CN111027555B - License plate recognition method and device and electronic equipment - Google Patents
License plate recognition method and device and electronic equipment Download PDFInfo
- Publication number
- CN111027555B CN111027555B CN201811174638.1A CN201811174638A CN111027555B CN 111027555 B CN111027555 B CN 111027555B CN 201811174638 A CN201811174638 A CN 201811174638A CN 111027555 B CN111027555 B CN 111027555B
- Authority
- CN
- China
- Prior art keywords
- license plate
- target
- model
- feature sequence
- features
- Prior art date
- Legal status (The legal status 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 status listed.)
- Active
Links
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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- General Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Character Discrimination (AREA)
- Traffic Control Systems (AREA)
- Image Analysis (AREA)
Abstract
The application provides a license plate recognition method, a license plate recognition device and electronic equipment, comprising the following steps: determining a target license plate feature sequence according to each license plate feature of a target license plate in a target license plate region, wherein the license plate features are license plate attribute features extracted from an image containing the target license plate by a convolutional neural network; inputting the target license plate feature sequence into an attention model, carrying out character string recognition on the target license plate feature sequence by the attention model according to model parameters trained by taking the editing distance as a loss function, and outputting the license plate number of the target license plate; and obtaining the license plate number of the target license plate output by the attention model. The method provided by the application can improve the accuracy of license plate recognition.
Description
Technical Field
The present application relates to the field of image processing, and in particular, to a license plate recognition method, device and electronic equipment.
Background
As image processing technologies are becoming mature, license plate recognition technologies are also evolving.
In the existing license plate recognition technology, character segmentation is generally performed on a license plate region to obtain a single character image and position information of the single character, then character recognition is performed on the single character image, and the position information of the single character image is combined to form the whole character string of the license plate region.
However, the method of firstly dividing characters and then identifying single characters is not ideal for identifying abnormal license plate images such as adhesion of license plate characters, overlarge or undersize character spacing and the like, and the identified license plate numbers are easy to generate the problems of word missing, word missing and the like.
Disclosure of Invention
In view of the above, the present application provides a license plate recognition method, device and electronic equipment for improving accuracy of license plate recognition.
Specifically, the application is realized by the following technical scheme:
according to a first aspect of the present application, there is provided a license plate recognition method, the method comprising:
determining a target license plate feature sequence according to each license plate feature of a target license plate in a target license plate region, wherein the license plate features are license plate attribute features extracted from an image containing the target license plate by a convolutional neural network;
inputting the target license plate feature sequence into an attention model, carrying out character string recognition on the target license plate feature sequence by the attention model according to model parameters trained by taking the editing distance as a loss function, and outputting the license plate number of the target license plate;
and obtaining the license plate number of the target license plate output by the attention model.
Optionally, the determining the target license plate feature sequence according to each license plate feature in the target license plate region includes:
inputting each license plate feature into a bidirectional LSTM network, and processing each license plate feature by the bidirectional LSTM network to output a target license plate feature sequence, wherein adjacent license plate features in the target license plate feature sequence are related on adjacent time sequences;
and acquiring the target license plate feature sequence output by the bidirectional LSTM network.
Optionally, the attention model performs character string recognition on the target license plate feature sequence according to model parameters trained by taking the editing distance as a loss function, and outputs license plate numbers in the target license plate region, including:
according to model parameters trained by taking the calculated edit distance between the predicted character string and the calibrated character string as a loss function, calculating the activity value of the hidden layer in the attention model at each moment, and determining the license plate number of the target license plate according to the calculated activity value of the hidden layer at each moment.
Optionally, the calculating the activity value of the hidden layer in the attention model includes:
calculating weight factors of license plate features in the target license plate feature sequence at each moment;
Calculating semantic codes of all moments according to weight factors of all license plate features at all moments and the target license plate feature sequence;
and calculating the activity value of the hidden layer of the attention model at each moment based on the target license plate feature sequence and semantic codes at each moment.
Optionally, the convolutional neural network, the bidirectional LSTM network and the attention model are cascaded in a target neural network model, and the target neural network model further comprises a classification model, and the classification model is cascaded with the output of the convolutional neural network; the method further comprises the steps of:
inputting the license plate features into a preset classification model so that the classification model calculates the confidence that the license plate features of the target license plate correspond to different license plate types;
determining the license plate type with the highest calculated confidence as the license plate type of the target license plate;
and outputting the license plate type and the license plate number as the recognition result of the target license plate.
According to a second aspect of the present application, there is provided a license plate recognition device, the method comprising:
the determining unit is used for determining a target license plate feature sequence according to each license plate feature of the target license plate in the target license plate region, wherein the license plate features are license plate attribute features extracted from an image containing the target license plate by the convolutional neural network;
The recognition unit is used for inputting the target license plate feature sequence into an attention model, carrying out character string recognition on the target license plate feature sequence by the attention model according to model parameters trained by taking the editing distance as a loss function, and outputting the license plate number of the target license plate;
and the acquisition unit is used for acquiring the license plate number of the target license plate output by the attention model.
Optionally, the determining unit is specifically configured to input each license plate feature to a bidirectional LSTM network, so that the bidirectional LSTM network processes each license plate feature to output a target license plate feature sequence, and adjacent license plate features in the target license plate feature sequence are associated in adjacent time sequences; and acquiring the target license plate feature sequence output by the bidirectional LSTM network.
Optionally, the identification unit is specifically configured to calculate an activity value of a hidden layer in the attention model at each moment according to a model parameter trained by using an edit distance between the calculated predicted character string and the calibrated character string as a loss function, and determine a license plate number of the target license plate according to the calculated activity value of the hidden layer at each moment.
Optionally, the identifying unit is specifically configured to calculate a weight factor of each license plate feature in the target license plate feature sequence at each time when calculating the activity value of the hidden layer in the attention model; calculating semantic codes of all moments according to weight factors of all license plate features at all moments and the target license plate feature sequence; and calculating the activity value of the hidden layer of the attention model at each moment based on the target license plate feature sequence and semantic codes at each moment.
Optionally, the convolutional neural network, the bidirectional LSTM network and the attention model are cascaded in a target neural network model, and the target neural network model further comprises a classification model, and the classification model is cascaded with the output of the convolutional neural network; the apparatus further comprises:
the classification unit is used for inputting the license plate features into a preset classification model so that the classification model calculates the confidence that the license plate features of the target license plate correspond to different license plate types; determining the license plate type with the highest calculated confidence as the license plate type of the target license plate; and outputting the license plate type and the license plate number as the recognition result of the target license plate.
According to a third aspect of the present application there is provided an electronic device comprising a processor and a machine-readable storage medium storing machine-executable instructions executable by the processor, the processor being caused by the machine-executable instructions to perform the method of the first aspect.
According to a fourth aspect of the present application there is provided a machine-readable storage medium storing machine-executable instructions which, when invoked and executed by a processor, cause the processor to perform the method of the first aspect.
On one hand, after each license plate feature with spatial association is extracted, time sequence association processing is carried out on each license plate feature to obtain a target license plate feature sequence with spatial association and time sequence relationship, and then the whole character string recognition is carried out on the target license plate feature sequence through the attention model to obtain the license plate number. Because each license plate feature in the target license plate feature sequence for character string recognition has a spatial relationship and a time sequence relationship, the character string recognition by adopting the target license plate feature sequence can alleviate the problems of missing characters, missing characters and the like of the identified license plate number.
On the other hand, the attention model adopted by the application is obtained by calculating the editing distance between the identified license plate number and the marked license plate number whole character string as a loss function, rather than obtaining by calculating the Euclidean distance between the identified single character and the marked single character as the loss function, so that the integrity of the license plate number identified by the trained attention model is better, and the problems of the identified license plate number of losing characters, missing characters and the like can be relieved to a certain extent.
Drawings
Fig. 1 is a schematic diagram of a CNN network according to an exemplary embodiment of the present application;
FIG. 2 is a schematic diagram of an overall license plate recognition scheme according to an exemplary embodiment of the present application;
FIG. 3 is a flow chart of a license plate recognition method according to an exemplary embodiment of the present application;
FIG. 4 is a schematic diagram of a bi-directional LSTM network in accordance with an exemplary embodiment of the present application;
FIG. 5 is a schematic diagram of an attention model calculation shown in an exemplary embodiment of the present application;
FIG. 6 is a flowchart illustrating another license plate recognition method according to an exemplary embodiment of the present application;
FIG. 7 is a schematic diagram of a license plate recognition method according to an exemplary embodiment of the present application;
FIG. 8 is a hardware configuration diagram of an electronic device according to an exemplary embodiment of the present application;
fig. 9 is a block diagram of a license plate recognition device according to an exemplary embodiment of the present application.
Detailed Description
Reference will now be made in detail to exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, the same numbers in different drawings refer to the same or similar elements, unless otherwise indicated. The implementations described in the following exemplary examples do not represent all implementations consistent with the application. Rather, they are merely examples of apparatus and methods consistent with aspects of the application as detailed in the accompanying claims.
The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. As used in this specification and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It should also be understood that the term "and/or" as used herein refers to and encompasses any or all possible combinations of one or more of the associated listed items.
It should be understood that although the terms first, second, third, etc. may be used herein to describe various information, these information should not be limited by these terms. These terms are only used to distinguish one type of information from another. For example, first information may also be referred to as second information, and similarly, second information may also be referred to as first information, without departing from the scope of the application. The word "if" as used herein may be interpreted as "at … …" or "at … …" or "responsive to a determination", depending on the context.
In the existing license plate recognition technology, a CNN (Convolutional Neural Networks, convolutional neural network) network is generally adopted to perform feature extraction on license plate images, and then character segmentation and character recognition are performed by using the extracted license plate features.
Referring to fig. 1, fig. 1 is a schematic diagram of a CNN network according to an exemplary embodiment of the present application;
CNN networks typically include multiple network layers with full connectivity between each network layer. But the nodes on each network layer are independent and not associated. Therefore, the license plate features extracted by the CNN usually have only spatial association relationship, but do not have time sequence association relationship, so that each license plate feature is independent in time sequence, i.e. the current feature does not keep memory of the previous feature or the next feature. Therefore, the CNN network is utilized to identify license plate numbers, and particularly when the license plates with adhered characters and excessively large and small character spacing are identified, the problems of missing characters, missing characters and the like of the identified license plate numbers are easy to occur.
In view of this, the present application proposes a license plate recognition scheme, and referring to fig. 2, the license plate recognition scheme provided by the present application will be generally described.
As shown in fig. 2, the whole scheme of license plate recognition mainly comprises the following five parts:
a first part: a vehicle image is acquired as shown at 201 in fig. 2.
A second part: license plate detection is performed on the vehicle image, as shown at 202 in fig. 2.
In detecting license plates of vehicle objects, it is generally involved that a license plate region is located in a vehicle image, and then each license plate feature of a target license plate in the license plate region is proposed. The license plate features in the license plate features are related in spatial positions.
Third section: character recognition, as shown at 203 in fig. 2.
In the third part, time sequence association processing can be carried out on each license plate feature to obtain a target license plate feature sequence. Adjacent license plate features in the target license plate feature sequence are associated on adjacent time sequences. And then, carrying out overall character string recognition on the target license plate feature sequence by using the attention model to obtain the license plate number. The attention model is obtained by training a predicted value of the license plate number and the integral editing distance of the marked license plate number as a loss function through identifying a sample.
Fourth part: license plate type determination is shown as 204 in fig. 2.
In the fourth section, each license plate feature described above may be input into the classification model so that the classification model determines the license plate type of the vehicle.
Fifth part: and outputting license plate recognition results, as shown in 205 in fig. 2.
The license plate recognition result comprises: the identified license plate number and license plate type.
According to the license plate recognition technical scheme, on one hand, after each license plate feature with spatial association of the target license plate is extracted, time sequence association processing is carried out on each license plate feature, a target license plate feature sequence with spatial association and time sequence relationship is obtained, and then the whole character string recognition is carried out on the target license plate feature sequence through the attention model, so that the license plate number is obtained. Because each license plate feature in the target license plate feature sequence for character string recognition has a spatial relationship and a time sequence relationship, the character string recognition by adopting the target license plate feature sequence can alleviate the problems of missing characters, missing characters and the like of the identified license plate number.
On the other hand, the attention model adopted by the application is obtained by calculating the editing distance between the identified license plate number and the marked license plate number whole character string as a loss function, rather than obtaining by calculating the Euclidean distance between the identified single character and the marked single character as the loss function, so that the integrity of the license plate number identified by the trained attention model is better, and the problems of the identified license plate number of losing characters, missing characters and the like can be relieved to a certain extent.
The application can be applied to image acquisition equipment such as cameras in scenes such as crossroads and intelligent monitoring, and can also be applied to a background server of a monitoring system.
The license plate recognition process proposed herein is described in detail below.
Referring to fig. 3, fig. 3 is a flowchart illustrating a license plate recognition method according to an exemplary embodiment of the present application;
step 301: and determining a target license plate feature sequence according to each license plate feature of the target license plate in the target license plate region.
Step 1: an image of a target vehicle is acquired.
The target vehicle image is an image obtained by shooting the target vehicle by the image acquisition device, and the target vehicle image contains the target vehicle.
In an alternative implementation, the image capture device may capture the target vehicle image when the license plate recognition method is applied to the image capture device. Alternatively, the image capturing device may receive the target vehicle image transmitted by another device (e.g., another image capturing device, a server, etc.).
In another alternative implementation, when the license plate recognition method is applied to a background server of the monitoring system, the background server may receive the target vehicle image acquired by the image acquisition device.
The acquisition of the target vehicle image is only exemplarily described here, and is not particularly limited.
The image capturing apparatus may be a camera, a video camera, or the like having an image capturing function, and is described here by way of example only, without being limited to the image capturing apparatus.
Step 2: and positioning a license plate region of the target vehicle in the target vehicle image.
When the method is realized, the license plate region can be positioned by adopting a neural network, and parameters such as coordinates, length and width of the license plate region are determined.
For example, a large number of candidate frames are drawn in the whole image in advance, and feature extraction is performed on all the candidate frames by using a preset neural network. For each candidate box, a confidence level of the candidate box relative to each object is calculated according to the features extracted from the candidate box. The candidate frame with the highest confidence coefficient relative to the license plate is determined to be the candidate frame containing the license plate, the area framed by the candidate frame is the license plate area, the coordinates of the candidate frame are the coordinates of the license plate area, and the length and the width of the candidate frame are the length and the width of the license plate area.
Of course, the whole image may be divided into sub-images according to pixel points or pixel small blocks composed of adjacent pixel points. And extracting the characteristics of each sub-image by using a preset neural network, calculating the confidence coefficient of each article corresponding to the sub-image, finding the sub-image with the highest confidence coefficient corresponding to the license plate, determining the region contained in the sub-image as the license plate region, wherein the coordinates of the sub-image are the coordinates of the license plate region, and the length and width of the sub-image are the length and width of the license plate region.
Here, this is an exemplary illustration of locating license plate regions, which are not specifically limited.
The neural network for locating the license plate Region may be an FRCNN (Fast Region-based Convolutional Neural Networks, region-based Fast convolution neural network) network or a YOLO ((You Only Look Once)) network, which is only exemplified herein and is not specifically limited thereto.
After the license plate area is determined and positioned, the license plate area image can be intercepted from the target vehicle image according to the determined coordinates, length, width and other parameters of the license plate area.
Step 3: extracting license plate features aiming at a target license plate in the license plate region; the license plate features have a spatial association with each other.
In order to reduce the calculation amount of the subsequent feature extraction and the subsequent character string recognition, the license plate region can be preprocessed, and then each license plate feature aiming at the target license plate is extracted from the preprocessed license plate region.
For example, the license plate region may be normalized to a certain size. The pretreatment of the license plate region is only exemplified here, and is not particularly limited.
In the embodiment of the application, each license plate feature aiming at the target license plate can be proposed in the preprocessed license plate area by utilizing the CNN network. The license plate features extracted by using the CNN network have spatial correlation with each other.
When the license plate is realized, the preprocessed license plate region can be input into a CNN network, the CNN network performs feature extraction on the license plate region, and various license plate features are output.
For example: after the license plate area is input into the CNN network, assuming that the license plate features extracted by the CNN are a plurality of character features, the character features contained in each formed license plate feature are character feature 'A', character feature '1', character feature '2', character feature '3', character feature '4' and character feature '5' in sequence from left to right according to the space position.
Of course, this is merely illustrative of license plate features, as well as individual license plate features. In practical applications, the license features may further include license frame features, texture features, edge features, etc., and each license feature formed is more complex, which is only exemplified herein and is not specifically limited thereto.
Step 4: processing the extracted license plate features to form a target license plate feature sequence; adjacent license plate features in the target license plate feature sequence are associated on adjacent time sequences.
In the embodiment of the application, in order to alleviate the problems of missing characters, missing characters and the like of the identified license plate numbers, the application also processes each license plate feature to obtain a target license plate feature sequence, so that adjacent features in the target license plate sequence are not only spatially associated, but also are associated on adjacent time sequences, the relevance of each license plate feature is stronger, and when the method is realized, the application can process each license plate feature by utilizing a bidirectional LSTM network to obtain the target license plate feature sequence.
Referring to fig. 4, fig. 4 is a schematic diagram of a bidirectional LSTM network according to an exemplary embodiment of the present application.
A bi-directional LSTM network typically includes an input layer, a hidden layer, and an output layer.
The hidden layers of the bidirectional LSTM network comprise two layers, namely a Forward layer and a Backward layer.
In the bidirectional LSTM network, the Forward layer calculates according to the sequence from time 1 to time t and stores the activity value of the Forward layer at each time, and the Backward layer calculates according to the sequence from time t to time 1 and stores the activity value of the Backward layer at each time. And the output layer combines the activity value of the Forward layer and the activity value of the Backward layer at each moment to obtain a final output result.
The two-way LSTM network Forward layer structure enables the post-processed features of the two-way LSTM network to have the memory of the pre-processed features, and the two-way LSTM network Backward layer structure enables the pre-processed features to have the memory of the post-processed features, so that adjacent license plate features of the target license plate features obtained through LSTM network processing also have the correlation on adjacent time sequences.
When the method is realized, each license plate feature can be input into a bidirectional LSTM model, each license plate feature of the bidirectional LSTM model is processed, then a target license plate feature sequence is output, and adjacent license plate features in the target license plate feature sequence are in adjacent time sequence association.
For example, the above license plate number a12345 is still taken as an example.
Assuming that the features in each license plate feature are character feature "A", character feature "1", character feature "2", character feature "3", character feature "4" and character feature "5" in sequence, inputting the six features into a bidirectional LSTM model for processing, and obtaining each license plate feature in the target license plate feature sequence, wherein the license plate features have time sequence correlation.
In the case of character feature "3", after the bidirectional LSTM network processing, it can be known that character feature "a", character feature "1", character feature "2" exist before character feature "3", and character feature "4" and character feature "5" exist after character feature "3". Of course, in practical applications, license features may also include license frame features, texture features, edge features, etc., so that the resulting target license feature sequence is more complex, which is only illustrated here by way of example.
It should be noted that, the present application may also use a unidirectional LSTM network, an RNN network formed by an activation function of the tanh algorithm, and other neural networks capable of processing time sequence association to process each license plate feature, so that each license plate feature in the output target license plate feature sequence has time sequence association with each other, but the bidirectional LSTM network may enable the current license plate feature to have the association of the previously processed license plate feature and the later processed license plate feature, so that the association is stronger, and therefore, the bidirectional LSTM network is preferred. The exemplary description of the processing of each license plate feature is not particularly limited.
Step 302: and inputting the target license plate feature sequence into an attention model, carrying out character string recognition on the target license plate feature sequence by the attention model according to model parameters trained by taking the editing distance as a loss function, and outputting the license plate number of the target license plate.
Step 303: and obtaining the license plate number of the target license plate output by the attention model.
Note that, the attention model used in the present application is different from the existing attention model:
the loss function adopted by the existing attention model in training is a softmax algorithm, and the softmax algorithm is used for calculating a loss value by identifying a Euclidean distance between each character and a calibration character by a sample.
For example, assuming that the license plate number is a12845, the character string obtained by the recognition sample is a12345, and the calibrated character string is a12845, the existing attention model calculates the difference loss by calculating the euclidean distance between the single recognition character and the calibration character, that is, calculating the euclidean distance between the recognition character a and the calibration character a, the euclidean distance between the recognition character 1 and the calibration character 1, the euclidean distance between the recognition character 2 and the calibration character 2, the euclidean distance between the recognition character 3 and the calibration character 8, the euclidean distance between the recognition character 4 and the calibration character 4, and the euclidean distance between the recognition character 5 and the calibration character 5.
Because the Euclidean distance between the single character obtained by calculation and recognition and the calibrated single character is used as a loss function to train the attention model, the recognized license plate number has poor integrity, and particularly, the recognized license plate number has the problems of word loss, word leakage and the like easily when the license plate number is used for recognizing license plates with adhered characters, overlarge character spacing and the like.
The attention model of the application adopts the edit distance of the whole character as a loss function during training, namely, the difference loss is calculated by calculating the edit distance of the whole character string (including all characters) obtained by recognition and the calibration character string.
For example, assuming that the license plate number is a12845, the character string obtained by identifying the sample is a12345, and the calibration character string is a12845. The mode of calculating the loss value by the attention model of the application is as follows: and calculating the integral editing distance between the character string obtained by the identification sample and the calibration character string, namely calculating the integral editing distance between the character string 'A12345' obtained by the identification and the calibration character string 'A12845'.
The attention model is trained by calculating the editing distance between the identification character string and the whole calibration character string as a loss function, so that the identified license plate number is good in integrity, and particularly, the problems of word loss, word leakage and the like of the identified license plate number can be effectively alleviated when the license plate number is identified, such as adhesion of license plate characters, overlarge character spacing and the like.
Step 302 is described in detail below in terms of both attention model training and license plate number recognition using an attention model.
1) Training of attention models
Firstly, a license plate region sample can be positioned in a vehicle image sample, then each license plate characteristic sample is extracted from the license plate region sample, and each license plate characteristic sample is input into a bidirectional LSTM network to obtain a target license plate characteristic sequence sample.
The target license plate feature sequence sample, and the license plate number calibrated from the vehicle image sample, are then input into the attention model to train model parameters of the attention model.
In training, two phases, forward propagation and backward propagation, may typically be included:
and (3) forward spreading, namely sequentially backward spreading the target license plate feature sequence sample from the first layer to the last layer by the attention model so as to perform overall character string recognition on the target license plate feature sequence sample by the attention model, and obtaining the target license plate number.
And (5) back-spreading, and calculating the integral editing distance between the target license plate number and the standard license plate number by using the attention model as a loss value. The differential loss is then propagated back from the last layer of the attention model, in turn, to the first layer to adjust the model parameters of the attention model until the loss value converges.
2) License plate number identification using attention model
Typically, the attention model is built from a unidirectional LSTM network, which typically has an input layer, a hidden layer, and an output layer. The attention model thus also has an input layer, a hidden layer and an output layer.
Referring to fig. 5, fig. 5 is a schematic diagram of an attention model calculation according to an exemplary embodiment of the present application.
After the attention model is trained, the target license plate feature sequence can be input into the trained attention model, the attention model can calculate the activity value of the hidden layer of the attention model through the model parameters obtained by training with the editing distance as a loss function, and the license plate number of the target license plate is determined according to the calculated activity value.
When calculating the hidden layer activity value, the attention model can calculate the weight factor corresponding to each license plate in the target license plate sequence.
For example, as shown in FIG. 5, X in FIG. 5 1 、X 2 、…、X M The sequence is the target license plate characteristic sequence, alpha in figure 5 t,1 Is X 1 Weight factor at time T, alpha t,2 Is X 2 Weight factor at time T, …, alpha t,M Is X M Weight factor at time T.
Then, the attention model can calculate semantic codes at all times according to the weight factors of the license plate features at all times and the target license plate feature sequence.
As shown in FIG. 5, the attention model may be based on license plate feature X 1 X is as follows 1 Weight factor alpha of (2) t,1 License plate feature X 2 X is as follows 2 Weight factor alpha of (2) t,2 … license plate feature X M X is as follows M Weight factor alpha of (2) t,M Performing operation to obtain semantic code C at time T t 。
The attention model may then calculate the activity value of the hidden layer of the attention model at each time instant based on the target license plate feature sequence and the semantic code at each time instant.
As shown in FIG. 5, the initial activity value of the hidden layer is a preset value S 0 . The attention model may be based on the target license plate feature sequence X 1 、X 2 、…、X M And the calculated 1-time semantic code C 1 And combine with S 0 Obtaining the activity value S of the hidden layer 1 1 And so on.
The attention model may be based on the target license plate feature sequence X 1 、X 2 、…、X M And the semantic code C of the calculated T moment t Calculating the activity value S of the hidden layer by combining the T-1 moment T-1 Obtaining the activity value S of the hidden layer at the moment T T 。
Finally, the attention model can determine the license plate number of the target license plate according to the calculated activity value of the hidden layer at each moment.
For example, as shown in FIG. 5, the attention model may decode according to the hidden layer activity values at various times to obtain decoding results, such as Y T The decoding result at time T. The decoding result contains the confidence coefficient of each candidate character, then the character with the highest confidence coefficient can be used as the recognized character, and then the recognized characters at each moment are combined according to the moment sequence to generate the recognized license plate number.
Note that, assuming that the license plate is a12345 including 6 characters, each of the above-described time points is 6 time points, and the decoding result of each of the above-described output time points is also a decoding result of 6 time points.
In addition, in the embodiment of the application, the license plate number of the target vehicle can be identified, and the license plate type of the target vehicle can also be identified.
Specifically, the convolutional neural network, the bidirectional LSTM network and the attention model are cascaded in a target neural network model, wherein the target neural network model further comprises a classification model, and the classification model is cascaded with the output of the convolutional neural network (namely, the CNN network).
In implementation, after a plurality of license plate features aiming at a target license plate are extracted from a license plate region by adopting a CNN network, each license plate feature can be input into a classification model, and the classification model can calculate the confidence that a set formed by each license plate feature corresponds to different types of license plates. The license plate type with the highest calculated confidence may then be determined as the license plate type of the target vehicle. The application can output the license plate number of the identified target vehicle and the license plate type of the target vehicle as the identification result of the target license plate.
The classification model may be a classification model built based on a softmax algorithm, and may also be other classifiers, such as a SSD (Single Shot MultiBox Detector) classifier, etc., which are only exemplified herein, and are not specifically limited thereto.
The license plate types can comprise a civil large-sized vehicle license plate, a civil small-sized vehicle license plate, a police vehicle license plate and the like. The license plate type is only exemplarily described herein, and is not particularly limited.
The license plate recognition method provided by the application is described in detail below by means of specific embodiments and with reference to fig. 6 and 7.
Referring to fig. 6, fig. 6 is a flowchart illustrating another license plate recognition method according to an exemplary embodiment of the present application, which may include the following steps.
It should be noted that, as shown in fig. 6, the present application designs a target neural network model including a convolutional neural network (CNN network), a bidirectional LSTM network, an attention model, and a classification model. The system comprises a bidirectional LSTM network, a convolutional neural network, an attention model, a classification model and a classification model, wherein the bidirectional LSTM network and the convolutional neural network are in output cascade connection.
Step 601: an image of a target vehicle is acquired.
The specific implementation may refer to step 301, and will not be described herein.
Step 602: and positioning a license plate region of the target vehicle in the target vehicle image.
The specific implementation may refer to step 301, and will not be described herein.
The target vehicle image may be input into a YOLO model or an FRCNN network, which may output a license plate region, as shown in fig. 7.
Step 603: and utilizing a CNN network to propose various license plate characteristics in the license plate region.
When the method is realized, the license plate region image can be input into a CNN network, and the CNN network performs feature extraction on the license plate region image to extract a plurality of license plate features. Each license plate feature extracted by the CNN network has spatial correlation with each other.
Step 604: and determining the license plate type of the target vehicle by using a softmax-classification model.
In implementation, the license plate features may be input into a softmax-classification model that computes the confidence that the license plate features correspond to different license plate types.
The license plate type with the highest calculated confidence may then be determined as the license plate type of the target vehicle.
Step 305: and processing the license plate features by using a bidirectional LSTM network to obtain a target license plate feature sequence.
When the method is realized, the license plate features can be input into a bidirectional LSTM network, and the bidirectional LSTM network can process the license plate features to obtain a target license plate feature sequence. Adjacent time sequence association is arranged between adjacent license plate features in the target license plate feature sequence.
Step 606: and carrying out overall character string recognition on the target license plate feature sequence by using the Attention model to obtain the license plate number.
Specific implementation can refer to step 302 and step 303, and will not be described herein.
The loss function adopted by the Attention model in training is as follows: the loss value is calculated by calculating the editing distance between the identified license plate number and the integral character string of the standard license plate number, and the integral character string is calculated, so that the identified license plate number has good integrity, and particularly, the problems of word loss, word leakage and the like of the identified license plate number can be effectively alleviated when the license plate number is identified with adhesion of license plate characters, overlarge character spacing and the like.
During recognition, the target license plate feature sequence can be input into an Attention model, and the Attention model can recognize the whole character string of the target license plate feature sequence to obtain a license plate number.
Step 607: and outputting the identified license plate number and license plate type.
As shown in FIG. 7, the final output is license plate number "LuA 88888", license plate type "Small vehicle".
According to the license plate recognition technical scheme, on one hand, after each license plate feature with spatial association of the target license plate is extracted, time sequence association processing is carried out on each license plate feature, a target license plate feature sequence with spatial association and time sequence relationship is obtained, and then the whole character string recognition is carried out on the target license plate feature sequence through the attention model, so that the license plate number is obtained. Because each license plate feature in the target license plate feature sequence for character string recognition has a spatial relationship and a time sequence relationship, the character string recognition by adopting the target license plate feature sequence can alleviate the problems of missing characters, missing characters and the like of the identified license plate number.
On the other hand, the attention model adopted by the application is obtained by calculating the editing distance between the identified license plate number and the marked license plate number whole character string as a loss function, rather than obtaining by calculating the Euclidean distance between the identified single character and the marked single character as the loss function, so that the integrity of the license plate number identified by the trained attention model is better, and the problems of the identified license plate number of losing characters, missing characters and the like can be relieved to a certain extent.
Referring to fig. 8, fig. 8 is a hardware configuration diagram of an electronic device according to an exemplary embodiment of the present application.
The electronic device includes: a communication interface 801, a processor 802, a machine-readable storage medium 803, and a bus 804; wherein the communication interface 801, the processor 802, and the machine-readable storage medium 803 communicate with each other via a bus 804. The processor 802 may perform the license plate recognition method described above by reading and executing machine-executable instructions in the machine-readable storage medium 803 corresponding to the license plate recognition control logic.
The machine-readable storage medium 803 referred to herein may be any electronic, magnetic, optical, or other physical storage device that may contain or store information, such as executable instructions, data, or the like. For example, a machine-readable storage medium may be: volatile memory, nonvolatile memory, or similar storage medium. In particular, the machine-readable storage medium 803 may be RAM (Radom Access Memory, random access memory), flash memory, a storage drive (e.g., hard drive), a solid state drive, any type of storage disk (e.g., optical disk, DVD, etc.), or a similar storage medium, or a combination thereof.
Referring to fig. 9, fig. 9 is a block diagram of a license plate recognition apparatus according to an exemplary embodiment of the present application, which may include the following units.
The determining unit 901 is configured to determine a target license plate feature sequence according to each license plate feature of a target license plate in a target license plate region, where the license plate feature is a license plate attribute feature extracted from an image including the target license plate by a convolutional neural network;
the recognition unit 902 is configured to input the target license plate feature sequence to an attention model, perform character string recognition on the target license plate feature sequence by using the attention model according to model parameters trained by using an edit distance as a loss function, and output a license plate number of the target license plate;
and the obtaining unit 903 is configured to obtain a license plate number of the target license plate output by the attention model.
Optionally, the determining unit is specifically configured to input each license plate feature to a bidirectional LSTM network, so that the bidirectional LSTM network processes each license plate feature to output a target license plate feature sequence, and adjacent license plate features in the target license plate feature sequence are associated in adjacent time sequences; and acquiring the target license plate feature sequence output by the bidirectional LSTM network.
Optionally, the identifying unit 902 is specifically configured to calculate an activity value of a hidden layer in the attention model at each moment according to a model parameter trained by using an edit distance between the calculated predicted character string and the calibrated character string as a loss function, and determine a license plate number of the target license plate according to the calculated activity value of the hidden layer at each moment.
Optionally, the identifying unit 902 is specifically configured to calculate a weight factor of each license plate feature in the target license plate feature sequence at each time when calculating the activity value of the hidden layer in the attention model; calculating semantic codes of all moments according to weight factors of all license plate features at all moments and the target license plate feature sequence; and calculating the activity value of the hidden layer of the attention model at each moment based on the target license plate feature sequence and semantic codes at each moment.
Optionally, the convolutional neural network, the bidirectional LSTM network, and the attention model are cascaded in a target neural network model, the target neural network model further including a classification model, the classification model being cascaded with an output of the convolutional neural network, the apparatus further comprising:
the classification unit 904 is configured to input the license plate features into a preset classification model, so that the classification model calculates confidence degrees of the license plate features of the target license plate corresponding to different license plate types; determining the license plate type with the highest calculated confidence as the license plate type of the target license plate; and outputting the license plate type and the license plate number as the recognition result of the target license plate.
The implementation process of the functions and roles of each unit in the above device is specifically shown in the implementation process of the corresponding steps in the above method, and will not be described herein again.
For the device embodiments, reference is made to the description of the method embodiments for the relevant points, since they essentially correspond to the method embodiments. The apparatus embodiments described above are merely illustrative, wherein the elements illustrated as separate elements may or may not be physically separate, and the elements shown as elements may or may not be physical elements, may be located in one place, or may be distributed over a plurality of network elements. Some or all of the modules may be selected according to actual needs to achieve the purposes of the present application. Those of ordinary skill in the art will understand and implement the present application without undue burden.
The foregoing description of the preferred embodiments of the application is not intended to be limiting, but rather to enable any modification, equivalent replacement, improvement or the like to be made within the spirit and principles of the application.
Claims (10)
1. A license plate recognition method, the method comprising:
Determining a target license plate feature sequence according to each license plate feature of a target license plate in a target license plate region, wherein the license plate features are license plate attribute features extracted from images containing the target license plate by a convolutional neural network, and adjacent license plate features in the target license plate feature sequence are related on adjacent time sequences;
inputting the target license plate feature sequence into an attention model, calculating the activity value of a hidden layer in the attention model at each moment by using the attention model according to model parameters trained by taking the edit distance between a calculated and predicted character string and a calibrated character string as a loss function, and determining the license plate number of the target license plate according to the calculated activity value of the hidden layer at each moment, wherein the calculated and predicted character string comprises all characters in the license plate number;
obtaining the license plate number of the target license plate output by the attention model;
the determining the license plate number of the target license plate according to the calculated activity values of the hidden layers at all moments comprises the following steps: the attention model decodes according to the hidden layer activity values at all times to obtain decoding results, the decoding results comprise confidence degrees of all candidate characters, the character with the highest confidence degree is used as the recognized character, and then the recognized characters at all times are combined according to the time sequence to generate the recognized license plate number.
2. The method of claim 1, wherein said determining a target license plate feature sequence from each license plate feature in the target license plate region comprises:
inputting the license plate features into a bidirectional LSTM network, and processing the license plate features by the bidirectional LSTM network to output a target license plate feature sequence;
and acquiring the target license plate feature sequence output by the bidirectional LSTM network.
3. The method of claim 1, wherein said calculating the activity value of the hidden layer in the attention model comprises:
calculating weight factors of license plate features in the target license plate feature sequence at each moment;
calculating semantic codes of all moments according to weight factors of all license plate features at all moments and the target license plate feature sequence;
and calculating the activity value of the hidden layer of the attention model at each moment based on the target license plate feature sequence and semantic codes at each moment.
4. The method of claim 2, wherein the convolutional neural network, the bi-directional LSTM network, and the attention model are cascaded in a target neural network model, the target neural network model further comprising a classification model, the classification model being cascaded with an output of the convolutional neural network; the method further comprises the steps of:
Inputting the license plate features into a preset classification model so that the classification model calculates the confidence that the license plate features of the target license plate correspond to different license plate types;
determining the license plate type with the highest calculated confidence as the license plate type of the target license plate;
and outputting the license plate type and the license plate number as the recognition result of the target license plate.
5. A license plate recognition device, the device comprising:
the determining unit is used for determining a target license plate feature sequence according to each license plate feature of the target license plate in the target license plate region, wherein the license plate features are license plate attribute features extracted from images containing the target license plate by the convolutional neural network, and adjacent license plate features in the target license plate feature sequence are associated in adjacent time sequences;
the recognition unit is used for inputting the target license plate feature sequence into an attention model, calculating the activity value of a hidden layer in the attention model at each moment according to model parameters trained by taking the edit distance between a calculated and predicted character string and a calibrated character string as a loss function by the attention model, and determining the license plate number of the target license plate according to the calculated activity value of the hidden layer at each moment, wherein the calculated and predicted character string comprises all characters in the license plate number;
The acquisition unit is used for acquiring the license plate number of the target license plate output by the attention model;
the determining the license plate number of the target license plate according to the calculated activity values of the hidden layers at all moments comprises the following steps: the attention model decodes according to the hidden layer activity values at all times to obtain decoding results, the decoding results comprise confidence degrees of all candidate characters, the character with the highest confidence degree is used as the recognized character, and then the recognized characters at all times are combined according to the time sequence to generate the recognized license plate number.
6. The apparatus of claim 5, wherein the determining unit is specifically configured to input each license plate feature to a bidirectional LSTM network, and process each license plate feature by the bidirectional LSTM network to output a target license plate feature sequence; and acquiring the target license plate feature sequence output by the bidirectional LSTM network.
7. The apparatus according to claim 5, wherein the identifying unit is configured to calculate a weight factor of each license plate feature in the target license plate feature sequence at each time when calculating the activity value of the hidden layer in the attention model; calculating semantic codes of all moments according to weight factors of all license plate features at all moments and the target license plate feature sequence; and calculating the activity value of the hidden layer of the attention model at each moment based on the target license plate feature sequence and semantic codes at each moment.
8. The apparatus of claim 6, wherein the convolutional neural network, the bidirectional LSTM network, and the attention model are cascaded in a target neural network model, the target neural network model further comprising a classification model, the classification model being cascaded with an output of the convolutional neural network;
the apparatus further comprises: the classification unit is used for inputting the license plate features into a preset classification model so that the classification model calculates the confidence that the license plate features of the target license plate correspond to different license plate types; determining the license plate type with the highest calculated confidence as the license plate type of the target license plate; and outputting the license plate type and the license plate number as the recognition result of the target license plate.
9. An electronic device comprising a processor and a machine-readable storage medium storing machine-executable instructions executable by the processor to cause the method of any one of claims 1 to 4 to be performed.
10. A machine-readable storage medium storing machine-executable instructions which, when invoked and executed by a processor, cause the processor to perform the method of any one of claims 1 to 4.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811174638.1A CN111027555B (en) | 2018-10-09 | 2018-10-09 | License plate recognition method and device and electronic equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811174638.1A CN111027555B (en) | 2018-10-09 | 2018-10-09 | License plate recognition method and device and electronic equipment |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111027555A CN111027555A (en) | 2020-04-17 |
| CN111027555B true CN111027555B (en) | 2023-09-26 |
Family
ID=70191036
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811174638.1A Active CN111027555B (en) | 2018-10-09 | 2018-10-09 | License plate recognition method and device and electronic equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111027555B (en) |
Families Citing this family (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113515985B (en) * | 2020-07-01 | 2022-07-22 | 阿里巴巴集团控股有限公司 | Self-service weighing system, weighing detection method, weighing detection equipment and storage medium |
| CN112132031B (en) * | 2020-09-23 | 2024-04-16 | 平安国际智慧城市科技股份有限公司 | Vehicle style identification method and device, electronic equipment and storage medium |
| CN112508018A (en) * | 2020-12-14 | 2021-03-16 | 北京澎思科技有限公司 | License plate recognition method and device and storage medium |
| CN113486885A (en) * | 2021-06-17 | 2021-10-08 | 杭州鸿泉物联网技术股份有限公司 | License plate recognition method and device, electronic equipment and storage medium |
| CN114529896A (en) * | 2022-02-24 | 2022-05-24 | 上海闪马智能科技有限公司 | License plate data processing method and device, storage medium and electronic device |
Citations (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8331621B1 (en) * | 2001-10-17 | 2012-12-11 | United Toll Systems, Inc. | Vehicle image capture system |
| CN106960206A (en) * | 2017-02-08 | 2017-07-18 | 北京捷通华声科技股份有限公司 | Character identifying method and character recognition system |
| CN107239778A (en) * | 2017-06-09 | 2017-10-10 | 中国科学技术大学 | The licence plate recognition method of efficiently and accurately |
| CN107368831A (en) * | 2017-07-19 | 2017-11-21 | 中国人民解放军国防科学技术大学 | English words and digit recognition method in a kind of natural scene image |
| CN108009543A (en) * | 2017-11-29 | 2018-05-08 | 深圳市华尊科技股份有限公司 | A kind of licence plate recognition method and device |
| CN108053653A (en) * | 2018-01-11 | 2018-05-18 | 广东蔚海数问大数据科技有限公司 | Vehicle behavior Forecasting Methodology and device based on LSTM |
| CN108122209A (en) * | 2017-12-14 | 2018-06-05 | 浙江捷尚视觉科技股份有限公司 | A kind of car plate deblurring method based on confrontation generation network |
| WO2018099194A1 (en) * | 2016-11-30 | 2018-06-07 | 杭州海康威视数字技术股份有限公司 | Character identification method and device |
| WO2018112900A1 (en) * | 2016-12-23 | 2018-06-28 | 深圳先进技术研究院 | License plate recognition method and apparatus, and user equipment |
| CN108229474A (en) * | 2017-12-29 | 2018-06-29 | 北京旷视科技有限公司 | Licence plate recognition method, device and electronic equipment |
| WO2018121006A1 (en) * | 2016-12-30 | 2018-07-05 | 杭州海康威视数字技术股份有限公司 | Method and device for license plate positioning |
| CN108388896A (en) * | 2018-02-09 | 2018-08-10 | 杭州雄迈集成电路技术有限公司 | A kind of licence plate recognition method based on dynamic time sequence convolutional neural networks |
| CN108615036A (en) * | 2018-05-09 | 2018-10-02 | 中国科学技术大学 | A kind of natural scene text recognition method based on convolution attention network |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9785855B2 (en) * | 2015-12-17 | 2017-10-10 | Conduent Business Services, Llc | Coarse-to-fine cascade adaptations for license plate recognition with convolutional neural networks |
| US10068146B2 (en) * | 2016-02-25 | 2018-09-04 | Conduent Business Services, Llc | Method and system for detection-based segmentation-free license plate recognition |
-
2018
- 2018-10-09 CN CN201811174638.1A patent/CN111027555B/en active Active
Patent Citations (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8331621B1 (en) * | 2001-10-17 | 2012-12-11 | United Toll Systems, Inc. | Vehicle image capture system |
| WO2018099194A1 (en) * | 2016-11-30 | 2018-06-07 | 杭州海康威视数字技术股份有限公司 | Character identification method and device |
| WO2018112900A1 (en) * | 2016-12-23 | 2018-06-28 | 深圳先进技术研究院 | License plate recognition method and apparatus, and user equipment |
| WO2018121006A1 (en) * | 2016-12-30 | 2018-07-05 | 杭州海康威视数字技术股份有限公司 | Method and device for license plate positioning |
| CN106960206A (en) * | 2017-02-08 | 2017-07-18 | 北京捷通华声科技股份有限公司 | Character identifying method and character recognition system |
| CN107239778A (en) * | 2017-06-09 | 2017-10-10 | 中国科学技术大学 | The licence plate recognition method of efficiently and accurately |
| CN107368831A (en) * | 2017-07-19 | 2017-11-21 | 中国人民解放军国防科学技术大学 | English words and digit recognition method in a kind of natural scene image |
| CN108009543A (en) * | 2017-11-29 | 2018-05-08 | 深圳市华尊科技股份有限公司 | A kind of licence plate recognition method and device |
| CN108122209A (en) * | 2017-12-14 | 2018-06-05 | 浙江捷尚视觉科技股份有限公司 | A kind of car plate deblurring method based on confrontation generation network |
| CN108229474A (en) * | 2017-12-29 | 2018-06-29 | 北京旷视科技有限公司 | Licence plate recognition method, device and electronic equipment |
| CN108053653A (en) * | 2018-01-11 | 2018-05-18 | 广东蔚海数问大数据科技有限公司 | Vehicle behavior Forecasting Methodology and device based on LSTM |
| CN108388896A (en) * | 2018-02-09 | 2018-08-10 | 杭州雄迈集成电路技术有限公司 | A kind of licence plate recognition method based on dynamic time sequence convolutional neural networks |
| CN108615036A (en) * | 2018-05-09 | 2018-10-02 | 中国科学技术大学 | A kind of natural scene text recognition method based on convolution attention network |
Non-Patent Citations (2)
| Title |
|---|
| Hui Li, Chunhua Shen.Reading Car License Plates Using Deep Convolutional Neural Networks and LSTMs.《CVPR》.2016,全文. * |
| 曹正凤.基于深度学习的端到端车牌检测识别系统.《中国交通信息化》.2018,全文. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111027555A (en) | 2020-04-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111027555B (en) | License plate recognition method and device and electronic equipment | |
| CN113065558B (en) | Lightweight small target detection method combined with attention mechanism | |
| Hausler et al. | Multi-process fusion: Visual place recognition using multiple image processing methods | |
| CN108805083B (en) | Single-stage video behavior detection method | |
| CN110263659B (en) | Finger vein recognition method and system based on triplet loss and lightweight network | |
| Noh et al. | A new framework for background subtraction using multiple cues | |
| CN109919977B (en) | Video motion person tracking and identity recognition method based on time characteristics | |
| CN111832484B (en) | Loop detection method based on convolution perception hash algorithm | |
| Weinzaepfel et al. | Learning to detect motion boundaries | |
| CN110717411A (en) | Pedestrian re-identification method based on deep layer feature fusion | |
| CN114972418A (en) | Maneuvering multi-target tracking method based on combination of nuclear adaptive filtering and YOLOX detection | |
| CN111104925B (en) | Image processing method, image processing apparatus, storage medium, and electronic device | |
| CN112541434B (en) | Face recognition method based on central point tracking model | |
| CN110909591A (en) | Self-adaptive non-maximum value inhibition processing method for pedestrian image detection by using coding vector | |
| EP3012779A1 (en) | Method and apparatus for extracting feature correspondences from multiple images | |
| CN109063549A (en) | High-resolution based on deep neural network is taken photo by plane video moving object detection method | |
| CN115937596A (en) | Target detection method, training method and device of model thereof, and storage medium | |
| CN115482523A (en) | Small object target detection method and system of lightweight multi-scale attention mechanism | |
| CN114742112A (en) | Object association method and device and electronic equipment | |
| CN115049731A (en) | Visual mapping and positioning method based on binocular camera | |
| CN113256683B (en) | Target tracking method and related equipment | |
| CN112115786B (en) | Monocular vision odometer method based on attention U-net | |
| CN113112479A (en) | Progressive target detection method and device based on key block extraction | |
| Kim et al. | Diminishing unwanted objects based on object detection using deep learning and image inpainting | |
| CN116824641A (en) | Gesture classification method, device, equipment and computer storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |