WO2022111355A1 - 车牌识别方法及装置、存储介质、终端 - Google Patents

车牌识别方法及装置、存储介质、终端 Download PDF

Info

Publication number
WO2022111355A1
WO2022111355A1 PCT/CN2021/131158 CN2021131158W WO2022111355A1 WO 2022111355 A1 WO2022111355 A1 WO 2022111355A1 CN 2021131158 W CN2021131158 W CN 2021131158W WO 2022111355 A1 WO2022111355 A1 WO 2022111355A1
Authority
WO
WIPO (PCT)
Prior art keywords
license plate
image
feature map
character
recognized
Prior art date
Application number
PCT/CN2021/131158
Other languages
English (en)
French (fr)
Inventor
陈圣卫
Original Assignee
展讯通信（上海）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 展讯通信（上海）有限公司 filed Critical 展讯通信（上海）有限公司
Publication of WO2022111355A1 publication Critical patent/WO2022111355A1/zh

Links

Images

Classifications

    • 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
    • G06V30/14—Image acquisition
    • G06V30/148—Segmentation of character regions
    • G06V30/153—Segmentation of character regions using recognition of characters or words
    • 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
    • G06N3/08—Learning 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
    • G06V10/20—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
    • G06V10/20—Image preprocessing
    • G06V10/30—Noise filtering
    • 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

Definitions

  • Embodiments of the present invention relate to the field of license plate recognition, and in particular, to a license plate recognition method and device, a storage medium, and a terminal.
  • the mainstream deep learning solution extracts the feature map through the convolutional neural network, sequentially identifies the text information on each text position of the license plate through the cyclic neural network, and finally uses the CTC loss function to calculate the corresponding text position according to each element in the feature matrix. attention weights for text recognition.
  • license plates in our country usually contain Chinese characters, numbers and letters. Due to the diversity of content contained in license plates, the accuracy of license plate recognition results is low.
  • the technical problem solved by the embodiments of the present invention is that the accuracy of the license plate recognition result is low.
  • an embodiment of the present invention provides a license plate recognition method, which includes: acquiring an image of a license plate to be recognized; inputting the image of the license plate to be recognized into a character segmentation network for character segmentation to obtain several character images, each of which is a There are corresponding character labels, and the character segmentation network is pre-trained; the character images are input into the license plate recognition network, weight parameters are selected according to the character labels of each character image, and the selected weight parameters are used to carry out the character image.
  • the license plate recognition network is pre-trained, including multiple groups of weight parameters, each group of weight parameters corresponds to the character label one-to-one; according to the recognition result of each character image, the license plate corresponding to the license plate image is obtained. No.
  • inputting the to-be-recognized license plate image into a character segmentation network for character segmentation to obtain several character images includes: down-sampling the to-be-recognized license plate image to obtain a first feature map; A feature map is upsampled to obtain a second feature map, wherein an attention mechanism is added during at least one of the upsampling, the second feature map has the same scale as the first feature map; according to the second feature Figure, predict the position of each character and the character label of the character; according to the predicted position of each character, perform character segmentation on the image of the license plate to be recognized to obtain the several character images.
  • the performing upsampling on the first feature map to obtain the second feature map includes: using a plurality of serially connected upsampling modules to perform upsampling on the first feature map for multiple times to obtain the The second feature map, the multiple up-sampling modules are used to receive the first feature map or the up-sampling feature map output by the upper-level up-sampling module, and perform channel transformation processing on the up-sampling feature map output by the last up-sampling module, The second feature map is obtained, and the number of channels of the second feature map is related to the total number of categories of the character labels; wherein, at least one upsampling module includes an attention mechanism module, and the attention mechanism module is used for The channels of the input feature map are weighted, and the feature map after the weighting process is used as the up-sampled feature map.
  • the attention mechanism module is used to perform global average pooling processing and maximum pooling processing on the input feature map respectively, and obtain a feature map after global average pooling processing and a maximum pooling processed feature map.
  • feature map ; convolve the feature map after the global average pooling process and the feature map after the maximum pooling process respectively, to obtain two feature maps after convolution processing; according to the two convolution processing
  • the Sigmoid activation function is used to determine the weight of each channel, and the input feature map is weighted by using the channel weight to obtain the up-sampling feature map.
  • the upsampling module includes: an interpolation unit, configured to perform interpolation processing on the input feature map to obtain an interpolated feature map, wherein the input feature map is the first feature map or the upper The up-sampling feature map output by the first-level up-sampling module; a plurality of second residual networks connected in series are used to receive the feature map to be up-sampled or the processed feature map output by the second residual network of the previous level, the said The feature map to be upsampled is an interpolated feature map output by the interpolation unit, and each second residual network includes: one or more convolution layers and one or more grouped convolution layers. It also includes: for each upsampling, fusing the feature maps output by the second residual networks in series with the downsampling feature maps of the same scale, and using the fused feature maps as the attention mechanism module input of.
  • the interpolation unit is configured to perform interpolation processing on the input feature map by using a bilinear interpolation method.
  • the character segmentation network includes a plurality of down-sampling modules connected in series, and the down-sampling module is used for down-sampling the to-be-recognized license plate image to obtain a first feature map, and each down-sampling module includes a plurality of The first residual network connected in series, wherein: a plurality of first residual networks connected in series are used to receive the feature map to be down-sampled or the processed feature map output by the first residual network of the previous level.
  • the sampled feature map is the license plate image to be recognized or the down-sampled feature map output by the upper-level down-sampling module, and each residual network includes one or more convolutional layers and one or more grouped convolutional layers.
  • the license plate recognition method further includes: after acquiring the license plate image to be recognized, correcting the to-be-recognized license plate image.
  • the rectifying the image of the license plate to be recognized includes: using a space transformation matrix to correct the image of the license plate to be recognized.
  • the performing denoising on the to-be-recognized license plate image includes: performing image feature enhancement on the to-be-recognized license plate image to obtain an image after image feature enhancement; processing to obtain a feature map after nonlinear processing; performing a convolution operation on the feature map after nonlinear processing to obtain a weight matrix; using the weight matrix to weight the license plate image to be recognized to obtain a noise feature map; According to the noise feature map and the to-be-recognized license plate image, a de-noised feature map is obtained, and the de-noised feature map is used as the to-be-recognized license plate image.
  • performing nonlinear processing based on the image after image enhancement to obtain a feature map after nonlinear processing includes: performing image fusion on the image of the license plate to be recognized and the image after image feature enhancement, Obtain a fused image; perform nonlinear processing on the fused image to obtain a feature map after nonlinear processing.
  • Tanh activation function is used to perform nonlinear processing on the fused image.
  • performing image feature enhancement on the to-be-recognized license plate image includes: convolving the to-be-recognized license plate image through a warp layer, and processing it through a BN layer to obtain the image feature enhancement. Image.
  • An embodiment of the present invention also provides a license plate recognition device, comprising: an acquisition unit for acquiring a license plate image to be recognized; a character segmentation unit for inputting the to-be-recognized license plate image into a character segmentation network for character segmentation to obtain several character images , each character image has a corresponding character label, and the character segmentation network is pre-trained; the license plate recognition unit is used to input the several character images into the license plate recognition network, and select the weight parameter according to the character label of each character image.
  • each group of weight parameters corresponds to the character label one-to-one.
  • An embodiment of the present invention further provides a storage medium on which a computer program is stored, and when the computer program is run by a processor, the steps of any of the above-mentioned license plate recognition methods are executed.
  • An embodiment of the present invention further provides a terminal, including a memory and a processor, the memory stores a computer program that can run on the processor, and the processor executes any of the above license plates when running the computer program Identify the steps of the method.
  • the weight parameter corresponding to the character label is selected from the pre-trained license plate recognition network.
  • the weight parameter of the training uses the weight parameter corresponding to the character label to recognize the character image, so the accuracy of the license plate recognition result can be improved.
  • At least one of the upsampling modules includes an attention mechanism module, and the attention mechanism module is used to perform weighting processing on the channels of the input feature map, and The feature map after weighting processing is used as an up-sampled feature map.
  • the license plate with a certain inclination angle in the image of the license plate to be recognized can be quickly recognized, which greatly reduces the limitation of the camera placement angle and improves the accuracy of the inclination angle of the license plate in the license plate image to be recognized. Tolerance and recognition range, and can improve the accuracy of license plate recognition results and improve the robustness of the overall recognition model.
  • the image of the license plate to be recognized is denoised
  • the image of the license plate to be recognized and the image after image feature enhancement are image fused to obtain the fused image, and the nonlinear processing is performed based on the fused image to obtain the non-linear processed image. feature map.
  • Fig. 1 is a flow chart of a license plate recognition method in an embodiment of the present invention
  • FIG. 3 is a schematic structural diagram of a character segmentation network in an embodiment of the present invention.
  • FIG. 4 is a schematic structural diagram of a downsampling module in an embodiment of the present invention.
  • FIG. 5 is a schematic structural diagram of an upsampling module in an embodiment of the present invention.
  • FIG. 6 is a flowchart of an image denoising in an embodiment of the present invention.
  • FIG. 7 is a schematic structural diagram of a license plate recognition device in an embodiment of the present invention.
  • character segmentation is performed on the license plate image to be recognized to obtain several character images, and according to the character label corresponding to each character image, the license plate recognition network corresponding to the character label is selected from the pre-trained license plate recognition network.
  • Weight parameter the character image is recognized based on the selected weight parameter. Since different types of character labels correspond to pre-trained weight parameters respectively, and the weight parameters corresponding to the character labels are used to recognize the character image, the accuracy of the license plate recognition result can be improved.
  • An embodiment of the present invention provides a method for recognizing a license plate.
  • a flowchart of a method for recognizing a license plate in an embodiment of the present invention is provided.
  • the method for recognizing a license plate may specifically include the following steps:
  • Step S11 acquiring an image of the license plate to be recognized.
  • Step S12 Input the license plate image to be recognized into a character segmentation network for character segmentation to obtain several character images, each character image has a corresponding character label, and the character segmentation network is pre-trained.
  • inputting the license plate image to be recognized into a character segmentation network for character segmentation may include the following steps S121 to S124:
  • Step S121 down-sampling the to-be-recognized license plate image to obtain a first feature map.
  • the character segmentation network 30 may include a plurality of down-sampling modules 31 , and the plurality of down-sampling modules 31 are connected in series.
  • the image of the license plate to be recognized is down-sampled by a plurality of down-sampling modules 31 to obtain a first feature map.
  • the input of the first-level down-sampling module 31 is the image of the license plate to be recognized. Starting from the second stage downsampling module 31 , its input is the output of the previous stage downsampling module 31 . The last stage downsampling module 31 outputs the first feature map.
  • the character segmentation network 30 illustrated in FIG. 3 includes four downsampling modules 31 .
  • the number of down-sampling modules 31 and the down-sampling multiple of each down-sampling module can be set according to the requirements of the actual character segmentation network 30 such as weight reduction and down-sampling multiple, which are not limited here.
  • each downsampling module 31 may include a plurality of first residual networks 311 connected in series.
  • a plurality of first residual networks 311 connected in series are used to receive the feature map to be downsampled or the processed feature map output by the first residual network 311 of the previous level, and the feature map to be downsampled is the feature map to be identified.
  • the license plate image or the down-sampling feature map output by the upper-level down-sampling module 31 is used to receive the feature map to be downsampled or the processed feature map output by the first residual network 311 of the previous level.
  • the input of the first-level first residual network 311 is the license plate image to be recognized.
  • the input of the first first residual network 311 in the downsampling module 31 is the output of the previous stage downsampling module 31.
  • the output of the last stage first residual network 311 in the last stage downsampling module 31 is the first feature map.
  • the first residual network 311 may be constructed based on a convolutional neural network.
  • Each first residual network 311 includes one or more convolutional layers, and one or more grouped convolutional layers.
  • the use of a plurality of first residual networks 311 can avoid the disappearance of gradients, so as to ensure fine-grained image features in the down-sampled image obtained after down-sampling.
  • each first residual network 311 includes two convolutional layers with a convolution kernel of 1 ⁇ 1, and one grouped convolutional layer with a convolution kernel of 3 ⁇ 3.
  • the convolution layer with the convolution kernel of 1 ⁇ 1 can be executed first, and the output of the convolution layer with the convolution kernel of 1 ⁇ 1 can be used as the input of the grouped convolution layer with the convolution kernel of 3 ⁇ 3.
  • the output of a grouped convolutional layer with a kernel of 3 ⁇ 3 is used as the input of another convolutional layer with a kernel of 1 ⁇ 1.
  • the output of another convolutional layer with a convolution kernel of 1 ⁇ 1 can be used as the input of the first residual network 311 in the next stage, or as the input of the downsampling module 31 in the next stage, or when the other convolution kernel is
  • the output of another convolutional layer with a 1 ⁇ 1 convolution kernel is the first feature map.
  • the downsampling module 31 when the downsampling module 31 implements double downsampling, among the first residual networks 311 in the downsampling module 31, one convolution has a stride of 2, and all other convolutions have a stride of 1. .
  • the convolution with a stride of 2 can be set in the convolution layer with the convolution kernel of 1 ⁇ 1, or in the grouped convolution layer with the convolution kernel of 3 ⁇ 3.
  • the step size of one convolution is set to 4, and the step size of all other convolutions is set to is 1, where the convolution with stride 2 can be set in the convolution layer with the convolution kernel of 1 ⁇ 1, or in the grouped convolution layer with the convolution kernel of 3 ⁇ 3.
  • set the stride of the two convolutions to 2 respectively, and set the stride of all other convolutions to 1.
  • the convolution with a stride of 2 can be set to a convolutional layer with a convolution kernel of 1 ⁇ 1. , which can also be set in a grouped convolutional layer with a convolution kernel of 3 ⁇ 3.
  • the convolution step size in the first residual network 311 when the convolution step size in the first residual network 311 is greater than 1, the feature map input to the first residual network 311 needs to be sampled with the same length.
  • the convolution operation is performed, and the convolution result is added to the output of the last convolution in the first residual network 311 as the output of the first residual network 311 .
  • the step size of the convolution is correspondingly different, which can be set according to the specific requirements, which will not be repeated here.
  • Step S122 up-sampling the first feature map to obtain a second feature map.
  • the first feature map when the first feature map is up-sampled to obtain a second feature map, an attention mechanism is added during at least one of the up-sampling, and the second feature map has the same scale as the first feature map. For example, increase the attention mechanism at least on the last upsampling.
  • the character segmentation network 30 may include a plurality of upsampling modules 32 connected in series. Multiple upsampling modules 32 connected in series may be used to upsample the first feature map multiple times to obtain the second feature map, and multiple upsampling modules 32 are used to receive the first feature map or the previous feature map.
  • the up-sampling feature map output by the stage up-sampling module 32 is subjected to channel transformation processing on the up-sampling feature map output by the last stage up-sampling module 32 to obtain the second feature map.
  • the number of channels of the second feature map is the same as the related to the total number of categories of the character labels.
  • the up-sampling feature map output by the last-stage up-sampling module 32 can be convolved to implement channel transformation processing, and the number of channels is transformed into a second feature map with 3 channels.
  • the number of channels of the input license plate image to be recognized is 3.
  • At least one upsampling module 32 includes an attention mechanism module 323 .
  • the attention mechanism module 323 is configured to perform weighting processing on the channels of the input feature maps, and use the feature maps after the weighting processing as the up-sampled feature maps. Processing by the attention mechanism module 323 can improve the accuracy of character location positioning, effectively reduce the probability of missing characters, and improve the reliability of the character segmentation network.
  • the attention mechanism module 323 is configured to perform global average pooling on the input feature map to obtain a feature map after global average pooling. Perform maximum pooling on the input feature map to obtain a feature map after maximum pooling.
  • the attention mechanism module 323 convolves the feature map after the global average pooling process and the feature map after the max pooling process respectively, to obtain two feature maps after the convolution process.
  • the weight of each channel can be determined by using a sigmoid activation function, and the input feature map is weighted by using the weight of each channel to obtain the up-sampling feature map.
  • each upsampling module 32 includes an attention mechanism module 323 .
  • upsampling may be performed in an interpolation manner.
  • each upsampling module 32 may include an interpolation unit 321 and a plurality of second residual networks 322 connected in series. in:
  • the interpolation unit 321 is configured to perform interpolation processing on the input feature map to obtain an interpolated feature map, wherein the input feature map is the first feature map or the upsampling output by the upper-level upsampling module 32 feature map.
  • the interpolation unit 321 uses a bilinear interpolation method to perform interpolation processing on the input feature map.
  • a plurality of second residual networks 322 connected in series are used to receive the feature map to be upsampled or the processed feature map output by the second residual network 322 of the previous stage, and the feature map to be upsampled is the interpolation The interpolated feature map output by the unit 321.
  • the second residual network 322 may be constructed based on a convolutional neural network.
  • Each second residual network 322 may include one or more convolutional layers, and one or more grouped convolutional layers.
  • each second residual network 322 includes two convolutional layers with a convolution kernel of 1 ⁇ 1 and one grouped convolutional layer with a convolutional kernel of 3 ⁇ 3.
  • a convolutional layer with a convolution kernel of 1 ⁇ 1 can be executed first, and the output of a convolutional layer with a convolution kernel of 1 ⁇ 1 is used as the input of a grouped convolutional layer with a convolution kernel of 3 ⁇ 3.
  • the convolution kernel is The output of the 3 ⁇ 3 grouped convolutional layer is used as the input of another convolutional layer with a 1 ⁇ 1 convolution kernel.
  • the upsampling module includes the attention mechanism module 323
  • the output of another convolutional layer with a convolution kernel of 1 ⁇ 1 can be used as the input of the attention mechanism module 323 .
  • the upsampling module to which the second residual network 322 belongs does not include the attention mechanism module 323 , the output of another convolutional layer with a convolution kernel of 1 ⁇ 1 is used as the input of the next-level second residual network 322 .
  • the second residual network 322 is the last level of the second residual network 322, when the upsampling module 32 is the last level upsampling module 32 and does not include the attention mechanism module 323, it can be 1 based on another convolution kernel
  • the output of the ⁇ 1 convolutional layer yields the second feature map.
  • the stride of each convolution in the second residual network 322 is 1.
  • the convolution layer with the convolution kernel of 1 ⁇ 1 is firstly convolved, and the feature of each pixel can be enhanced.
  • the grouped convolution layer is configured in the second residual network 322 to ensure the convolution effect to reduce the amount of convolution parameters, make the network lighter, and speed up the network inference speed.
  • the process of downsampling is usually accompanied by loss of information, and the lost information is irreversible and will not be restored with upsampling, the granularity of information in the feature map obtained by upsampling is large.
  • the feature maps output by the second residual networks connected in series are compared with the same scale of the feature maps.
  • the sampled feature map is fused, and the fused feature map is used as the input of the attention mechanism module.
  • Step S123 according to the second feature map, predict the position of each character and the character label of the character.
  • the category of each pixel can be predicted according to the second feature map.
  • the position of each character and the character label of each character are determined according to the category prediction result of each pixel and the feature value of the pixel.
  • the category of the pixel corresponds to the character label.
  • Character labels can include three categories: labels corresponding to Chinese characters, labels corresponding to numbers, and labels corresponding to letters.
  • the categories of pixels are Chinese characters, numbers or letters.
  • Step S124 Perform character segmentation on the license plate image to be recognized according to the predicted position of each character to obtain the several character images.
  • the area where each character is located may be marked with a matrix frame. According to the marking result of the matrix box, the image corresponding to each character is cut out to obtain several corresponding character images.
  • Step S13 input the several character images into the license plate recognition network, select weight parameters according to the character labels of each character image, and use the selected weight parameters to recognize the character images.
  • the license plate recognition network is obtained by pre-training, and the license plate recognition network may include multiple sets of weight parameters, each set of weight parameters corresponding to character labels one-to-one.
  • the categories of character labels can include three categories, namely Chinese characters, numbers and letters.
  • the training samples of different types of character labels are used to train the license plate recognition network, and the weight parameters corresponding to each character label are obtained. For example, using the training samples whose character labels are Chinese characters to train the license plate recognition network, the corresponding weight parameters when the character labels are Chinese characters are obtained.
  • the license plate recognition model is trained by using the training samples whose character labels are numbers, and the corresponding weight parameters when the character labels are numbers are obtained.
  • the license plate recognition model is trained by using the training samples whose character labels are letters, and the corresponding weight parameters when the character labels are letters are obtained.
  • Step S14 obtaining the license plate number corresponding to the license plate image according to the recognition result of each character image.
  • the recognition results of all characters can be integrated according to the position of each character on the license plate image to be recognized to obtain the license plate number corresponding to the license plate image.
  • the information related to the position of the character on the license plate image to be recognized can be obtained when characters are segmented on the license plate image to be recognized.
  • the weight parameter corresponding to the character label is selected from the pre-trained license plate recognition network. Based on the selected weight parameter The character image is recognized. Since different types of character labels correspond to pre-trained weight parameters respectively, and the weight parameters corresponding to the character labels are used to recognize the character image, the accuracy of the license plate recognition result can be improved.
  • both the character segmentation network and the license plate recognition network can adopt a lightweight network structure such as a limited-layer convolutional layer or a grouped convolutional layer, which can make the character segmentation network and the license plate recognition network lightweight and have better real-time performance. End-to-end license plate recognition.
  • the to-be-recognized license plate image is corrected.
  • a spatial transformation matrix may be used to correct the image of the license plate to be recognized.
  • the spatial transformation matrix can be obtained by pre-training using a spatial transformation network (Spatial Transformer Network, STN).
  • STN Spatial Transformer Network
  • the STN network may be composed of two convolution kernels of 3 ⁇ 3 convolution layers and one fully connected layer, so as to realize the lightweight of the STN network.
  • the image is interfered by different types of noise, which makes it difficult to recognize the license plate, resulting in a low accuracy rate of the license plate character recognition.
  • the space transformation matrix is used to correct the to-be-recognized license plate image.
  • the corrected image is used as the image of the license plate to be recognized, and image denoising is performed.
  • the image denoising method may include the following steps S61 to S65:
  • Step S61 performing image feature enhancement on the license plate image to be recognized to obtain an image with enhanced image features.
  • image denoising may be implemented based on a convolutional neural network.
  • the to-be-recognized license plate image can be convolved through a warp layer and processed through a BN layer to obtain an image with enhanced image features.
  • the denoising module for image denoising can be obtained based on the convolutional neural network.
  • the denoising module can use a convolutional layer with a convolution kernel of 3 ⁇ 3 and a BN layer to reduce the weight of the denoising module.
  • an image with enhanced image features can be obtained.
  • Step S62 performing nonlinear processing based on the image after image enhancement to obtain a feature map after nonlinear processing.
  • the image of the license plate to be recognized and the image after image feature enhancement are image fused to obtain a fused image. Perform nonlinear processing on the fused image to obtain a feature map after nonlinear processing.
  • the Tanh activation function is used to perform nonlinear processing on the fused image. It can improve the efficiency of nonlinear processing while taking into account the lightweight of the denoising module.
  • Step S63 performing a convolution operation on the feature map after the nonlinear processing to obtain a weight matrix.
  • the denoising module may perform convolution on the feature map after nonlinear processing to compress the acquired features into a weight matrix (also referred to as a vector), where the convolution kernel of the convolution is 1 ⁇ 1.
  • Step S64 using the weight matrix to weight the license plate image to be recognized to obtain a noise feature map.
  • the obtained weight matrix is weighted with the license plate image to be recognized, that is, the weight matrix is multiplied by the license plate image to be recognized to obtain a noise feature map.
  • Step S65 Obtain a denoised feature map according to the noise feature map and the to-be-recognized license plate image, and use the de-noised feature map as the to-be-recognized license plate image.
  • the noise feature map can be subtracted from the to-be-recognized license plate image to obtain a denoised feature map.
  • the license plate area samples are obtained, and the coordinate information of the four corners of the license plate in each license plate area sample image is corrected using the perspective transformation method, and the obtained license plate image is used as the license plate area sample label.
  • a license plate correction network STN is constructed.
  • the STN network consists of two 3x3 convolutional layers and one fully connected layer.
  • the spatial transformation matrix is obtained through training and learning.
  • the spatial transformation matrix is used to correct the license plate image to be recognized during license plate recognition.
  • the license plate correction network is trained using license plate region samples and license plate sample labels
  • the denoising module is trained using the corrected license plate samples and the noised license plate samples.
  • the denoising process in the training process of the denoising module may refer to the descriptions in steps S61 to S65. Different from the denoising method used in the actual license plate recognition method, in the denoising training, the effect of the denoised feature map is checked after step S65, if the effect of the denoised feature map is not satisfactory. Set the conditions, and continue iterative training until the set conditions are met, and the training of the denoising module is completed.
  • each license plate sample has a license plate character label image.
  • the license plate character label picture divides letters, characters and Chinese characters into three categories.
  • the character segmentation network is trained by using license plate sample pictures whose character types are letters, characters, and Chinese characters.
  • the segmentation network model may include several downsampling modules and several upsampling modules.
  • Each downsampling module includes several residual networks, consisting of two convolutional layers with 1x1 convolution kernel and one grouped convolutional layer with 3x3 convolution kernel, and BN layer, Relu layer are added to each convolutional layer. activation function.
  • Each upsampling module also has several residual networks, and an interpolation module is set before several residual networks.
  • An attention mechanism module is set after several residual networks.
  • the attention mechanism module can make the character segmentation model pay more attention to meaningful and important channels to improve the accuracy of character segmentation.
  • the training process of the character segmentation network reference may be made to the descriptions in steps S121 to S124 in the foregoing embodiment.
  • Different from the actual license plate recognition process, in the training process of the character segmentation model it is necessary to compare the obtained character images with the license plate character label pictures to determine whether the output results of the character segmentation model meet the requirements, and if not, continue. Training until the obtained character images meet the requirements, and a character segmentation model is obtained.
  • the constructed license plate recognition network can be composed of a grouped convolutional residual network.
  • the grouped convolutional residual network module includes two 1x1 convolutional layers and a 3x3 grouped convolutional layer, and adds a BN layer and a Relu activation function to each convolutional layer. Grouped convolution makes the network more lightweight and speeds up network inference.
  • the license plate recognition network is trained using character sample images and character labels, and three sets of weight parameters are obtained.
  • the license plate recognition network is trained by using Chinese character sample pictures and Chinese character labels, and the weight parameters corresponding to Chinese characters are obtained.
  • the license plate recognition network is trained with letter sample pictures and Chinese character labels, and the weight parameters corresponding to the numbers are obtained.
  • the license plate recognition network is trained by using digital sample pictures and Chinese character labels, and the weight parameters corresponding to the numbers are obtained. So that in practical applications, the corresponding weight parameters can be selected according to the input of Chinese characters, characters and letters, and the recognition accuracy of various characters can be enhanced.
  • An embodiment of the present invention also provides a license plate recognition device.
  • a schematic structural diagram of a license plate recognition device 70 in an embodiment of the present invention is given.
  • the license plate recognition device 70 may include:
  • an acquisition unit 71 configured to acquire an image of the license plate to be recognized
  • the character segmentation unit 72 is configured to input the license plate image to be recognized into a character segmentation network for character segmentation, to obtain several character images, each character image has a corresponding character label, and the character segmentation network is pre-trained;
  • the license plate recognition unit 73 is used for inputting the several character images into the license plate recognition network, selecting weight parameters according to the character labels of each character image, and using the selected weight parameters to recognize the character images, according to the recognition result of each character image , to obtain the license plate number corresponding to the license plate image, wherein the license plate recognition network is pre-trained, and includes multiple sets of weight parameters, each set of weight parameters corresponding to character labels one-to-one.
  • the license plate recognition device 70 may correspond to a terminal (also referred to as user equipment) or a chip with an optimized function of license plate recognition; or a chip with a data processing function, such as a baseband chip; or a user
  • the device includes a chip module with a license plate recognition function chip; or corresponds to a chip module with a data processing function chip, or corresponds to a user equipment.
  • each module/unit included in each device and product described in the above embodiments it may be a software module/unit, a hardware module/unit, or a part of a software module/unit, a part of which is a software module/unit. is a hardware module/unit.
  • each module/unit included therein may be implemented by hardware such as circuits, or at least some of the modules/units may be implemented by a software program.
  • Running on the processor integrated inside the chip the remaining (if any) part of the modules/units can be implemented by hardware such as circuits; for each device and product applied to or integrated in the chip module, the modules/units contained therein can be They are all implemented by hardware such as circuits, and different modules/units can be located in the same component of the chip module (such as chips, circuit modules, etc.) or in different components, or at least some of the modules/units can be implemented by software programs.
  • the software program runs on the processor integrated inside the chip module, and the remaining (if any) part of the modules/units can be implemented by hardware such as circuits; for each device and product applied to or integrated in the terminal, each module contained in it
  • the units/units may all be implemented in hardware such as circuits, and different modules/units may be located in the same component (eg, chip, circuit module, etc.) or in different components in the terminal, or at least some of the modules/units may be implemented by software programs Realization, the software program runs on the processor integrated inside the terminal, and the remaining (if any) part of the modules/units can be implemented in hardware such as circuits.
  • An embodiment of the present invention further provides a storage medium, which is a non-volatile storage medium or a non-transitory storage medium, and stores a computer program thereon, and when the computer program is run by a processor, executes the above-mentioned any of the present invention.
  • the steps of the license plate recognition method provided in an embodiment.
  • An embodiment of the present invention further provides a terminal, including a memory and a processor, where the memory stores a computer program that can run on the processor, and when the processor runs the computer program, any one of the foregoing embodiments of the present invention is executed Steps of the License Plate Recognition Method Provided in the Embodiments

Landscapes

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

Abstract

一种车牌识别方法及装置、存储介质、终端，所述车牌识别方法，包括：获取待识别车牌图像；将所述待识别车牌图像输入字符分割网络进行字符分割，得到若干字符图像，每一字符图像均具有对应的字符标签，所述字符分割网络是预先训练得到的；将所述若干字符图像输入车牌识别网络，根据各字符图像的字符标签选择权重参数，并采用所选择的权重参数对字符图像进行识别，其中，所述车牌识别网络是预先训练得到的，包括多组权重参数，每组权重参数分别与字符标签一一对应；根据每一字符图像的识别结果，得到所述车牌图像对应的车牌号。上述方案，能够提高车牌识别结果的准确度。

Description

车牌识别方法及装置、存储介质、终端
本申请要求2020年11月30日提交中国专利局、申请号为2020113734668、发明名称为“车牌识别方法及装置、存储介质、终端”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本发明实施例涉及车牌识别领域，尤其涉及一种车牌识别方法及装置、存储介质、终端。
背景技术
图像处理技术和深度学习的快速发展，车牌识别技术在停车系统以及电子警察系统得到快速普及应用。这些系统往往要求低延时、高准确性，从而对模型算法提出了很高的标准。主流的深度学习解决方案通过卷积神经网络提取特征图、通过循环神经网络以顺序识别出车牌各文本位置上的文本信息、最后使用CTC损失函数，根据所述特征矩阵中各元素对相应文本位置的注意力权重，进行文本识别。
然而，我国车牌通常包含有汉字、数字和字母，鉴于车牌中所包含的内容的多样性，导致车牌识别结果的准确率较低。
发明内容
本发明实施例解决的技术问题是车牌识别结果的准确率较低。
为解决上述技术问题，本发明实施例提供一种车牌识别方法，包括：获取待识别车牌图像；将所述待识别车牌图像输入字符分割网络进行字符分割，得到若干字符图像，每一字符图像均具有对应的字符标签，所述字符分割网络是预先训练得到的；将所述若干字符图像输 入车牌识别网络，根据各字符图像的字符标签选择权重参数，并采用所选择的权重参数对字符图像进行识别，其中，所述车牌识别网络是预先训练得到的，包括多组权重参数，每组权重参数分别与字符标签一一对应；根据每一字符图像的识别结果，得到所述车牌图像对应的车牌号。
可选的，所述将所述待识别车牌图像输入至字符分割网络进行字符分割，得到若干字符图像，包括：对所述待识别车牌图像进行下采样，得到第一特征图；对所述第一特征图进行上采样，得到第二特征图，其中，至少在其中一次上采样时增加注意力机制，所述第二特征图与所述第一特征图的尺度相同；根据所述第二特征图，预测每一字符的位置以及字符的字符标签；根据预测的每一字符的位置，对所述待识别车牌图像进行字符分割，得到所述若干字符图像。
可选的，所述对所述第一特征图进行上采样，得到第二特征图，包括：分别采用多个串联的上采样模块对所述第一特征图进行多次上采样，得到所述第二特征图，多个上采样模块用于接收所述第一特征图或者上一级上采样模块输出的上采样特征图，对最后一个上采样模块输出的上采样特征图进行通道变换处理，得到所述第二特征图，所述第二特征图的通道数目与所述字符标签的总类别数目相关；其中，至少一个上采样模块包括注意力机制模块，所述注意力机制模块，用于对输入的特征图的通道进行加权处理，将加权处理之后的特征图作为上采样特征图。
可选的，所述注意力机制模块，用于对所述输入的特征图分别进行全局平均池化处理和最大池化处理，得到全局平均池化处理后的特征图以及最大池化处理后的特征图；分别将所述全局平均池化处理后的特征图以及所述最大池化处理后的特征图进行卷积，得到两个卷积处理后的特征图；根据所述两个卷积处理后的特征图，采用Sigmoid激活函数确定各通道权重，采用所述各通道权重对所述输入的特征图进行加权处理，得到所述上采样特征图。
可选的，所述上采样模块包括：插值单元，用于对输入的特征图进行插值处理，得到插值处理后的特征图，其中，所述输入的特征图为所述第一特征图或者上一级上采样模块输出的上采样特征图；多个串联的第二残差网络，用于接收待上采样的特征图或者上一级第二残差网络输出的处理后的特征图，所述待上采样的特征图为所述插值单元输出的插值处理后的特征图，每个第二残差网络包括：一个或多个卷积层，以及一个或多个分组卷积层。还包括：对于每次上采样时，将所述多个串联的第二残差网络输出的特征图与相同尺度的下采样特征图进行融合，将融合后的特征图作为所述注意力机制模块的输入。
可选的，所述插值单元，用于采用双线性插值方法对所述输入的特征图进行插值处理。
可选的，每个第二残差网络中除第一个卷积层之外的其他卷积层均具有BN层且采用的激活函数为Relu激活函数。
可选的，所述字符分割网络包括多个串联的下采样模块，所述下采样模块用于对所述待识别车牌图像进行下采样，得到第一特征图，每个下采样模块包括多个串联的第一残差网络，其中：多个串联的第一残差网络，用于接收待下采样的特征图或者上一级第一残差网络输出的处理后的特征图，所述待下采样的特征图为所述待识别车牌图像或者上一级下采样模块输出的下采样特征图，每个残差网络包括一个或多个卷积层，以及一个或多个分组卷积层。
可选的，每个第一残差网络中除第一个卷积层之外的其他卷积层均具有BN层且采用的激活函数为Relu激活函数。
可选的，所述车牌识别方法还包括：在获取到待识别车牌图像之后，对所述待识别车牌图像进行矫正。
可选的，所述对所述待识别车牌图像进行矫正，包括：采用空间变换矩阵对所述待识别车牌图像进行矫正。
可选的，所述对所述待识别车牌图像进行去噪，包括：对所述待识别车牌图像进行图像特征增强，得到图像特征增强后的图像；基于所述图像增强后的图像进行非线性处理，得到非线性处理之后的特征图；对所述非线性处理之后的特征图进行卷积操作，得到权重矩阵；采用所述权重矩阵对所述待识别车牌图像进行加权，得到噪声特征图；根据所述噪声特征图与所述待识别车牌图像，得到去噪处理后的特征图，将所述去噪处理后的特征图作为所述待识别车牌图像。
可选的，所述基于所述图像增强后的图像进行非线性处理，得到非线性处理之后的特征图，包括：将所述待识别车牌图像与所述图像特征增强后的图像进行图像融合，得到融合后的图像；对所述融合后的图像进行非线性处理，得到非线性处理之后的特征图。
可选的，采用Tanh激活函数对所述融合后的图像进行非线性处理。
可选的，所述对所述待识别车牌图像进行图像特征增强，包括：将所述待识别车牌图像经一个卷经层进行卷积，并经过一个BN层处理，得到所述图像特征增强后的图像。
本发明实施例还提供一种车牌识别装置，包括：获取单元，用于获取待识别车牌图像；字符分割单元，用于将所述待识别车牌图像输入字符分割网络进行字符分割，得到若干字符图像，每一字符图像均具有对应的字符标签，所述字符分割网络是预先训练得到的；车牌识别单元，用于将所述若干字符图像输入车牌识别网络，根据各字符图像的字符标签选择权重参数，并采用所选择的权重参数对字符图像进行识别，根据每一字符图像的识别结果，得到所述车牌图像对应的车牌号，其中，所述车牌识别网络是预先训练得到的，包括多组权重参数，每组权重参数分别与字符标签一一对应。
本发明实施例还提供一种存储介质，其上存储有计算机程序，所述计算机程序被处理器运行时执行上述任一种车牌识别方法的步骤。
本发明实施例还提供一种终端，包括存储器和处理器，所述存储器上存储有能够在所述处理器上运行的计算机程序，所述处理器运行所述计算机程序时执行上述任一种车牌识别方法的步骤。
与现有技术相比，本发明实施例的技术方案具有以下有益效果：
在对待识别车牌图像进行字符分割，得到若干字符图像，根据每一字符图像对应的字符标签，从预先训练的车牌识别网络中选择与字符标签对应的权重参数，不同类型的字符标签分别对应有预训练的权重参数，采用与字符标签对应的权重参数来对字符图像进行识别，故可以提高车牌识别结果的准确度。
进一步，在对第一特征图进行上采样得到第二特征图时，至少一个上采样模块包括注意力机制模块，所述注意力机制模块，用于对输入的特征图的通道进行加权处理，将加权处理之后的特征图作为上采样特征图。使字符分割模型更关注有意义、重要的通道以提升字符分割准确率，从而可以提高字符位置定位的准确性，有效地降低漏检字符的情况，提高字符分割网络的可靠性。
进一步，通过对待识别车牌图像进行矫正，可以实现对待识别车牌图像中具有一定倾斜角度下的车牌进行快速识别，大大降低了摄像机摆放角度的限制，提高对待识别车牌图像中的车牌的倾斜角度的容忍度以及识别范围，并可以提高车牌识别结果的准确度和提高整体识别模型的鲁棒性。
进一步，通过对待识别车牌图像进行去噪，可以有效的去除夜晚、雨雪天气或其他高噪声的情况下噪声影响，提高车牌识别的准确率以及使得整个车牌识别模型具有更好的鲁棒性。
进一步，在对待识别车牌图像进行去噪时，将待识别车牌图像与图像特征增强后的图像进行图像融合，得到融合后的图像，基于融合后的图像进行非线性处理，得到非线性处理之后的特征图。通过待识别车牌图像与图像特征增强后的图像进行图像融合，可以较好的保留 待识别车牌图像中的信息，增强融合图像中噪声特征，以提高去噪效果。此外，通过对融合后的图像进行非线性归一化可以增强变换能力，有助于较好的提取出噪声特征。
附图说明
图1是本发明实施例中的一种车牌识别方法的流程图；
图2是本发明实施例中的一种字符分割的流程图；
图3是本发明实施例中的一种字符分割网络的结构示意图；
图4是本发明实施例中的一种下采样模块的结构示意图；
图5是本发明实施例中的一种上采样模块的结构示意图；
图6是本发明实施例中的一种图像去噪的流程图；
图7是本发明实施例中的一种车牌识别装置的结构示意图。
具体实施方式
如上所述，现有技术中，对车牌的识别结果的准确率较低。
为了解决上述问题，在本发明实施例中，在对待识别车牌图像进行字符分割，得到若干字符图像，根据每一字符图像对应的字符标签，从预先训练的车牌识别网络中选择与字符标签对应的权重参数，基于选择的权重参数对该字符图像进行识别。由于不同类型的字符标签分别对应有预训练的权重参数，采用与字符标签对应的权重参数来对字符图像进行识别，故可以提高车牌识别结果的准确度。
为使本发明实施例的上述目的、特征和有益效果能够更为明显易懂，下面结合附图对本发明的具体实施例做详细的说明。
本发明实施例提供一种车牌识别方法，参照图1，给出了本发明实施例中的一种车牌识别方法的流程图，车牌识别方法具体可以包括如下步骤：
步骤S11，获取待识别车牌图像。
步骤S12，将所述待识别车牌图像输入字符分割网络进行字符分割，得到若干字符图像，每一字符图像均具有对应的字符标签，所述字符分割网络是预先训练得到的。
参照图2，给出了本发明实施例中的一种字符分割的流程图。在具体实施中，将所述待识别车牌图像输入字符分割网络进行字符分割可以包括如下步骤S121至步骤S124：
步骤S121，对所述待识别车牌图像进行下采样，得到第一特征图。
参照图3，给出了本发明实施例中的一种字符分割网络的结构示意图。在具体实施中，结合图2及图3，所述字符分割网络30可以包括多个下采样模块31，多个下采样模块31串联。通过多个下采样模块31对所述待识别车牌图像进行下采样，得到第一特征图。
其中，第一级下采样模块31的输入为待识别车牌图像。自第二级下采样模块31开始，其输入为上一级下采样模块31的输出。最后一级下采样模块31输出第一特征图。
需要说明的是，图3中示意的字符分割网络30包括四个下采样模块31。在实际应用中，下采样模块31的数目以及每一下采样模块的下采样倍数可以根据实际字符分割网络30的轻量化以及下采样倍数等需求进行设定，此处不做限定。
参照图4，给出了本发明实施例中的一种下采样模块的结构示意图。在一个实施例中，每个下采样模块31可以包括多个串联的第一残差网络311。多个串联的第一残差网络311用于接收待下采样的特征图或者上一级第一残差网络311输出的处理后的特征图，所述待下采样的特征图为所述待识别车牌图像或者上一级下采样模块31输出的下采样特征图。
具体而言，针对第一级下采样模块31，第一级第一残差网络311的输入为待识别车牌图像。自第二级下采样模块31起，下采样模块 31中的第一个第一残差网络311的输入为上一级下采样模块31的输出。针对每个下采样模块31，自第二级第一残差网络311起，其输入为上一级第一残差网络311的输出。针对最后一级下采样模块31中的最后一级第一残差网络311的输出为第一特征图。
在具体实施中，第一残差网络311可以基于卷积神经网络构建。每一第一残差网络311包括一个或多个卷积层，以及一个或多个分组卷积层。在对待识别车牌图像进行下采样时，采用多个第一残差网络311可以避免梯度的消失，以确保下采样后得到的下采样图像中的图像特征的细粒度。
在一个非限制性实施例中，每个第一残差网络311包括两个卷积核为1×1的卷积层，一个卷积核为3×3的分组卷积层。可以先执行卷积核为1×1的卷积层，将卷积核为1×1的卷积层的输出作为卷积核为3×3的分组卷积层的输入。卷积核为3×3的分组卷积层的输出作为另一个卷积核为1×1的卷积层的输入。另一个卷积核为1×1的卷积层的输出可以作为下一级第一残差网络311的输入，或者，作为下一级下采样模块31的输入，或者当另一个卷积核为1×1的卷积层所属的下采样模块31为最后一级下采样模块31时，另一个卷积核为1×1的卷积层的输出即为第一特征图。
在具体实施中，每个第一残差网络311中除第一个卷积层之外的其他卷积层均具有批量标准化(Batch Normalization，BN)层且采用的激活函数为Relu激活函数。
在具体实施中，当下采样模块31实现两倍下采样时，在下采样模块31中的数个第一残差网络311中，有一卷积的步长为2，其余所有卷积的步长为1。其中，步长为2的卷积可以设置于卷积核为1×1的卷积层，也可以设置于卷积核为3×3的分组卷积层。
相应地，当下采样模块31实现四倍下采样时，在下采样模块31中的数个第一残差网络311中，将一卷积的步长设置为4，其余所有卷积的步长均设置为1，其中，步长为2的卷积可以设置于卷积核为 1×1的卷积层，也可以设置于卷积核为3×3的分组卷积层。或者，将两个卷积的步长分别设置为2，其余所有卷积的步长均设置为1，其中，步长为2的卷积可以设置于卷积核为1×1的卷积层，也可以设置于卷积核为3×3的分组卷积层。
需要说明的是，针对每个下采样模块31，当其中的第一残差网络311中的卷积步长大于1时，则需要对输入第一残差网络311的特征图进行相同步长的卷积操作，并将卷积结果与该第一残差网络311中最后一次卷积的输出相加，作为该第一残差网络311的输出。
需要说明的是，根据下采样模块31实现的下采样倍数不同，卷积的步长相应地不同，具体根据需求进行设定即可，此处不再赘述。
步骤S122，对所述第一特征图进行上采样，得到第二特征图。
在具体实施中，在第一特征图进行上采样，得到第二特征图时，至少在其中一次上采样时增加注意力机制，所述第二特征图与所述第一特征图的尺度相同。例如，至少在最后一次上采样时增加注意力机制。
在具体实施中，参照图3，字符分割网络30可以包括多个串联的上采样模块32。可以分别采用多个串联的上采样模块32对所述第一特征图进行多次上采样，得到所述第二特征图，多个上采样模块32用于接收所述第一特征图或者上一级上采样模块32输出的上采样特征图，对最后一级上采样模块32输出的上采样特征图进行通道变换处理，得到所述第二特征图，所述第二特征图的通道数目与所述字符标签的总类别数目相关。
在具体实施中，可以对最后一级上采样模块32输出的上采样特征图进行卷积，以实现通道变换处理，将通道的数目变换成3，通道数目为3的第二特征图。其中，输入的待识别车牌图像的通道数目为3。
参照图5，给出了本发明实施例中的一种上采样模块的结构示意 图。在具体实施中，至少一个上采样模块32包括注意力机制模块323。所述注意力机制模块323用于对输入的特征图的通道进行加权处理，将加权处理之后的特征图作为上采样特征图。通过注意力机制模块323处理可以提高字符位置定位的准确性，有效地降低漏检字符的概率，提高字符分割网络的可靠性。
在具体实施中，所述注意力机制模块323用于对所述输入的特征图进行全局平均池化处理，得到全局平均池化处理后的特征图。对所述输入的特征图进行最大池化处理，得到最大池化处理后的特征图。注意力机制模块323分别将所述全局平均池化处理后的特征图以及所述最大池化处理后的特征图进行卷积，得到两个卷积处理后的特征图。根据所述两个卷积处理后的特征图，采用Sigmoid激活函数可以确定各通道权重，采用所述各通道权重对所述输入的特征图进行加权处理，得到所述上采样特征图。
在一个非限制性实施例中，每个上采样模块32均包括注意力机制模块323。
在本发明实施例中，可以采用插值方式进行上采样。
具体而言，每个上采样模块32可以包括插值单元321以及多个串联的第二残差网络322。其中：
插值单元321，用于对输入的特征图进行插值处理，得到插值处理后的特征图，其中，所述输入的特征图为所述第一特征图或者上一级上采样模块32输出的上采样特征图。
在一个非限制性实施例中，插值单元321采用双线性插值方法对所述输入的特征图进行插值处理。
多个串联的第二残差网络322，用于接收待上采样的特征图或者上一级第二残差网络322输出的处理后的特征图，所述待上采样的特征图为所述插值单元321输出的插值处理后的特征图。
在具体实施中，第二残差网络322可以基于卷积神经网络构建。 每个第二残差网络322可以包括：一个或多个卷积层，以及一个或多个分组卷积层。当上采样模块32的数目与下采样模块31的数目相同时，每一上采样模块32的上采样倍数和对应的下采样模块31的下采样倍数相同。
在一个非限制性实施例中，每个第二残差网络322包括两个卷积核为1×1的卷积层，一个卷积核为3×3的分组卷积层。可以先执行卷积核为1×1的卷积层，将一个卷积核为1×1的卷积层的输出作为卷积核为3×3的分组卷积层的输入，卷积核为3×3的分组卷积层的输出作为另一个卷积核为1×1的卷积层的输入。
当所述的上采样模块包括注意力机制模块323时，另一个卷积核为1×1的卷积层的输出可以作为注意力机制模块323的输入。
当第二残差网络322所属的上采样模块不包括注意力机制模块323时，另一个卷积核为1×1的卷积层的输出作为下一级第二残差网络322的输入。
若第二残差网络322为最后一级第二残差网络322，当上采样模块32为最后一级上采样模块32且不包含注意力机制模块323时，可以基于另一个卷积核为1×1的卷积层的输出得到第二特征图。
在具体实施中，第二残差网络322中每个卷积的步长均为1。
对于输入至第二残差网络322内的特征图，先经过卷积核为1×1的卷积层进行卷积，可以对每个像素的特征进行增强。在第二残差网络322中配置分组卷积层在确保卷积效果以减少卷积参数量，使网络更轻量化，加快了网络推理速度。
由于在下采样的过程中通常伴随着信息的丢失，且丢失的信息是不可逆的，并不会随着上采样而复原，导致上采样得到的特征图中信息的粒度较大。为了提高所得到的第二特征图中的细粒度特征，在本发明实施例中，对于每次上采样时，将所述多个串联的第二残差网络输出的特征图与相同尺度的下采样特征图进行融合，将融合后的特征 图作为所述注意力机制模块的输入。
在具体实施中，每一第二残差网络322中除第一个卷积层之外的其他卷积层均具有BN层且采用的激活函数为Relu激活函数，以增强第二残差网络322的特征图中的特征，提高第二残差网络322的处理效果。
步骤S123，根据所述第二特征图，预测每一字符的位置以及字符的字符标签。
在具体实施中，可以根据第二特征图对每个像素的类别进行预测。根据每一像素的类别预测结果结合像素的特征值，确定每一字符的位置以及每一字符的字符标签。其中，像素的类别与字符标签对应。字符标签可以包括三种类别：分别为汉字对应的标签、数字对应的标签以及字母对应的标签。相应地，像素的类别为汉字类别、数字类别或者字母类别。
步骤S124，根据预测的每一字符的位置，对所述待识别车牌图像进行字符分割，得到所述若干字符图像。
在具体实施中，根据预测的每一字符的位置，可以对每一字符所在的区域进行标矩阵框。根据矩阵框的标示结果，抠出每一字符对应的图像，得到对应的若干字符图像。
步骤S13，将所述若干字符图像输入车牌识别网络，根据各字符图像的字符标签选择权重参数，并采用所选择的权重参数对字符图像进行识别。
在具体实施中，所述车牌识别网络是预先训练得到的，车牌识别网络可以包括多组权重参数，每组权重参数分别与字符标签一一对应。
在车牌识别网络训练时，针对车牌识别场景，字符标签的类别可以包括三类，分别为汉字、数字和字母。分别采用不同类别的字符标签的训练样本进行车牌识别网络的训练，得到各字符标签对应的权重 参数。例如，采用字符标签为汉字的训练样本进行车牌识别网络的训练，得到字符标签为汉字时对应的权重参数。又如，采用字符标签为数字的训练样本进行车牌识别模型的训练，得到字符标签为数字时对应的权重参数。再如，采用字符标签为字母的训练样本进行车牌识别模型的训练，得到字符标签为字母时对应的权重参数。
步骤S14，根据每一字符图像的识别结果，得到所述车牌图像对应的车牌号。
在具体实施中，得到每一字符对应的识别结果时，可以按照各个字符在待识别车牌图像上的位置，将所有字符的识别结果进行整合，得到所述车牌图像对应的车牌号。其中，字符在待识别车牌图像上的位置相关的信息可以在对待识别车牌图像进行字符分割时得到。
由上可知，在对待识别车牌图像进行字符分割，得到若干字符图像，根据每一字符图像对应的字符标签，从预先训练的车牌识别网络中选择与字符标签对应的权重参数，基于选择的权重参数对该字符图像进行识别。由于不同类型的字符标签分别对应有预训练的权重参数，采用与字符标签对应的权重参数来对字符图像进行识别，故可以提高车牌识别结果的准确度。
此外，字符分割网络、车牌识别网络均可以采用有限层的卷积层或者分组卷积层等轻量化网络结构，可以使得字符分割网络、车牌识别网络轻量化，以具有较好的实时性，实现端到端的车牌识别。
在具体实施中，为了进一步提高图像识别效果，在获取到待识别车牌图像之后，对所述待识别车牌图像进行矫正。
在本发明实施例中，可以采用空间变换矩阵对所述待识别车牌图像进行矫正。
其中，空间变换矩阵可以采用空间变换网络(Spatial Transformer Network，STN)进行预先训练得到。
在本发明一实施例中，STN网络可以由两个卷积核为3x3卷积 层，1个全连接层组成，以实现STN网络的轻量化。
在具体实施中，由于受天气环境影响，比如大雨、大雾、沙尘暴等，图像受到不同类型的噪声干扰，造成车牌难以识别，导致车牌字符识别的准确率较低。
为了进一步提高车牌识别准确度，在本发明实施例中，采用所述空间变换矩阵对所述待识别车牌图像进行矫正。在一些实施例中，当对待识别车牌图像进行矫正时，则将矫正后的图像作为待识别车牌图像，进行图像去噪。
参照图6，给出了本发明实施例中的一种图像去噪的流程图。图像去噪方法可以包括如下步骤S61至步骤S65：
步骤S61，对所述待识别车牌图像进行图像特征增强，得到图像特征增强后的图像。
在本发明实施例中，可以基于卷积神经网络实现对图像的去噪。具体而言，可以将所述待识别车牌图像经一个卷经层进行卷积，并经过一个BN层处理，得到所述图像特征增强后的图像。也即用于图像去噪的去噪模块可以基于卷积神经网络得到。例如，去噪模块可以采用卷积核为3×3的卷积层以及BN层，以实现去噪模块的轻量化。待识别车牌图像经卷积核为3×3的卷积层以及BN层之后，可以得到图像特征增强后的图像。
步骤S62，基于所述图像增强后的图像进行非线性处理，得到非线性处理之后的特征图。
为了提高去噪效果，在本发明实施例中，将所述待识别车牌图像与所述图像特征增强后的图像进行图像融合，得到融合后的图像。对所述融合后的图像进行非线性处理，得到非线性处理之后的特征图。通过将待识别车牌图像与所述图像特征增强后的图像进行图像融合，基于融合后的图像进行后续胡处理，以提高去噪模块的噪声表示能力，提高噪声特征提取效果。
在具体实施中，采用Tanh激活函数对所述融合后的图像进行非线性处理。可以提高非线性处理效率的同时，兼顾去噪模块的轻量化。
步骤S63，对所述非线性处理之后的特征图进行卷积操作，得到权重矩阵。
在具体实施中，去噪模块可以将非线性处理之后的特征图进行卷积将获取的特征压缩为权重矩阵(也可称为向量)，其中卷积的卷积核为1×1。
步骤S64，采用所述权重矩阵对所述待识别车牌图像进行加权，得到噪声特征图。
将得到的权重矩阵与待识别车牌图像进行加权，也即将权重矩阵与待识别车牌图像相乘，得到噪声特征图。
步骤S65，根据所述噪声特征图与所述待识别车牌图像，得到去噪处理后的特征图，将所述去噪处理后的特征图作为所述待识别车牌图像。
在本发明实施例中，可以将噪声特征图与所述待识别车牌图像相减，得到去噪处理后的特征图。
为了便于本领域技术人员更好的理解和实现本发明实施例，下面分别对上述实施例中的车牌识别方法中所采用的车牌矫正网络、字符分割网络以及车牌识别网络的训练过程进行说明。
首先获取车牌区域样本，通过每个车牌区域样本图片中车牌的四个角点的坐标信息使用透视变换方法进行矫正，获得的车牌图片作为车牌区域样本标签。
构建车牌矫正网络STN，STN网络由两个3x3卷积层，1个全连接层组成，通过训练学习得到空间变换矩阵，空间变换矩阵在车牌识别时对待识别车牌图像进行矫正。
具体而言，使用车牌区域样本和车牌样本标签训练车牌矫正网 络，使用矫正后车牌样本与加噪的车牌样本训练去噪模块。其中，去噪模块的训练过程中去噪处理可以参考步骤S61至步骤S65中的描述。与实际车牌识别方法中使用的去噪方法不同的是，在去噪的训练过中，在步骤S65之后检验去噪处理后的特征图的效果，若是去噪处理后的特征图的效果不满足设定条件，在继续迭代训练，直至满足设定条件，完成去噪模块的训练。
在字符分割网络的训练过程中，每个车牌样本具有一个车牌字符标签图片。车牌字符标签图片将字母、字符、汉字分为三类。分别采用字符类型为字母、字符、汉字的车牌样本图片进行字符分割网络的训练。分割网络模型可以包括若干下采样模块和若干上采样模块。每个下采样模块包括数个残差网络，由两个卷积核为1x1的卷积层，一个卷积核为3x3的分组卷积层组成，并在每个卷积层加入BN层、Relu激活函数。每个上采样模块也有数个残差网络，在数个残差网络之前设置有插值模块。在数个残差网络之后设置有注意力机制模块。注意力机制模块可以使字符分割模型更关注有意义、重要的通道以提升字符分割准确率。其中，字符分割网络的训练过程可以参考上述实施例中步骤S121至步骤S124中的描述。与实际车牌识别过程不同的是，在字符分割模型的训练过程中，需要将得到所述若干字符图像与车牌字符标签图片进行比较，判断字符分割模型的输出结果是否满足要求，若是不满足则继续训练，直至得到的字符图像满足要求，得到字符分割模型。
在车牌识别网络的训练过程中，构建的车牌识别网络可以由分组卷积残差网络构成。分组卷积残差网络模块包括两个1x1卷积层以及一个3x3分组卷积层，并在每个卷积层加入BN层、Relu激活函数。分组卷积使网络更轻量化，加快了网络推理速度。使用字符样本图片和字符标签训练车牌识别网络，得到三组权重参数。采用汉字样本图片与汉字标签训练车牌识别网络，得到汉字对应的权重参数。采用字母样本图片与汉字标签训练车牌识别网络，得到数字对应的权重参数。采用数字样本图片与汉字标签训练车牌识别网络，得到数字对应 的权重参数。以使得在实际应用中，可以根据输入为汉字、字符、字母的不同，选择对应的权重参数，增强各种字符的识别准确率。
本发明实施例还提供一种车牌识别装置，参照图7，给出了本发明实施例中的一种车牌识别装置70的结构示意图，车牌识别装置70可以包括：
获取单元71，用于获取待识别车牌图像；
字符分割单元72，用于将所述待识别车牌图像输入字符分割网络进行字符分割，得到若干字符图像，每一字符图像均具有对应的字符标签，所述字符分割网络是预先训练得到的；
车牌识别单元73，用于将所述若干字符图像输入车牌识别网络，根据各字符图像的字符标签选择权重参数，并采用所选择的权重参数对字符图像进行识别，根据每一字符图像的识别结果，得到所述车牌图像对应的车牌号，其中，所述车牌识别网络是预先训练得到的，包括多组权重参数，每组权重参数分别与字符标签一一对应。
在具体实施中，车牌识别装置70的具体工作流程及原理可以参考本发明上述任一实施例中提供的车牌识别方法的描述，此处不再赘述。
在具体实施中，车牌识别装置70可以对应于终端(也可称为用户设备)或者具有车牌识别功能的优化功能的芯片；或者对应于具有数据处理功能的芯片，如基带芯片；或者对应于用户设备中包括车牌识别功能的芯片的芯片模组；或者对应于具有数据处理功能芯片的芯片模组，或者对应于用户设备。
在具体实施中，关于上述实施例中描述的各个装置、产品包含的各个模块/单元，其可以是软件模块/单元，也可以是硬件模块/单元，或者也可以部分是软件模块/单元，部分是硬件模块/单元。
例如，对于应用于或集成于芯片的各个装置、产品，其包含的各个模块/单元可以都采用电路等硬件的方式实现，或者，至少部分模 块/单元可以采用软件程序的方式实现，该软件程序运行于芯片内部集成的处理器，剩余的(如果有)部分模块/单元可以采用电路等硬件方式实现；对于应用于或集成于芯片模组的各个装置、产品，其包含的各个模块/单元可以都采用电路等硬件的方式实现，不同的模块/单元可以位于芯片模组的同一组件(例如芯片、电路模块等)或者不同组件中，或者，至少部分模块/单元可以采用软件程序的方式实现，该软件程序运行于芯片模组内部集成的处理器，剩余的(如果有)部分模块/单元可以采用电路等硬件方式实现；对于应用于或集成于终端的各个装置、产品，其包含的各个模块/单元可以都采用电路等硬件的方式实现，不同的模块/单元可以位于终端内同一组件(例如，芯片、电路模块等)或者不同组件中，或者，至少部分模块/单元可以采用软件程序的方式实现，该软件程序运行于终端内部集成的处理器，剩余的(如果有)部分模块/单元可以采用电路等硬件方式实现。
本发明实施例还提供一种存储介质，所述存储介质为非易失性存储介质或非瞬态存储介质，其上存储有计算机程序，所述计算机程序被处理器运行时执行本发明上述任一实施例中提供的车牌识别方法的步骤。
本发明实施例还提供一种终端，包括存储器和处理器，所述存储器上存储有能够在所述处理器上运行的计算机程序，所述处理器运行所述计算机程序时执行本发明上述任一实施例中提供的车牌识别方法的步骤
本领域普通技术人员可以理解上述实施例的各种方法中的全部或部分步骤是可以通过程序来指令相关的硬件来完成，该程序可以存储于任一计算机可读存储介质中，存储介质可以包括：ROM、RAM、磁盘或光盘等。
虽然本发明披露如上，但本发明并非限定于此。任何本领域技术人员，在不脱离本发明的精神和范围内，均可作各种更动与修改，因此本发明的保护范围应当以权利要求所限定的范围为准。

Claims (20)

  1. 一种车牌识别方法，其特征在于，包括：
    获取待识别车牌图像；
    将所述待识别车牌图像输入字符分割网络进行字符分割，得到若干字符图像，每一字符图像均具有对应的字符标签，所述字符分割网络是预先训练得到的；
    将所述若干字符图像输入车牌识别网络，根据各字符图像的字符标签选择权重参数，并采用所选择的权重参数对字符图像进行识别，其中，所述车牌识别网络是预先训练得到的，包括多组权重参数，每组权重参数分别与字符标签一一对应；
    根据每一字符图像的识别结果，得到所述车牌图像对应的车牌号。
  2. 如权利要求1所述的车牌识别方法，其特征在于，所述将所述待识别车牌图像输入至字符分割网络进行字符分割，得到若干字符图像，包括：
    对所述待识别车牌图像进行下采样，得到第一特征图；
    对所述第一特征图进行上采样，得到第二特征图，其中，至少在其中一次上采样时增加注意力机制，所述第二特征图与所述第一特征图的尺度相同；
    根据所述第二特征图，预测每一字符的位置以及字符的字符标签；
    根据预测的每一字符的位置，对所述待识别车牌图像进行字符分割，得到所述若干字符图像。
  3. 如权利要求2所述的车牌识别方法，其特征在于，所述对所述第一特征图进行上采样，得到第二特征图，包括：
    分别采用多个串联的上采样模块对所述第一特征图进行多次上采样，得到所述第二特征图，多个上采样模块用于接收所述第一特征图或者上一级上采样模块输出的上采样特征图，对最后一个上采样模块输出的上采样特征图进行通道变换处理，得到所述第二特征图，所述第二特征图的通道数目与所述字符标签的总类别数目相关；
    其中，至少一个上采样模块包括注意力机制模块，所述注意力机制模块，用于对输入的特征图的通道进行加权处理，将加权处理之后的特征图作为上采样特征图。
  4. 如权利要求3所述的车牌识别方法，其特征在于，所述注意力机制模块，用于对所述输入的特征图分别进行全局平均池化处理和最大池化处理，得到全局平均池化处理后的特征图以及最大池化处理后的特征图；
    分别将所述全局平均池化处理后的特征图以及所述最大池化处理后的特征图进行卷积，得到两个卷积处理后的特征图；
    根据所述两个卷积处理后的特征图，采用Sigmoid激活函数确定各通道权重，采用所述各通道权重对所述输入的特征图进行加权处理，得到所述上采样特征图。
  5. 如权利要求4所述的车牌识别方法，其特征在于，所述上采样模块包括：
    插值单元，用于对输入的特征图进行插值处理，得到插值处理后的特征图，其中，所述输入的特征图为所述第一特征图或者上一级上采样模块输出的上采样特征图；
    多个串联的第二残差网络，用于接收待上采样的特征图或者上一级第二残差网络输出的处理后的特征图，所述待上采样的特征图为所述插值单元输出的插值处理后的特征图，每个第二残差网络包括：一个或多个卷积层，以及一个或多个分组卷积层。
  6. 如权利要求5所述的车牌识别方法，其特征在于，还包括：对 于每次上采样时，将所述多个串联的第二残差网络输出的特征图与相同尺度的下采样特征图进行融合，将融合后的特征图作为所述注意力机制模块的输入。
  7. 如权利要求5所述的车牌识别方法，其特征在于，所述插值单元，用于采用双线性插值方法对所述输入的特征图进行插值处理。
  8. 如权利要求5所述的车牌识别方法，其特征在于，每个第二残差网络中除第一个卷积层之外的其他卷积层均具有BN层且采用的激活函数为Relu激活函数。
  9. 如权利要求2所述的车牌识别方法，其特征在于，所述字符分割网络包括多个串联的下采样模块，所述下采样模块用于对所述待识别车牌图像进行下采样，得到第一特征图，每个下采样模块包括多个串联的第一残差网络，其中：
    多个串联的第一残差网络，用于接收待下采样的特征图或者上一级第一残差网络输出的处理后的特征图，所述待下采样的特征图为所述待识别车牌图像或者上一级下采样模块输出的下采样特征图，每个残差网络包括一个或多个卷积层，以及一个或多个分组卷积层。
  10. 如权利要求9所述的车牌识别方法，其特征在于，每个第一残差网络中除第一个卷积层之外的其他卷积层均具有BN层且采用的激活函数为Relu激活函数。
  11. 如权利要求1至10任一项所述的车牌识别方法，其特征在于，还包括：
    在获取到待识别车牌图像之后，对所述待识别车牌图像进行矫正。
  12. 如权利要求11所述的车牌识别方法，其特征在于，所述对所述待识别车牌图像进行矫正，包括：
    采用空间变换矩阵对所述待识别车牌图像进行矫正。
  13. 如权利要求1至10任一项所述的车牌识别方法，其特征在于，还包括：
    在获取到待识别车牌图像之后，对所述待识别车牌图像进行去噪。
  14. 如权利要求13所述的车牌识别方法，其特征在于，所述对所述待识别车牌图像进行去噪，包括：
    对所述待识别车牌图像进行图像特征增强，得到图像特征增强后的图像；
    基于所述图像增强后的图像进行非线性处理，得到非线性处理之后的特征图；
    对所述非线性处理之后的特征图进行卷积操作，得到权重矩阵；
    采用所述权重矩阵对所述待识别车牌图像进行加权，得到噪声特征图；
    根据所述噪声特征图与所述待识别车牌图像，得到去噪处理后的特征图，将所述去噪处理后的特征图作为所述待识别车牌图像。
  15. 如权利要求14所述的车牌识别方法，其特征在于，所述基于所述图像增强后的图像进行非线性处理，得到非线性处理之后的特征图，包括：
    将所述待识别车牌图像与所述图像特征增强后的图像进行图像融合，得到融合后的图像；
    对所述融合后的图像进行非线性处理，得到非线性处理之后的特征图。
  16. 如权利要求15所述的车牌识别方法，其特征在于，采用Tanh激活函数对所述融合后的图像进行非线性处理。
  17. 如权利要求14所述的车牌识别方法，其特征在于，所述对所 述待识别车牌图像进行图像特征增强，包括：
    将所述待识别车牌图像经一个卷经层进行卷积，并经过一个BN层处理，得到所述图像特征增强后的图像。
  18. 一种车牌识别装置，其特征在于，包括：
    获取单元，用于获取待识别车牌图像；
    字符分割单元，用于将所述待识别车牌图像输入字符分割网络进行字符分割，得到若干字符图像，每一字符图像均具有对应的字符标签，所述字符分割网络是预先训练得到的；
    车牌识别单元，用于将所述若干字符图像输入车牌识别网络，根据各字符图像的字符标签选择权重参数，并采用所选择的权重参数对字符图像进行识别，根据每一字符图像的识别结果，得到所述车牌图像对应的车牌号，其中，所述车牌识别网络是预先训练得到的，包括多组权重参数，每组权重参数分别与字符标签一一对应。
  19. 一种存储介质，其上存储有计算机程序，其特征在于，所述计算机程序被处理器运行时执行权利要求1至17任一项所述的车牌识别方法的步骤。
  20. 一种终端，包括存储器和处理器，所述存储器上存储有能够在所述处理器上运行的计算机程序，其特征在于，所述处理器运行所述计算机程序时执行权利要求1至17中任一项所述的车牌识别方法的步骤。
PCT/CN2021/131158 2020-11-30 2021-11-17 车牌识别方法及装置、存储介质、终端 WO2022111355A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202011373466.8 2020-11-30
CN202011373466.8A CN112446383B (zh) 2020-11-30 2020-11-30 车牌识别方法及装置、存储介质、终端

Publications (1)

Publication Number Publication Date
WO2022111355A1 true WO2022111355A1 (zh) 2022-06-02

Family

ID=74737505

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/131158 WO2022111355A1 (zh) 2020-11-30 2021-11-17 车牌识别方法及装置、存储介质、终端

Country Status (2)

Country Link
CN (1) CN112446383B (zh)
WO (1) WO2022111355A1 (zh)

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114973230A (zh) * 2022-06-27 2022-08-30 成都佳华物链云科技有限公司 一种车牌检测方法、电子设备及计算机可读存储介质
CN115035526A (zh) * 2022-06-21 2022-09-09 亿嘉和科技股份有限公司 基于深度学习的led字符自动定位和识别方法
CN115100639A (zh) * 2022-08-08 2022-09-23 松立控股集团股份有限公司 一种利用单个字符矫正的车牌识别方法
CN115331310A (zh) * 2022-08-23 2022-11-11 广州紫为云科技有限公司 一种多人手势识别方法、装置及介质
CN115546879A (zh) * 2022-11-29 2022-12-30 城云科技（中国）有限公司 用于表情识别的细粒度识别模型及方法
CN115601549A (zh) * 2022-12-07 2023-01-13 山东锋士信息技术有限公司（Cn） 基于可变形卷积和自注意力模型的河湖遥感图像分割方法
CN115601744A (zh) * 2022-12-14 2023-01-13 松立控股集团股份有限公司（Cn） 一种车身与车牌颜色相近的车牌检测方法
CN116153091A (zh) * 2023-01-10 2023-05-23 成都旋极历通信息技术有限公司 一种基于智慧防疫的车辆快速通行方法
CN116386023A (zh) * 2023-05-29 2023-07-04 松立控股集团股份有限公司 基于时空扩散的高位相机车牌识别方法、系统及电子设备
CN117765539A (zh) * 2023-12-25 2024-03-26 暗物质(北京)智能科技有限公司 一种基于深度学习的字帖图像自动矫正方法及相关产品
CN118351524A (zh) * 2024-04-19 2024-07-16 北京科技大学 一种基于特征洗牌的低资源车牌识别方法及装置

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112446383B (zh) * 2020-11-30 2022-09-02 展讯通信（上海）有限公司 车牌识别方法及装置、存储介质、终端
CN112967295B (zh) * 2021-03-10 2024-04-05 中国科学院深圳先进技术研究院 一种基于残差网络和注意力机制的图像处理方法及系统
CN113554030B (zh) * 2021-07-27 2022-08-16 上海大学 基于单字符注意力的多类型车牌识别方法及系统
CN114612893B (zh) * 2022-03-31 2024-07-19 重庆紫光华山智安科技有限公司 车牌遮挡检测方法、系统、电子设备及可读存储介质
CN116311213B (zh) * 2023-05-18 2023-08-22 珠海亿智电子科技有限公司 基于全局信息整合的车牌识别方法、装置、设备及介质

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6026177A (en) * 1995-08-29 2000-02-15 The Hong Kong University Of Science & Technology Method for identifying a sequence of alphanumeric characters
CN105975968A (zh) * 2016-05-06 2016-09-28 西安理工大学 一种基于Caffe框架的深度学习车牌字符识别方法
CN109447074A (zh) * 2018-09-03 2019-03-08 中国平安人寿保险股份有限公司 一种车牌识别方法及终端设备
CN111079763A (zh) * 2019-12-05 2020-04-28 北京嘉楠捷思信息技术有限公司 训练样本生成、模型训练、字符识别方法及其装置
CN111563504A (zh) * 2020-07-16 2020-08-21 平安国际智慧城市科技股份有限公司 车牌识别方法及相关设备
CN112446383A (zh) * 2020-11-30 2021-03-05 展讯通信（上海）有限公司 车牌识别方法及装置、存储介质、终端

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106845478B (zh) * 2016-12-30 2019-09-10 同观科技（深圳）有限公司 一种字符置信度的二次车牌识别方法及装置
CN107704857B (zh) * 2017-09-25 2020-07-24 北京邮电大学 一种端到端的轻量级车牌识别方法及装置
CN110532955B (zh) * 2019-08-30 2022-03-08 中国科学院宁波材料技术与工程研究所 基于特征注意力和子上采样的实例分割方法和装置
CN111667489B (zh) * 2020-04-30 2022-04-05 华东师范大学 基于双分支注意力深度学习的癌症高光谱图像分割方法及系统

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6026177A (en) * 1995-08-29 2000-02-15 The Hong Kong University Of Science & Technology Method for identifying a sequence of alphanumeric characters
CN105975968A (zh) * 2016-05-06 2016-09-28 西安理工大学 一种基于Caffe框架的深度学习车牌字符识别方法
CN109447074A (zh) * 2018-09-03 2019-03-08 中国平安人寿保险股份有限公司 一种车牌识别方法及终端设备
CN111079763A (zh) * 2019-12-05 2020-04-28 北京嘉楠捷思信息技术有限公司 训练样本生成、模型训练、字符识别方法及其装置
CN111563504A (zh) * 2020-07-16 2020-08-21 平安国际智慧城市科技股份有限公司 车牌识别方法及相关设备
CN112446383A (zh) * 2020-11-30 2021-03-05 展讯通信（上海）有限公司 车牌识别方法及装置、存储介质、终端

Cited By (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115035526A (zh) * 2022-06-21 2022-09-09 亿嘉和科技股份有限公司 基于深度学习的led字符自动定位和识别方法
CN114973230A (zh) * 2022-06-27 2022-08-30 成都佳华物链云科技有限公司 一种车牌检测方法、电子设备及计算机可读存储介质
CN115100639B (zh) * 2022-08-08 2024-04-09 松立控股集团股份有限公司 一种利用单个字符矫正的车牌识别方法
CN115100639A (zh) * 2022-08-08 2022-09-23 松立控股集团股份有限公司 一种利用单个字符矫正的车牌识别方法
CN115331310A (zh) * 2022-08-23 2022-11-11 广州紫为云科技有限公司 一种多人手势识别方法、装置及介质
CN115546879A (zh) * 2022-11-29 2022-12-30 城云科技（中国）有限公司 用于表情识别的细粒度识别模型及方法
CN115546879B (zh) * 2022-11-29 2023-02-17 城云科技（中国）有限公司 用于表情识别的细粒度识别模型及方法
CN115601549A (zh) * 2022-12-07 2023-01-13 山东锋士信息技术有限公司（Cn） 基于可变形卷积和自注意力模型的河湖遥感图像分割方法
CN115601744A (zh) * 2022-12-14 2023-01-13 松立控股集团股份有限公司（Cn） 一种车身与车牌颜色相近的车牌检测方法
CN116153091B (zh) * 2023-01-10 2024-02-06 成都旋极历通信息技术有限公司 一种基于智慧防疫的车辆快速通行方法
CN116153091A (zh) * 2023-01-10 2023-05-23 成都旋极历通信息技术有限公司 一种基于智慧防疫的车辆快速通行方法
CN116386023A (zh) * 2023-05-29 2023-07-04 松立控股集团股份有限公司 基于时空扩散的高位相机车牌识别方法、系统及电子设备
CN116386023B (zh) * 2023-05-29 2023-08-25 松立控股集团股份有限公司 基于时空扩散的高位相机车牌识别方法、系统及电子设备
CN117765539A (zh) * 2023-12-25 2024-03-26 暗物质(北京)智能科技有限公司 一种基于深度学习的字帖图像自动矫正方法及相关产品
CN118351524A (zh) * 2024-04-19 2024-07-16 北京科技大学 一种基于特征洗牌的低资源车牌识别方法及装置

Also Published As

Publication number Publication date
CN112446383A (zh) 2021-03-05
CN112446383B (zh) 2022-09-02

Similar Documents

Publication Publication Date Title
WO2022111355A1 (zh) 车牌识别方法及装置、存储介质、终端
CN110232394B (zh) 一种多尺度图像语义分割方法
CN109711481B (zh) 用于画作多标签识别的神经网络、相关方法、介质和设备
CN109299274B (zh) 一种基于全卷积神经网络的自然场景文本检测方法
CN108492271B (zh) 一种融合多尺度信息的自动图像增强系统及方法
CN110782420A (zh) 一种基于深度学习的小目标特征表示增强方法
CN108960261B (zh) 一种基于注意力机制的显著物体检测方法
CN111860398B (zh) 遥感图像目标检测方法、系统及终端设备
CN111553867B (zh) 一种图像去模糊方法、装置、计算机设备及存储介质
CN111523553A (zh) 一种基于相似度矩阵的中心点网络多目标检测方法
CN115222946B (zh) 一种单阶段实例图像分割方法、装置以及计算机设备
CN111967464B (zh) 一种基于深度学习的弱监督目标定位方法
CN110781980B (zh) 目标检测模型的训练方法、目标检测方法及装置
CN116645592B (zh) 一种基于图像处理的裂缝检测方法和存储介质
CN111461213A (zh) 一种目标检测模型的训练方法、目标快速检测方法
CN112700460B (zh) 图像分割方法及系统
CN114048822A (zh) 一种图像的注意力机制特征融合分割方法
CN116863194A (zh) 一种足溃疡图像分类方法、系统、设备及介质
CN115546640A (zh) 遥感影像的云检测方法、装置、电子设备及存储介质
CN116071748A (zh) 一种基于频域全局滤波的无监督视频目标分割方法
CN117952883A (zh) 一种基于双边网格和显著性引导的逆光图像增强方法
CN117542075A (zh) 基于注意力机制的小样本图像分类方法及装置
CN117115616A (zh) 一种基于卷积神经网络的实时低照度图像目标检测方法
CN116704206A (zh) 图像处理方法、装置、计算机设备和存储介质
CN116311290A (zh) 基于深度学习的手写及打印文本检测方法和装置

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: 21896856

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: 21896856

Country of ref document: EP

Kind code of ref document: A1