidate areas are filtered by setting thresholds. Through wavelet scale decomposition, edge subgraphs with clear texture, different spatial resolutions and different directions are proposed. Then the sub-image extraction is realized by using the characteristics of low frequency in horizontal direction and high frequency in vertical direction in the daily mark area of license plate. Finally, the mathematical morphology method is used to perform a series of morphological operations on the detail image after wavelet decomposition to further eliminate useless information and noise, so as to determine the location of the license plate.

License plate location method based on genetic algorithm: the main feature of license plate daily mark area is that the background color of license plate is often quite different from the color of car body and text; In addition, the length ratio of license plate changes in a certain range, and there is a maximum and minimum length-width ratio. According to these features, corresponding features can be extracted from gray-scale images. In addition, the interval between characters in the license plate is relatively uniform, and the characters and the background color of the license plate jump in gray value, while the characters themselves and the gray level inside the bottom of the license plate are relatively uniform. Because the license plate has a continuous or discontinuous border due to wear, and there are many characters in the license plate, which are basically arranged horizontally, the rectangular area of the license plate has rich edges and presents regular texture characteristics. Therefore, in practical application, we only need to gray and binarize the color image first, and filter the binary image horizontally by using one-dimensional filter banks reflecting different densities, and then we can get the texture feature vector of the license plate image. Then the image of the local area is filtered to get its feature vector, and compared with the feature description vector of the license plate, the possibility of this area as the license plate area is obtained.

Using neural network to realize the license plate location algorithm: it can be realized by directly sensing the gray image, that is, using the sliding window as the sampling window (long or narrow sliding window can be selected according to the characteristics of the license plate), moving on the gray image in turn, and taking the image blocks covered by the window as the input of the neural network. The BP network is a three-layer fully connected feedforward network, and its input layer neurons are the size of a sliding window for several days, and its output layer neurons are L for several days. When the output is close to half, it means that the image block under the sliding window belongs to the license plate area; When the output is close to half, it means that the image block under the sliding window belongs to the background area. The selection of sample set and the search strategy of the algorithm are very important, which will affect the positioning effect. Therefore, the license plate and Beijing should be sampled alternately and repeatedly, and the selected images should contain different lighting conditions, background complexity and license plate color as much as possible, which is beneficial to the generalization of the network and can strengthen the fault tolerance of the network. As for the search strategy, because the license plate is generally located in the middle and lower part of the image, it is generally traversed from bottom to top, so it is not easy to mistakenly locate the logo as the license plate area, and when there are multiple candidate license plate areas after traversing the image, the lowest candidate license plate area should also be given priority. In addition, because of the fault tolerance of neural network, it is necessary to horizontally correct the license plate with large inclination angle before neural network processing.

Research on Character Segmentation Method

The task of character segmentation is to cut each character in a multi-line or multi-character image into a single character from the whole image. There are many algorithms for character segmentation, and usually different algorithms are adopted according to different processing objects. Common methods mainly include: template matching method, horizontal projection method, clustering analysis method, image segmentation method based on adaptive degradation morphological characteristics and so on. Here we elaborate on the first three methods.

Template matching method: this method first calculates the smooth curve of vertical integral projection on binary image, searches for the local minimum value of the smooth curve, and obtains a series of trough positions; Then, taking two adjacent valleys as the left and right boundaries respectively, a group of rectangular regions are extracted; Finally, according to certain rules, the rectangular area is deleted, divided, merged and resized, so as to realize the single-character segmentation of license plate area.

Horizontal projection method: this method scans the image of license plate area line by line from bottom to top and then from white to top, finds and records the position of the scanned 1 th white pixel respectively, and determines the approximate height range of the image; Within this height range, scan line by line from left to right. When the 1 th white pixel is encountered, it is considered as the starting position of character segmentation, and then scanning continues until a column without white pixels is encountered, which is considered as the result of this character segmentation. Repeat the above process until the right end of the image, and get a more accurate width range of each character: within the known more accurate width range of each character, scan line by line from top to bottom and from bottom to top respectively, and determine the accurate height range of each character.

Cluster analysis: this method is based on the principle that pixels belonging to the same character form a connected domain, combined with prior knowledge, the fixed ratio of the height and spacing of characters, etc. , the characters in the license plate area are segmented one by one.