CN118053148A - New energy vehicle identification method and system - Google Patents

Abstract

The present application discloses a new energy vehicle identification method, which can detect license plates accurately and quickly. It can not only locate the license plate position and classify the detection target, but also correct the distorted license plate to improve the subsequent license plate recognition accuracy. It can also identify the specific license plate characters of single-layer license plates or double-layer license plates, and can also identify the color of the license plate to help improve the recognition accuracy of the system. Finally, by judging the length of the license plate characters or the color of the license plate, the vehicle type is distinguished, which further improves the recognition accuracy of new energy vehicles.

Description

A new energy vehicle identification method and system

Technical Field

The present invention relates to the technical field of new energy vehicles, and in particular to a new energy vehicle identification method and system.

Background technique

New energy vehicles are cars that replace traditional cars and use energy from new energy technologies, such as electric cars, plug-in hybrid cars, hydrogen fuel cell cars, etc. Such cars often have advanced power and energy recovery technologies, high operating efficiency, can greatly reduce environmental pollution, and help humans create a cleaner environment.

However, the appearance of cars is complex and objective, so how to objectively and efficiently identify new energy vehicles has become a difficult problem. As we all know, the license plate is one of the important signs to identify the type of vehicle, and the license plates of different types of vehicles will also have clear and different regulations on the encoding rules. Therefore, by identifying the characters on the license plate, it can be determined whether the vehicle is a new energy vehicle. Therefore, how to efficiently and accurately obtain the license plate number has become the key to automatically identifying a large number of motor vehicle types.

The existing mature license plate detection and recognition algorithms include EasyPR, HyperLPR, etc. Among them, the algorithm flow of EasyPR is: first, the license plate position is detected through the information features of the license plate, and then each character on the license plate is detected and segmented separately, and finally each segmented character is matched with the meaning of the character template to identify the character. The algorithm flow of HyperLPR is: the first step is to use opencv's HAAR Cascade to detect the approximate position of the license plate, the second step is to expand the rectangular area of the detected approximate position, the third step is to use multi-level binarization and RANSAC similar to MSER to fit the upper and lower boundaries of the license plate, the fourth step is to use a convolutional neural network to regress the left and right boundaries of the license plate, the fifth step is to use an algorithm based on texture field to correct the tilt of the license plate, and the sixth step is to use a convolutional neural network sliding window to cut characters and use a convolutional neural network to recognize characters.

However, the above algorithms need to separate each character of the license plate for recognition, and none of them corrects the distorted license plate, cannot recognize double-layer license plates, and cannot recognize the color of the license plate.

Summary of the invention

In view of the above-mentioned defects or deficiencies in the prior art, it is desired to provide a new energy vehicle identification method and system.

In a first aspect, an embodiment of the present application provides a method for identifying a new energy vehicle, the method comprising:

S1: Input video or image including license plate;

S2: Detect the location of the license plate in the video or image;

S3: cropping the image corresponding to the license plate according to a preset size according to the detected position of the license plate;

S4: Correct the cropped image, and recognize the license plate information in the corrected image to obtain the license plate characters and license plate color;

S5: Determine whether the vehicle is a new energy vehicle based on the license plate characters and license plate color.

In one embodiment, the input includes a video or image of a license plate, including:

S11: Read the image data of the original license plate;

S12: Preprocessing the image data, wherein the preprocessing includes expanding the data and increasing the data diversity;

S13: Adjust the image data to obtain image data with high recognition.

In one embodiment, detecting the location of the license plate in the video or image includes:

S21: extracting size information of the license plate data in the image;

S22: adjusting the extracted size information according to a preset ratio so that the ratio of the extracted size information to the original license plate size information is within a preset value.

In one of the embodiments, the correcting the cropped image includes: converting an arbitrarily tilted and distorted quadrilateral image into a rectangular image by perspective transformation;

Crop a rectangular image.

In one embodiment, after correcting the cropped image, the method further comprises:

Determine whether the license plate characters in the corrected and cropped image are single-layer or double-layer. If they are single-layer, directly recognize the license plate information in the corrected image; if they are double-layer, convert the double-layer characters into single-layer characters and then recognize the license plate information in the corrected image.

In one embodiment, the identifying the license plate information in the rectified image includes:

Normalize the size of the rectified image to meet the input requirements of the CRDPNet network;

The processed image is input into the Backbone network in the CRDPNet network to extract the license plate character information in the image.

In one embodiment, after extracting the license plate character information in the image, the method further includes:

Calculate the probability of each character appearing, and use the character with the highest probability of appearing as the license plate character sequence.

In one embodiment, judging whether a vehicle is a new energy vehicle based on the license plate characters and the license plate color includes:

When the number of characters in the license plate meets eight digits or the color of the license plate is green, the vehicle is judged to be a new energy vehicle.

In a second aspect, an embodiment of the present application provides a new energy vehicle identification system, the system comprising:

An input module, used to input a video or image including a license plate;

A detection module is used to detect the location of the license plate in the video or image;

A cropping module, used to crop the image corresponding to the license plate according to a preset size based on the detected position of the license plate;

A recognition module is used to correct the cropped image and recognize the license plate information in the corrected image to obtain the license plate characters and license plate color;

The judgment module is used to judge whether the vehicle is a new energy vehicle based on the license plate characters and license plate color.

The beneficial effects of this application include:

The new energy vehicle identification method provided by the present application can detect license plates accurately and quickly. It can not only locate the license plate position and classify the detection target, but also correct the distorted license plate to improve the subsequent license plate recognition accuracy. It can also identify the specific license plate characters of single-layer license plates or double-layer license plates, and can also identify the color of the license plate to help improve the recognition accuracy of the system. Finally, by judging the length of the license plate characters or the color of the license plate, the vehicle type is identified, which further improves the recognition accuracy of new energy vehicles.

BRIEF DESCRIPTION OF THE DRAWINGS

Other features, objects and advantages of the present application will become more apparent by reading the detailed description of non-limiting embodiments made with reference to the following drawings:

FIG1 is a schematic diagram showing a flow chart of a new energy vehicle identification method provided in an embodiment of the present application;

FIG2 is a schematic diagram showing a flow chart of another new energy vehicle identification method provided in an embodiment of the present application;

FIG3 shows a schematic diagram of the network structure of the YOLO8Plate algorithm model provided in an embodiment of the present application;

FIG4 shows a schematic diagram of the structure of a CBS module provided in an embodiment of the present application;

FIG5 shows a schematic diagram of a C2f module network structure provided in an embodiment of the present application;

FIG6 shows a network structure of an SPPF module provided in an embodiment of the present application;

FIG7 shows feature graphs of different scales of Neck learning provided by an embodiment of the present application;

FIG8 shows a Neck network architecture diagram of the YOLO8Plate provided in an embodiment of the present application;

FIG9 shows a flowchart of the post-processing part provided in an embodiment of the present application;

FIG10 shows a schematic diagram of a license plate distortion correction by perspective transformation provided by an embodiment of the present application;

FIG11 is a schematic diagram showing a double-row license plate processed into a single-row license plate according to an embodiment of the present application;

FIG12 is a schematic diagram of the existing GA36-2018 standard license plate style;

FIG. 13 is a schematic diagram of the existing standard license plate style for new energy vehicles.

Detailed ways

The present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant application, rather than to limit the application. It is also necessary to explain that, for ease of description, only the parts related to the application are shown in the accompanying drawings.

It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

As shown in Figure 12, Figure 12 is the standard license plate style specified by GA36-2018. my country's current automobile license plates follow the national standard of the People's Republic of China for automobile license plates, namely GA36-2018, and the current license plates include both single-layer license plates and double-layer license plates. Domestic conventional license plates can be divided into traditional gasoline-powered small car license plates with white characters on a blue background, traditional gasoline-powered medium and large car license plates with black characters on a yellow background, and new energy power vehicle license plates with black characters on a green background, etc., according to their power mode. In addition, special vehicles are also divided into police, military, agricultural, and embassy vehicles. According to the national standard GA36-2018, the domestic standard license plate size is required to be fixed at 440mm×140mm, and the spacing between each character in the license plate is also fixed. Generally, the license plate characters are 7 digits, and the characters are composed of Chinese characters representing provinces, A-Z capital letters (excluding O and I) and 0-9 digital identifiers that mark different license plates. In addition, the number of characters in special vehicle license plates may not be 7 digits.

As shown in Figure 13, Figure 13 is a standard style of new energy vehicle license plate, and the new energy vehicle license plate consists of 8 license plate characters, and the overall color is green. Therefore, by the length of the license plate characters and the underlying color of the license plate, it is possible to determine whether the vehicle to which the license plate belongs is a new energy vehicle. In order to more accurately determine whether the vehicle to which the license plate belongs is a new energy vehicle, this application provides the following solution.

Please refer to FIG. 1 and FIG. 2 , FIG. 1 shows a new energy vehicle identification method provided by an embodiment of the present application, the method comprising:

Step 110: Input a video or image including a license plate;

Step 120: Detect the location of the license plate in the video or image;

Step 130: cropping the image corresponding to the license plate according to a preset size based on the detected position of the license plate;

Step 140: Correcting the cropped image, and recognizing the license plate information in the corrected image to obtain the license plate characters and license plate color;

Step 150: Determine whether the vehicle is a new energy vehicle based on the license plate characters and license plate color.

Exemplarily, the new energy vehicle identification method in this application mainly adopts the YOLO8Plate algorithm model network structure shown in Figure 3. The algorithm mainly includes five parts: input layer (Input), backbone network (Backbone), fusion layer (Neck), detection head (Detection Head) and post-processing part (PostProcessing).

Among them, the main function of step 110 is to read the original license plate image data and pre-process the data to lay the foundation for further feature extraction. The image pre-processing process mainly consists of three parts: data enhancement, anchor adaptive adjustment, and image size adjustment.

YOLO8Plate uses the Mosaic data enhancement method. The main idea is to stitch multiple images together to form a larger picture, thereby expanding the input license plate image dataset and increasing data diversity, thereby improving the generalization and robustness of the model. In specific implementation, the Mosaic data enhancement method stitches four randomly selected images together according to a certain ratio, random cropping, and random arrangement, and then adjusts the relative position of the target frame coordinates to adapt to the size and position of the new image. This method improves the recognition ability of small target objects and well meets the requirements for car license plates in this paper.

Anchor (anchor box) adaptive adjustment is a technology in target detection. Its main idea is to automatically extract information such as the scale and aspect ratio of objects in the data set to generate the best anchor box. In network training, the network outputs a predicted box based on the anchor, and then compares it with the real box GT and calculates the loss, thereby continuously updating and iterating the network parameters in reverse. In the YOLO8Plate algorithm, the initial anchor can be set for the data set, and the algorithm itself will continuously calculate, adjust and update the optimal anchor value after each training.

The final image resizing refers to resizing the input original image to a uniform size so that it can be further trained in the model network. The specific process is divided into three steps: the first step is to determine the scaling ratio and scale it according to the ratio; the second step is to fill the image with black, white or gray borders if the image is not large enough after scaling; the last step is to send the scaled and filled image to the network.

The backbone network (Backbone) consists of the CBS module in Figure 4, the C2f module in Figure 5 and the SPPF module in Figure 6, and its main function is to extract the feature information of the image.

Among them, as shown in Figure 4, the CBS module is a common basic module of the convolutional neural network, which consists of a convolution layer, a BatchNorm layer, and a SiLu activation function. Among them, the main task of the convolution layer is to perform convolution calculations, so as to use the convolution operation to extract the spatial feature information of the local area in the image. The BatchNorm layer is a normalization layer, which is often located after the convolution layer in the network. Its function is to standardize the feature quantities in the network, thereby speeding up the training speed and enhancing the generalization of the network. In the implementation process, it first inputs a batch of feature maps, then calculates the mean and variance of the feature values on each channel, and normalizes the feature values on each channel, and finally linearly transforms the normalized results to obtain the output results. The SiLu activation function is a nonlinear function that introduces nonlinear transformation capabilities, which enables the network to better adapt to various data change distributions.

The C2f module shown in Figure 5 adds more parallel branches and enriches the gradient branches, allowing YOLO8Plate to obtain more branch gradient flow information while having the advantage of being lightweight.

The SPPF (Spatial Pyramid Pooling-Fast) module is shown in Figure 6. The SPPF module is a convolutional neural network module improved based on the idea of spatial pyramid pooling (SPP), which achieves the goal of adaptive size output and is used in this model. In addition, the SPPF module solves two problems: First, in order to make the image size meet the designed network layer input specifications, the target image needs to undergo operations such as scaling and cropping, but the above operations often cause image distortion. In contrast, SPPF can solve the image distortion problem. Secondly, when the network needs to process multiple scale targets, it cannot use a fixed-size fully connected layer. The SPPF module can replace the role of the fully connected layer in the network model.

The SPPF module is used to extract multi-scale features and then convert them into fixed-size feature vector outputs to detect and identify objects of different sizes. Specifically, the SPPF module first divides the input feature map into grids of different sizes, and then performs a pooling operation on each grid to obtain a feature vector of fixed dimension. These feature vectors are finally spliced together in a certain order to form an output vector of fixed size. This makes it possible to handle objects of different scales, and the dimensions of the fully connected layer input will not be different due to different feature map sizes.

The Neck part of the YOLO8Plate algorithm in this application uses a bottom-up path based on the top-down FPN (Feature Pyramid Network) architecture. It builds high-, medium-, and low-resolution feature pyramids, fuses the low-resolution high-level feature pyramid with the high-resolution low-level feature pyramid, performs multi-scale feature fusion on feature maps of different sizes, and passes these features to the prediction layer to achieve effective detection of targets of different scales.

As can be seen from Figure 7, the left side of the Neck part is the feature map of different scales obtained by extracting features from the Backbone part. Through the Backbone part, the algorithm network can obtain five groups of feature maps of different scales {P1, P2, P3, P4, P5}, and their sizes are {1/2, 1/4, 1/8, 1/16, 1/32} of the original image.

Next is the key part of Neck. The specific network of Neck is shown in Figure 8. The left half of Neck upsamples the feature map obtained on the feature pyramid structure, and then fuses the feature map {P3, P4, P5} obtained by downsampling from Backbone with the feature maps of different scales obtained in the upsampling process, thereby increasing the scale of the feature map and fusing feature maps of different scales. The right side of Neck continues to downsample in a similar way, and also obtains feature maps of different scales and gradually decreasing, and fuses the shallow graphic features and deep semantic features to obtain more complete image features.

The Detection Head part of the YOLO8Plate algorithm consists of a loss function and non-maximum suppression (NMS). The Detection Head part takes the feature map output by the Neck part as input and outputs the predicted license plate category (single-row license plate or double-row license plate), the predicted key box, the predicted key point coordinates of the four right angles of the license plate, and the predicted category confidence.

The training mainly includes several aspects of loss: prediction key rectangular box loss Loss _box , confidence loss Loss _obj , classification loss Loss _cls , key point loss Loss _kpts , key point confidence loss function Loss _{kpts_conf} . The overall loss function formula of the YOLO8Plate algorithm is as follows:

The predicted key rectangular box loss used by YOLO8Plate is CIOU_LOSS, and the confidence loss Loss _obj is BCE_LOSS loss function.

The confidence of each prediction box represents the reliability of the prediction box, and the value range is 0 to 1. The larger the value, the more credible the prediction box is, that is, the closer it is to the actual minimum enclosing key box of the target. Taking the 3*80*80 image matrix as an example (the same applies to the 40*40 and 20*20 image matrices), assuming that the confidence label is the matrix L, the prediction confidence is the matrix P, and the prediction is correct as mask = true, then the confidence loss function formula of the 3*80*80 feature image matrix is:

The classification loss Loss _cls used by YOLO8Plate is also the BCE_LOSS loss function. Taking the 3*80*80 image matrix as an example (the same applies to the 3*40*40 and 3*20*20 image matrices), the neural network predicts 3 prediction boxes for each pixel grid of the 80*80 matrix, and the prediction information of each prediction box contains N classification probabilities. N is the total number of categories. For example, the license plate data set in this article has 2 categories, namely single-row license plates and double-row license plates, so N is 2. Therefore, for the license plate data set in this article, each prediction box has N classification probabilities of 0 to 1, so the neural network predicts a total of 3*80*80*80 classification probabilities, forming the prediction probability matrix P. Assuming that the label probability is set as the matrix L, and the correct prediction is expressed as mask=true, the classification loss function formula of the 3*80*80 feature image matrix is:

The key point loss Loss _kpts of the YOLO8Plate algorithm solves the regression problem of key point coordinates, and the loss function it uses is WING_LOSS. WING_LOSS is a piecewise composite loss function. When the error is large in the early stage of training, the L1 loss function that is not overly sensitive to outliers is used to promote the rapid recovery of key points with positioning errors from large errors. In the later stage of training, the error is relatively small, and a logarithmic function with an offset is used to make the training of the neural network pay more attention to samples with small or medium errors, restoring the balance between errors of different sizes. The WINGLOSS function formula is as follows:

Among them, the positive number w limits the range of the nonlinear part to the interval [-w, w], ∈ represents the curvature of the constrained nonlinear region, and C is a constant to smoothly connect the linear and nonlinear parts of the segment. The value of ∈ is a very small value to prevent the network training from becoming unstable, that is, the gradient explosion problem caused by small errors.

For each key point, the YOLO8Plate algorithm learns a confidence parameter, which can indicate whether the target has the key point. The key point confidence loss function Loss _{kpts_conf} used by the YOLO8Plate algorithm is also BCE_LOSS. Assume represents the prediction confidence of the nth key point, δ(v _n ) represents the visibility flag of each key point, and the key point confidence loss function formula is as follows:

In the final training stage, the YOLO8Plate algorithm predicts and generates a large number of target prediction key boxes, which may cause overlap and redundancy in the candidate boxes of the same object. Therefore, in this case, the YOLO8Plate algorithm uses the NMS algorithm to determine which candidate boxes are invalid and only retains the best detection results. The specific process is to find the prediction box with the highest score as the output box, and then delete all other bounding boxes whose IOU (intersection-over-union ratio) is greater than the threshold. Then find the prediction box with the highest score from the remaining prediction boxes as the output box, and then delete other bounding boxes whose IOU is greater than the threshold. Repeat the above steps until all prediction boxes have been processed.

The main task of the post-processing part of the YOLO8Plate algorithm is to crop and correct the license plate images detected in the image or video, and convert all detected license plates into single-row license plates, which are then input into the subsequent license plate character recognition network to improve the accuracy of license plate character recognition.

The flowchart of the post-processing part is shown in Figure 9. Through the detection head (DetectionHead), the YOLO8Plate algorithm can obtain output results such as license plate category and key point coordinates. Next, the post-processing part uses the perspective transformation method to transform any tilted and distorted quadrilateral into a rectangle through the four key point coordinates, thereby cropping and correcting the small license plate image in the image.

Perspective transformation is an image processing technology based on the projection principle, which can reproject the image to a new plane. Applying perspective transformation technology to the system in this paper can convert any quadrilateral distorted license plate into a normal rectangular license plate. The specific formula is as follows:

Among them, (U, V, W) is the homogeneous coordinate corresponding to the coordinate (u, v) in the original image, (x, y) represents the coordinate in the transformed image, and T represents the perspective transformation matrix. In order to obtain the perspective transformation matrix T, at least 4 sets of coordinate points of the original image and the coordinate points corresponding to the transformed image are required. Through the perspective transformation matrix T, the system in this paper can realize the correction of distorted license plates.

Specifically, the post-processing part is implemented in three steps: First, through the four key point coordinates output by the detection head, a quadrilateral is selected in the image before transformation, and it is set to a rectangle in the transformed image. Second, the getPerspectiveTransform() function in OpenCV is used to calculate the perspective transformation matrix based on the vertex coordinates of the quadrilateral and camera parameters. Finally, the warpPerspective() function in OpenCV is used to apply the perspective transformation matrix to the original image to obtain the transformed image. The image effect after perspective transformation is shown in Figure 10.

After completing the perspective transformation, the post-processing part will then transform the double-row license plate. The PostProcessing part will first identify the type of license plate corresponding to the license plate. If the license plate is a single-row license plate, the post-processing part will not process it and directly input it into the subsequent license plate character recognition network. If the license plate is a double-row license plate, the post-processing part will convert it into a single-row license plate after correction as the input of the subsequent license plate character recognition network. Specifically, after the license plate thumbnail is cropped and perspective-corrected, the post-processing part will first calculate the size of the thumbnail, then crop the upper 5/12 height of the license plate as upper_img, and crop the lower 2/3 height of the license plate image as lower_img. Finally, resize the upper_img to the size of lower_img, and horizontally splice the upper_img and lower_img from left to right to obtain a single-row license plate. The specific effect is shown in Figure 11.

CRDPNet is a deep learning algorithm for license plate character recognition. It can automatically identify license plate character numbers from license plate images without segmenting the license plate character sequence to obtain single character pairing recognition. Among them, the network structure of CRDPNet includes a preprocessing part (Preprocessing), a backbone network (Backbone), and a prediction network part (Prediction). CRDPNet normalizes the input data of the network through the preprocessing module, then extracts features from the image through the backbone network, and finally predicts the license plate character sequence through the prediction network module.

In addition, the CRDPNet network uses CTC (Connectionist Temporal Classification) loss for end-to-end model training, so that the model automatically learns the starting and ending positions of the license plate number, and predicts the probability vector and label sequence that are most likely to identify the correct license plate characters, thereby identifying license plate character sequences of indefinite length.

The main task of the preprocessing module of the CRDPNet network is to adjust the size of the input license plate image, normalize the image matrix and adjust the image matrix channels to meet the input requirements of the CRDPNet network.

The backbone of the CRDPNet network is mainly composed of convolutional layers containing convolutional neural networks, which are responsible for extracting the image features of the license plate image. It will eventually output a feature tensor of size (1, 27, 78) and send it to the subsequent network structure for decoding and calculation to obtain the license plate characters.

The detailed backbone network structure is shown in Table 1. Each convolutional layer module will perform BatchNorm (batch normalization) operation and use Relu activation function after the convolution operation. In addition, the maximum pooling layer will also use the ceil_mod strategy in the pooling process, that is, each pooling operation will automatically fill the part of the image that does not meet the pooling kernel size, so as to ensure that the pooling output tensor size meets the subsequent input requirements.

Table 1 Backbone network structure

The Prdiction part consists of a character prediction branch and a color prediction branch.

The main function of the character prediction branch is to classify and predict the features extracted by Backbone, obtain a tensor that can predict the probability of each character in the character sequence, and finally calculate each character with the highest prediction probability, that is, obtain the license plate character sequence that is most likely to be recognized correctly. The network structure of the character prediction branch is shown in Table 2:

Table 2 Character prediction branch network structure

After the character prediction branch, the network will eventually output a prediction result tensor of size (1, 21, 78). The value 78 represents all possible license plate character types. The character index with the highest probability and the comparison character table are then selected. The CRDPNet algorithm network can recognize the predicted license plate character sequence results.

The color prediction branch performs color prediction on the extracted license plate background color features and finally obtains the color prediction probability tensor. The color prediction branch is shown in Table 3.

Table 3 Color prediction branch network structure

After the color prediction branch, the network finally outputs a prediction result tensor of only 5 elements. The value 5 represents the possible license plate colors, including blue, green, black, white, and yellow. The subsequent color prediction branch selects the color index with the highest prediction probability and compares it with the color table to identify the predicted license plate color result.

In some embodiments, the input includes a video or image of a license plate, including:

S11: Read the image data of the original license plate;

S12: preprocessing the image data, including expanding the data and increasing the data diversity;

S13: Adjust the image data to obtain image data with high recognition.

In some embodiments, detecting the location of a license plate in a video or image includes:

S21: extracting size information of the license plate data in the image;

S22: adjusting the extracted size information according to a preset ratio so that the ratio of the extracted size information to the original license plate size information is within a preset value.

In some embodiments, correcting the cropped image includes: converting the arbitrarily tilted and distorted quadrilateral image into a rectangular image by perspective transformation;

Crop a rectangular image.

In some embodiments, after correcting the cropped image, the method also includes: determining whether the license plate characters in the corrected cropped image are single-layer or double-layer. If it is single-layer, directly identifying the license plate information in the corrected image; if it is double-layer, converting the double-layer characters into single-layer characters and then identifying the license plate information in the corrected image.

In some embodiments, identifying the license plate information in the rectified image includes:

Normalize the size of the rectified image to meet the input requirements of the CRDPNet network;

The processed image is input into the Backbone network in the CRDPNet network to extract the license plate character information in the image.

In some embodiments, after extracting the license plate character information in the image, the method further includes:

Calculate the probability of each character appearing, and use the character with the highest probability of appearing as the license plate character sequence.

In some embodiments, judging whether a vehicle is a new energy vehicle based on the license plate characters and the license plate color includes:

When the number of characters in the license plate meets eight digits or the color of the license plate is green, the vehicle is judged to be a new energy vehicle.

The above description is only a preferred embodiment of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope of application involved in the present application is not limited to the technical solution formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the aforementioned application concept. For example, the above features are replaced with the technical features with similar functions disclosed in this application (but not limited to) by each other to form a technical solution.

Claims (9)

1. A new energy vehicle identification method, characterized in that the method comprises:

S1: inputting a video or image including a license plate;

s2: detecting the position of a license plate in a video or an image;

s3: cutting an image corresponding to the license plate according to a preset size according to the detected position of the license plate;

S4: correcting the cut image, and identifying license plate information in the corrected image to obtain license plate characters and license plate colors;

S5: and judging whether the vehicle is a new energy vehicle according to the license plate characters and the license plate colors.

2. The method of claim 1, wherein the input comprises a video or image of a license plate, comprising:

s11: reading image data of an original license plate;

s12: preprocessing image data, wherein the preprocessing comprises expanding the data and increasing the diversity of the data;

s13: and adjusting the image data to obtain the image data with strong identification.

3. The method for recognizing a new energy vehicle according to claim 1, wherein the detecting the position of the license plate in the video or the image includes:

s21: extracting size information of license plate data in the image;

S22: and adjusting the extracted size information according to a preset proportion so that the proportion of the extracted size information to the original license plate size information is within a preset value.

4. The new energy vehicle recognition method according to claim 1, wherein the correcting the clipped image includes:

converting the quadrilateral image with any oblique distortion into a rectangular image in a perspective transformation mode;

And cutting the rectangular image.

5. The new energy vehicle recognition method according to claim 1, wherein after correcting the clipped image, the method further comprises:

judging whether license plate characters in the corrected and cut image are single-layer or double-layer, if the license plate characters are single-layer, directly identifying license plate information in the corrected image; if the double-layer character is double-layer, converting the double-layer character into a single-layer character, and then correcting license plate information in the image.

6. The method for recognizing a new energy vehicle according to claim 1, wherein the recognizing license plate information in the corrected image includes:

Normalizing the corrected image size to enable the corrected image size to meet the input requirement of CRDPNet networks;

Inputting the processed image into a backup network in CRDPNet networks, and extracting license plate character information in the image.

7. The method for recognizing a new energy vehicle according to claim 6, further comprising, after extracting license plate character information in the image:

And calculating the occurrence probability of each character, and taking the character with the largest occurrence probability as a license plate character sequence.

8. The method for recognizing a new energy vehicle according to claim 1, wherein the step of judging whether the vehicle is a new energy vehicle based on license plate characters and license plate colors comprises:

And when the number of the license plate characters meets eight digits or the license plate color is green, judging that the vehicle is a new energy vehicle.

9. A new energy vehicle identification system, the system comprising:

The input module is used for inputting videos or images comprising license plates;

The detection module is used for detecting the position of the license plate in the video or the image;

the cutting module is used for cutting the image corresponding to the license plate according to the preset size according to the detected position of the license plate;

The recognition module is used for correcting the cut image, recognizing license plate information in the corrected image and obtaining license plate characters and license plate colors;

and the judging module is used for judging whether the vehicle is a new energy vehicle according to the license plate characters and the license plate colors.