WO2021008018A1 - 基于人工智能的车辆识别方法、装置、程序及存储介质 - Google Patents

基于人工智能的车辆识别方法、装置、程序及存储介质 Download PDF

Info

Publication number
WO2021008018A1
WO2021008018A1 PCT/CN2019/116552 CN2019116552W WO2021008018A1 WO 2021008018 A1 WO2021008018 A1 WO 2021008018A1 CN 2019116552 W CN2019116552 W CN 2019116552W WO 2021008018 A1 WO2021008018 A1 WO 2021008018A1
Authority
WO
WIPO (PCT)
Prior art keywords
vehicle
artificial intelligence
recognition model
images
data set
Prior art date
Application number
PCT/CN2019/116552
Other languages
English (en)
French (fr)
Inventor
戴广宇
王晶晶
Original Assignee
平安科技（深圳）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 平安科技（深圳）有限公司 filed Critical 平安科技（深圳）有限公司
Publication of WO2021008018A1 publication Critical patent/WO2021008018A1/zh

Links

Images

Classifications

    • 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
    • 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
    • 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
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/08—Detecting or categorising vehicles

Definitions

  • This application relates to the field of computer technology, and in particular to an artificial intelligence-based vehicle identification method, device, program, and computer-readable storage medium.
  • This application provides an artificial intelligence-based vehicle identification method, device, program, and computer-readable storage medium, the main purpose of which is to provide an automatic identification solution for vehicle types.
  • an artificial intelligence-based vehicle identification method includes:
  • a vehicle image mark the vehicle image with a category label, store the vehicle image with the category label, establish a vehicle image data set, and perform preprocessing operations on the images in the vehicle image data set, wherein the vehicle
  • the images include vehicle images obtained from the Internet using a web crawler program and images of vehicles entering and exiting vehicles captured in a preset area using a camera;
  • Obtain the vehicle image to be recognized use the trained vehicle recognition model to recognize the vehicle type in the vehicle image to be recognized, and use one-hot encoding to encode and output the vehicle category to be recognized.
  • the present application also provides an artificial intelligence-based vehicle identification device, which includes a memory and a processor, and the memory stores an artificial intelligence-based vehicle identification that can run on the processor.
  • a program when the artificial intelligence-based vehicle recognition program is executed by the processor, the following steps are implemented:
  • a vehicle image mark the vehicle image with a category label, store the vehicle image with the category label, establish a vehicle image data set, and perform preprocessing operations on the images in the vehicle image data set, wherein the vehicle
  • the images include vehicle images obtained from the Internet using a web crawler program and images of vehicles entering and exiting vehicles captured in a preset area using a camera;
  • Construct a vehicle recognition model of a deep residual network and use the vehicle image data set to train the vehicle recognition model to recognize the type of vehicle in the image, and in the vehicle recognition model training, use a random inactivation method to enhance the vehicle
  • the generalization ability of the recognition model uses a batch standardization algorithm in the vehicle recognition model training process so that the input of each layer of neural network maintains the same distribution.
  • this application also provides an artificial intelligence-based vehicle recognition program
  • the artificial intelligence-based vehicle recognition program includes:
  • Data set establishment module used to obtain vehicle images, label the vehicle images with category tags, store the vehicle images with category tags, establish a vehicle image data set, and preprocess the images in the vehicle image data set operating;
  • Model training module used to construct a vehicle recognition model, train the vehicle recognition model using the vehicle image data set, and use the vehicle recognition model to recognize the type of vehicle in the image;
  • Vehicle category recognition module used to obtain the vehicle image to be recognized, use the above-trained vehicle recognition model to recognize the vehicle type in the vehicle image to be recognized, and use one-hot encoding to encode and output the vehicle category .
  • the present application also provides a computer-readable storage medium that stores an artificial intelligence-based vehicle recognition program.
  • the artificial intelligence-based vehicle recognition program can be used by one or A plurality of processors execute to implement the steps of the artificial intelligence-based vehicle identification method as described above.
  • the artificial intelligence-based vehicle identification method, device, program, and computer-readable storage medium proposed in this application acquire vehicle images, mark the vehicle images with category tags, store the vehicle images with category tags, and establish a vehicle image data set , And perform preprocessing operations on the images in the vehicle image data set; construct a vehicle recognition model, and use the vehicle image data set to train the vehicle recognition model, and use the vehicle recognition model to recognize the type of vehicle in the image; use
  • the camera obtains the vehicle image to be recognized, uses the trained vehicle recognition model to recognize the vehicle type in the vehicle image to be recognized, and uses one-hot encoding to encode and output the vehicle category. Therefore, the application can automatically identify the vehicle category, so that different fees can be charged according to different models.
  • FIG. 1 is a schematic flowchart of an artificial intelligence-based vehicle identification method provided by an embodiment of the application
  • FIG. 2 is a schematic diagram of a median filter algorithm in an artificial intelligence-based vehicle identification method provided by an embodiment of the application;
  • FIG. 3 is a schematic diagram of the internal structure of an artificial intelligence-based vehicle identification device provided by an embodiment of the application;
  • FIG. 4 is a schematic diagram of modules of an artificial intelligence-based vehicle recognition program provided by an embodiment of the application.
  • This application provides a vehicle identification method based on artificial intelligence.
  • FIG. 1 it is a schematic flowchart of an artificial intelligence-based vehicle recognition method provided by an embodiment of this application.
  • the method can be executed by a device, and the device can be implemented by software and/or hardware.
  • the vehicle identification method based on artificial intelligence includes:
  • the preferred embodiment of the present application uses a web crawler program to obtain vehicle images from the Internet.
  • Web crawlers also known as web spiders, web robots, are programs or scripts that automatically crawl Internet information according to certain rules.
  • a preferred embodiment of the present application proposes a method for obtaining images of various types of vehicles from the Internet using a crawler program , And manually mark each vehicle image with a category label to generate a basic vehicle image data set.
  • a camera may also be used to obtain images of vehicles entering and exiting a preset area, such as a parking lot, a highway intersection, etc.
  • this application is for cameras, such as parking lots.
  • the actual vehicle picture acquired by the camera also uses manual labeling to label the image of the vehicle category.
  • This application combines the basic vehicle image data set and the vehicle image data set in the actual application obtained by the camera into a vehicle image data set.
  • the advantage of this method is that, first, it is guaranteed that the deep residual network has enough training set for training to achieve convergence. Second, because the vehicle images in the actual application scenarios of parking lots and high-speed entrances are added to the basic vehicle image data set, the distribution of the training set and the test set are more similar, which will have a positive impact on the classification effect of the deep residual network model .
  • the preprocessing operation includes:
  • Median filtering is a non-linear digital filter technology used to remove noise in images or other signals.
  • the main idea of median filtering is to check the samples in the input signal and determine whether it represents the signal, and use an observation window composed of an odd number of samples to achieve this function. The values in the observation window are sorted, and the median value in the middle of the observation window is output. Then, the oldest value is discarded, a new sample is obtained, and the above calculation process is repeated.
  • Median filtering is especially useful for speckle noise and salt and pepper noise in images.
  • the specific operation process of the median filtering is shown in FIG. 2.
  • the filter has a size of N*N, and N is a positive integer.
  • the filter with a size of 3*3 is used in FIG. 2.
  • the median filter method will use preset pixels, such as a pixel with a pixel value of 1, as the original center value of the filter, and for all pixel values in a 3*3 window (ie, in Figure 2 (Shaded area) for sorting, and replace the original center value of the filter with the sorted intermediate value in the N*N size filter window.
  • the pixel value of 1 is replaced by the pixel value of 23, and the result after median filtering is shown in the right part of FIG. 2.
  • the method of acquiring the image of the vehicle to be identified may be different from the vehicle image data set.
  • the shooting direction/position/angle has changed, which will cause the sample distribution in the training set and the test set to change.
  • this case performs data enhancement processing on the samples in the data set.
  • the content of the data enhancement includes: horizontal flipping, vertical flipping, rotation transformation and affine transformation processing of the vehicle image .
  • the data enhancement method described in this application can enhance the robustness of the deep residual network.
  • the preferred embodiment of the present application uses a deep residual network (ResNet) as a vehicle recognition model to classify and recognize vehicle types.
  • ResNet deep residual network
  • ResNet The general idea of the ResNet model is to represent the layer as a learning residual function according to the input. Experiments show that deep residual networks are easier to optimize, and can improve accuracy by increasing considerable depth.
  • the core of ResNet is to solve the side effect of increasing depth, which is the degradation of deep neural networks. This can improve network performance by simply increasing the depth of the network.
  • arbitrary functions can be fitted, so some layers can be used to fit functions in ResNet.
  • the preferred embodiment of the present application uses the vehicle image data set to train the vehicle recognition model, it further includes: using dropout technology to enhance the generalization ability of the recognition model, and/or using Batch Normalization to speed up Model training speed and reduce the distribution difference between the training set and the data set.
  • Dropout means that in the training process of the deep residual network, the neural network unit is temporarily dropped from the network according to a certain probability. For stochastic gradient descent, because it is randomly discarded, each mini-batch is essentially training a different network.
  • the dropout technology can effectively prevent the deep neural network from overfitting and improve the generalization performance of the recognition model. In the iterative process of each batch of batches, dropout will force a neural unit to work with other randomly selected neural units. This eliminates the joint adaptability between neuron nodes and improves the generalization performance of the model.
  • this application introduces Batch Normalization technology in the deep residual network.
  • the recognition model is based on a very important assumption that the samples in the training set and the test set are independent and identically distributed.
  • This embodiment uses the Batch Normalization technology to solve these problems. Batch Normalization is to keep the input of each layer of neural network in the same distribution during the training of the deep residual network.
  • the reason for the slow training convergence is generally that the overall distribution gradually moves to a nonlinear function
  • the upper and lower limits of the value interval are close to each other, so this causes the gradient of the underlying neural network to disappear during back propagation.
  • This is the essential reason for the slower and slower convergence of the training deep residual network, and Batch Normalization is through a certain standardization method . To force the distribution of the input value of any neuron in each layer of the neural network to a standard normal distribution with a mean of 0 and a variance of 1, essentially constraining the increasingly skewed distribution back to a more standard distribution. This makes the activation input value fall in an area where the nonlinear function is more sensitive to the input, so a small change in the input will lead to a larger change in the loss function.
  • the preferred embodiment of the present application uses the softmax regression function as the output layer of the deep residual network to output the network model recognition result.
  • the preferred embodiment of the present application uses the softmax layer as the output layer of the deep residual network to output the vehicle recognition result.
  • Softmax has a very wide range of applications in machine learning and deep learning. Especially in dealing with multi-classification problems, the final output unit of the classifier requires the softmax function for numerical processing.
  • V i is the output of the previous stage classifier output unit.
  • i represents the category index
  • the total number of categories is C
  • e is an infinite non-cyclic decimal
  • S i represents the ratio of the index of the current element to the sum of all element indices.
  • Softmax converts the output values of multiple categories into relative probabilities, making it easier to understand and compare. The output of softmax characterizes the relative probability between different categories. In the vehicle identification task, assuming that the probability value corresponding to the first dimension is the largest, it means that the vehicle to be identified is more likely to belong to the first type of vehicle. Softmax converts continuous values into relative probabilities, making the selection of results more intuitive.
  • the one-hot encoding is also called one-bit effective encoding, which mainly uses N-bit status registers to encode N states, each state has its own independent register bit, and only one bit is valid at any time.
  • One-hot encoding is the representation of category variables as binary vectors. This method requires mapping the classification labels to integer values. Then, each integer value is represented as a binary vector, except that the index position of the integer is 1, and the remaining positions are 0. For vehicle identification tasks. Suppose we have two vehicle labels,'AA Brand A Series' and'BB Brand B Series', we can assign the integer value of'AA Brand A Series' to 0 and the integer value of'BB Brand B Series' to 1 .
  • the current allocation method is called integer encoding.
  • the ‘AA Brand A Series’ coded as 0 will be represented by a binary vector [1,0], where the 0th index is marked as 1.
  • the ‘BB Brand B Series’ label coded as 1 will be represented by a binary vector [0,1], where the first index is marked as 1.
  • One-hot encoding is more accurate for data classification, because deep neural networks cannot be directly used for data classification.
  • the categories of the data must be converted into numbers, and the input and output variables for the classification are the same. Integer encoding is only applicable when there is a comparative relationship between category tags, such as ‘cold’, ‘warm’, and ‘hot’. But for the task of vehicle recognition, since there is no comparative relationship between categories, in this task, it is only suitable for using one-hot encoding.
  • this application also includes S4. According to the type of the vehicle, a preset fee rule is used to calculate and output the corresponding fee standard.
  • the application also provides a vehicle identification device based on artificial intelligence.
  • FIG. 3 it is a schematic diagram of the internal structure of an artificial intelligence-based vehicle identification device provided by an embodiment of this application.
  • the vehicle identification device 1 based on artificial intelligence may be a PC (Personal Computer, personal computer), or a terminal device such as a smart phone, a tablet computer, or a portable computer.
  • the artificial intelligence-based vehicle identification device 1 at least includes a memory 11, a processor 12, a communication bus 13, and a network interface 14.
  • the memory 11 includes at least one type of readable storage medium, and the readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory, etc.), magnetic memory, magnetic disk, optical disk, etc.
  • the memory 11 may be an internal storage unit of the artificial intelligence-based vehicle identification device 1, for example, the hard disk of the artificial intelligence-based vehicle identification device 1.
  • the memory 11 may also be an external storage device of the vehicle identification device 1 based on artificial intelligence, such as a plug-in hard disk equipped on the vehicle identification device 1 based on artificial intelligence, or a smart media card (SMC). ), Secure Digital (SD) card, Flash Card, etc.
  • the memory 11 may also include both an internal storage unit of the vehicle identification device 1 based on artificial intelligence and an external storage device.
  • the memory 11 can be used not only to store application software and various data installed in the artificial intelligence-based vehicle identification device 1, such as the code of the artificial intelligence-based vehicle identification program 01, etc., but also to temporarily store the output or to be output The data.
  • the processor 12 may be a central processing unit (CPU), controller, microcontroller, microprocessor or other data processing chip in some embodiments, and is used to run the program code or processing stored in the memory 11 Data, such as execution of vehicle recognition program 01 based on artificial intelligence.
  • CPU central processing unit
  • controller microcontroller
  • microprocessor or other data processing chip in some embodiments, and is used to run the program code or processing stored in the memory 11 Data, such as execution of vehicle recognition program 01 based on artificial intelligence.
  • the communication bus 13 is used to realize the connection and communication between these components.
  • the network interface 14 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface), and is usually used to establish a communication connection between the device 1 and other electronic devices.
  • the device 1 may also include a user interface.
  • the user interface may include a display (Display) and an input unit such as a keyboard (Keyboard).
  • the optional user interface may also include a standard wired interface and a wireless interface.
  • the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode, organic light emitting diode) touch device, etc.
  • the display can also be appropriately called a display screen or a display unit, which is used to display the information processed in the artificial intelligence-based vehicle identification device 1 and to display a visualized user interface.
  • Figure 3 only shows the artificial intelligence-based vehicle identification device 1 with components 11-14 and the artificial intelligence-based vehicle identification program 01. Those skilled in the art can understand that the structure shown in Figure 1 does not constitute a The limitation of the artificial intelligence vehicle identification device 1 may include fewer or more components than shown, or a combination of certain components, or a different component arrangement.
  • the memory 11 stores an artificial intelligence-based vehicle recognition program 01; when the processor 12 executes the artificial intelligence-based vehicle recognition program 01 stored in the memory 11, the following steps are implemented:
  • Step 1 Obtain a vehicle image, mark the vehicle image with a category label, store the vehicle image with the category label, establish a vehicle image data set, and perform preprocessing operations on the images in the vehicle image data set.
  • the preferred embodiment of the present application uses a web crawler program to obtain vehicle images from the Internet.
  • Web crawlers also known as web spiders, web robots, are programs or scripts that automatically crawl Internet information according to certain rules.
  • a preferred embodiment of the present application proposes a method for obtaining images of various types of vehicles from the Internet using a crawler program , And manually mark each vehicle image with a category label to generate a basic vehicle image data set.
  • a camera may also be used to obtain images of vehicles entering and exiting a preset area, such as a parking lot, a highway intersection, etc.
  • this application is for cameras, such as parking lots.
  • the actual vehicle picture acquired by the camera also uses manual labeling to label the image of the vehicle category.
  • This application combines the basic vehicle image data set and the vehicle image data set in the actual application obtained by the camera into a vehicle image data set.
  • the advantage of this method is that, first, it is guaranteed that the deep residual network has enough training set for training to achieve convergence. Second, because the vehicle images in the actual application scenarios of parking lots and high-speed entrances are added to the basic vehicle image data set, the distribution of the training set and the test set are more similar, which will have a positive impact on the classification effect of the deep residual network model .
  • the preprocessing operation includes:
  • Median filtering is a non-linear digital filter technology used to remove noise in images or other signals.
  • the main idea of median filtering is to check the samples in the input signal and determine whether it represents the signal, and use an observation window composed of an odd number of samples to achieve this function. The values in the observation window are sorted, and the median value in the middle of the observation window is output. Then, the oldest value is discarded, a new sample is obtained, and the above calculation process is repeated.
  • Median filtering is especially useful for speckle noise and salt and pepper noise in images.
  • the specific operation process of the median filtering is shown in FIG. 2.
  • the filter has a size of N*N, and N is a positive integer.
  • the filter with a size of 3*3 is used in FIG. 2.
  • the median filter method will use preset pixels, such as a pixel with a pixel value of 1, as the original center value of the filter, and for all pixel values in a 3*3 window (ie, in Figure 2 (Shaded area) for sorting, and replace the original center value of the filter with the sorted intermediate value in the N*N size filter window.
  • the pixel value of 1 is replaced by the pixel value of 23, and the result after median filtering is shown in the right part of FIG. 2.
  • the method of acquiring the image of the vehicle to be identified may be different from that of the vehicle image data set.
  • the shooting direction/position/angle has changed, which will cause the sample distribution in the training set and the test set to change.
  • this case performs data enhancement processing on the samples in the data set.
  • the content of the data enhancement includes: horizontal flipping, vertical flipping, rotation transformation and affine transformation processing of the vehicle image .
  • the data enhancement method described in this application can enhance the robustness of the deep residual network.
  • Step 2 Construct a vehicle recognition model, train the vehicle recognition model using the vehicle image data set, and use the vehicle recognition model to recognize the type of vehicle in the image.
  • the preferred embodiment of the present application uses a deep residual network (ResNet) as a vehicle recognition model to classify and recognize vehicle types.
  • ResNet deep residual network
  • ResNet The general idea of the ResNet model is to represent the layer as a learning residual function according to the input. Experiments show that deep residual networks are easier to optimize, and can improve accuracy by increasing considerable depth.
  • the core of ResNet is to solve the side effect of increasing depth, which is the degradation of deep neural networks. This can improve network performance by simply increasing the depth of the network.
  • arbitrary functions can be fitted, so some layers can be used to fit functions in ResNet.
  • the preferred embodiment of the present application uses the vehicle image data set to train the vehicle recognition model, it further includes: using dropout technology to enhance the generalization ability of the recognition model, and/or using Batch Normalization to speed up Model training speed and reduce the distribution difference between the training set and the data set.
  • Dropout means that in the training process of the deep residual network, the neural network unit is temporarily dropped from the network according to a certain probability. For stochastic gradient descent. Since it is randomly discarded, each mini-batch is essentially training a different network.
  • the dropout technology can effectively prevent the deep neural network from overfitting and improve the generalization performance of the recognition model. In the iterative process of each batch of batches, dropout will force a neural unit to work with other randomly selected neural units. This eliminates the joint adaptability between neuron nodes and improves the generalization performance of the model.
  • this application introduces Batch Normalization technology in the deep residual network.
  • the recognition model is based on a very important assumption that the samples in the training set and the test set are independent and identically distributed.
  • This embodiment uses the Batch Normalization technology to solve these problems. Batch Normalization is to keep the input of each layer of neural network in the same distribution during the training of the deep residual network.
  • the reason for the slow training convergence is generally that the overall distribution gradually moves to a nonlinear function
  • the upper and lower limits of the value interval are close to each other, so this causes the gradient of the underlying neural network to disappear during back propagation.
  • This is the essential reason for the slower and slower convergence of the training deep residual network, and Batch Normalization is through a certain standardization method . To force the distribution of the input value of any neuron in each layer of the neural network to a standard normal distribution with a mean of 0 and a variance of 1, essentially constraining the increasingly skewed distribution back to a more standard distribution. This makes the activation input value fall in an area where the nonlinear function is more sensitive to the input, so a small change in the input will lead to a larger change in the loss function.
  • the preferred embodiment of the present application uses the softmax regression function as the output layer of the deep residual network to output the network model recognition result.
  • the preferred embodiment of the present application uses the softmax layer as the output layer of the deep residual network to output the vehicle recognition result.
  • Softmax has a very wide range of applications in machine learning and deep learning. Especially in dealing with multi-classification problems, the final output unit of the classifier requires the softmax function for numerical processing.
  • V i is the output of the previous stage classifier output unit.
  • i represents the category index
  • S i represents the ratio of the index of the current element to the sum of the indices of all elements.
  • Softmax converts the output values of multiple categories into relative probabilities, making it easier to understand and compare. The output of softmax characterizes the relative probability between different categories. In the vehicle identification task, assuming that the probability value corresponding to the first dimension is the largest, it means that the vehicle to be identified is more likely to belong to the first type of vehicle. Softmax converts continuous values into relative probabilities, making the selection of results more intuitive.
  • Step 3 Use the camera to obtain the image of the vehicle to be recognized, use the above-mentioned trained vehicle recognition model to recognize the type of vehicle in the image, and use the one-hot encoding to encode and output the vehicle category.
  • the one-hot encoding is also called one-bit effective encoding, which mainly uses N-bit status registers to encode N states, each state has its own independent register bit, and only one bit is valid at any time.
  • One-hot encoding is the representation of category variables as binary vectors. This method requires mapping the classification labels to integer values. Then, each integer value is represented as a binary vector, except that the index position of the integer is 1, and the remaining positions are 0. For vehicle identification tasks. Suppose we have two vehicle labels,'AA Brand A Series' and'BB Brand B Series', we can assign the integer value of'AA Brand A Series' to 0 and the integer value of'BB Brand B Series' to 1 .
  • the current allocation method is called integer encoding.
  • the ‘AA Brand A Series’ coded as 0 will be represented by a binary vector [1,0], where the 0th index is marked as 1.
  • the ‘BB Brand B Series’ label coded as 1 will be represented by a binary vector [0,1], where the first index is marked as 1.
  • One-hot encoding is more accurate for data classification, because deep neural networks cannot be directly used for data classification.
  • the categories of the data must be converted into numbers, and the input and output variables for the classification are the same. Integer encoding is only applicable when there is a comparative relationship between category tags, such as ‘cold’, ‘warm’, and ‘hot’. But for the task of vehicle recognition, since there is no comparative relationship between categories, in this task, it is only suitable for using one-hot encoding.
  • this application also includes step four.
  • the preset fee rule is used to calculate and output the corresponding fee standard.
  • the artificial intelligence-based vehicle recognition program can also be divided into one or more modules, and the one or more modules are stored in the memory 11 and run by one or more processors (this The embodiment is executed by the processor 12) to complete this application.
  • the module referred to in this application refers to a series of computer program instruction segments that can complete specific functions, and is used to describe the artificial intelligence-based vehicle recognition program in the artificial intelligence-based vehicle Identify the execution process in the device.
  • FIG. 4 is a schematic diagram of program modules of an embodiment of the artificial intelligence-based vehicle recognition program of this application.
  • the artificial intelligence-based vehicle recognition program can be divided into a data set establishment module 10 and model training.
  • the data set establishment module 10 is used to: obtain vehicle images, mark the vehicle images with category tags, store the vehicle images with category tags, establish a vehicle image data set, and preprocess the images in the vehicle image data set Processing operation.
  • the preprocessing operation on the images in the vehicle image data set includes:
  • the noise reduction processing includes:
  • the original center value of the filter is replaced with the sorted intermediate value in the N*N filter window.
  • the model training module 20 is configured to construct a vehicle recognition model, train the vehicle recognition model using the vehicle image data set, and use the vehicle recognition model to recognize the type of vehicle in the image.
  • the training of the vehicle recognition model using the vehicle image data set further includes:
  • a regression function is used as the output layer of the vehicle recognition model, and the recognition result of the vehicle recognition model is output.
  • the vehicle category recognition module 30 is used to obtain a vehicle image to be recognized by using a camera, recognize the vehicle type in the vehicle image to be recognized in the above-mentioned trained vehicle recognition model, and use one-hot encoding for the vehicle category Perform encoding output.
  • the artificial intelligence-based vehicle recognition program 01 further includes: a cost calculation module 40, exemplarily:
  • the fee calculation module 40 is used for calculating and outputting corresponding charging standards by using preset fee rules according to the type of the vehicle.
  • an embodiment of the present application also proposes a computer-readable storage medium that stores an artificial intelligence-based vehicle recognition program, and the artificial intelligence-based vehicle recognition program can be processed by one or more Executed to achieve the following operations:
  • a vehicle image mark the vehicle image with a category label, store the vehicle image with the category label, establish a vehicle image data set, and perform preprocessing operations on the images in the vehicle image data set, wherein the vehicle
  • the images include vehicle images obtained from the Internet using a web crawler program and images of vehicles entering and exiting vehicles captured in a preset area using a camera;
  • Construct a vehicle recognition model of a deep residual network and use the vehicle image data set to train the vehicle recognition model to recognize the type of vehicle in the image, and in the vehicle recognition model training, use a random inactivation method to enhance the vehicle
  • To identify the generalization ability of the recognition model use batch standardization algorithms in the vehicle recognition model training process, so that the input of each layer of neural network maintains the same distribution; to obtain the vehicle image to be recognized, use the above-mentioned trained vehicle recognition model to recognize The vehicle type in the vehicle image to be recognized, and the one-hot encoding is used to encode and output the vehicle type to be recognized.

Landscapes

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

Abstract

一种基于人工智能的车辆识别方法、装置、程序及计算机可读存储介质，涉及人工智能技术领域。所述方法包括：获取车辆图像，对所述车辆图像标记类别标签，存储所述带有类别标签的车辆图像，建立车辆图像数据集，并对所述车辆图像数据集中的图像进行预处理操作（S1）；构建车辆识别模型，并利用所述车辆图像数据集训练所述车辆识别模型，使用所述车辆识别模型识别图像中车辆的种类（S2）；利用摄像头获取待识别的车辆图像，使用上述已训练好的车辆识别模型中识别所述待识别的车辆图像中的车辆种类，并使用独热编码对所述车辆类别进行编码输出（S3）。自动识别车辆类别，从而实现按照不同车型收取不同的费用。

Description

基于人工智能的车辆识别方法、装置、程序及存储介质
本申请基于巴黎公约申明享有2019年7月18日递交的申请号为CN201910651863.8、名称为“基于人工智能的车辆识别方法、装置及存储介质”的中国专利申请的优先权，该中国专利申请的整体内容以参考的方式结合在本申请中。
技术领域
本申请涉及计算机技术领域，尤其涉及一种基于人工智能的车辆识别方法、装置、程序及计算机可读存储介质。
背景技术
随着社会经济的发展，机动车保有量迅速增长，居民乘小汽车出行的次数增多，许多城市“停车难，难停车”的问题日渐显著，供需紧张、乱收费、违章停车、进出停车场不易、寻找停车场困难等现象普遍存在。为了城市交通的更好发展以及更好的利用有限的停车位资源，很多城市逐渐开始实行按照不同车型收取不同的停车费用的收费方法。对于绝大部分停车场来说，每日的车辆出入量巨大，单凭人工方式判断车辆类型极易发生混乱。
发明内容
本申请提供一种基于人工智能的车辆识别方法、装置、程序及计算机可读存储介质，其主要目的在于提供一种车辆种类的自动识别方案。
为实现上述目的，本申请提供的一种基于人工智能的车辆识别方法，包括：
获取车辆图像，对所述车辆图像标记类别标签，存储所述带有类别标签的车辆图像，建立车辆图像数据集，并对所述车辆图像数据集中的图像进行预处理操作，其中，所述车辆图像包括使用网络爬虫程序从互联网中获取的车辆图像以及使用摄像机在预设区域捕获的出入车辆的车辆图像；
构建深度残差网络的车辆识别模型，并利用所述车辆图像数据集训练所述车辆识别模型识别图像中车辆的种类，并在所述车辆识别模型训练中，使 用随机失活方法增强所述车辆识别模型的泛化能力，使用批标准化算法在所述车辆识别模型训练过程，使得每一层神经网络的输入保持相同分布；
获取待识别的车辆图像，使用上述已训练好的车辆识别模型识别所述待识别的车辆图像中的车辆种类，并使用独热编码对所述待识别的车辆类别进行编码输出。
此外，为实现上述目的，本申请还提供一种基于人工智能的车辆识别装置，该装置包括存储器和处理器，所述存储器中存储有可在所述处理器上运行的基于人工智能的车辆识别程序，所述基于人工智能的车辆识别程序被所述处理器执行时实现如下步骤：
获取车辆图像，对所述车辆图像标记类别标签，存储所述带有类别标签的车辆图像，建立车辆图像数据集，并对所述车辆图像数据集中的图像进行预处理操作，其中，所述车辆图像包括使用网络爬虫程序从互联网中获取的车辆图像以及使用摄像机在预设区域捕获的出入车辆的车辆图像；
构建深度残差网络的车辆识别模型，并利用所述车辆图像数据集训练所述车辆识别模型识别图像中车辆的种类，并在所述车辆识别模型训练中，使用随机失活方法增强所述车辆识别模型的泛化能力，使用批标准化算法在所述车辆识别模型训练过程，使得每一层神经网络的输入保持相同分布。
此外，为实现上述目的，本申请还提供一种基于人工智能的车辆识别程序，所述基于人工智能的车辆识别程序包括：
数据集建立模块：用于获取车辆图像，对所述车辆图像标记类别标签，存储所述带有类别标签的车辆图像，建立车辆图像数据集，并对所述车辆图像数据集中的图像进行预处理操作；
模型训练模块：用于构建车辆识别模型，并利用所述车辆图像数据集训练所述车辆识别模型，使用所述车辆识别模型识别图像中车辆的种类；及
车辆类别识别模块：用于获取待识别的车辆图像，使用上述已训练好的车辆识别模型中识别所述待识别的车辆图像中的车辆种类，并使用独热编码对所述车辆类别进行编码输出。
此外，为实现上述目的，本申请还提供一种计算机可读存储介质，所述计算机可读存储介质上存储有基于人工智能的车辆识别程序，所述基于人工智能的车辆识别程序可被一个或者多个处理器执行，以实现如上所述的基于 人工智能的车辆识别方法的步骤。
本申请提出的基于人工智能的车辆识别方法、装置、程序及计算机可读存储介质获取车辆图像，对所述车辆图像标记类别标签，存储所述带有类别标签的车辆图像，建立车辆图像数据集，并对所述车辆图像数据集中的图像进行预处理操作；构建车辆识别模型，并利用所述车辆图像数据集训练所述车辆识别模型，使用所述车辆识别模型识别图像中车辆的种类；利用摄像头获取待识别的车辆图像，使用上述已训练好的车辆识别模型中识别所述待识别的车辆图像中的车辆种类，并使用独热编码对所述车辆类别进行编码输出。因此，本申请可以自动识别车辆类别，从而可以实现按照不同车型收取不同的费用。
附图说明
图1为本申请一实施例提供的基于人工智能的车辆识别方法的流程示意图；
图2为本申请一实施例提供的基于人工智能的车辆识别方法中中值滤波算法的示意图；
图3为本申请一实施例提供的基于人工智能的车辆识别装置的内部结构示意图；
图4为本申请一实施例提供的基于人工智能的车辆识别程序的模块示意图。
本申请目的的实现、功能特点及优点将结合实施例，参照附图做进一步说明。
具体实施方式
为了使本申请的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本申请进行进一步详细说明。应当理解，此处所描述的具体实施例仅用以解释本申请，并不用于限定本申请。基于本申请中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本申请保护的范围。
本申请的说明书和权利要求书及上述附图中的术语“第一”、“第二”、“第三”、“第四”等(如果存在)是用于区别类似的对象，而不必用于描述特定的顺序或先后次序。应该理解这样使用的数据在适当情况下可以互换，以便这里描述的实施例能够以除了在这里图示或描述的内容以外的顺序实施。此外，所述“第一”、“第二”等的描述仅用于描述目的，而不能理解为指示或暗示其相对重要性或者隐含指明所指示的技术特征的数量。由此，限定有“第一”、“第二”的特征可以明示或者隐含地包括至少一个该特征。
进一步地，术语“包括”和“具有”以及他们的任何变形，意图在于覆盖不排他的包含，例如，包含了一系列步骤或单元的过程、方法、***、产品或设备不必限于清楚地列出的那些步骤或单元，而是可包括没有清楚地列出的或对于这些过程、方法、产品或设备固有的其它步骤或单元。
另外，各个实施例之间的技术方案可以相互结合，但是必须是以本领域普通技术人员能够实现为基础，当技术方案的结合出现相互矛盾或无法实现时应当认为这种技术方案的结合不存在，也不在本申请要求的保护范围之内。
本申请提供一种基于人工智能的车辆识别方法。参照图1所示，为本申请一实施例提供的基于人工智能的车辆识别方法的流程示意图。该方法可以由一个装置执行，该装置可以由软件和/或硬件实现。
在本实施例中，基于人工智能的车辆识别方法包括：
S1、获取车辆图像，对所述车辆图像标记类别标签，存储所述带有类别标签的车辆图像，建立车辆图像数据集，并对所述车辆图像数据集中的图像进行预处理操作。
本申请较佳实施例使用网络爬虫程序从互联网中获取车辆图像。
网络爬虫，又被称为网页蜘蛛，网络机器人，是一种按照一定的规则，自动地抓取互联网信息的程序或者脚本。
由于训练深度残差网络需要大量有标记的样本，而且对各类型车辆进行图像采集会耗费大量人力物力，所以本申请较佳实施例提出一种使用爬虫程序从互联网上获取各类型车辆图像的方法，并通过人工标记的方式对每幅车辆图像标记类别标签，生成基础车辆图像数据集。
进一步地，本申请较佳实施例还可以使用摄像机获取预设区域，如停车场、高速路口等，出入车辆的车辆图像。
在机器学习领域，只有当训练集与测试集的样本服从同一分布时，识别模型才会达到较好的分类效果，因此，在基础车辆图像数据集的基础上，本申请对于摄像头，例如停车场摄像头获取的实际的车辆图片，同样使用人工标注的方式对图像进行车辆类别的标注。
本申请将基础车辆图像数据集与摄像头获取的实际应用中的车辆图像数据集合并成一个车辆图像数据集。这种方法的优势在于，第一，保证深度残差网络有足够的训练集进行训练可以达到收敛。第二，由于在基础车辆图像数据集中加入了停车场、高速出入口实际应用场景中的车辆图像，使得训练集与测试集的分布更加相似，这将对深度残差网络模型的分类效果产生正面影响。
由于互联网中的图片质量良莠不齐，因此需要对车辆图像数据集的车辆图像进行预处理操作。
本申请较佳实施例中，所述预处理操作包括：
(1)使用中值滤波算法对车辆图像进行降噪处理。
中值滤波是一种非线性数字滤波器技术，用于去除图像或者其他信号中的噪声。中值滤波的主要思想是检查输入信号中的采样并判断它是否代表了信号，使用奇数个采样组成的观察窗实现这项功能。观察窗口中的数值进行排序，位于观察窗中间的的中值作为输出，然后，丢弃最早的值，取得新的采样，重复上面的计算过程。中值滤波对于图像中的斑点噪声和椒盐噪声来说尤其有用。
本申请较佳实施例中，所述中值滤波的具体操作过程如图2所示。其中，所述滤波器为N*N大小，N为正整数，图2中采用的是3*3大小的滤波器。如图2所示，中值滤波法将以预设像素点，如像素值为1的像素点，为滤波器原始中心值，对3*3大小的窗内的所有像素值(即图2中阴影区域)进行排序，用所述N*N大小的滤波器窗内排序后的中间值代替所述滤波器原始中心值。在发明实施例中，值为1的像素值被像素值23代替，中值滤波后的结果如图2的右侧部分所示。
(2)使用翻转旋转方法对车辆图像进行数据增强处理。
实际应用中待识别车辆图像获取的方式可能与车辆图像数据集中不同，比如说拍摄的方向/位置/角度等发生了变化，这会引起训练集与测试集中的样 本分布发生变化。为了提高深度残差网络模型的识别准确性，本案对数据集中的样本进行数据增强处理，所述数据增强的内容包括：将所述车辆图像进行水平翻转，垂直翻转，旋转变换与仿射变换处理。
本申请所述数据增强方法可以增强深度残差网络的鲁棒性。
S2、构建车辆识别模型，并利用所述车辆图像数据集训练所述车辆识别模型，使用所述车辆识别模型识别图像中车辆的种类。
本申请较佳实施例使用深度残差网络(ResNet)作为车辆识别模型，进行车辆种类的分类识别。
所述ResNet模型的总体思想是根据输入将层表示为学习残差函数。实验表明，深度残差网络更容易优化，并且能够通过增加相当的深度来提高准确率。ResNet的核心是解决了增加深度带来的副作用，也就是深度神经网络的退化问题。这样能够通过单纯地增加网络的深度来提高网络性能。根据多层的神经网络理论上可以拟合任意函数，那么在ResNet中可以利用一些层来拟合函数。
可选地，本申请较佳实施例在利用所述车辆图像数据集训练所述车辆识别模型时还包括：使用随机失活(dropout)技术增强识别模型泛化能力，及/或使用Batch Normalization加快模型训练速度以及缩小训练集与数据集之间的分布差异。
本申请使用dropout技术来增加深度残差网络的泛化性能。dropout是指在深度残差网络的训练过程中，对神经网络单元按照一定的概率将其暂时从网络中丢弃。对于随机梯度下降来说，由于是随机丢弃，故而每一个mini-batch实质上都在训练一个不同的网络。dropout技术可以有效地防止深度神经网络发生过拟合，提高识别模型的泛化性能。在每一轮batch的迭代过程中dropout都会强迫一个神经单元和随机挑选出来的其他神经单元共同工作。这消除了神经元节点之间的联合适应性，使得模型的泛化性能提升。
进一步地，本申请在深度残差网络中引入了批标准化(Batch Normalization)技术。对于车辆识别任务，识别模型基于一条非常重要的假设，那就是训练集与测试集中的样本是独立同分布的。但在车辆识别的实际应用中，保证训练集与测试集严格的服从同一分布是比较困难的。除此之外，由于训练深度残差网络需要耗费的时间较多。本实施例使用Batch Normalization技术来解决 这些问题。Batch Normalization就是在深度残差网络训练过程中使得每一层神经网络的输入保持相同分布的。因为深度残差网络在做非线性变换前的激活输入值随着网络深度加深或者在训练过程中，其分布逐渐发生偏移或者变动，之所以训练收敛慢，一般是整体分布逐渐往非线性函数的取值区间的上下限两端靠近，所以这导致反向传播时底层神经网络的梯度消失，这是训练深度残差网络收敛越来越慢的本质原因，而Batch Normalization就是通过一定的规范化手段，把每层神经网络任意神经元的输入值的分布强行约束到均值为0方差为1的标准正态分布，实质上是将越来越偏的分布约束回比较标准的分布。这使得激活输入值落在非线性函数对输入比较敏感的区域，这样输入的小变化就会导致损失函数较大的变化。
进一步地，本申请较佳实施例使用softmax回归函数作为深度残差网络的输出层输出网络模型识别结果。
本申请较佳实施例使用softmax层作为深度残差网络的输出层输出车辆识别结果。Softmax在机器学习和深度学习中有着非常广泛的应用。尤其是在处理多分类问题方面，分类器最后的输出单元需要softmax函数进行数值处理。
关于softmax函数的定义如下所示：
Figure PCTCN2019116552-appb-000001
其中，V i是分类器前级输出单元的输出。i表示类别索引，总的类别个数为C，e为无限不循环小数。S i表示当前元素的指数与所有元素指数和的比值。softmax将多分类的输出数值转化为相对概率，更容易理解与比较。softmax的输出表征了不同类别之间的相对概率。在车辆识别任务当中，假设第一维对应的概率值最大，则代表待识别车辆属于第一类车辆类型的可能性更大。softmax将连续数值转化为相对概率使得结果选取更加直观。
S3、利用摄像头获取待识别的车辆图像，使用上述已训练好的车辆识别模型识别图像中车辆的种类，并使用独热编码对车辆类别进行编码输出。
所述独热编码又称一位有效编码，主要是采用N位状态寄存器来对N个状态进行编码，每个状态都有其独立的寄存器位，并且在任意时候只有一位有效。独热编码是类目变量作为二进制向量的表示。该方法要求将分类标签映射到整数值。然后，每个整数值被表示为二进制向量，除了整数的索引位置为1外，其余的位置都为0。对于车辆识别任务来说。假设我们有‘AA品牌 A系’和‘BB品牌B系’这两种车辆标签，我们可以将‘AA品牌A系’的整数值分配为0，‘BB品牌B系’的整数值分配为1。当前的分配方式叫做整数编码。那么根据整数编码我们可以得到车辆的类别标签。编码为0的‘AA品牌A系’将用二进制向量[1,0]表示，其中第0个索引被标记为1。编码为1的‘BB品牌B系’标签将用一个二进制向量[0,1]表示，其中第一个索引被标记为1。独热编码进行数据的分类更加准确，由于深度神经网络无法直接用于数据分类。数据的类别必须转换成数字，对于分类的输入和输出变量都是一样的。整数编码只适用于类别标签之间存在比较关系的情况中，比如说‘冷’，‘温暖’，‘热’。但是对于车辆识别任务来说，由于类别之间没有比较关系的存在，所以在这个任务中，只适用于使用独热编码。
进一步地，本申请还包括S4、根据所述车辆的种类，利用预设的费用规则，计算并输出对应的收费标准。
本申请还提供一种基于人工智能的车辆识别装置。参照图3所示，为本申请一实施例提供的基于人工智能的车辆识别装置的内部结构示意图。
在本实施例中，基于人工智能的车辆识别装置1可以是PC(Personal Computer，个人电脑)，也可以是智能手机、平板电脑、便携计算机等终端设备。该基于人工智能的车辆识别装置1至少包括存储器11、处理器12，通信总线13，以及网络接口14。
其中，存储器11至少包括一种类型的可读存储介质，所述可读存储介质包括闪存、硬盘、多媒体卡、卡型存储器(例如，SD或DX存储器等)、磁性存储器、磁盘、光盘等。存储器11在一些实施例中可以是基于人工智能的车辆识别装置1的内部存储单元，例如该基于人工智能的车辆识别装置1的硬盘。存储器11在另一些实施例中也可以是基于人工智能的车辆识别装置1的外部存储设备，例如基于人工智能的车辆识别装置1上配备的插接式硬盘，智能存储卡(Smart Media Card,SMC)，安全数字(Secure Digital,SD)卡，闪存卡(Flash Card)等。进一步地，存储器11还可以既包括基于人工智能的车辆识别装置1的内部存储单元也包括外部存储设备。存储器11不仅可以用于存储安装于基于人工智能的车辆识别装置1的应用软件及各类数据，例如基于人工智能的车辆识别程序01的代码等，还可以用于暂时地存储已经输出 或者将要输出的数据。
处理器12在一些实施例中可以是一中央处理器(Central Processing Unit,CPU)、控制器、微控制器、微处理器或其他数据处理芯片，用于运行存储器11中存储的程序代码或处理数据，例如执行基于人工智能的车辆识别程序01等。
通信总线13用于实现这些组件之间的连接通信。
网络接口14可选的可以包括标准的有线接口、无线接口(如WI-FI接口)，通常用于在该装置1与其他电子设备之间建立通信连接。
可选地，该装置1还可以包括用户接口，用户接口可以包括显示器(Display)、输入单元比如键盘(Keyboard)，可选的用户接口还可以包括标准的有线接口、无线接口。可选地，在一些实施例中，显示器可以是LED显示器、液晶显示器、触控式液晶显示器以及OLED(Organic Light-Emitting Diode，有机发光二极管)触摸器等。其中，显示器也可以适当的称为显示屏或显示单元，用于显示在基于人工智能的车辆识别装置1中处理的信息以及用于显示可视化的用户界面。
图3仅示出了具有组件11-14以及基于人工智能的车辆识别程序01的基于人工智能的车辆识别装置1，本领域技术人员可以理解的是，图1示出的结构并不构成对基于人工智能的车辆识别装置1的限定，可以包括比图示更少或者更多的部件，或者组合某些部件，或者不同的部件布置。
在图3所示的装置1实施例中，存储器11中存储有基于人工智能的车辆识别程序01；处理器12执行存储器11中存储的基于人工智能的车辆识别程序01时实现如下步骤：
步骤一、获取车辆图像，对所述车辆图像标记类别标签，存储所述带有类别标签的车辆图像，建立车辆图像数据集，并对所述车辆图像数据集中的图像进行预处理操作。
本申请较佳实施例使用网络爬虫程序从互联网中获取车辆图像。
网络爬虫，又被称为网页蜘蛛，网络机器人，是一种按照一定的规则，自动地抓取互联网信息的程序或者脚本。
由于训练深度残差网络需要大量有标记的样本，而且对各类型车辆进行图像采集会耗费大量人力物力，所以本申请较佳实施例提出一种使用爬虫程 序从互联网上获取各类型车辆图像的方法，并通过人工标记的方式对每幅车辆图像标记类别标签，生成基础车辆图像数据集。
进一步地，本申请较佳实施例还可以使用摄像机获取预设区域，如停车场、高速路口等，出入车辆的车辆图像。
在机器学习领域，只有当训练集与测试集的样本服从同一分布时，识别模型才会达到较好的分类效果，因此，在基础车辆图像数据集的基础上，本申请对于摄像头，例如停车场摄像头获取的实际的车辆图片，同样使用人工标注的方式对图像进行车辆类别的标注。
本申请将基础车辆图像数据集与摄像头获取的实际应用中的车辆图像数据集合并成一个车辆图像数据集。这种方法的优势在于，第一，保证深度残差网络有足够的训练集进行训练可以达到收敛。第二，由于在基础车辆图像数据集中加入了停车场、高速出入口实际应用场景中的车辆图像，使得训练集与测试集的分布更加相似，这将对深度残差网络模型的分类效果产生正面影响。
由于互联网中的图片质量良莠不齐，因此需要对车辆图像数据集的车辆图像进行预处理操作。
本申请较佳实施例中，所述预处理操作包括：
(2)使用中值滤波算法对车辆图像进行降噪处理。
中值滤波是一种非线性数字滤波器技术，用于去除图像或者其他信号中的噪声。中值滤波的主要思想是检查输入信号中的采样并判断它是否代表了信号，使用奇数个采样组成的观察窗实现这项功能。观察窗口中的数值进行排序，位于观察窗中间的的中值作为输出，然后，丢弃最早的值，取得新的采样，重复上面的计算过程。中值滤波对于图像中的斑点噪声和椒盐噪声来说尤其有用。
本申请较佳实施例中，所述中值滤波的具体操作过程如图2所示。其中，所述滤波器为N*N大小，N为正整数，图2中采用的是3*3大小的滤波器。如图2所示，中值滤波法将以预设像素点，如像素值为1的像素点，为滤波器原始中心值，对3*3大小的窗内的所有像素值(即图2中阴影区域)进行排序，用所述N*N大小的滤波器窗内排序后的中间值代替所述滤波器原始中心值。在发明实施例中，值为1的像素值被像素值23代替，中值滤波后的结 果如图2的右侧部分所示。
(2)使用翻转旋转方法对车辆图像进行数据增强处理。
实际应用中待识别车辆图像获取的方式可能与车辆图像数据集中不同，比如说拍摄的方向/位置/角度等发生了变化，这会引起训练集与测试集中的样本分布发生变化。为了提高深度残差网络模型的识别准确性，本案对数据集中的样本进行数据增强处理，所述数据增强的内容包括：将所述车辆图像进行水平翻转，垂直翻转，旋转变换与仿射变换处理。
本申请所述数据增强方法可以增强深度残差网络的鲁棒性。
步骤二、构建车辆识别模型，并利用所述车辆图像数据集训练所述车辆识别模型，使用所述车辆识别模型识别图像中车辆的种类。
本申请较佳实施例使用深度残差网络(ResNet)作为车辆识别模型，进行车辆种类的分类识别。
所述ResNet模型的总体思想是根据输入将层表示为学习残差函数。实验表明，深度残差网络更容易优化，并且能够通过增加相当的深度来提高准确率。ResNet的核心是解决了增加深度带来的副作用，也就是深度神经网络的退化问题。这样能够通过单纯地增加网络的深度来提高网络性能。根据多层的神经网络理论上可以拟合任意函数，那么在ResNet中可以利用一些层来拟合函数。
可选地，本申请较佳实施例在利用所述车辆图像数据集训练所述车辆识别模型时还包括：使用随机失活(dropout)技术增强识别模型泛化能力，及/或使用Batch Normalization加快模型训练速度以及缩小训练集与数据集之间的分布差异。
本申请使用dropout技术来增加深度残差网络的泛化性能。dropout是指在深度残差网络的训练过程中，对神经网络单元按照一定的概率将其暂时从网络中丢弃。对于随机梯度下降来说。由于是随机丢弃，故而每一个mini-batch实质上都在训练一个不同的网络。dropout技术可以有效地防止深度神经网络发生过拟合，提高识别模型的泛化性能。在每一轮batch的迭代过程中dropout都会强迫一个神经单元和随机挑选出来的其他神经单元共同工作。这消除了神经元节点之间的联合适应性，使得模型的泛化性能提升。
进一步地，本申请在深度残差网络中引入了批标准化(Batch Normalization) 技术。对于车辆识别任务，识别模型基于一条非常重要的假设，那就是训练集与测试集中的样本是独立同分布的。但在车辆识别的实际应用中，保证训练集与测试集严格的服从同一分布是比较困难的。除此之外，由于训练深度残差网络需要耗费的时间较多。本实施例使用Batch Normalization技术来解决这些问题。Batch Normalization就是在深度残差网络训练过程中使得每一层神经网络的输入保持相同分布的。因为深度残差网络在做非线性变换前的激活输入值随着网络深度加深或者在训练过程中，其分布逐渐发生偏移或者变动，之所以训练收敛慢，一般是整体分布逐渐往非线性函数的取值区间的上下限两端靠近，所以这导致反向传播时底层神经网络的梯度消失，这是训练深度残差网络收敛越来越慢的本质原因，而Batch Normalization就是通过一定的规范化手段，把每层神经网络任意神经元的输入值的分布强行约束到均值为0方差为1的标准正态分布，实质上是将越来越偏的分布约束回比较标准的分布。这使得激活输入值落在非线性函数对输入比较敏感的区域，这样输入的小变化就会导致损失函数较大的变化。
进一步地，本申请较佳实施例使用softmax回归函数作为深度残差网络的输出层输出网络模型识别结果。
本申请较佳实施例使用softmax层作为深度残差网络的输出层输出车辆识别结果。Softmax在机器学习和深度学习中有着非常广泛的应用。尤其是在处理多分类问题方面，分类器最后的输出单元需要softmax函数进行数值处理。
关于softmax函数的定义如下所示：
Figure PCTCN2019116552-appb-000002
其中，V i是分类器前级输出单元的输出。i表示类别索引，总的类别个数为C。S i表示的是当前元素的指数与所有元素指数和的比值。softmax将多分类的输出数值转化为相对概率，更容易理解与比较。softmax的输出表征了不同类别之间的相对概率。在车辆识别任务当中，假设第一维对应的概率值最大，则代表待识别车辆属于第一类车辆类型的可能性更大。softmax将连续数值转化为相对概率使得结果选取更加直观。
步骤三、利用摄像头获取待识别的车辆图像，使用上述已训练好的车辆识别模型识别图像中车辆的种类，并使用独热编码对车辆类别进行编码输出。
所述独热编码又称一位有效编码，主要是采用N位状态寄存器来对N个 状态进行编码，每个状态都有其独立的寄存器位，并且在任意时候只有一位有效。独热编码是类目变量作为二进制向量的表示。该方法要求将分类标签映射到整数值。然后，每个整数值被表示为二进制向量，除了整数的索引位置为1外，其余的位置都为0。对于车辆识别任务来说。假设我们有‘AA品牌A系’和‘BB品牌B系’这两种车辆标签，我们可以将‘AA品牌A系’的整数值分配为0，‘BB品牌B系’的整数值分配为1。当前的分配方式叫做整数编码。那么根据整数编码我们可以得到车辆的类别标签。编码为0的‘AA品牌A系’将用二进制向量[1,0]表示，其中第0个索引被标记为1。编码为1的‘BB品牌B系’标签将用一个二进制向量[0,1]表示，其中第一个索引被标记为1。独热编码进行数据的分类更加准确，由于深度神经网络无法直接用于数据分类。数据的类别必须转换成数字，对于分类的输入和输出变量都是一样的。整数编码只适用于类别标签之间存在比较关系的情况中，比如说‘冷’，‘温暖’，‘热’。但是对于车辆识别任务来说，由于类别之间没有比较关系的存在，所以在这个任务中，只适用于使用独热编码。
进一步地，本申请还包括步骤四、根据所述车辆的种类，利用预设的费用规则，计算并输出对应的收费标准。
可选地，在其他实施例中，基于人工智能的车辆识别程序还可以被分割为一个或者多个模块，一个或者多个模块被存储于存储器11中，并由一个或多个处理器(本实施例为处理器12)所执行以完成本申请，本申请所称的模块是指能够完成特定功能的一系列计算机程序指令段，用于描述基于人工智能的车辆识别程序在基于人工智能的车辆识别装置中的执行过程。
例如，参照图4所示，为本申请基于人工智能的车辆识别程序一实施例的程序模块示意图，该实施例中，基于人工智能的车辆识别程序可以被分割为数据集建立模块10、模型训练模块20、车辆类别识别模块30，示例性地：
数据集建立模块10用于：获取车辆图像，对所述车辆图像标记类别标签，存储所述带有类别标签的车辆图像，建立车辆图像数据集，并对所述车辆图像数据集中的图像进行预处理操作。
可选地，所述对所述车辆图像数据集中的图像进行预处理操作包括：
使用中值滤波算法对车辆图像进行降噪处理；
使用翻转旋转方法对车辆图像进行数据增强处理。
可选地，所述降噪处理包括：
取N*N大小的滤波器，其中，N为正整数；
以预设像素点为滤波器原始中心值，对N*N大小的滤波器窗内的所有像素值进行排序；
用所述N*N大小的滤波器窗内排序后的中间值代替所述滤波器原始中心值。
模型训练模块20用于：构建车辆识别模型，并利用所述车辆图像数据集训练所述车辆识别模型，使用所述车辆识别模型识别图像中车辆的种类。
可选地，所述利用所述车辆图像数据集训练所述车辆识别模型，还包括：
使用随机失活技术增强所述车辆识别模型的泛化能力；
使用Batch Normalization技术在所述车辆识别模型训练过程，使得每一层神经网络的输入保持相同分布；及
使用回归函数作为所述车辆识别模型的输出层，输出所述车辆识别模型的识别结果。
车辆类别识别模块30用于：利用摄像头获取待识别的车辆图像，使用上述已训练好的车辆识别模型中识别所述待识别的车辆图像中的车辆种类，并使用独热编码对所述车辆类别进行编码输出。
进一步地，所述基于人工智能的车辆识别程序01还包括：费用计算模块40，示例性地：
费用计算模块40用于：根据所述车辆的种类，利用预设的费用规则，计算并输出对应的收费标准。
上述数据集建立模块10、模型训练模块20、车辆类别识别模块30、费用计算模块40等程序模块被执行时所实现的功能或操作步骤与上述实施例大体相同，在此不再赘述。
此外，本申请实施例还提出一种计算机可读存储介质，所述计算机可读存储介质上存储有基于人工智能的车辆识别程序，所述基于人工智能的车辆识别程序可被一个或多个处理器执行，以实现如下操作：
获取车辆图像，对所述车辆图像标记类别标签，存储所述带有类别标签 的车辆图像，建立车辆图像数据集，并对所述车辆图像数据集中的图像进行预处理操作，其中，所述车辆图像包括使用网络爬虫程序从互联网中获取的车辆图像以及使用摄像机在预设区域捕获的出入车辆的车辆图像；
构建深度残差网络的车辆识别模型，并利用所述车辆图像数据集训练所述车辆识别模型识别图像中车辆的种类，并在所述车辆识别模型训练中，使用随机失活方法增强所述车辆识别模型的泛化能力，使用批标准化算法在所述车辆识别模型训练过程，使得每一层神经网络的输入保持相同分布；获取待识别的车辆图像，使用上述已训练好的车辆识别模型识别所述待识别的车辆图像中的车辆种类，并使用独热编码对所述待识别的车辆类别进行编码输出。
本申请计算机可读存储介质具体实施方式与上述基于人工智能的车辆识别装置和方法各实施例基本相同，在此不作累述。
需要说明的是，上述本申请实施例序号仅仅为了描述，不代表实施例的优劣。并且本文中的术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、装置、物品或者方法不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、装置、物品或者方法所固有的要素。在没有更多限制的情况下，由语句“包括一个......”限定的要素，并不排除在包括该要素的过程、装置、物品或者方法中还存在另外的相同要素。
通过以上的实施方式的描述，本领域的技术人员可以清楚地了解到上述实施例方法可借助软件加必需的通用硬件平台的方式来实现，当然也可以通过硬件，但很多情况下前者是更佳的实施方式。基于这样的理解，本申请的技术方案本质上或者说对现有技术做出贡献的部分可以以软件产品的形式体现出来，该计算机软件产品存储在如上所述的一个存储介质(如ROM/RAM、磁碟、光盘)中，包括若干指令用以使得一台终端设备(可以是手机，计算机，服务器，或者网络设备等)执行本申请各个实施例所述的方法。
以上仅为本申请的优选实施例，并非因此限制本申请的专利范围，凡是利用本申请说明书及附图内容所作的等效结构或等效流程变换，或直接或间接运用在其他相关的技术领域，均同理包括在本申请的专利保护范围内。

Claims (20)

  1. 一种基于人工智能的车辆识别方法，其特征在于，所述方法包括：
    获取车辆图像，对所述车辆图像标记类别标签，存储所述带有类别标签的车辆图像，建立车辆图像数据集，并对所述车辆图像数据集中的图像进行预处理操作，其中，所述车辆图像包括使用网络爬虫程序从互联网中获取的车辆图像以及使用摄像机在预设区域捕获的出入车辆的车辆图像；
    构建深度残差网络的车辆识别模型，并利用所述车辆图像数据集训练所述车辆识别模型识别图像中车辆的种类，并在所述车辆识别模型训练中，使用随机失活方法增强所述车辆识别模型的泛化能力，使用批标准化算法在所述车辆识别模型训练过程，使得每一层神经网络的输入保持相同分布；及
    获取待识别的车辆图像，使用上述已训练好的车辆识别模型识别所述待识别的车辆图像中的车辆种类，并使用独热编码对所述待识别的车辆类别进行编码输出。
  2. 如权利要求1所述的基于人工智能的车辆识别方法，其特征在于，所述对所述车辆图像数据集中的图像进行预处理操作包括：
    使用中值滤波算法对车辆图像进行降噪处理；及
    使用翻转旋转方法对车辆图像进行数据增强处理。
  3. 如权利要求2所述的基于人工智能的车辆识别方法，其特征在于，所述降噪处理包括：
    取N*N大小的滤波器，其中，N为正整数；
    以预设像素点为滤波器原始中心值，对N*N大小的滤波器窗内的所有像素值进行排序；及
    用所述N*N大小的滤波器窗内排序后的中间值代替所述滤波器原始中心值。
  4. 如权利要求1所述的基于人工智能的车辆识别方法，其特征在于，所述车辆识别模型的输出层使用回归函数输出所述车辆识别模型的识别结果，其中，所述函数的定义如下所示：
    Figure PCTCN2019116552-appb-100001
    其中，V i是分类器前级输出单元的输出，i表示类别索引，e为无限不循 环小数，C为总的类别个数，S i表示当前元素的指数与所有元素指数和的比值。
  5. 如权利要求1至4中任意一项所述的基于人工智能的车辆识别方法，其特征在于，该方法还包括：
    根据所述车辆的种类，利用预设的费用规则，计算并输出对应的收费标准。
  6. 一种基于人工智能的车辆识别装置，其特征在于，所述装置包括存储器和处理器，所述存储器上存储有可在所述处理器上运行的基于人工智能的车辆识别程序，所述基于人工智能的车辆识别程序被所述处理器执行时实现如下步骤：
    获取车辆图像，对所述车辆图像标记类别标签，存储所述带有类别标签的车辆图像，建立车辆图像数据集，并对所述车辆图像数据集中的图像进行预处理操作，其中，所述车辆图像包括使用网络爬虫程序从互联网中获取的车辆图像以及使用摄像机在预设区域捕获的出入车辆的车辆图像；
    构建深度残差网络的车辆识别模型，并利用所述车辆图像数据集训练所述车辆识别模型识别图像中车辆的种类，并在所述车辆识别模型训练中，使用随机失活方法增强所述车辆识别模型的泛化能力，使用批标准化算法在所述车辆识别模型训练过程，使得每一层神经网络的输入保持相同分布。
  7. 如权利要求6所述的基于人工智能的车辆识别装置，其特征在于，所述对所述车辆图像数据集中的图像进行预处理操作包括：
    使用中值滤波算法对车辆图像进行降噪处理；及
    使用翻转旋转方法对车辆图像进行数据增强处理。
  8. 如权利要求7所述的基于人工智能的车辆识别装置，其特征在于，所述降噪处理包括：
    取N*N大小的滤波器，其中，N为正整数；
    以预设像素点为滤波器原始中心值，对N*N大小的滤波器窗内的所有像素值进行排序；及
    用所述N*N大小的滤波器窗内排序后的中间值代替所述滤波器原始中心值。
  9. 如权利要求6所述的基于人工智能的车辆识别装置，其特征在于，所述基于人工智能的车辆识别程序被所述处理器执行时实现如下步骤：获取待 识别的车辆图像，使用上述已训练好的车辆识别模型识别所述待识别的车辆图像中的车辆种类，并使用独热编码对所述待识别的车辆类别进行编码输出。
  10. 如权利要求6所述的基于人工智能的车辆识别装置，其特征在于，所述车辆识别模型的输出层使用回归函数输出所述车辆识别模型的识别结果，其中，所述函数的定义如下所示：
    Figure PCTCN2019116552-appb-100002
    其中，V i是分类器前级输出单元的输出，i表示类别索引，e为无限不循环小数，C为总的类别个数，S i表示当前元素的指数与所有元素指数和的比值。
  11. 如权利要求6至10中任意一项所述的基于人工智能的车辆识别装置，其特征在于，所述基于人工智能的车辆识别程序被所述处理器执行时实现如下步骤：根据所述车辆的种类，利用预设的费用规则，计算并输出对应的收费标准。
  12. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质上存储有基于人工智能的车辆识别程序，所述基于人工智能的车辆识别程序可被一个或者多个处理器执行，以实现如下步骤：
    获取车辆图像，对所述车辆图像标记类别标签，存储所述带有类别标签的车辆图像，建立车辆图像数据集，并对所述车辆图像数据集中的图像进行预处理操作，其中，所述车辆图像包括使用网络爬虫程序从互联网中获取的车辆图像以及使用摄像机在预设区域捕获的出入车辆的车辆图像；
    构建深度残差网络的车辆识别模型，并利用所述车辆图像数据集训练所述车辆识别模型识别图像中车辆的种类，并在所述车辆识别模型训练中，使用随机失活方法增强所述车辆识别模型的泛化能力，使用批标准化算法在所述车辆识别模型训练过程，使得每一层神经网络的输入保持相同分布；及
    获取待识别的车辆图像，使用上述已训练好的车辆识别模型识别所述待识别的车辆图像中的车辆种类，并使用独热编码对所述待识别的车辆类别进行编码输出。
  13. 如权利要求12所述的计算机可读存储介质，其特征在于，所述对所述车辆图像数据集中的图像进行预处理操作包括：
    使用中值滤波算法对车辆图像进行降噪处理；及
    使用翻转旋转方法对车辆图像进行数据增强处理。
  14. 如权利要求13所述的计算机可读存储介质，其特征在于，所述降噪处理包括：
    取N*N大小的滤波器，其中，N为正整数；
    以预设像素点为滤波器原始中心值，对N*N大小的滤波器窗内的所有像素值进行排序；及
    用所述N*N大小的滤波器窗内排序后的中间值代替所述滤波器原始中心值。
  15. 如权利要求12所述的计算机可读存储介质，其特征在于，所述车辆识别模型的输出层使用回归函数输出所述车辆识别模型的识别结果，其中，所述函数的定义如下所示：
    Figure PCTCN2019116552-appb-100003
    其中，V i是分类器前级输出单元的输出，i表示类别索引，e为无限不循环小数，C为总的类别个数，S i表示当前元素的指数与所有元素指数和的比值。
  16. 如权利要求12至15中任意一项所述的计算机可读存储介质，其特征在于，所述基于人工智能的车辆识别程序被所述处理器执行时实现如下步骤：根据所述车辆的种类，利用预设的费用规则，计算并输出对应的收费标准。
  17. 一种基于人工智能的车辆识别程序，其特征在于，所述基于人工智能的车辆识别程序包括：
    数据集建立模块：用于获取车辆图像，对所述车辆图像标记类别标签，存储所述带有类别标签的车辆图像，建立车辆图像数据集，并对所述车辆图像数据集中的图像进行预处理操作；
    模型训练模块：用于构建车辆识别模型，并利用所述车辆图像数据集训练所述车辆识别模型，使用所述车辆识别模型识别图像中车辆的种类；及
    车辆类别识别模块：用于获取待识别的车辆图像，使用上述已训练好的车辆识别模型中识别所述待识别的车辆图像中的车辆种类，并使用独热编码对所述车辆类别进行编码输出。
  18. 如权利要求17所述的基于人工智能的车辆识别程序，其特征在于，所述对所述车辆图像数据集中的图像进行预处理操作包括：
    使用中值滤波算法对车辆图像进行降噪处理；及
    使用翻转旋转方法对车辆图像进行数据增强处理；
    其中，所述降噪处理包括：
    取N*N大小的滤波器，其中，N为正整数；
    以预设像素点为滤波器原始中心值，对N*N大小的滤波器窗内的所有像素值进行排序；及
    用所述N*N大小的滤波器窗内排序后的中间值代替所述滤波器原始中心值。
  19. 如权利要求17所述的基于人工智能的车辆识别程序，其特征在于，所述车辆识别模型的输出层使用回归函数输出所述车辆识别模型的识别结果，其中，所述函数的定义如下所示：
    Figure PCTCN2019116552-appb-100004
    其中，V i是分类器前级输出单元的输出，i表示类别索引，e为无限不循环小数，C为总的类别个数，S i表示当前元素的指数与所有元素指数和的比值。
  20. 如权利要求17至19中任意一项所述的基于人工智能的车辆识别程序，其特征在于，所述基于人工智能的车辆识别程序还包括：
    费用计算模块：用于根据所述车辆的种类，利用预设的费用规则，计算并输出对应的收费标准。
PCT/CN2019/116552 2019-07-18 2019-11-08 基于人工智能的车辆识别方法、装置、程序及存储介质 WO2021008018A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910651863.8 2019-07-18
CN201910651863.8A CN110516703A (zh) 2019-07-18 2019-07-18 基于人工智能的车辆识别方法、装置及存储介质

Publications (1)

Publication Number Publication Date
WO2021008018A1 true WO2021008018A1 (zh) 2021-01-21

Family

ID=68622843

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/116552 WO2021008018A1 (zh) 2019-07-18 2019-11-08 基于人工智能的车辆识别方法、装置、程序及存储介质

Country Status (2)

Country Link
CN (1) CN110516703A (zh)
WO (1) WO2021008018A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113822292A (zh) * 2021-02-05 2021-12-21 深圳信息职业技术学院 车辆特征信息存储方法、装置、计算机设备及存储介质
CN115601593A (zh) * 2022-10-18 2023-01-13 苏州魔视智能科技有限公司（Cn） 一种图像分类方法、装置、设备及介质

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109740735B (zh) 2018-12-29 2020-12-29 百度在线网络技术（北京）有限公司 多神经网络输出方法及装置、服务器、计算机可读介质
CN111126224A (zh) * 2019-12-17 2020-05-08 成都通甲优博科技有限责任公司 车辆检测方法及分类识别模型训练方法
CN111145365A (zh) * 2019-12-17 2020-05-12 北京明略软件***有限公司 一种实现分类处理的方法、装置、计算机存储介质及终端
KR102208690B1 (ko) * 2020-07-23 2021-01-28 주식회사 어반베이스 데이터 증강 기반 스타일 분석 모델 학습 장치 및 방법
KR102208685B1 (ko) * 2020-07-23 2021-01-28 주식회사 어반베이스 데이터 증강 기반 공간 분석 모델 학습 장치 및 방법
CN112566117B (zh) * 2020-11-06 2023-12-08 厦门大学 基于度量学习的车辆节点身份识别方法及装置
CN112598084B (zh) * 2021-03-02 2021-06-29 深圳金三立视频科技股份有限公司 一种基于图像处理的车辆类型识别方法及终端
CN113343881A (zh) * 2021-06-21 2021-09-03 浪潮云信息技术股份公司 基于深度学习的车辆品牌型号细粒度分类***及方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108171134A (zh) * 2017-12-20 2018-06-15 中车工业研究院有限公司 一种操作动作辨识方法及装置
US20180253622A1 (en) * 2017-03-06 2018-09-06 Honda Motor Co., Ltd. Systems for performing semantic segmentation and methods thereof
CN108875803A (zh) * 2018-05-30 2018-11-23 长安大学 一种基于视频图像的危化品运输车辆检测与识别方法
CN109190444A (zh) * 2018-07-02 2019-01-11 南京大学 一种基于视频的收费车道车辆特征识别***的实现方法

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10423861B2 (en) * 2017-10-16 2019-09-24 Illumina, Inc. Deep learning-based techniques for training deep convolutional neural networks
CN109063719B (zh) * 2018-04-23 2022-02-01 湖北工业大学 一种联合结构相似性和类信息的图像分类方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180253622A1 (en) * 2017-03-06 2018-09-06 Honda Motor Co., Ltd. Systems for performing semantic segmentation and methods thereof
CN108171134A (zh) * 2017-12-20 2018-06-15 中车工业研究院有限公司 一种操作动作辨识方法及装置
CN108875803A (zh) * 2018-05-30 2018-11-23 长安大学 一种基于视频图像的危化品运输车辆检测与识别方法
CN109190444A (zh) * 2018-07-02 2019-01-11 南京大学 一种基于视频的收费车道车辆特征识别***的实现方法

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113822292A (zh) * 2021-02-05 2021-12-21 深圳信息职业技术学院 车辆特征信息存储方法、装置、计算机设备及存储介质
CN115601593A (zh) * 2022-10-18 2023-01-13 苏州魔视智能科技有限公司（Cn） 一种图像分类方法、装置、设备及介质

Also Published As

Publication number Publication date
CN110516703A (zh) 2019-11-29

Similar Documents

Publication Publication Date Title
WO2021008018A1 (zh) 基于人工智能的车辆识别方法、装置、程序及存储介质
WO2021238281A1 (zh) 一种神经网络的训练方法、图像分类***及相关设备
CN106845487B (zh) 一种端到端的车牌识别方法
CN111476309A (zh) 图像处理方法、模型训练方法、装置、设备及可读介质
CN108009543A (zh) 一种车牌识别方法及装置
WO2022179587A1 (zh) 一种特征提取的方法以及装置
CN113239924A (zh) 一种基于迁移学习的弱监督目标检测方法及***
CN110826429A (zh) 一种基于景区视频的旅游突发事件自动监测的方法及***
Ye et al. A two-stage real-time YOLOv2-based road marking detector with lightweight spatial transformation-invariant classification
CN107545263A (zh) 一种物体检测方法及装置
CN112749653A (zh) 行人检测方法、装置、电子设备及存储介质
CN112069884A (zh) 一种暴力视频分类方法、***和存储介质
CN114169381A (zh) 图像标注方法、装置、终端设备及存储介质
Fan Research and realization of video target detection system based on deep learning
Wang et al. Deep leaning-based ultra-fast stair detection
Rehman et al. An efficient approach for vehicle number plate recognition in Pakistan
Boppana et al. Comparative analysis of single-stage yolo algorithms for vehicle detection under extreme weather conditions
CN117218622A (zh) 路况检测方法、电子设备及存储介质
Chan Deep learning-based scene-awareness approach for intelligent change detection in videos
CN112330697B (zh) 一种图像分割方法、装置、电子设备和可读存储介质
Zhang et al. A hybrid convolutional architecture for accurate image manipulation localization at the pixel-level
CN116680580A (zh) 基于多模态训练的信息匹配方法、装置、电子设备及介质
CN115240163A (zh) 一种基于一阶段检测网络的交通标志检测方法及***
CN112580505A (zh) 网点开关门状态识别方法、装置、电子设备及存储介质
Wang et al. Multi-branch detection network based on trigger attention for pedestrian detection under occlusion

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: 19937442

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: 19937442

Country of ref document: EP

Kind code of ref document: A1