Attribute information determination method and apparatus, storage medium, and electronic device

Abstract

Embodiments of the present application provide a kind of attribute information determination method, device, storage medium and electronic device, wherein the method comprises: using the target network model of training target image to determine the target attribute information of target object included in target image;Wherein, target network model is obtained by training initial network model using multiple sets of training data in the following way: by analyzing the training image included in each set of training data through initial network model, to determine the predicted position information of training object included in training image;Determine the predicted polar coordinate parameter based on predicted position information;Determine the label polar coordinate parameter based on the label position information included in the label information of training object included in each set of training data;Determine the position loss value of initial network model based on predicted polar coordinate parameter and label polar coordinate parameter, update the network parameters of initial network model based on position loss value, obtain target network model.

Description

TECHNICAL FIELD

[0001] Embodiments of the present application relate to the field of communications, and in particular, to a method and apparatus for determining attribute information, a storage medium, and an electronic device. BACKGROUND

[0002] When the camera equipment is installed at a high position, or there is a certain angle with the object being photographed, the object in the image obtained by the photographing is usually not directly facing the camera equipment, and the object in the image usually has a certain angle. The following takes a motor vehicle as an example for illustration:

[0003] The rapid growth of the number of vehicles has brought great pressure to the traffic management of cities, and also caused various traffic problems to occur frequently. In order to regulate the parking order and relieve the pressure on the traffic management of cities caused by random parking, the traffic management department will set some parking spaces on the roadside to facilitate the parking of motor vehicles. In order to uniformly manage the parking behavior of motor vehicles in the delineated parking area, monitoring equipment will be set on the same side or different side of the area, and the information of vehicles entering and leaving the parking space will be monitored and managed from real-time video and images. Due to the limitation of height and angle, the license plates of vehicles in the captured scene will have various rotating angles.

[0004] In the related art, a general license plate detection scheme is to generate a rectangular region parallel to the obtained license plate image to locate the position of the license plate. However, in the roadside parking scene, due to the limitation of the installation position and height of the monitoring equipment, the license plates of many vehicles in the scene have a certain angle with the currently captured scene. The traditional detection cannot well locate the real position of the license plate with a large angle, so that the extracted features contain more noise, which is not conducive to the subsequent recognition of license plate characters. The license plate image detected in the related art is shown in FIG. 1. Figure 1 .

[0005] Therefore, the related art has the problem of low accuracy of recognizing attribute information of an object.

[0006] In view of the above problems in the related art, no effective solution has been proposed so far. SUMMARY

[0007] Embodiments of the present application provide a method and apparatus for determining attribute information, a storage medium, and an electronic device, to at least solve the problem of low accuracy of recognizing attribute information of an object in the related art.

[0008] According to one embodiment of the present application, a method for determining attribute information is provided, comprising: using a target image to determine target attribute information of a target object included in the target image by using a trained target network model; wherein the target network model is obtained by training an initial network model using a plurality of sets of training data in the following manner: analyzing a training image included in each set of the training data by using the initial network model to determine predicted position information of a training object included in the training image; determining predicted polar coordinate parameters based on the predicted position information; determining label polar coordinate parameters based on label position information included in label information of the training object included in each set of the training data; determining a position loss value of the initial network model based on the predicted polar coordinate parameters and the label polar coordinate parameters, updating network parameters of the initial network model based on the position loss value to obtain the target network model.

[0009] According to another embodiment of the present application, a device for determining attribute information is provided, comprising: an analysis module configured to analyze a target image to determine target attribute information of a target object included in the target image by using a trained target network model; wherein the target network model is obtained by training an initial network model using a plurality of sets of training data in the following manner: analyzing a training image included in each set of the training data by using the initial network model to determine predicted position information of a training object included in the training image; determining predicted polar coordinate parameters based on the predicted position information; determining label polar coordinate parameters based on label position information included in label information of the training object included in each set of the training data; determining a position loss value of the initial network model based on the predicted polar coordinate parameters and the label polar coordinate parameters, updating network parameters of the initial network model based on the position loss value to obtain the target network model.

[0010] According to still another embodiment of the present application, a computer readable storage medium is provided, wherein the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the method described in any of the above embodiments.

[0011] According to still another embodiment of the present application, an electronic device is provided, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to execute the computer program to perform the steps of any of the above method embodiments.

[0012] By the present application, a target image is analyzed using a trained target network model to determine target attribute information of a target object included in the target image. The target network model is a model obtained by training an initial network model using a plurality of sets of training data. Each set of training data includes a training image, and the initial network model is used to analyze the training image to determine predicted position information of a training object included in the training image. The predicted position information is used to determine predicted polar coordinate parameters. Label position information included in label information of the training object included in each set of training data is used to determine label polar coordinate parameters. The position loss value of the initial network model is determined based on the predicted polar coordinate parameters and the label polar coordinate parameters. The network parameters of the initial network model are updated based on the position loss value to obtain the target network model. Since the target network model is obtained by training the initial network model and converging using polar coordinate parameters, the convergence training using polar coordinate parameters can make the distribution of parameters consistent, and the target attribute information determined by the target network model is more accurate. Therefore, the problem of low accuracy of attribute information of an identified object in related technologies is solved, and the accuracy of attribute information of an identified object is improved. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 is a schematic diagram of a license plate image detected in related technologies;

[0014] Figure 2 is a hardware structure block diagram of a mobile terminal of an attribute information determination method according to an embodiment of the present application;

[0015] Figure 3 is a flowchart of an attribute information determination method according to an embodiment of the present application;

[0016] Figure 4 is a schematic diagram of an image obtained according to an exemplary embodiment of the present application;

[0017] Figure 5 is a schematic diagram of a target network model detection process according to an exemplary embodiment of the present application;

[0018] Figure 6 is a schematic diagram of a target object image determined by a target network model according to an exemplary embodiment of the present application;

[0019] Figure 7 is a schematic diagram of polar coordinate parameters according to an exemplary embodiment of the present application;

[0020] Figure 8 is a schematic diagram of the difference between a predicted target and a real target position under polar coordinate parameters according to an exemplary embodiment of the present application;

[0021] Figure 9 is a schematic diagram of a training process of an initial network model according to an exemplary embodiment of the present application;

[0022] Figure 10 This is a schematic diagram of the target area according to an exemplary embodiment of the present invention;

[0023] Figure 11 This is an overall flowchart of the network test inference phase according to an exemplary embodiment of the present invention;

[0024] Figure 12 This is a flowchart of a method for determining attribute information according to a specific embodiment of the present invention;

[0025] Figure 13 This is a structural block diagram of an attribute information determination device according to an embodiment of the present invention. Detailed Implementation

[0026] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings and examples.

[0027] It should be noted that the terms "first," "second," etc., in the specification, claims, and drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0028] The methods and embodiments provided in this application can be executed on a mobile terminal, a computer terminal, or a similar computing device. Taking running on a mobile terminal as an example, Figure 2 This is a hardware structure block diagram of a mobile terminal for a method of determining attribute information according to an embodiment of the present invention. Figure 2 As shown, a mobile terminal may include one or more ( Figure 2 Only one is shown in the diagram. A processor 102 (which may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.) and a memory 104 for storing data are also shown. The mobile terminal may further include a transmission device 106 for communication functions and an input / output device 108. Those skilled in the art will understand that... Figure 2 The structure shown is for illustrative purposes only and does not limit the structure of the mobile terminal described above. For example, the mobile terminal may also include components that are more... Figure 2 The more or fewer components shown, or having the same Figure 2 The different configurations shown.

[0029] The memory 104 can be used to store computer programs, such as software programs of application software and modules, such as a computer program corresponding to the attribute information determination method in the embodiments of the present application. The processor 102 can execute various functional applications and data processing, i.e., implement the above method, by running the computer programs stored in the memory 104. The memory 104 can include a high-speed random access memory, and can further include a non-volatile memory, such as one or more magnetic storage devices, flash memories, or other non-volatile solid-state memories. In some examples, the memory 104 can further include memories remotely arranged with respect to the processor 102, which can be connected to the mobile terminal through a network. Examples of the above network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0030] The transmission device 106 is configured to receive or send data via a network. Specific examples of the above network can include a wireless network provided by a communication provider of the mobile terminal. In one example, the transmission device 106 includes a network adapter (NIC), which can be connected to other network devices through a base station so as to be able to communicate with the Internet. In one example, the transmission device 106 can be a radio frequency (RF) module, which is configured to communicate with the Internet in a wireless manner.

[0031] In the embodiments, an attribute information determination method is provided, Figure 3 The flowchart of the attribute information determination method according to the embodiments of the present application is shown in FIG. 2, which includes the following steps: Figure 3

[0032] In step S302, the target network model is used to analyze the target image, so as to determine the target attribute information of the target object included in the target image.

[0033] In the embodiments, the target network model is obtained by training an initial network model using a plurality of sets of training data in the following manner: the initial network model is used to analyze training images included in each set of training data, so as to determine predicted position information of training objects included in the training images; label polar coordinate parameters are determined based on label position information included in label information of the training objects included in each set of training data; a position loss value of the initial network model is determined based on the predicted polar coordinate parameters and the label polar coordinate parameters, and network parameters of the initial network model are updated based on the position loss value, so as to obtain the target network model.

[0034] ​In the above embodiments, the target object can be a license plate, a person, an animal, an article, etc. The image can be an image captured by a camera device or an image frame obtained from a monitoring video. When the target object is a license plate, the target attribute information can include the type of the license plate, the license plate number, etc. When the target object is a person, the target attribute information can include facial features (such as facial feature), wearing features (such as whether wearing glasses, wearing a mask, etc.). When the target object is an animal, the target attribute information can include external features of the animal, such as color, type, whether tied with a rope, etc. When the target object is an article, the target attribute information can include article features, such as the shape and type of the article, etc.

[0035] In the above embodiments, when the target object is a license plate, for the image frame obtained from the monitoring video, the vehicle in the frame can be detected in real time, the entry, exit and parking in the parking area of the vehicle are determined according to the position of the vehicle and the parking area delineated in the current scene, and the vehicle is captured, the license plate is detected and recognized, and the state of the vehicle is determined. Further detection and recognition are performed on the captured image after state determination. The obtained image diagram can be seen in FIG. 1. Figure 4 For the obtained image, vehicle detection can be performed, and Faster-RCNN or YOLO model can be used for vehicle detection to determine the image of the vehicle.

[0036] In the above embodiments, after the image of the vehicle is determined, the target network model can be used to analyze the target image to determine the target attribute information of the target object. The target network model can be a model obtained by training an initial network model with multiple sets of training data. The structure of the initial network model is the same as that of the target network model, each set of training data is used to train the initial network model multiple times, and the network parameters of the initial network model are updated with the loss value obtained each time. When the change between the loss values obtained by training for a predetermined number of times is less than a predetermined threshold, the training is stopped, and the initial network model with updated network parameters after the last training is determined as the target network model. The predetermined number of times can be a predetermined number of times, such as 5 times (this value is only an example, and can also be 2 times, 3 times, 6 times, 8 times, etc., which are not limited by the present application). The predetermined threshold is also a predetermined threshold, which is not limited by the present application.

[0037] In the above embodiment, a plurality of sets of training data can be acquired, each set of training data including a training image and label information of a training object in the training image, the label information including label position information of the training object. In each training process, the initial network model can predict prediction position information of the training image, and the prediction polar coordinate parameters can be determined according to the prediction position information. The label polar coordinate parameters can also be determined according to the label position information, and the position loss value of the initial network model can be determined according to the prediction polar coordinate parameters and the label polar coordinate parameters. The position loss value is used for reverse update iteration to update the network parameters, and the training is stopped when the change between the position loss values obtained by continuous training is less than a predetermined threshold.

[0038] Optionally, the execution subject of the above steps can be a background processor or other device with similar processing capability, and can also be a machine integrated with at least an image acquisition device and a data processing device, wherein the image acquisition device can include a camera and other image acquisition modules, and the data processing device can include a computer, a mobile phone and other terminals, but is not limited thereto.

[0039] According to the application, the target image is analyzed by using the trained target network model to determine the target attribute information of the target object included in the target image. The target network model is a model obtained by training an initial network model by using a plurality of sets of training data. The training image included in each set of training data is analyzed by using the initial network model to determine the prediction position information of the training object included in the training image, the prediction polar coordinate parameters are determined according to the prediction position information, the label polar coordinate parameters are determined according to the label position information included in the label information of the training object included in each set of training data, the position loss value of the initial network model is determined according to the prediction polar coordinate parameters and the label polar coordinate parameters, and the network parameters of the initial network model are updated according to the position loss value to obtain the target network model. Since the target network model is obtained by training the initial network model and converging by using the polar coordinate parameters, the consistent distribution of the parameters can be achieved by converging the polar coordinate parameters for training, the target attribute information determined by the target network model is more accurate, and therefore, the problem of low accuracy of the attribute information of the recognized object in the related art can be solved, and the effect of improving the accuracy of the attribute information of the recognized object is achieved.

[0040] In an example embodiment, the analyzing the target image using the trained target network model to determine the target attribute information of the target object included in the target image comprises: determining a target feature of the target image using the target network model; performing sampling processing on the target feature using the target network model to obtain a plurality of features; fusing the plurality of features to obtain a fused feature; and determining the target attribute information based on the fused feature. In this embodiment, when analyzing the target image using the target network model, the target feature of the target image can be determined using the target network model, the sampling processing can be performed on the target feature using the target network model to obtain a plurality of features, the plurality of features can be fused to obtain a fused feature, and the target attribute information can be determined according to the fused feature.

[0041] In the above embodiment, the target network model can include a feature extraction network and a feature fusion network. The target feature of the target image can be determined by the feature extraction network included in the target network model, the sampling processing can be performed on the target feature using the feature fusion network, and the plurality of features obtained can be fused to obtain a fused feature.

[0042] In an example embodiment, determining the target attribute information based on the fused feature comprises: determining an image region of the target object in the target image based on the fused feature; and determining the target attribute information based on the image region and the fused feature. In this embodiment, the image region of the target object in the target image can be determined according to the fused feature, and the target attribute information can be determined according to the image region and the fused feature.

[0043] In the above embodiment, the target network model can further include a detection network and an identification network. The image region of the target object can be determined by analyzing the fused feature using the detection network. The target attribute information can be determined by analyzing the fused feature and the image region using the identification network. This implementation enables the sharing of the features learned by the fusion network in the detection network and the identification network, while enabling the two tasks to learn cooperatively. The shared features can be output together with the detection and identification of the object using one forward pass, and this coupled learning method can enable the network to learn certain context information, and the two tasks can constrain the learning of image features by the network together.

[0044] In an example embodiment, the sampling processing of the target feature using the target network model includes: performing N times of down-sampling processing on the target feature to obtain N+1 first features of different scales, and performing N times of up-sampling processing on the last feature included in the first features to obtain N+1 second features of different scales; and the obtaining of the fusion feature by fusing a plurality of the features includes: fusing the features of the same scale in the first features and the second features to obtain the fusion feature. In this embodiment, when the target object is a license plate, the image and label information of vehicle detection can be sent to the feature extraction network for extraction of high-dimensional image features to obtain the target feature. The target feature can be the last layer of convolutional features. The target feature can be processed by twice down-sampling to obtain three scale feature maps. Then, the last down-sampled feature map is processed by three times of up-sampling, and the feature maps in the previous down-sampling process are fused in the up-sampling process. The fusion mode can be a feature map channel splicing mode, so that the fusion mode of first top-down and then bottom-up can enhance the detection effect of small license plate targets. The schematic diagram of the detection process of the target network model can be referred to in FIG. 8, and the schematic diagram of the image of the target object determined by the target network model can be referred to in FIG. 9. Figure 5 The schematic diagram of the image of the target object determined by the target network model can be referred to in FIG. 9. Figure 6

[0045] In the above embodiment, in the roadside parking scene, many vehicles are far away from the monitoring device, and the license plate is relatively small, so that the license plate target becomes a small target, and the phenomenon of missed detection and false detection is prone to occur. In order to alleviate this phenomenon, a top-down and bottom-up combined feature fusion mode is adopted to enhance the learning of the detection network on the position information of small targets.

[0046] In an example embodiment, the determining of the predicted polar coordinate parameters based on the predicted position information includes: determining a predicted center point coordinate and a predicted vertex coordinate of the training object included in the predicted position information; and determining the predicted polar coordinate parameters of the training object based on the predicted center point coordinate and the predicted vertex coordinate. The determining of the label polar coordinate parameters based on the label position information of the training object includes: determining a label center point coordinate and a label vertex coordinate of the training object included in the label position information; and determining the label polar coordinate parameters of the training object based on the label center point coordinate and the label vertex coordinate. In this embodiment, when the convergence training is performed, the predicted polar coordinate parameters can be determined according to the predicted center point coordinate and the predicted vertex coordinate, and the label polar coordinate parameters can be determined according to the label center point coordinate and the label vertex coordinate.

[0047] ​In the related art, for the regression parameters of the network, the corner point parameter regression of (x1, y1, x2, y2) or (x, y, w, h) is adopted, and the predicted rectangular region generated in this way is parallel to the edge of the acquired image. In order to better locate the target with a certain angle, some target detection methods with rotation adopt the regression mode of (x, y, w, h, θ), which does not consider the consistency of the distribution of the scale parameter and the angle parameter; or adopt the four corner point regression mode of (x1, y1, x2, y2, x3, y3, x4, y4), which has too many parameters and is not conducive to the convergence and learning of the network.

[0048] In the present embodiment, in order to enable the network to better learn the position information of the rotated target, the polar coordinate parameter mode of (x, y, ρ, θ1, θ2) can be adopted, wherein the polar coordinate parameter schematic diagram can be referred to as FIG. 1. Figure 7 , wherein x and y represent the geometric center of the predicted region, i.e. the predicted center point coordinates, ρ is the polar radius, and ρ can be calculated by the formula , wherein x i and y i represent the coordinates of the four corner points of the predicted rectangular frame, i.e. the predicted vertex coordinates, , and (x, y) represent the predicted center point coordinates.

[0049] In the above embodiment, for θ1 and θ2, since x and y are the geometric center of the detected rectangle, θ1 = θ2 = θ can be obtained, and then the parameters that the network needs to learn can be converted into x, y, ρ, θ four parameters. Then the position loss value can be determined according to the predicted polar coordinate parameters and the label polar coordinate parameters, the predicted center point coordinates and the label center point coordinates. The polar coordinate parameters include the polar radius and the angle parameter.

[0050] In the above embodiment, the corner point parameters of the predicted position are polarized, so that the distribution of the parameters is consistent, which is conducive to the learning and convergence of the network.

[0051] In one example embodiment, determining the position loss value of the initial network model based on the predicted polar coordinate parameters and the label polar coordinate parameters includes: determining a center point loss value based on the predicted center point coordinates and the label center point coordinates; determining the polar coordinate parameter loss value based on the predicted polar coordinate parameters and the label polar coordinate parameters; and determining the sum of the center point loss value and the polar coordinate parameter loss value as the position loss value. In the present embodiment, the center point loss value can be determined according to the predicted center point coordinates and the label center point coordinates. The center point loss value can be represented as L reg ((x,y), , wherein (x, y) represents the predicted center point coordinates.

[0052] In the above embodiments, for learning the ρ and θ parameters, in order to balance the inconsistency in the distribution of the scale parameter ρ and the angle parameter θ, the difference in the area of ​​the fan-shaped region formed by the difference between the predicted target region and the real target region due to the difference of ρ and θ can be used to measure the difference in scale and angle between the two rectangular regions. A schematic diagram of the difference in position between the predicted target and the real target under polar coordinate parameters can be found in the appendix. Figure 8 ,like Figure 8 As shown, the solid-lined box represents the target label rectangle, i.e., the location of the training object indicated in the label information; the dashed-lined box represents the target rectangle predicted by the network, i.e., the location of the training object predicted by the initial network model. The distance between the midpoint of the predicted rectangle and the target ground truth region is used... To constrain this, when the points in two regions coincide, only the polar radius ρ and rotation angle θ remain as differences. Due to the difference in the length of ρ and the difference in θ, the resulting sector regions will differ, such as... Figure 8 As shown in the shaded area, minimizing the shaded area is the learning objective for the scale and angle tasks after the midpoint coordinates coincide. The regression loss function (polar coordinate parameter loss value) for ρ and θ can be expressed as: in, These represent the polar radius of the target region calculated based on the label information and the angle of the polar radius relative to the image edge, respectively. The final regression loss function (location loss value) is the sum of these two values, which can be expressed as:

[0053] In an exemplary embodiment, training an initial network model using multiple sets of training data further includes: determining the regional features of the target region corresponding to the label location information; inputting the regional features into the initial network model to identify the predicted attribute information included in the regional features; determining the attribute loss value of the initial network model based on the predicted attribute information and the label attribute information of the training object included in the label information; and adjusting the network parameters of the initial network model based on the attribute loss value. In this embodiment, the target network model may also include a recognition model, and training the recognition model is also included when training the initial network model. The regional features of the target region corresponding to the label location information can be determined, and the regional features can be input into the initial recognition network to identify the predicted attribute information included in the regional features. The attribute loss value of the initial network model can be determined based on the predicted attribute information and the label attribute information included in the label information, and the network parameters of the initial network model can be adjusted based on the attribute loss value. During training, the network parameters of the initial network model can be adjusted based on the attribute loss value and the location loss value to determine the target network model. The regional features may be the top-level features included in the fused features.

[0054] In the above embodiment, in the process of model training, in order to prevent the early detection and positioning effect from being inaccurate, the features of the error area are sent to the recognition network, which affects the learning and convergence of the recognition network. The target position information on the label is directly used to cut the fusion features of the uppermost layer, which are used as the feature map of object recognition and sent to the recognition network for recognition, and the CTCLoss is used to constrain the learning of the character recognition network. Wherein, the CTCLoss can be expressed as Wherein, is the license plate character label of s, s is the sequence decoded by the character recognition network, and N is the number of character recognition regions. Wherein, the training flowchart of the initial network model can be referred to the attached Figure 9 .

[0055] In an example embodiment, determining the region feature of the target region corresponding to the label position information comprises: determining the label type information included in the label information; in the case that the label type information indicates that the training object is of a first type, determining the label region corresponding to the label position information, expanding the label region by a predetermined size to obtain the target region, and determining the region feature from the training fusion feature, wherein the training fusion feature is a feature obtained by fusing a plurality of training features by the initial network model, and the plurality of training features are features obtained by sampling the features of the training image. In the case that the label type indicates that the training object is of a second type, the indication line corresponding to the label position information is determined, the training image is divided according to the indication line to obtain the target region, and the region feature is determined from the training fusion feature. In this embodiment, when determining the target region, the label type information included in the label information can be determined first. In the case that the label type information indicates that the training object is of a first type, the label region corresponding to the label position information can be determined, the label region is expanded by a predetermined size to obtain the target region. Wherein, the first type can be a single-layer license plate type, and the target region schematic diagram can be referred to the attached Figure 10 .

[0056] In the above embodiment, in the training process, the initial recognition network is fed with the features obtained by segmenting the network-shared features with the label information, so as to alleviate the influence of the inaccurate target positioning result on the learning of character recognition in the early stage of network training. For the test stage of the model, it is a network forward inference process, and the input is only the obtained original image, and the detection result output by the detection network is needed as the input of the recognition network. In the test stage, in order to better include all characters of the license plate, the positioning result of the license plate of the detection network is expanded by 1 / 10 in the width and height two scale directions (the value is only an exemplary description, and the predetermined size can also be 1 / 8, 1 / 12, etc., and the present application does not limit this), and then the shared features are cut and fed into the recognition network, and finally the recognition of the license plate characters is realized through decoding.

[0057] In the above embodiment, when the label type indicates that the training object is of the second type, the indication line corresponding to the label position information can be determined, the training image can be segmented according to the indication line to obtain two upper and lower regions, and the two upper and lower regions are the target regions. The second type can be a double-layer license plate type.

[0058] In the above embodiment, the overall flowchart of the network test inference stage can be seen from the accompanying Figure 11 As shown in the accompanying Figure 11 The flowchart includes: inputting the training sample into the feature extraction network to extract the target feature, performing fusion processing on the target feature by the feature fusion network to obtain the fusion feature, inputting the fusion feature into the detection network to determine the prediction position information of the target object, inputting the prediction position information and the fusion feature into the character recognition network (corresponding to the above recognition network) to determine the license plate character recognition result (corresponding to the above target label information) through the CTE decoder.

[0059] In the above embodiment, when the target object is a person, the target network model can also be used to identify the target attribute information (such as facial features) of the person. For example, in the kitchen of a restaurant or the like, the monitoring device in the kitchen can have a certain angle with the position of the cooking range, so that the image including the chef captured has a certain angle. When the target network model is used for face recognition of the collected image, the problem of inaccurate recognition caused by the shooting angle can be effectively avoided, and the recognition accuracy is improved.

[0060] Next, taking the target object as a license plate as an example, the method for determining the attribute information will be described in combination with the specific embodiments:

[0061] Figure 12 The method for determining the attribute information according to the specific embodiments of the present application is shown in the accompanying Figure 12 The method includes:

[0062] Step S1202, according to the vehicle position, the vehicle driving into, driving out and parking space in three state judgment and capture.

[0063] Step S1204, get the scene snapshot image.

[0064] Step S1206, vehicle detection.

[0065] Step S1208, rotate the license plate detection output license plate position and type information and keep the high-level features of the license plate.

[0066] Step S1210, judge the license plate type. When the license plate is a single-layer license plate, step S1212 is executed, and when the license plate is a double-layer license plate, step S1214 is executed.

[0067] Step S1212, the single-layer license plate value needs to expand the license plate detection area by 1 / 8 of the original length.

[0068] Step S1214, the double-layer license plate is divided into upper and lower regions according to the detection frame.

[0069] Step S1216. License plate character recognition model.

[0070] In the foregoing embodiment, in order to enhance the detection effect of the license plate with angle, the polar coordinate parameter regression method is introduced, the difference between the prediction region and the target real region midpoint position is constrained by using a loss function, and in the case of midpoint coincidence, the scale and angle difference is constrained by using the minimum difference area, which can solve the problem that the position parameter, the scale parameter and the angle parameter are difficult to converge due to different distributions. In order to couple the license plate recognition and license plate detection tasks, the shared feature method is adopted, the features extracted by the convolutional neural network are used for common learning of the detection and recognition network, and the top-down and then bottom-up method is used for feature fusion, the top-down process selects the downsampling method, the bottom-up process selects the encoding method, and the two processes are transversely connected, and the loss functions of the two tasks can constrain the network learning of the image features together.

[0071] Through the description of the above embodiments, those skilled in the art can clearly understand that the method according to the above embodiments can be realized by means of software and necessary general hardware platform, of course, it can also be realized by hardware, but in many cases, the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes a plurality of instructions for making a terminal device (which can be a mobile phone, computer, server or network device) execute the method described in each embodiment of the present application.

[0072] Also provided in the embodiments is a device for determining attribute information, which is configured to implement the above-described embodiments and preferred embodiments, and will not be described again. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, implementation in hardware, or a combination of software and hardware, is also possible and contemplated.

[0073] Figure 13 is a structural block diagram of a device for determining attribute information according to an embodiment of the application, as shown in Figure 13 The device comprises:

[0074] an analysis module 1302 configured to analyze a target image using a trained target network model to determine target attribute information of a target object included in the target image;

[0075] The target network model is obtained by training an initial network model using a plurality of sets of training data in the following manner: analyzing a training image included in each set of training data using the initial network model to determine predicted position information of a training object included in the training image; determining predicted polar coordinate parameters based on the predicted position information; determining label polar coordinate parameters based on label position information included in label information of the training object included in each set of training data; determining a position loss value of the initial network model based on the predicted polar coordinate parameters and the label polar coordinate parameters, updating network parameters of the initial network model based on the position loss value, and obtaining the target network model.

[0076] In one exemplary embodiment, the analysis module 1302 can be configured to analyze a target image using a trained target network model to determine target attribute information of a target object included in the target image in the following manner: determining target features of the target image using the target network model; performing sampling processing on the target features using the target network model to obtain a plurality of features; fusing the plurality of features to obtain fused features; and determining the target attribute information based on the fused features.

[0077] In one exemplary embodiment, the analysis module 1302 can be configured to determine the target attribute information based on the fused features in the following manner: determining an image region of the target object in the target image based on the fused features; and determining the target attribute information based on the image region and the fused features.

[0078] In an example embodiment, the analysis module 1302 can implement sampling the target feature using the target network model to obtain a plurality of features by: performing N times down-sampling processing on the target feature to obtain N+l first features of different scales, performing N times up-sampling processing on the last feature included in the first features to obtain N+l second features of different scales; and the analysis module 1302 can implement fusing a plurality of the features to obtain a fused feature by: fusing the features of the same scale included in the first features and the second features to obtain the fused feature.

[0079] In an example embodiment, the device can be configured to determine a predicted polar coordinate parameter based on the predicted position information: determine a predicted center point coordinate of the training object included in the predicted position information, and a predicted vertex coordinate, and determine the predicted polar coordinate parameter of the training object based on the predicted center point coordinate and the predicted vertex coordinate; and the device can implement determining a label polar coordinate parameter based on the label position information of the training object by: determining a label center point coordinate of the training object included in the label position information and a label vertex coordinate, and determining the label polar coordinate parameter of the training object based on the label center point coordinate and the label vertex coordinate.

[0080] In an example embodiment, the device can implement determining a position loss value of the initial network model based on the predicted polar coordinate parameter and the label polar coordinate parameter by: determining a center point loss value based on the predicted center point coordinate and the label center point coordinate; determining the polar coordinate parameter loss value based on the predicted polar coordinate parameter and the label polar coordinate parameter; and determining the sum of the center point loss value and the polar coordinate parameter loss value as the position loss value.

[0081] In an example embodiment, the device can also be configured to train the initial network model using a plurality of sets of training data: determine a region feature of a target region corresponding to the label position information; input the region feature into the initial network model to identify predicted attribute information included in the region feature; determine an attribute loss value of the initial network model based on the predicted attribute information and label attribute information of the training object included in the label information; and adjust the network parameter of the initial network model based on the attribute loss value.

[0082] In an example embodiment, the device can determine the region feature of the target region corresponding to the label position information by: determining label type information included in the label information; in a case where the label type information indicates that the training object is of a first type, determining a label region corresponding to the label position information, expanding the label region by a predetermined size to obtain the target region, and determining the region feature from a training fusion feature, wherein the training fusion feature is a feature obtained by fusing a plurality of training features by the initial network model, and the plurality of training features are features obtained by sampling a feature of a training image; and in a case where the label type indicates that the training object is of a second type, determining an indication line corresponding to the label position information, dividing the training image according to the indication line to obtain the target region, and determining the region feature from the training fusion feature.

[0083] It should be noted that each of the above modules can be implemented by software or hardware, and for the latter, the implementation can be achieved by the following ways, but is not limited thereto: all the above modules are located in the same processor; or the above modules are located in different processors in any combination.

[0084] Embodiments of the present application also provide a computer readable storage medium, which stores a computer program, wherein the computer program is executed by a processor to implement the steps of the method described in any of the above embodiments.

[0085] In an example embodiment, the computer readable storage medium can include, but is not limited to, a U disk, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk, and various media that can store computer programs.

[0086] Embodiments of the present application also provide an electronic device, which comprises a memory storing a computer program and a processor configured to execute the computer program to perform the steps of any of the above method embodiments.

[0087] In an example embodiment, the electronic device can further comprise a transmission device connected to the processor and an input / output device connected to the processor.

[0088] The specific examples in the present embodiment can refer to the examples described in the above embodiments and example embodiments, and the present embodiment will not be described here again.

[0089] It should be apparent to those skilled in the art that the modules or steps of the application described above can be implemented with general computing devices, which can be centralized on a single computing device or distributed on a network of multiple computing devices, which can be implemented with program codes executable by the computing devices, so that they can be stored in storage devices and executed by the computing devices, and in some cases, the steps shown or described can be executed in different orders than shown, or made into individual integrated circuit modules, or made into a single integrated circuit module. Thus, the application is not limited to any particular combination of hardware and software.

[0090] The preferred embodiments of the application described above are intended to be merely exemplary and those skilled in the art will readily suggest modifications and variations to the specific embodiments disclosed without departing from the principles of the application. Any and all such modifications and variations are intended to be included herein within the scope of the present application.

Claims

1. A method for determining attribute information, characterized in that, include: The trained target network model is used to analyze the target image to determine the target attribute information of the target object included in the target image. The target network model is obtained by training an initial network model using multiple sets of training data in the following manner: analyzing the training images included in each set of training data using the initial network model to determine the predicted location information of the training objects included in the training images; determining predicted polar coordinate parameters based on the predicted location information; determining label polar coordinate parameters based on the label location information included in the label information of the training objects included in each set of training data; determining the location loss value of the initial network model based on the predicted polar coordinate parameters and the label polar coordinate parameters; updating the network parameters of the initial network model based on the location loss value to obtain the target network model. Analyzing a target image using a trained target network model to determine the target attribute information of the target object included in the target image includes: using the target network model to determine the target features of the target image; using the target network model to sample the target features to obtain multiple features; fusing the multiple features to obtain a fused feature; and determining the target attribute information based on the fused feature. Determining the target attribute information based on the fusion features includes: determining the image region of the target object in the target image based on the fusion features; and determining the target attribute information based on the image region and the fusion features. The target network model is used to sample the target features to obtain multiple features, including: performing N downsampling operations on the target features to obtain N+1 first features of different scales, and performing N upsampling operations on the last feature included in the first features to obtain N+1 second features of different scales; fusing multiple features to obtain fused features includes: fusing the first features with features of the same scale included in the second features to obtain the fused features.

2. The method according to claim 1, characterized in that, Determining the predicted polar coordinate parameters based on the predicted position information includes: determining the predicted center point coordinates and predicted vertex coordinates of the training object included in the predicted position information, and determining the predicted polar coordinate parameters of the training object based on the predicted center point coordinates and the predicted vertex coordinates. Determining the label polar coordinate parameters based on the label position information of the training object includes: determining the coordinates of the label center point and the label vertex of the training object included in the label position information, and determining the label polar coordinate parameters of the training object based on the label center point coordinates and the label vertex coordinates.

3. The method according to claim 2, characterized in that, Determining the location loss value of the initial network model based on the predicted polar coordinate parameters and the label polar coordinate parameters includes: The center point loss value is determined based on the predicted center point coordinates and the label center point coordinates. The polar coordinate parameter loss value is determined based on the predicted polar coordinate parameters and the label polar coordinate parameters; The sum of the center point loss value and the polar coordinate parameter loss value is determined as the position loss value.

4. The method according to claim 1, characterized in that, Training an initial network model using multiple sets of training data also includes: Determine the regional features of the target region corresponding to the label location information; The region features are input into the initial network model to identify the predicted attribute information included in the region features; The attribute loss value of the initial network model is determined based on the predicted attribute information and the label attribute information of the training object included in the label information; The network parameters of the initial network model are adjusted based on the attribute loss value.

5. The method according to claim 4, characterized in that, The regional features for determining the target region corresponding to the label location information include: Determine the label type information included in the label information; When the label type information indicates that the training object is of the first type, the label region corresponding to the label location information is determined, the label region is expanded by a predetermined size to obtain the target region, and the region features are determined from the training fusion features, wherein the training fusion features are features obtained by fusing multiple training features through the initial network model, and the multiple training features are features obtained by sampling the features of the training image; When the label type indicates that the training object is of the second type, the indicator line corresponding to the label position information is determined, the training image is segmented according to the indicator line to obtain the target region, and the region features are determined from the training fusion features.

6. A device for determining attribute information, characterized in that, include: The analysis module is used to analyze the target image using the trained target network model to determine the target attribute information of the target object included in the target image. The target network model is obtained by training an initial network model using multiple sets of training data in the following manner: analyzing the training images included in each set of training data using the initial network model to determine the predicted location information of the training objects included in the training images; determining predicted polar coordinate parameters based on the predicted location information; determining label polar coordinate parameters based on the label location information included in the label information of the training objects included in each set of training data; determining the location loss value of the initial network model based on the predicted polar coordinate parameters and the label polar coordinate parameters; updating the network parameters of the initial network model based on the location loss value to obtain the target network model. The analysis module analyzes the target image using a trained target network model to determine the target attribute information of the target object included in the target image in the following manner: using the target network model to determine the target features of the target image; using the target network model to sample the target features to obtain multiple features; fusing the multiple features to obtain a fused feature; and determining the target attribute information based on the fused feature. The analysis module determines the target attribute information based on the fusion features in the following manner: determining the image region of the target object in the target image based on the fusion features; and determining the target attribute information based on the image region and the fusion features. The analysis module performs sampling processing on the target features using the target network model to obtain multiple features in the following manner: the target features are downsampled N times to obtain N+1 first features of different scales; the last feature included in the first features is upsampled N times to obtain N+1 second features of different scales; and multiple features are fused to obtain fused features by fusing the first features with features of the same scale included in the second features to obtain the fused features.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein the computer program, when executed by a processor, implements the steps of the method described in any one of claims 1 to 5.

8. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method as described in any one of claims 1 to 5.

Images