WO2021196873A1 - 车牌字符识别方法、装置、电子设备和存储介质 - Google Patents

车牌字符识别方法、装置、电子设备和存储介质 Download PDF

Info

Publication number
WO2021196873A1
WO2021196873A1 PCT/CN2021/074915 CN2021074915W WO2021196873A1 WO 2021196873 A1 WO2021196873 A1 WO 2021196873A1 CN 2021074915 W CN2021074915 W CN 2021074915W WO 2021196873 A1 WO2021196873 A1 WO 2021196873A1
Authority
WO
WIPO (PCT)
Prior art keywords
license plate
network
layer
image
network structure
Prior art date
Application number
PCT/CN2021/074915
Other languages
English (en)
French (fr)
Inventor
黄艳庭
王凯
徐红祥
陆文涛
Original Assignee
京东方科技集团股份有限公司
合肥京东方显示技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 京东方科技集团股份有限公司, 合肥京东方显示技术有限公司 filed Critical 京东方科技集团股份有限公司
Priority to US17/914,993 priority Critical patent/US20230154210A1/en
Publication of WO2021196873A1 publication Critical patent/WO2021196873A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/60—Type of objects
    • G06V20/62—Text, e.g. of license plates, overlay texts or captions on TV images
    • G06V20/63—Scene text, e.g. street names
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/60—Type of objects
    • G06V20/62—Text, e.g. of license plates, overlay texts or captions on TV images
    • G06V20/625—License plates
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/24—Aligning, centring, orientation detection or correction of the image
    • G06V10/242—Aligning, centring, orientation detection or correction of the image by image rotation, e.g. by 90 degrees
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
    • G06V10/449—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
    • G06V10/451—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters with interaction between the filter responses, e.g. cortical complex cells
    • G06V10/454—Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/7715—Feature extraction, e.g. by transforming the feature space, e.g. multi-dimensional scaling [MDS]; Mappings, e.g. subspace methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
    • G06V30/10—Character recognition

Definitions

  • This application relates to the field of image technology, and in particular to a method, device, electronic device, and storage medium for recognizing license plate characters.
  • the method of recognizing the license plate in the related intelligent license plate recognition system is generally as follows: firstly, it is necessary to perform character segmentation on the located license plate, then extract the character characteristics of a single character, and finally perform the license plate character recognition.
  • the above-mentioned license plate is based on the character segmentation. In the way of recognition, the accuracy of character segmentation is difficult to guarantee, resulting in low accuracy of license plate recognition.
  • This application proposes a license plate character recognition method, device, electronic equipment and storage medium, which does not require character segmentation of the license plate, and the license plate characters on the license plate can be obtained by directly recognizing the entire license plate, avoiding segmentation and separation of the characters of the license plate Recognition improves the technical effect of recognition speed and recognition accuracy.
  • An embodiment of the present application proposes a license plate character recognition method, including: acquiring a vehicle image collected by an image acquisition device; positioning the license plate area of the vehicle image to obtain a license plate image; and analyzing the license plate through a convolutional neural network.
  • Image feature extraction is performed to obtain the feature information of the license plate image, wherein the convolutional neural network includes a residual network structure; the feature information is parsed through a bidirectional cyclic neural network model to obtain the license plate corresponding to the license plate image character.
  • the license plate character recognition method of the embodiment of the present application locates the vehicle area in the license plate image after acquiring the vehicle image collected by the image acquisition device to obtain the license plate image, and passes the convolutional neural network including the residual network structure
  • the feature extraction of the license plate image can effectively avoid the gradient disappearance and reduce the feature loss in the convolution process of the convolutional neural network, so that the bidirectional cyclic neural network model can accurately identify the license plate characters of the license plate image based on the feature information of the license plate image. Therefore, there is no need to perform character segmentation on the license plate, and the license plate characters on the license plate can be obtained by directly recognizing the entire license plate, avoiding segmentation and separate recognition of the characters of the license plate, and improving the recognition speed and recognition accuracy.
  • a license plate character recognition device which includes: an image acquisition module for acquiring a vehicle image collected by an image acquisition device; a license plate positioning module for locating a license plate area on the vehicle image to Obtain a license plate image; a feature extraction module for feature extraction of the license plate image through a convolutional neural network to obtain feature information of the license plate image, wherein the convolutional neural network includes a residual network structure; a recognition module , Used to analyze the feature information through the bidirectional cyclic neural network model to obtain the license plate characters corresponding to the license plate image.
  • the license plate character recognition device of the embodiment of the application after acquiring the vehicle image collected by the image acquisition device, locates the vehicle area in the license plate image to obtain the license plate image, and passes the convolutional neural network including the residual network structure
  • the feature extraction of the license plate image can effectively avoid the gradient disappearance and reduce the feature loss in the convolution process of the convolutional neural network, so that the bidirectional cyclic neural network model can accurately identify the license plate characters of the license plate image based on the feature information of the license plate image. Therefore, there is no need to perform character segmentation on the license plate, and the license plate characters on the license plate can be obtained by directly recognizing the entire license plate, avoiding segmentation and separate recognition of the characters of the license plate, and improving the recognition speed and recognition accuracy.
  • Another embodiment of the present application proposes an electronic device, including: an electronic device including: a memory and a processor; the memory stores computer instructions, and when the computer instructions are executed by the processor, Realize the license plate character recognition method of the embodiment of the present application.
  • Another embodiment of the present application proposes a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the license plate character recognition method disclosed in the embodiments of the present application.
  • Fig. 1 is a schematic flowchart of a method for recognizing characters on a license plate according to an embodiment of the present application
  • FIG. 2 is a schematic diagram of a network structure of a convolutional neural network including a first residual network structure and a second residual network structure;
  • FIG. 3 is a schematic diagram of the network structure of the first residual network structure
  • Figure 4 is a schematic diagram of the network structure of the bidirectional cyclic neural network model
  • FIG. 5 is a schematic flowchart of a method for recognizing characters on a license plate according to another embodiment of the present application.
  • Fig. 6 is a schematic structural diagram of a license plate character recognition device according to an embodiment of the present application.
  • Fig. 7 is a schematic structural diagram of a license plate character recognition device according to another embodiment of the present application.
  • Fig. 8 is a block diagram of an electronic device used to implement the method for recognizing license plate characters in an embodiment of the present application.
  • Fig. 1 is a schematic flowchart of a method for recognizing license plate characters according to an embodiment of the present application.
  • the execution body of the license plate character recognition method provided in this embodiment is a license plate character recognition device.
  • the license plate character recognition device can be configured in a smart license plate recognition system, and the smart license plate recognition system can be installed in an electronic device.
  • the electronic device may be a hardware device such as a terminal device and a server.
  • the license plate character recognition method may include:
  • Step 101 Acquire a vehicle image collected by an image collection device.
  • the image acquisition device may be a camera.
  • step 102 the license plate area is located on the vehicle image to obtain the license plate image.
  • the YOLOV3 network in order to improve the ability to quickly locate the license plate image in the license plate image, after obtaining the license plate image, the YOLOV3 network can be used to locate the license plate area of the vehicle image to obtain the license plate image.
  • YOLOV3 network is a kind of deep learning image detection algorithm, which is mainly based on the regression idea.
  • the two steps of target recognition and target positioning are carried out at the same time, which can improve the speed of target detection and meet the requirements of real-time detection.
  • the YOLOV3 network uses a 53-layer convolutional network and uses a softmax layer.
  • the YOLOV3 network can also use three-scale feature fusion for object detection and positioning, and use K-means clustering to generate nine a priori boxes as bounding boxes (that is, a compact rectangle containing license plates).
  • the initial size of the frame can thus achieve the effect of accelerating the convergence speed and improving the detection accuracy.
  • the Yolov3 network can accurately locate the license plates of vehicles of various colors in various complex environments, and the recognition time is about 20 ms.
  • Step 103 Perform feature extraction on the license plate image through a convolutional neural network to obtain feature information of the license plate image, where the convolutional neural network includes a residual network structure.
  • the convolutional layer in the convolutional neural network usually performs feature extraction on the license plate image to obtain the feature image corresponding to the license plate image.
  • the convolutional layer performs feature extraction on the license plate image, it is easy to cause the characteristics of the license plate image Loss, and when the depth of the convolutional neural network increases, the problem of gradient disappearance occurs.
  • a residual network structure can be set in the convolutional neural network, so that the residual network structure can effectively prevent the disappearance of the gradient caused by the increase in depth in the convolutional neural network. And reduce the problem of feature loss in the convolution process.
  • the aforementioned convolutional neural network may also include a convolutional layer and a pooling layer.
  • the convolutional neural network in this embodiment may be connected in sequence according to the convolutional layer, the residual network structure, and the pooling layer.
  • the convolutional neural network of this embodiment includes a convolutional layer, a residual network structure, and a pooling layer that are sequentially connected.
  • the convolutional layer in the convolutional neural network will perform feature extraction on the license plate image to obtain the first feature image of the license plate image.
  • the first feature can be The image is input to the residual network structure, and the characteristic image output by the residual structure is input to the pooling layer for corresponding processing.
  • three convolutional layers may be set in the convolutional neural network to perform feature extraction on the license plate image.
  • the first convolutional layer, the residual network structure, the second convolutional layer, the third convolutional layer, and the pooling layer are sequentially connected to form the convolutional neural network of this embodiment, thereby minimizing the feature loss of the convolution process.
  • a batch normalization layer is added after the first convolutional layer in the convolutional neural network, thereby effectively improving network performance and speeding up convergence.
  • the convolutional neural network in this embodiment may use Leaky ReLU as the activation function.
  • the residual network structure of this embodiment includes the first residual network structure and The second residual network structure.
  • the convolutional neural network includes a first convolutional layer, a first residual network structure, a second residual network structure, a second convolutional layer, a third convolutional layer, and a pooling layer that are sequentially connected.
  • the network structure diagram of the convolutional neural network including the first residual network structure and the second residual network structure is shown in FIG. 2.
  • the network structure of the convolutional neural network in FIG. 2 is specifically: a first convolutional layer, a first residual network structure connected to the first convolutional layer, and a second residual network structure connected to the first residual network structure , A second convolutional layer connected to the second residual network structure, a third convolutional layer connected to the second convolutional layer, and a maximum pooling layer connected to the third convolutional layer.
  • the convolutional neural network parameters are shown in Table 1, where Filters is the number of convolution kernels, K is the size of the convolution kernel, S is the step size, and P is the padding parameter.
  • Residual-1 in Table 1 represents the first residual network structure
  • Residual-2 represents the second residual network structure
  • ConvolutionN represents the Nth convolutional layer, where N is any integer from 1 to 3.
  • a Batch Normalization layer may be set between the second residual network structure in the convolutional neural network and Leaky ReLU.
  • the network structures of the first residual network structure and the second residual network structure are the same, and the network structures both include the first network substructure, the second network substructure and the adder, and the first network substructure
  • the structure includes the largest pooling layer, the first convolution sublayer, the second convolution sublayer, the third convolution sublayer, and the fourth convolution sublayer that are sequentially connected.
  • the second network substructure includes the fifth convolution sublayer. Network, the input ends of the maximum pooling layer and the fifth convolution sub-network are all connected to the output end of the output layer of the first convolutional layer, and the output ends of the fourth convolution sub-layer and the fifth convolution sub-network are both connected to the addition The input terminal of the device is connected.
  • FIG. 3 the network structure diagram of the first residual network structure is shown in FIG. 3.
  • the characteristic image can be input into the first residual network structure, and the first residual network structure The output is input into the second residual network structure again, and then the characteristic information output by the second residual network structure is input to the corresponding processing layer connected to the second residual network structure for subsequent processing through the corresponding processing layer.
  • the network structures of the first residual network structure and the second residual network structure in this embodiment are the same, the network parameters used by the two are different.
  • the residual network structure of this embodiment can also make the forward and backward propagation of information smoother.
  • the number of convolution kernels corresponding to the convolution layer in the first residual network structure and the second residual network structure are different.
  • Convolution in the table represents the convolutional layer
  • ConvolutionN represents the Nth convolutional sublayer, where N is any integer from 1 to 5.
  • Step 104 Analyze the feature information through the bidirectional cyclic neural network model to obtain license plate characters corresponding to the license plate image.
  • the above-mentioned bidirectional cyclic neural network model may include a bidirectional cyclic neural network and a Connectionist Temporal Classification (CTC) network.
  • CTC Connectionist Temporal Classification
  • the bidirectional cyclic neural network of this embodiment may include a first layer of bidirectional long-term memory (BLSTM) sequentially connected. And the second layer two-way long and short-term memory network.
  • BLSTM bidirectional long-term memory
  • each LSTM network in the bidirectional long and short-term memory network in this embodiment contains 512 hidden units. Therefore, the output depth of each layer of BLSTM is 1024, and the output of the last layer of BLSTM is connected to two fully connected layers. The depth of the full connection of the layer is 1024, and the depth of the full connection of the second layer is the number of license plate character types (for example, 66).
  • a dropout layer is provided between the two fully connected layers in this embodiment. That is, the second layer of the bidirectional long-short-term memory network of this embodiment includes: a first fully connected layer, a dropout layer connected to the first fully connected layer, and a second fully connected layer connected to the dropout layer.
  • FIG. 4 the schematic diagram of the network structure of the bidirectional cyclic neural network model in this embodiment is shown in FIG. 4.
  • the feature extraction of the entire license plate image is performed directly through the convolutional neural network, and the extracted feature information is analyzed through the bidirectional cyclic neural network model to Get the license plate characters in the license plate image. From this, it can be seen that this embodiment does not need to perform character segmentation on the license plate image, so the time used for character segmentation in the license plate recognition process can be reduced, and the license plate recognition speed can be improved.
  • the license plate character recognition method of the embodiment of the present application locates the vehicle area in the license plate image after acquiring the vehicle image collected by the image acquisition device to obtain the license plate image, and passes the convolutional neural network including the residual network structure
  • the feature extraction of the license plate image can effectively avoid the gradient disappearance and reduce the feature loss in the convolution process of the convolutional neural network, so that the bidirectional cyclic neural network model can accurately identify the license plate characters of the license plate image based on the feature information of the license plate image. Therefore, there is no need to perform character segmentation on the license plate, and the license plate characters on the license plate can be obtained by directly recognizing the entire license plate, avoiding segmentation and separate recognition of the characters of the license plate, and improving the recognition speed and recognition accuracy.
  • Fig. 5 is a schematic flowchart of a method for recognizing license plate characters according to another embodiment of the present application. Among them, it should be noted that this embodiment is a further refinement or optimization of the above-mentioned embodiment.
  • the license plate character recognition method may include:
  • Step 501 Acquire a vehicle image collected by an image collection device.
  • Step 502 Use YOLOV3 network to locate the license plate area of the vehicle image to obtain the license plate image.
  • Step 503 Use the spatial transformation network to perform tilt correction on the license plate image.
  • the spatial transform network may include a local network, a grid generator, and a sampler.
  • the license plate image can be input into the local network to obtain the affine transformation parameters used to correct the license plate image; the affine transformation parameters are input into the grid generator to obtain the corresponding affine Transformation matrix: A sampler is used to perform affine transformation on each pixel coordinate in the license plate image based on the affine transformation matrix to correct the license plate image.
  • the above-mentioned local network is a network used to regress and transform the parameter ⁇ .
  • the specific parameters of the network are shown in Table 4, where Filters is the number of convolution kernels, K is the size of the convolution kernel, S is the step size, and P is The padding parameter; its input is a grayscale image, and then it passes through a series of hidden network layers, and finally outputs the spatial transformation parameters.
  • the form of ⁇ can be various, if 2D affine transformation is required, ⁇ is the output of a 6-dimensional (2x3) vector.
  • the size of ⁇ depends on the type of transformation.
  • the input is a 2D image, and 2D affine transformation needs to be implemented.
  • the network outputs a 2*3 matrix as shown in formula 2-1, expressed as A ⁇ .
  • the aforementioned grid generator constructs a sampling grid based on predicted transformation parameters, which is an output obtained after sampling and transformation of points in a group of input images. What the grid generator actually gets is a mapping relationship T ⁇ . Assume that the coordinates of each pixel of the input image are The coordinate of each pixel output by STN is The space transformation function T ⁇ is a two-dimensional affine transformation function, then with The corresponding relationship can be expressed as 2-2:
  • the sampler uses the sampling grid and the input image as input at the same time to obtain the transformed result of the input image.
  • the above-mentioned sampler samples the input image features on the basis of the mapping relationship obtained by the network generator, and obtains the image features that have undergone spatial transformation to achieve a correction effect.
  • the sampling method used in this embodiment is bilinear interpolation.
  • the noise interference around the license plate image can also be filtered through the local network to make the license plate easier to recognize.
  • Step 504 Perform feature extraction on the license plate image through a convolutional neural network to obtain feature information of the license plate image, where the convolutional neural network includes a residual network structure.
  • Step 505 Analyze the feature information through the bidirectional cyclic neural network model to obtain license plate characters corresponding to the license plate image.
  • the license plate characters can include Chinese characters, letters, and numbers.
  • the vehicle area in the license plate image is located to obtain the license plate image, and the license plate image is tilted and corrected.
  • the convolutional neural network of the difference network structure extracts the characteristics of the corrected license plate image, effectively avoiding the disappearance of the gradient and reducing the feature loss in the convolution process of the convolutional neural network, so that the bidirectional cyclic neural network model can be based on the characteristics of the license plate image
  • the information accurately recognizes the license plate characters of the license plate image.
  • this embodiment does not need to perform character segmentation on the license plate after directly locating the license plate in the picture vehicle, and directly input the license plate image obtained after the positioning into the convolutional neural network including the residual network structure Carry out feature extraction, and use the bidirectional cyclic neural network model to recognize the license plate characters based on the feature information of the license plate image, which can directly recognize the entire license plate, avoid segmentation and separate recognition of the license plate characters, and improve the recognition speed.
  • This embodiment conducts experiments on the license plate recognition method using STN and the license plate recognition method without STN. According to the experimental results, it is determined that the accuracy rate of the license plate recognition with the STN network removed is 94.6%, which is lower than the accuracy of the license plate recognition model with the STN network. The rate is 96.1%. The average loss of license plate recognition that removes the STN network is 0.53, which is higher than the average loss of 0.40 for license plate recognition that joins the STN network. From this, it can be seen that performing tilt correction before inputting the license plate image to the convolutional neural network can further improve the recognition accuracy of the license plate characters.
  • Table 5 shows that when the network of this application has no STN network, the number of hidden units of one-layer BLSTM, two-layer BLSTM and three-layer BLSTM is At 128, 256, 512, 1024, the accuracy of license plate recognition and the average loss of license plate recognition. Since the necessity of the STN network has been verified above, in order to improve the training convergence speed, the STN network will no longer be added during network verification.
  • the obtained experimental parameters reflect the rationality of the BLSTM parameters selected in this application.
  • the number before the slash in the above table is the value of the license plate recognition accuracy rate ACC
  • the number after the slash represents the value of the average loss loss of the license plate recognition.
  • 94.6/0.53 indicates that the license plate recognition is accurate.
  • the ACC rate is 94.6%
  • the average loss of license plate recognition is 0.53.
  • this embodiment also analyzes the fully connected layer FC_layer.
  • Table 6 shows the license plate recognition accuracy rate and the average number of license plates recognized by the model when there is no STN network in the first-layer fully-connected, two-layer fully-connected, and three-layer fully-connected Loss, the experimental parameters obtained through this can reflect the rationality of the fully connected parameters selected by the network of this application.
  • ACC represents the accuracy of license plate recognition under the corresponding model
  • loss represents the average loss of license plate recognition under the corresponding model
  • F1 represents the output depth of the first layer of fully connected
  • F2 represents the output depth of the second layer of fully connected
  • F3 represents the third The output depth of the layer fully connected.
  • an embodiment of the present application also provides a license plate character recognition device, because the license plate character recognition device provided by the embodiment of the application is different from the license plate character recognition device provided by the foregoing embodiments.
  • the license plate character recognition method corresponds, so the implementation of the license plate character recognition method is also applicable to the license plate character recognition device provided in this embodiment, and will not be described in detail in this embodiment.
  • Fig. 6 is a schematic structural diagram of a license plate character recognition device according to an embodiment of the present application.
  • the license plate character recognition device 600 includes an image acquisition module 110, a license plate location module 120, a feature extraction module 130, and a recognition module 140, wherein:
  • the image acquisition module 110 is used to acquire the vehicle image collected by the image acquisition device.
  • the license plate location module 120 is used to locate the license plate area of the vehicle image to obtain the license plate image.
  • the feature extraction module 130 is configured to perform feature extraction on the license plate image through a convolutional neural network to obtain feature information of the license plate image, where the convolutional neural network includes a residual network structure.
  • the recognition module 140 is used to analyze the feature information through the bidirectional cyclic neural network model to obtain the license plate characters corresponding to the license plate image.
  • the convolutional neural network includes a convolutional layer, a residual network structure, and a pooling layer that are sequentially connected.
  • the convolutional neural network includes a first convolutional layer, a residual network structure, a second convolutional layer, a third convolutional layer, and a pooling layer that are sequentially connected.
  • the convolutional layer includes 3 layers
  • the residual network structure includes a first residual network structure and a second residual network structure
  • the convolutional neural network includes sequentially connected The first convolutional layer, the first residual network structure, the second residual network structure, the second convolutional layer, the third convolutional layer, and the pooling layer.
  • the first residual network structure and the second residual network structure have the same network structure, and the network structure includes a first network substructure, a second network substructure, and an adder.
  • the first network substructure Including the largest pooling layer, the first convolution sublayer, the second convolution sublayer, the third convolution sublayer, and the fourth convolution sublayer that are sequentially connected.
  • the second network substructure includes the fifth convolution subnetwork. , The input ends of the maximum pooling layer and the fifth convolution subnetwork are both connected to the output ends of the output layer of the first convolution layer, and the output ends of the fourth convolution sublayer and the fifth convolution subnetwork are both connected to the adder The input terminal is connected.
  • the device further includes:
  • the tilt correction module 150 is configured to use the spatial transformation network to perform tilt correction on the license plate image.
  • the spatial transformation network includes a local network, a grid generator, and a sampler.
  • the above-mentioned tilt correction module 150 is specifically used to: input the license plate image into the local network to be used to correct the license plate image Affine transformation parameters; input the affine transformation parameters to the grid generator to obtain the corresponding affine transformation matrix; use the sampler to perform affine transformation on each pixel coordinate in the license plate image based on the affine transformation matrix to Correct the license plate image.
  • the bidirectional cyclic neural network model includes a bidirectional cyclic neural network and a CTC network for connection timing classification.
  • the bidirectional cyclic neural network includes a first layer of bidirectional long and short-term memory network and a second layer of bidirectional long and short-term memory network that are sequentially connected.
  • the second-layer bidirectional long-short-term memory network includes: a first fully connected layer; a dropout layer connected to the first fully connected layer; and a second fully connected layer connected to the dropout layer.
  • the license plate character recognition device of the embodiment of the application after acquiring the vehicle image collected by the image acquisition device, locates the vehicle area in the license plate image to obtain the license plate image, and passes the convolutional neural network including the residual network structure
  • the feature extraction of the license plate image can effectively avoid the gradient disappearance and reduce the feature loss in the convolution process of the convolutional neural network, so that the bidirectional cyclic neural network model can accurately identify the license plate characters of the license plate image based on the feature information of the license plate image. Therefore, there is no need to perform character segmentation on the license plate, and the license plate characters on the license plate can be obtained by directly recognizing the entire license plate, avoiding segmentation and separate recognition of the characters of the license plate, and improving the recognition speed and recognition accuracy.
  • the present application also provides an electronic device and a readable storage medium.
  • FIG. 8 it is a block diagram of an electronic device of a method for recognizing license plate characters according to an embodiment of the present application.
  • the electronic device includes:
  • the processor 1002 implements the database management method provided in the foregoing embodiment when executing instructions.
  • the electronic equipment also includes:
  • the communication interface 1003 is used for communication between the memory 1001 and the processor 1002.
  • the memory 1001 is used to store computer instructions that can run on the processor 1002.
  • the memory 1001 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), for example, at least one disk memory.
  • the processor 1002 is used to implement the license plate character recognition method of the foregoing embodiment when executing a program.
  • the bus may be an Industry Standard Architecture (ISA) bus, Peripheral Component (PCI) bus, or Extended Industry Standard Architecture (EISA) bus, etc.
  • ISA Industry Standard Architecture
  • PCI Peripheral Component
  • EISA Extended Industry Standard Architecture
  • the bus can be divided into address bus, data bus, control bus and so on. For ease of representation, only one thick line is used in FIG. 8, but it does not mean that there is only one bus or one type of bus.
  • the memory 1001, the processor 1002, and the communication interface 1003 are integrated on a single chip, the memory 1001, the processor 1002, and the communication interface 1003 can communicate with each other through internal interfaces.
  • the processor 1002 may be a central processing unit (Central Processing Unit, referred to as CPU), or a specific integrated circuit (Application Specific Integrated Circuit, referred to as ASIC), or configured to implement one or more of the embodiments of the present application integrated circuit.
  • CPU Central Processing Unit
  • ASIC Application Specific Integrated Circuit
  • first and second are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined with “first” and “second” may explicitly or implicitly include at least one of the features. In the description of the present application, "a plurality of” means at least two, such as two, three, etc., unless specifically defined otherwise.
  • a "computer-readable medium” can be any device that can contain, store, communicate, propagate, or transmit a program for use by an instruction execution system, device, or device or in combination with these instruction execution systems, devices, or devices.
  • computer readable media include the following: electrical connections (electronic devices) with one or more wiring, portable computer disk cases (magnetic devices), random access memory (RAM), Read only memory (ROM), erasable and editable read only memory (EPROM or flash memory), fiber optic devices, and portable compact disk read only memory (CDROM).
  • the computer-readable medium may even be paper or other suitable medium on which the program can be printed, because it can be used, for example, by optically scanning the paper or other medium, followed by editing, interpretation, or other suitable media if necessary. The program is processed in a manner to obtain the program electronically, and then stored in the computer memory.
  • each part of this application can be implemented by hardware, software, firmware, or a combination thereof.
  • multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system.
  • Discrete logic gate circuits with logic functions for data signals Logic circuits, application specific integrated circuits with suitable combinational logic gates, programmable gate array (PGA), field programmable gate array (FPGA), etc.
  • a person of ordinary skill in the art can understand that all or part of the steps carried in the method of the foregoing embodiments can be implemented by a program instructing relevant hardware to complete.
  • the program can be stored in a computer-readable storage medium. When executed, it includes one of the steps of the method embodiment or a combination thereof.
  • the functional units in the various embodiments of the present application may be integrated into one processing module, or each unit may exist alone physically, or two or more units may be integrated into one module.
  • the above-mentioned integrated modules can be implemented in the form of hardware or software functional modules. If the integrated module is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer readable storage medium.
  • the aforementioned storage medium may be a read-only memory, a magnetic disk or an optical disk, etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Health & Medical Sciences (AREA)
  • Multimedia (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Molecular Biology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Computational Linguistics (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Biodiversity & Conservation Biology (AREA)
  • Image Analysis (AREA)
  • Character Discrimination (AREA)
  • Traffic Control Systems (AREA)

Abstract

一种车牌字符识别方法、装置、电子设备和存储介质，其中，该方法包括：在获取图像采集设备所采集到的车辆图像后，对车牌图像中的车辆区域进行定位，以得到车牌图像，并通过包括残差网络结构的卷积神经网络对车牌图像进行特征提取，有效避免梯度消失和减少了卷积神经网络卷积过程中的特征损失，从而使得双向循环神经网络模型可基于车牌图像的特征信息准确识别出车牌图像的车牌字符。由此，不需要对车牌进行字符分割，通过直接识别整个车牌即可得到车牌上的车牌字符，避免对车牌的字符进行切分和分开识别，提高了识别速度和识别准确度。

Description

车牌字符识别方法、装置、电子设备和存储介质
相关申请的交叉引用
本申请要求合肥京东方显示技术有限公司，京东方科技集团股份有限公司于2020年03月20日提交的、发明名称为“车牌字符识别方法、装置、电子设备和存储介质”的、中国专利申请号“202010234926.2”的优先权。
技术领域
本申请涉及图像技术领域，尤其涉及车牌字符识别方法、装置、电子设备和存储介质。
背景技术
随着中国汽车保有量的不断增加，智能车牌识别与管理系统的研究应用越来越重要。目前，智能车牌识别系统在高速公路收费站、小区停车场等区域的应用已经越来越广泛。
相关技术中，相关智能车牌识别系统中识别车牌的方式一般为：首先需要将定位到的车牌进行字符分割，然后对单个字符提取字符特征，最后进行车牌字符识别，然而，上述基于字符分割进行车牌识别的方式，字符分割的准确性难以保证，从而导致车牌识别准确性较低。
发明内容
本申请提出一种车牌字符识别方法、装置、电子设备和存储介质，不需要对车牌进行字符分割，通过直接识别整个车牌即可得到车牌上的车牌字符，避免对车牌的字符进行切分和分开识别，提高了识别速度和识别准确度的技术效果。
本申请一方面实施例提出了一种车牌字符识别方法，包括：获取图像采集设备所采集到的车辆图像；对车辆图像进行车牌区域定位，以得到车牌图像；通过卷积神经网络对所述车牌图像进行特征提取，以得到所述车牌图像的特征信息，其中，所述卷积神经网络包括残差网络结构；通过双向循环神经网络模型解析所述特征信息，以得到所述车牌图像对应的车牌字符。
本申请实施例的车牌字符识别方法，在获取图像采集设备所采集到的车辆图像后，对车牌图像中的车辆区域进行定位，以得到车牌图像，并通过包括残差网络结构的卷积神经网络对车牌图像进行特征提取，有效避免梯度消失和减少了卷积神经网络卷积过程中的特征损失，从而使得双向循环神经网络模型可基于车牌图像的特征信息准确 识别出车牌图像的车牌字符。由此，不需要对车牌进行字符分割，通过直接识别整个车牌即可得到车牌上的车牌字符，避免对车牌的字符进行切分和分开识别，提高了识别速度和识别准确度。
本申请另一方面实施例提出了一种车牌字符识别装置，包括：图像采集模块，用于获取图像采集设备所采集到的车辆图像；车牌定位模块，用于对车辆图像进行车牌区域定位，以得到车牌图像；特征提取模块，用于通过卷积神经网络对所述车牌图像进行特征提取，以得到所述车牌图像的特征信息，其中，所述卷积神经网络包括残差网络结构；识别模块，用于通过双向循环神经网络模型解析所述特征信息，以得到所述车牌图像对应的车牌字符。
本申请实施例的车牌字符识别装置，在获取图像采集设备所采集到的车辆图像后，对车牌图像中的车辆区域进行定位，以得到车牌图像，并通过包括残差网络结构的卷积神经网络对车牌图像进行特征提取，有效避免梯度消失和减少了卷积神经网络卷积过程中的特征损失，从而使得双向循环神经网络模型可基于车牌图像的特征信息准确识别出车牌图像的车牌字符。由此，不需要对车牌进行字符分割，通过直接识别整个车牌即可得到车牌上的车牌字符，避免对车牌的字符进行切分和分开识别，提高了识别速度和识别准确度。
本申请另一方面实施例提出了一种电子设备，包括：一种电子设备，包括：存储器，处理器；所述存储器中存储有计算机指令，当所述计算机指令被所述处理器执行时，实现本申请实施例的车牌字符识别方法。
本申请另一方面实施例提出了一种存储有计算机指令的非瞬时计算机可读存储介质，所述计算机指令用于使所述计算机执行本申请实施例公开的车牌字符识别方法。
上述可选方式所具有的其他效果将在下文中结合具体实施例加以说明。
附图说明
附图用于更好地理解本方案，不构成对本申请的限定。其中：
图1是根据本申请一个实施例的车牌字符识别方法的流程示意图；
图2是包括第一残差网络结构和第二残差网络结构的卷积神经网络的网络结构示意图；
图3是第一残差网络结构的网络结构示意图；
图4是双向循环神经网络模型的网络结构的示意图；
图5是根据本申请另一个实施例的车牌字符识别方法的流程示意图；
图6是根据本申请一个实施例的车牌字符识别装置的结构示意图；
图7是根据本申请另一个实施例的车牌字符识别装置的结构示意图；
图8是用来实现本申请实施例的车牌字符识别方法的电子设备的框图。
具体实施方式
下面详细描述本申请的实施例，所述实施例的示例在附图中示出，其中自始至终相同或类似的标号表示相同或类似的元件或具有相同或类似功能的元件。下面通过参考附图描述的实施例是示例性的，旨在用于解释本申请，而不能理解为对本申请的限制。
下面参考附图描述本申请实施例的车牌字符识别方法、装置和电子设备。
图1是根据本申请一个实施例的车牌字符识别方法的流程示意图。其中，需要说明的是，本实施例提供的车牌字符识别方法的执行主体为车牌字符识别装置，该车牌字符识别装置可以配置在智能车牌识别系统中，该智能车牌识别系统可以设置在电子设备中，例如，电子设备可以为终端设备、服务器等硬件设备。
如图1所示，该车牌字符识别方法可以包括：
步骤101，获取图像采集设备所采集到的车辆图像。
其中，图像采集设备可以为摄像头。
步骤102，对车辆图像进行车牌区域定位，以得到车牌图像。
在本实施例中，为了提高可快速定位出车牌图像中的车牌图像，在获取车牌图像后，可采用YOLOV3网络对车辆图像进行车牌区域定位，以得到车牌图像。
其中，YOLOV3网络是一种深度学习图像检测算法，其主要是基于回归思想，将目标识别和目标定位两个步骤同时进行，从而可提高目标检测的速度，达到了实时检测的要求。
其中，YOLOV3网络为使用一个53层的卷积网络，并且使用了softmax层。
在本实施例的一个实施例，该YOLOV3网络还可采用三个尺度特征融合进行物体检测定位，并使用K-means聚类产生九个先验框作为bounding box(即包含车牌的一个紧致矩形框)的初始尺寸，从而可达到加快收敛速度，提高检测精度的效果。
在本实施例，通过Yolov3网络能够对各种复杂环境下的各种颜色车辆的车牌进行准确的定位，识别时间在20ms左右。
步骤103，通过卷积神经网络对车牌图像进行特征提取，以得到车牌图像的特征信息，其中，该卷积神经网络包括残差网络结构。
在本实施例中，卷积神经网络中的卷积层通常对车牌图像进行特征提取，以得到对应车牌图像的特征图像，然而，卷积层对车牌图像进行特征提取时容易造成车牌图像的特征损失，并且在卷积神经网络深度增加时，出现了梯度消失问题。在本实施例 中，为了提高后续车牌字符识别的准确性，可在卷积神经网络中设置残差网络结构，从而通过残差网络结构，有效防止卷积神经网络中增加深度带来的梯度消失和减少在卷积过程中的特征损失的问题。
在本申请的一个实施例中，上述卷积神经网络除了包括残差网络结构，还可以包括卷积层和池化层。在本实施实施例中，为了避免梯度消失和减少卷积过程的特征损失，本实施例中的卷积神经网络可以按照卷积层、残差网络结构和池化层顺次连接。也就是说，本实施例的卷积神经网络包括顺次连接的卷积层、残差网络结构和池化层。
具体地，在将车牌图像输入到卷积神经网络后，卷积神经网络中的卷积层将对车牌图像进行特征提取，以得到车牌图像的第一特征图像，对应地，可将第一特征图像输入到残差网络结构，并将残差结构输出的特征图像输入到池化层中进行相应处理。
在本申请的一个实施例中，为了在保证识别效率和识别准确的情况下，可在卷积神经网络中设置三层卷积层，对车牌图像进行特征提取。
在本申请的一个实施例中，在上述卷积神经网络使用三层卷积层时，为了避免梯度消失和减少卷积过程中的特征损失，在本实施例中可将第一卷积层、残差网络结构、第二卷积层、第三卷积层和池化层顺次相连，以形成本实施例的卷积神经网络，从而最大程度的减少卷积过程的特征损失。
在本实施中，为了进一步提高车牌识别准确度，可将第一卷积层设置为1*1卷积层，即，第一卷积层的P＝1x1，其中，P为padding参数。
在本申请的一个实施例中，在卷积神经网络中第一卷积层后加上批标准化Batch Normalization层，从而可有效提升网络性能，加快收敛速度。
在本申请的一个实施例中，为了能有效防止梯度消失问题，本实施例中的卷积神经网络可使用Leaky ReLU作为激活函数。
在本实施例中，为了在减少计算量的同时，有效防止卷积神经网络的梯度消失和减少在卷积过程中的特征损失，本实施例的残差网络结构包括第一残差网络结构和第二残差网络结构。对应地，卷积神经网络包括顺次相连的第一卷积层、第一残差网络结构、第二残差网络结构、第二卷积层、第三卷积层和池化层。
其中，包括第一残差网络结构和第二残差网络结构的卷积神经网络的网络结构示意图，如图2所示。图2中的卷积神经网络的网络结构具体为：第一卷积层、与第一卷积层连接的第一残差网络结构、与第一残差网络结构连接的第二残差网络结构、与第二残差网络结构连接的第二卷积层、与第二卷积层连接的第三卷积层，以及与第三卷积层连接的最大池化层。
在本申请的一个实施例中，卷积神经网络参数示意，如表1所示，其中，Filters为 卷积核数，K为卷积核大小，S为步长，P为padding参数。
表2卷积神经网络参数
Figure PCTCN2021074915-appb-000001
其中，表1中的Residual-1表示第一残差网络结构，Residual-2表示第二残差网络结构，ConvolutionN表示第N卷积层，其中，N为1到3中的任意一个整数。
在本申请的一个实施例中，为了进一步有效提升网络性能，加快收敛速度，可在卷积神经网络中的第二残差网络结构和Leaky ReLU之间设置Batch Normalization层。
在本申请的一个实施例中，第一残差网络结构和第二残差网络结构的网络结构相同，网络结构均包括第一网络子结构、第二网络子结构和加法器，第一网络子结构包括顺次相连的最大池化层、第一卷积子层、第二卷积子层、第三卷积子层和第四卷积子层，第二网络子结构包括第五卷积子网络，最大池化层和第五卷积子网络的输入端均与第一卷积层的输出层的输出端连接，第四卷积子层和第五卷积子网络的输出端均与加法器的输入端相连。
其中，第一残差网络结构的网络结构示意图，如图3所示。
在本实施例中，在卷积神经网络中的第一卷积层输出车牌图像的特征图像后，可将该特征图像输入到第一残差网络结构中，并将第一残差网络结构的输出再次输入到第二残差网络结构中，然后，将第二残差网络结构输出的特征信息输入到与第二残差 网络结构连接的相应处理层，以通过相应处理层进行后续处理。
其中，本实施例中的第一残差网络结构和第二残差网络结构的网络结构虽然相同，但是两者所使用的网络参数是不同的。
本实施例中，通过以跳层连接的形式实现残差网络结构，有效防止卷积神经网络的梯度消失和减少在卷积过程中的特征损失，并且，还可以在一定程度上缓解梯度弥散问题。另外，本实施例的残差网络结构还可以使得信息前后向传播更加顺畅。
具体而言，第一残差网络结构和第二残差网络结构中卷积层所对应的卷积核数是不同的。
其中，本实施例中第一残差网络结构对应的细节参数，如表2所示，其中，Filters为卷积核数，K为卷积核大小，S为步长，P为padding参数。
表2第一残差网络结构的细节参数
Figure PCTCN2021074915-appb-000002
其中，表中的Convolution表示卷积层，ConvolutionN表示第N卷积子层，其中，N为1到5中的任意整数。
其中，第二残差网络结构的细节参数，如表3所示。
表3第二残差网络结构的细节参数
Figure PCTCN2021074915-appb-000003
Figure PCTCN2021074915-appb-000004
步骤104，通过双向循环神经网络模型解析特征信息，以得到车牌图像对应的车牌字符。
在本申请实施例中，上述双向循环神经网络模型可以包括双向循环神经网络和连接时序分类(Connectionist Temporal Classification，CTC)网络。
在本申请的一个实施例中，为了提高识别车牌字符的准确性，本实施例的双向循环神经网络可以包括顺次连接的第一层双向长短时记忆网络(bidirectional long short-term memory，BLSTM)和第二层双向长短时记忆网络。
其中，本实施例中的双向长短时记忆网络中的每一个LSTM网络含有512个隐藏单元，因此，每层BLSTM的输出深度为1024，最后一层BLSTM的输出接两层全连接层，第一层全连接的深度为1024，第二层全连接的深度为车牌字符类别数(例如，66)。
在本申请的一个实施例中，为了防止过拟合情况的发生，本实施例的两层全连接层之间设置有dropout层。也就是说，本实施例的第二层双向长短时记忆网络包括：第一全连接层，与第一全连接层连接的dropout层，以及与dropout层连接的第二全连接层。
其中，本实施例中的双向循环神经网络模型的网络结构的示意图，如图4所示。
在本实施例中，在从所采集到的车辆图像中得到车牌图像后，直接通过卷积神经网络对整个车牌图像进行特征提取，并通过双向循环神经网络模型解析所提取到的特征信息，以得到车牌图像中的车牌字符。由此，可以看出，该实施例由于无需对车牌图像进行字符分割，因此，可减少车牌识别过程中的字符分割所使用的时间，进而可提高车牌识别速度。
本申请实施例的车牌字符识别方法，在获取图像采集设备所采集到的车辆图像后，对车牌图像中的车辆区域进行定位，以得到车牌图像，并通过包括残差网络结构的卷积神经网络对车牌图像进行特征提取，有效避免梯度消失和减少了卷积神经网络卷积过程中的特征损失，从而使得双向循环神经网络模型可基于车牌图像的特征信息准确识别出车牌图像的车牌字符。由此，不需要对车牌进行字符分割，通过直接识别整个车牌即可得到车牌上的车牌字符，避免对车牌的字符进行切分和分开识别，提高了识别速度和识别准确度。
图5是根据本申请另一个实施例的车牌字符识别方法的流程示意图。其中，需要说明的是，该实施例是对上述实施例的进一步细化或者优化。
如图5所示，该车牌字符识别方法可以包括：
步骤501，获取图像采集设备所采集到的车辆图像。
步骤502，采用YOLOV3网络对车辆图像进行车牌区域定位，以得到车牌图像。
步骤503，利用空间变换网络，对车牌图像进行倾斜校正。
在本实施例中，空间变换网络(Spatial Transformer Network，STN)可以包括本地网络、网格生成器和采样器。
具体地，在获取车牌图像后，可将车牌图像输入到本地网络中，以得用来校正车牌图像的仿射变换参数；将仿射变换参数输入至网格生成器，以得到对应的仿射变换矩阵；通过采样器基于仿射变换矩阵对车牌图像中的每个像素坐标进行仿射变换，以校正车牌图像。
其中，上述本地网络是一个用来回归变换参数θ的网络，该网络的具体参数表示，如表4，其中，Filters为卷积核数，K为卷积核大小，S为步长，P为padding参数；它的输入是灰度图像，然后经过一系列的隐藏网络层，最后输出空间变换参数。θ的形式可以多样，如需实现2D仿射变换，θ就是一个6维(2x3)向量的输出。θ的尺寸大小依赖于变换的类型，输入是2D图像，需实现2D仿射变换。在这里，该网络输出一个2*3的矩阵如公式2-1所示，表示为A θ。
Figure PCTCN2021074915-appb-000005
表4本地网络参数
类型 参数
Input GrayScaleImage,W:220,H:32
Convolution1 Filters:8,K＝3,S＝1,P＝1x1
Leaky ReLU -
Max pooling1 K＝2,S＝2,P＝0*0
Convolution1 Filters:16,K＝3,S＝1,P＝1x1
Leaky ReLU -
Convolution1 Filters:32,K＝3,S＝1,P＝1x1
Leaky ReLU -
FC Layer Filters:64
Leaky ReLU -
Dropout -
FC Layer Filters:6
上述网格生成器(Grid Generator)是依据预测的变换参数来构建一个采样网格，它是将一组输入图像中的点经过采样变换后得到的输出。网格生成器其实得到的是一种映射关系T θ。假设输入图像的每个像素的坐标为
Figure PCTCN2021074915-appb-000006
STN输出的每个像素坐标为
Figure PCTCN2021074915-appb-000007
空间变换函数T θ为二维仿射变换函数，那么
Figure PCTCN2021074915-appb-000008
和
Figure PCTCN2021074915-appb-000009
的对应关系可以表示为2-2：
Figure PCTCN2021074915-appb-000010
采样器利用采样网格和输入的图像同时作为输入，得到了输入图像经过变换之后的结果。
上述采样器在网络生成器所得到的映射关系的基础上对输入图像特征进行采样，得到经过空间变换达到校正效果的图像特征。
其中，本实施例中所用的采样方法为双线性插值法。
本实施例通过STN在对车牌图像进行校正的同时，还可以通过本地网络过滤车牌图像周围的噪声干扰，使车牌更易于识别。
步骤504，通过卷积神经网络对车牌图像进行特征提取，以得到车牌图像的特征信息，其中，卷积神经网络包括残差网络结构。
步骤505，通过双向循环神经网络模型解析特征信息，以得到车牌图像对应的车牌字符。
其中，车牌字符可以包括汉字、字母、数字。
本申请实施例的车牌字符识别方法，在获取图像采集设备所采集到的车辆图像后，对车牌图像中的车辆区域进行定位，以得到车牌图像，并对车牌图像进行倾斜校正，并通过包括残差网络结构的卷积神经网络对校正后的车牌图像进行特征提取，有效避免梯度消失和减少了卷积神经网络卷积过程中的特征损失，从而使得双向循环神经网络模型可基于车牌图像的特征信息准确识别出车牌图像的车牌字符。
基于上述描述，可以看出，本实施例在对图片车辆中的车牌直接定位后，不需要对车牌进行字符分割，直接将定位后得到的车牌图像输入到包括残差网络结构的卷积神经网络进行特征提取，并采用双向循环神经网络模型基于车牌图像的特征信息进行车牌字符识别，可以直接识别整个车牌，避免对车牌的字符进行切分和分开识别，提高了识别速度。
本实施例对采用STN的车牌识别方法和不采用STN的车牌识别方法进行了实验，根据实验结果，确定移除STN网络的车牌识别准确率为94.6％，低于加入STN网络的车牌识别模型准确率96.1％。移除STN网络的车牌识别平均损失为0.53，高于加入STN网络的车牌平均损失为0.40。由此，可以看出，在对车牌图像输入卷积神经网络之前进行倾斜校正，可以进一步提高车牌字符的识别准确度。
为了表明本申请BLSTM的层数以及隐藏元(Hidden units)数的合理性，表5给出了本申请网络在无STN网络时，一层BLSTM、两层BLSTM和三层BLSTM在隐藏元数为128、256、512、1024时车牌识别准确率及车牌识别的平均损失，由于前面已经验证了STN网络的必要性，为了提升训练收敛速度，此处网络验证时将不再加入STN网络，通过以此获取的实验参数反映本申请所选取BLSTM参数的合理性，三层BLSTM时，车牌识别网络的收敛速度远慢于两层BLSTM和一层BLSTM，结果表明本申请网络所选择的两层BLSTM、512隐藏元是最合理的方案。
表5 BLSTM分析
Figure PCTCN2021074915-appb-000011
其中，需要说明的是，上述表中在斜杠之前的数字为车牌识别准确率ACC的取值，斜杠之后的数字表示车牌识别平均损失loss的取值，例如，94.6/0.53表示车牌识别准确率ACC为94.6％，车牌识别平均损失为0.53。
另外，本实施例还对全连接层FC_layer进行了分析。
为了选取最佳全连接的层数，表6给出了本申请网络在无STN网络时在一层全连接、两层全连接、三层全连接时的车牌识别准确率以及模型识别车牌的平均损失，通过以此获取的实验参数即可反映本申请网络所选取的全连接参数的合理。ACC表示在相应模型下车牌识别的准确率，loss表示在相应模型下车牌识别的平均损失，F1表示第一层全连接的输出深度，F2表示第二层全连接的输出深度，F3表示第三层全连接的输出深度。结果证明本申请模型拥有最优性能。
表6全连接层分析
Figure PCTCN2021074915-appb-000012
与上述几种实施例提供的车牌字符识别方法相对应，本申请的一种实施例还提供一种车牌字符识别装置，由于本申请实施例提供的车牌字符识别装置与上述几种实施例提供的车牌字符识别方法相对应，因此在车牌字符识别方法的实施方式也适用于本实施例提供的车牌字符识别装置，在本实施例中不再详细描述。
图6是根据本申请一个实施例的车牌字符识别装置的结构示意图。
如图6所示，该车牌字符识别装置600包括图像采集模块110、车牌定位模块120、特征提取模块130和识别模块140，其中：
图像采集模块110，用于获取图像采集设备所采集到的车辆图像。
车牌定位模块120，用于对车辆图像进行车牌区域定位，以得到车牌图像。
特征提取模块130，用于通过卷积神经网络对车牌图像进行特征提取，以得到车牌图像的特征信息，其中，卷积神经网络包括残差网络结构。
识别模块140，用于通过双向循环神经网络模型解析特征信息，以得到车牌图像对应的车牌字符。
在本申请的一个实施例中，卷积神经网络包括顺次连接的卷积层、残差网络结构和池化层。
在本申请的一个实施例中，卷积神经网络包括顺次相连的第一卷积层、残差网络结构、第二卷积层、第三卷积层和池化层。
在本申请的一个实施例中，所述卷积层包括3层，所述残差网络结构包括第一残差网络结构和第二残差网络结构，所述卷积神经网络包括顺次相连的第一卷积层、第一残差网络结构、第二残差网络结构、第二卷积层、第三卷积层和池化层。
在本申请的一个实施例中，第一残差网络结构和第二残差网络结构的网络结构相同，网络结构包括第一网络子结构、第二网络子结构和加法器，第一网络子结构包括顺次相连的最大池化层、第一卷积子层、第二卷积子层、第三卷积子层和第四卷积子层，第二网络子结构包括第五卷积子网络，最大池化层和第五卷积子网络的输入端均与第一卷积层的输出层的输出端连接，第四卷积子层和第五卷积子网络的输出端均与加法器的输入端相连。
在本申请的一个实施例中，在图6所示的实施例基础上，如图7所示，该装置还包括：
倾斜校正模块150，用于利用空间变换网络，对车牌图像进行倾斜校正。
在本申请的一个实施例中，空间变换网络包括本地网络、网格生成器和采样器，上述倾斜校正模块150，具体用于：将车牌图像输入到本地网络中，以得用来校正车牌图像的仿射变换参数；将仿射变换参数输入至网格生成器，以得到对应的仿射变换矩阵；通过采样器基于仿射变换矩阵对车牌图像中的每个像素坐标进行仿射变换，以校正车牌图像。
在本申请的一个实施中，双向循环神经网络模型包括双向循环神经网络和连接时序分类CTC网络。
在本申请的一个实施中，双向循环神经网络包括顺次连接的第一层双向长短时记忆网络和第二层双向长短时记忆网络。
在本申请的一个实施中第二层双向长短时记忆网络包括：第一全连接层；与第一全连接层连接的dropout层；与dropout层连接的第二全连接层。
本申请实施例的车牌字符识别装置，在获取图像采集设备所采集到的车辆图像后，对车牌图像中的车辆区域进行定位，以得到车牌图像，并通过包括残差网络结构的卷积神经网络对车牌图像进行特征提取，有效避免梯度消失和减少了卷积神经网络卷积过程中的特征损失，从而使得双向循环神经网络模型可基于车牌图像的特征信息准确识别出车牌图像的车牌字符。由此，不需要对车牌进行字符分割，通过直接识别整个车牌即可得到车牌上的车牌字符，避免对车牌的字符进行切分和分开识别，提高了识别速度和识别准确度。
根据本申请的实施例，本申请还提供了一种电子设备和一种可读存储介质。
如图8所示，是根据本申请实施例的车牌字符识别方法的电子设备的框图。
如图8所示，该电子设备该电子设备包括：
存储器1001、处理器1002及存储在存储器1001上并可在处理器1002上运行的计算机指令。
处理器1002执行指令时实现上述实施例中提供的数据库管理方法。
进一步地，电子设备还包括：
通信接口1003，用于存储器1001和处理器1002之间的通信。
存储器1001，用于存放可在处理器1002上运行的计算机指令。
存储器1001可能包含高速RAM存储器，也可能还包括非易失性存储器(non-volatile memory)，例如至少一个磁盘存储器。
处理器1002，用于执行程序时实现上述实施例的车牌字符识别方法。
如果存储器1001、处理器1002和通信接口1003独立实现，则通信接口1003、存储器1001和处理器1002可以通过总线相互连接并完成相互间的通信。总线可以是工业标准体系结构(Industry Standard Architecture，简称为ISA)总线、外部设备互连(Peripheral Component，简称为PCI)总线或扩展工业标准体系结构(Extended Industry Standard Architecture，简称为EISA)总线等。总线可以分为地址总线、数据总线、控制总线等。为便于表示，图8中仅用一条粗线表示，但并不表示仅有一根总线或一种类型的总线。
可选的，在具体实现上，如果存储器1001、处理器1002及通信接口1003，集成在一块芯片上实现，则存储器1001、处理器1002及通信接口1003可以通过内部接口完成相互间的通信。
处理器1002可能是一个中央处理器(Central Processing Unit，简称为CPU)，或者是特定集成电路(Application Specific Integrated Circuit，简称为ASIC)，或者是被配置成实施本申请实施例的一个或多个集成电路。
在本说明书的描述中，参考术语“一个实施例”、“一些实施例”、“示例”、“具体示例”、或“一些示例”等的描述意指结合该实施例或示例描述的具体特征、结构、材料或者特点包含于本申请的至少一个实施例或示例中。在本说明书中，对上述术语的示意性表述不必须针对的是相同的实施例或示例。而且，描述的具体特征、结构、材料或者特点可以在任一个或多个实施例或示例中以合适的方式结合。此外，在不相互矛盾的情况下，本领域的技术人员可以将本说明书中描述的不同实施例或示例以及不同实施例或示例的特征进行结合和组合。
此外，术语“第一”、“第二”仅用于描述目的，而不能理解为指示或暗示相对重要性或 者隐含指明所指示的技术特征的数量。由此，限定有“第一”、“第二”的特征可以明示或者隐含地包括至少一个该特征。在本申请的描述中，“多个”的含义是至少两个，例如两个，三个等，除非另有明确具体的限定。
流程图中或在此以其他方式描述的任何过程或方法描述可以被理解为，表示包括一个或更多个用于实现定制逻辑功能或过程的步骤的可执行指令的代码的模块、片段或部分，并且本申请的优选实施方式的范围包括另外的实现，其中可以不按所示出或讨论的顺序，包括根据所涉及的功能按基本同时的方式或按相反的顺序，来执行功能，这应被本申请的实施例所属技术领域的技术人员所理解。
在流程图中表示或在此以其他方式描述的逻辑和/或步骤，例如，可以被认为是用于实现逻辑功能的可执行指令的定序列表，可以具体实现在任何计算机可读介质中，以供指令执行系统、装置或设备(如基于计算机的系统、包括处理器的系统或其他可以从指令执行系统、装置或设备取指令并执行指令的系统)使用，或结合这些指令执行系统、装置或设备而使用。就本说明书而言，"计算机可读介质"可以是任何可以包含、存储、通信、传播或传输程序以供指令执行系统、装置或设备或结合这些指令执行系统、装置或设备而使用的装置。计算机可读介质的更具体的示例(非穷尽性列表)包括以下：具有一个或多个布线的电连接部(电子装置)，便携式计算机盘盒(磁装置)，随机存取存储器(RAM)，只读存储器(ROM)，可擦除可编辑只读存储器(EPROM或闪速存储器)，光纤装置，以及便携式光盘只读存储器(CDROM)。另外，计算机可读介质甚至可以是可在其上打印所述程序的纸或其他合适的介质，因为可以例如通过对纸或其他介质进行光学扫描，接着进行编辑、解译或必要时以其他合适方式进行处理来以电子方式获得所述程序，然后将其存储在计算机存储器中。
应当理解，本申请的各部分可以用硬件、软件、固件或它们的组合来实现。在上述实施方式中，多个步骤或方法可以用存储在存储器中且由合适的指令执行系统执行的软件或固件来实现。如，如果用硬件来实现和在另一实施方式中一样，可用本领域公知的下列技术中的任一项或他们的组合来实现：具有用于对数据信号实现逻辑功能的逻辑门电路的离散逻辑电路，具有合适的组合逻辑门电路的专用集成电路，可编程门阵列(PGA)，现场可编程门阵列(FPGA)等。
本技术领域的普通技术人员可以理解实现上述实施例方法携带的全部或部分步骤是可以通过程序来指令相关的硬件完成，所述的程序可以存储于一种计算机可读存储介质中，该程序在执行时，包括方法实施例的步骤之一或其组合。
此外，在本申请各个实施例中的各功能单元可以集成在一个处理模块中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个模块中。上述集成的模块既 可以采用硬件的形式实现，也可以采用软件功能模块的形式实现。所述集成的模块如果以软件功能模块的形式实现并作为独立的产品销售或使用时，也可以存储在一个计算机可读取存储介质中。
上述提到的存储介质可以是只读存储器，磁盘或光盘等。尽管上面已经示出和描述了本申请的实施例，可以理解的是，上述实施例是示例性的，不能理解为对本申请的限制，本领域的普通技术人员在本申请的范围内可以对上述实施例进行变化、修改、替换和变型。

Claims (15)

  1. 一种车牌字符识别方法，其特征在于，所述方法包括：
    获取图像采集设备所采集到的车辆图像；
    对车辆图像进行车牌区域定位，以得到车牌图像；
    通过卷积神经网络对所述车牌图像进行特征提取，以得到所述车牌图像的特征信息，其中，所述卷积神经网络包括残差网络结构；
    通过双向循环神经网络模型解析所述特征信息，以得到所述车牌图像对应的车牌字符。
  2. 根据权利要求1所述的方法，其特征在于，所述卷积神经网络包括顺次连接的卷积层、所述残差网络结构和池化层。
  3. 根据权利要求2所述的方法，其特征在于，所述卷积层包括3层，所述残差网络结构包括第一残差网络结构和第二残差网络结构，所述卷积神经网络包括顺次相连的第一卷积层、第一残差网络结构、第二残差网络结构、第二卷积层、第三卷积层和池化层。
  4. 根据权利要求3所述的方法，其特征在于，所述第一残差网络结构和所述第二残差网络结构的网络结构相同，所述网络结构包括第一网络子结构、第二网络子结构和加法器，所述第一网络子结构包括顺次相连的最大池化层、第一卷积子层、第二卷积子层、第三卷积子层和第四卷积子层，所述第二网络子结构包括第五卷积子网络，所述最大池化层和所述第五卷积子网络的输入端均与所述第一卷积层的输出层的输出端连接，所述第四卷积子层和所述第五卷积子网络的输出端均与加法器的输入端相连。
  5. 根据权利要求1-4中任一项所述的方法，其特征在于，在所述通过卷积神经网络对所述车牌图像进行特征提取，以得到所述车牌图像的特征信息之前，所述方法还包括：
    利用空间变换网络，对所述车牌图像进行倾斜校正。
  6. 根据权利要求1-5中任一项所述的方法，其特征在于，所述双向循环神经网络模型包括双向循环神经网络和连接时序分类CTC网络。
  7. 根据权利要求6所述的方法，其特征在于，所述双向循环神经网络包括顺次连接的第一层双向长短时记忆网络和第二层双向长短时记忆网络。
  8. 一种车牌字符识别装置，其特征在于，所述装置包括：
    图像采集模块，用于获取图像采集设备所采集到的车辆图像；
    车牌定位模块，用于对车辆图像进行车牌区域定位，以得到车牌图像；
    特征提取模块，用于通过卷积神经网络对所述车牌图像进行特征提取，以得到所述车牌图像的特征信息，其中，所述卷积神经网络包括残差网络结构；
    识别模块，用于通过双向循环神经网络模型解析所述特征信息，以得到所述车牌图像 对应的车牌字符。
  9. 根据权利要求8所述的装置，其特征在于，所述卷积神经网络包括顺次连接的卷积层、所述残差网络结构和池化层。
  10. 根据权利要求9所述的装置，其特征在于，所述卷积层包括3层，所述残差网络结构包括第一残差网络结构和第二残差网络结构，所述卷积神经网络包括顺次相连的第一卷积层、第一残差网络结构、第二残差网络结构、第二卷积层、第三卷积层和池化层。
  11. 根据权利要求10所述的装置，其特征在于，所述第一残差网络结构和所述第二残差网络结构的网络结构相同，所述网络结构包括第一网络子结构、第二网络子结构和加法器，所述第一网络子结构包括顺次相连的最大池化层、第一卷积子层、第二卷积子层、第三卷积子层和第四卷积子层，所述第二网络子结构包括第五卷积子网络，所述最大池化层和所述第五卷积子网络的输入端均与所述第一卷积层的输出层的输出端连接，所述第四卷积子层和所述第五卷积子网络的输出端均与加法器的输入端相连。
  12. 根据权利要求8-11中任一项所述的装置，其特征在于，所述双向循环神经网络模型包括双向循环神经网络和连接时序分类CTC网络。
  13. 根据权利要求12所述的装置，其特征在于，所述双向循环神经网络包括顺次连接的第一层双向长短时记忆网络和第二层双向长短时记忆网络。
  14. 一种电子设备，包括：存储器，处理器；所述存储器中存储有计算机指令，当所述计算机指令被所述处理器执行时，实现如权利要求1-7中任一项所述的车牌字符识别方法。
  15. 一种存储有计算机指令的非瞬时计算机可读存储介质，其特征在于，所述计算机指令用于使所述计算机执行权利要求1-7中任一项所述的车牌字符识别方法。
PCT/CN2021/074915 2020-03-30 2021-02-02 车牌字符识别方法、装置、电子设备和存储介质 WO2021196873A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/914,993 US20230154210A1 (en) 2020-03-30 2021-02-02 Method for recognizing license plate characters, electronic device and storage medium

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010234926.2A CN111444917A (zh) 2020-03-30 2020-03-30 车牌字符识别方法、装置、电子设备和存储介质
CN202010234926.2 2020-03-30

Publications (1)

Publication Number Publication Date
WO2021196873A1 true WO2021196873A1 (zh) 2021-10-07

Family

ID=71654027

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/074915 WO2021196873A1 (zh) 2020-03-30 2021-02-02 车牌字符识别方法、装置、电子设备和存储介质

Country Status (3)

Country Link
US (1) US20230154210A1 (zh)
CN (1) CN111444917A (zh)
WO (1) WO2021196873A1 (zh)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114092926A (zh) * 2021-10-20 2022-02-25 杭州电子科技大学 一种复杂环境下的车牌定位和识别方法
CN114299486A (zh) * 2021-12-30 2022-04-08 郑州信大先进技术研究院 嵌入式端实时高效的车牌识别方法、设备及介质
CN114495106A (zh) * 2022-04-18 2022-05-13 电子科技大学 一种应用于dfb激光器芯片的深度学习mocr方法
CN116958954A (zh) * 2023-07-27 2023-10-27 匀熵智能科技(无锡)有限公司 基于关键点与旁路矫正的车牌识别方法、装置及存储介质
CN117523543A (zh) * 2024-01-08 2024-02-06 成都大学 一种基于深度学习的金属压印字符识别方法

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111444917A (zh) * 2020-03-30 2020-07-24 合肥京东方显示技术有限公司 车牌字符识别方法、装置、电子设备和存储介质
CN112149661A (zh) * 2020-08-07 2020-12-29 珠海欧比特宇航科技股份有限公司 车牌识别方法、装置及介质
CN112016432A (zh) * 2020-08-24 2020-12-01 高新兴科技集团股份有限公司 基于深度学习的车牌字符识别方法、存储介质及电子设备
CN112215245A (zh) * 2020-11-05 2021-01-12 中国联合网络通信集团有限公司 一种图像识别方法及装置
CN112966631A (zh) * 2021-03-19 2021-06-15 浪潮云信息技术股份公司 一种无限制安防场景下的车牌检测识别系统及其方法
CN113705561A (zh) * 2021-09-02 2021-11-26 北京云蝶智学科技有限公司 特殊符号的识别方法及装置
CN117409298B (zh) * 2023-12-15 2024-04-02 西安航空学院 针对路面车辆识别的多尺寸目标精确识别方法及设备

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109815956A (zh) * 2018-12-28 2019-05-28 安徽清新互联信息科技有限公司 一种基于自适应位置分割的车牌字符识别方法
CN110321755A (zh) * 2018-03-28 2019-10-11 中移(苏州)软件技术有限公司 一种识别方法及装置
CN110414451A (zh) * 2019-07-31 2019-11-05 深圳市捷顺科技实业股份有限公司 一种基于端对端的车牌识别方法、装置、设备及存储介质
CN111444917A (zh) * 2020-03-30 2020-07-24 合肥京东方显示技术有限公司 车牌字符识别方法、装置、电子设备和存储介质

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110363049B (zh) * 2018-04-10 2024-01-12 阿里巴巴集团控股有限公司 图形元素检测识别和类别确定的方法及装置
CN109934864B (zh) * 2019-03-14 2023-01-20 东北大学 面向机械臂抓取位姿估计的残差网络深度学习方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110321755A (zh) * 2018-03-28 2019-10-11 中移(苏州)软件技术有限公司 一种识别方法及装置
CN109815956A (zh) * 2018-12-28 2019-05-28 安徽清新互联信息科技有限公司 一种基于自适应位置分割的车牌字符识别方法
CN110414451A (zh) * 2019-07-31 2019-11-05 深圳市捷顺科技实业股份有限公司 一种基于端对端的车牌识别方法、装置、设备及存储介质
CN111444917A (zh) * 2020-03-30 2020-07-24 合肥京东方显示技术有限公司 车牌字符识别方法、装置、电子设备和存储介质

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114092926A (zh) * 2021-10-20 2022-02-25 杭州电子科技大学 一种复杂环境下的车牌定位和识别方法
CN114299486A (zh) * 2021-12-30 2022-04-08 郑州信大先进技术研究院 嵌入式端实时高效的车牌识别方法、设备及介质
CN114299486B (zh) * 2021-12-30 2024-04-16 郑州信大先进技术研究院 嵌入式端实时高效的车牌识别方法、设备及介质
CN114495106A (zh) * 2022-04-18 2022-05-13 电子科技大学 一种应用于dfb激光器芯片的深度学习mocr方法
CN116958954A (zh) * 2023-07-27 2023-10-27 匀熵智能科技(无锡)有限公司 基于关键点与旁路矫正的车牌识别方法、装置及存储介质
CN116958954B (zh) * 2023-07-27 2024-03-22 匀熵智能科技(无锡)有限公司 基于关键点与旁路矫正的车牌识别方法、装置及存储介质
CN117523543A (zh) * 2024-01-08 2024-02-06 成都大学 一种基于深度学习的金属压印字符识别方法
CN117523543B (zh) * 2024-01-08 2024-03-19 成都大学 一种基于深度学习的金属压印字符识别方法

Also Published As

Publication number Publication date
CN111444917A (zh) 2020-07-24
US20230154210A1 (en) 2023-05-18

Similar Documents

Publication Publication Date Title
WO2021196873A1 (zh) 车牌字符识别方法、装置、电子设备和存储介质
US9965719B2 (en) Subcategory-aware convolutional neural networks for object detection
US20220092351A1 (en) Image classification method, neural network training method, and apparatus
WO2021238281A1 (zh) 一种神经网络的训练方法、图像分类系统及相关设备
CN108898086B (zh) 视频图像处理方法及装置、计算机可读介质和电子设备
WO2020164270A1 (zh) 基于深度学习的行人检测方法、系统、装置及存储介质
CN109241982B (zh) 基于深浅层卷积神经网络的目标检测方法
WO2021244621A1 (zh) 基于全局引导选择性上下文网络的场景语义解析方法
WO2019201035A1 (zh) 对图像中的对象节点的识别方法、装置、终端及计算机可读存储介质
JP2022515895A (ja) 物体認識方法及び装置
CN109766805B (zh) 一种基于深度学习的双层车牌字符识别方法
US20220148291A1 (en) Image classification method and apparatus, and image classification model training method and apparatus
CN111428664B (zh) 一种基于深度学习技术的计算机视觉的实时多人姿态估计方法
WO2021047587A1 (zh) 手势识别方法、电子设备、计算机可读存储介质和芯片
CN113537211B (zh) 一种基于非对称iou的深度学习车牌框定位方法
CN112307853A (zh) 航拍图像的检测方法、存储介质和电子装置
WO2024001123A1 (zh) 基于神经网络模型的图像识别方法、装置及终端设备
CN109710788A (zh) 图像样本标注和管理方法及设备
WO2024077781A1 (zh) 基于卷积神经网络模型的图像识别方法、装置及终端设备
CN115862045B (zh) 基于图文识别技术的病例自动识别方法、系统、设备及存储介质
CN111209873A (zh) 一种基于深度学习的高精度人脸关键点定位方法及系统
US20220198224A1 (en) Face recognition method, terminal device using the same, and computer readable storage medium
CN116229406B (zh) 车道线检测方法、系统、电子设备及存储介质
JP7246104B2 (ja) テキスト行識別に基づくナンバープレート識別方法
CN114067186B (zh) 一种行人检测方法、装置、电子设备及存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21779976

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21779976

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 15.05.2023)

122 Ep: pct application non-entry in european phase

Ref document number: 21779976

Country of ref document: EP

Kind code of ref document: A1