A blockchain light node face recognition method based on a PF-PR mode

Abstract

The application belongs to the technical field of face recognition methods, and particularly relates to a blockchain light node face recognition method based on a PF-PR mode, comprising the following steps: S1, constructing a PF data set; S2, constructing a PR data set; S3, training a PF module; S4, training a PR model; and S5, constructing a recognition system. The application recognizes faces through two stages, the first stage eliminates most of the background information, helps to more effectively and quickly obtain face feature information in the subsequent recognition process, and improves the recognition accuracy. Moreover, the application takes a complete face and a local face as the input of the recognition network, provides more feature information for the network, and organically fuses the two through feature fusion, so that the face recognition accuracy is guaranteed.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of face recognition methods, and particularly relates to a blockchain light node face recognition method based on a PF-PR mode. BACKGROUND

[0002] In a modern epidemic prevention and control system, the blockchain technology plays a great role and provides protection for the establishment of an epidemic information channel. Among the epidemic related information, personnel information is the most important part. Accurate collection of personnel information and provision of real and effective data can provide the greatest help for the accurate implementation of epidemic prevention measures. At present, face recognition is an important means of personnel information collection. The accuracy of face recognition technology directly affects the effect of information collection. Accurate face recognition technology is an important requirement for personnel information collection in an epidemic.

[0003] However, the existing face recognition technology has the following two problems: 1. The collected personnel image information contains a large amount of invalid information. When face recognition is performed, the recognition result is easily disturbed, and the recognition accuracy is low. 2. The existing face recognition technology takes single complete face information as input. When feature extraction is performed, the model cannot effectively focus attention on the features with obvious class information, resulting in slow model convergence speed, high sample requirement, and insufficient accuracy. SUMMARY

[0004] In view of the above technical problem of low recognition accuracy of the existing face recognition technology, the application provides a blockchain light node face recognition method based on a PF-PR mode, which has high recognition accuracy, high efficiency, and wide application range.

[0005] In order to solve the above technical problem, the technical scheme adopted by the application is as follows:

[0006] A blockchain light node face recognition method based on a PF-PR mode comprises the following steps:

[0007] S1, constructing a PF data set;

[0008] S2, constructing a PR data set;

[0009] S3, training the PF module;

[0010] S4, training the PR model;

[0011] S5, constructing a recognition system.

[0012] The method for constructing the PF data set in S1 is: constructing two module data sets, using target detection frames to mark the face data in the PF module data set, and the marking information includes: complete face, local face information, and the local face information includes eyes, ears, nose, and mouth; saving the label information as a json file, and the label file name is consistent with the face data file name.

[0013] The method for constructing the PR data set in S2 is: saving the complete face data and local face information in the same folder in the PR module data set, marking the corresponding personnel identity information, saving it as a csv file, and the label file name is consistent with the data folder name; dividing the two data sets according to the ratio of 7:1:2 to construct the training set, the verification set, and the test set.

[0014] The method for training the PF module in S3 is: using the training set of the PF data set to train the YoloV3 network, using SGD as the training optimizer, training 300 epochs at a learning rate of 0.001, verifying once every 5 epochs using the verification set, stopping if the model loss does not decrease for 15 consecutive epochs, and saving the model.

[0015] The method for training the PR model in S4 is: using the training set of the PR data set to train the feature fusion face recognition network, using Adam as the training optimizer, training 500 epochs at an initial learning rate of 0.001, verifying once every 10 epochs using the verification set, stopping if the model loss does not decrease for 20 consecutive epochs, and saving the model.

[0016] The feature fusion face recognition network structure is as follows: the model of the RP part contains two inputs, which are complete face information and local face information respectively, the complete face information is resized to a 3-channel rgb feature map with a size of 224*224 before input, then it is input into the backbone-A constructed by the ResNet50 network for feature extraction, and a feature matrix with a size of 7*7*2048 is output; the local face information is resized to 224*224 before input, and the 4 local face information data are added to the RGB three channels in the order of eyes, ears, nose and mouth to form an input feature of 224*224*3, then it is input into the backbone-B constructed by the ResNet18 network for feature extraction, and a feature matrix with a size of 7*7*512 is output; the feature maps output by the backbone-A and the backbone-B are combined in a Concate manner, and a 3-layer convolution layer with a convolution kernel size of 3 and a step of 1 is used for feature fusion, the convolution is performed in a SAME manner, the channel numbers of the 3-layer convolution layer are 2048, 1024 and 512 respectively, and finally a feature vector matrix with a size of 7*7*512 is obtained, the obtained feature matrix is subjected to GlobalAverage Pooling to obtain a feature vector with a size of 1*1*512, then the prediction result is output through a fully connected layer and a Softmax function.

[0017] The method for constructing the recognition system in S5 is as follows: the trained PF model and the PR model are combined, after the edge collects the face data, the face data is input into the PF model to complete the target recognition of the face information and the local information, according to the target recognition result, the data is cropped, and the cropped data is input into the PR model to complete the face recognition.

[0018] Compared with the prior art, the present application has the beneficial effects that:

[0019] The present application identifies the face through two stages, the first stage eliminates most of the background information, helps the subsequent identification process to obtain the face feature information more effectively and quickly, and improves the recognition accuracy. And the present application takes the complete face and the local face as the input of the recognition network, provides more feature information for the network, and organically fuses the two through feature fusion, so that the face recognition accuracy is guaranteed. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only exemplary, and for those skilled in the art, other embodiments can be derived from the provided drawings without creative labor.

[0021] The structures, proportions, sizes, etc. shown in the specification are merely used to cooperate with the content disclosed in the specification for understanding and reading by those skilled in the art, and are not used to limit the conditions that the application can be implemented, and therefore do not have technical significance. Any modification of the structure, change of the proportional relationship, or adjustment of the size, without affecting the effects that the application can produce and the purposes that the application can achieve, should still fall within the scope of the technology disclosed by the application.

[0022] Figure 1 Flowchart of the identification step of the application;

[0023] Figure 2 Blockchain light node diagram of the application. DETAILED DESCRIPTION

[0024] In order to make the purposes, technical solutions and advantages of the embodiments of the application more clear, the technical solutions in the embodiments of the application will be described clearly and completely below. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. These descriptions are only for further illustrating the features and advantages of the application, and are not limiting on the claims of the application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0025] The specific embodiments of the application will be further described in detail below with reference to the accompanying drawings and embodiments. The following embodiments are used to illustrate the application, but not to limit the scope of the application.

[0026] The terms "first", "second" are only used for description purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features limited by "first", "second" can explicitly or implicitly include one or more features. In the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise specified and limited.

[0027] In the description of the present application, it should be noted that, unless otherwise specified and limited, the terms "mounting", "connecting", "connection" should be understood broadly, for example, it can be fixed connection, or detachable connection, or integral connection; it can be mechanical connection, or electrical connection; it can be directly connected, or indirectly connected through an intermediate medium, or the communication inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0028] In the embodiment, the system is composed of two parts, the first part is a local information extraction module (PF module) based on Yolov3, and the second part is a face recognition module (PR module). A feature fusion face recognition network is designed, as shown in Figure 1 、 Figure 2 , and the specific construction is as follows:

[0029] Step 1: PF dataset construction. The dataset of the two modules of the application is constructed. The face data is labeled using the target detection frame in the PF module dataset, and the labeling information includes: complete face, local face information (eye, ear, nose, mouth). Save the label information as a json file, and the label file name is consistent with the face data file name.

[0030] Step 2: PR dataset construction. The complete face data and local face information in the PR module dataset are saved in the same folder, and the corresponding personnel identity information is labeled and saved as a csv file. The label file name is consistent with the data folder name. The two datasets are divided according to the ratio of 7:1:2 to construct the training set, the validation set and the test set.

[0031] Step 3: PF module training. The training set of the PF dataset is used to train the YoloV3 network. SGD is used as the training optimizer, and the learning rate is 0.001. Train for 300 epochs, validate every 5 epochs using the validation set, and stop if the model loss does not decrease for 15 consecutive epochs. Save the model.

[0032] Step 4: PR model training. The training set of the PR dataset is used to train the feature fusion face recognition network. Adam is used as the training optimizer, and the initial learning rate is 0.001. Train for 500 epochs, validate every 10 epochs using the validation set, and stop if the model loss does not decrease for 20 consecutive epochs. Save the model.

[0033] The feature fusion face recognition network structure is as follows: The RP part of the model contains two inputs: complete face information and partial face information. Before input, the complete face information is resized to a 224*224 three-channel RGB feature map, and then input into the backbone-A constructed by the ResNet50 network for feature extraction, outputting a 7*7*2048 feature matrix. Before input, the partial face information is resized to a 224*224 size, and the four partial face information data are summed in the order of eyes, ears, nose, and mouth, forming a 224*224*3 input feature, which is then input into the backbone-B constructed by the ResNet18 network for feature extraction, outputting a 7*7*512 feature matrix. The feature maps output from backbone-A and backbone-B are concatenated and then fused using three convolutional layers with a kernel size of 3 and a stride of 1. The convolution is performed using the SAME method, with the number of channels in the three convolutional layers being 2048, 1024, and 512, respectively, resulting in a 7*7*512 feature vector matrix. The obtained feature matrix is ​​then subjected to Global Average Pooling to obtain a 1*1*512 feature vector. Finally, the prediction result is output through a fully connected layer and the Softmax function.

[0034] Step 5: Building the Recognition System. The trained PF model and PR model are combined. After capturing face data at the edges, the face data is input into the PF model to complete target recognition of face and local information. Based on the target recognition results, the data is cropped, and the cropped data is input into the PR model to complete face recognition.

[0035] The above description only illustrates the preferred embodiments of the present invention. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention, and all such changes should be included within the protection scope of the present invention.

Claims

1. A blockchain light node face recognition method based on the PF-PR mode, characterized in that: Includes the following steps: S1. Construct the PF dataset; S2. Construct the PR dataset; S3. Train the PF module; The YOLOv3 network was trained using the training set of the PF dataset. SGD was used as the training optimizer, and the training was performed for 300 epochs with a learning rate of 0.

001. The model was validated using the validation set every 5 epochs. If the model loss did not decrease for 15 consecutive epochs, the training was stopped and the model was saved. S4. Train the PR model; Use the training set of the PR dataset to train the feature fusion face recognition network. Use Adam as the training optimizer, train for 500 epochs with an initial learning rate of 0.001, and validate the model with the validation set every 10 epochs. If the model loss does not decrease for 20 consecutive epochs, stop and save the model. The RP part of the model contains two inputs: complete face information and partial face information. Before input, the complete face information is resized to a 224*224 three-channel RGB feature map, which is then fed into the backbone-A constructed from ResNet50 for feature extraction, outputting a 7*7*2048 feature matrix. Before input, the partial face information is resized to a 224*224 size, and the four partial face data points are summed across their RGB channels in the order of eyes, ears, nose, and mouth, forming a 224*224*3 input. The features are then input into the ResNet18 network-constructed backbone-B for feature extraction, outputting a 7*7*512 feature matrix. The feature maps output from backbone-A and backbone-B are concatenated and fused using three convolutional layers with a kernel size of 3 and a stride of 1. The convolution is performed using the SAME method, with the three convolutional layers having 2048, 1024, and 512 channels respectively, resulting in a 7*7*512 feature vector matrix. The obtained feature matrix is ​​then subjected to Global Average Pooling to obtain a 1*1*512 feature vector, which is then passed through a fully connected layer and the Softmax function to output the prediction result. S5. Construct a recognition system.

2. The blockchain light node face recognition method based on the PF-PR mode according to claim 1, characterized in that: The method for constructing the PF dataset in S1 is as follows: construct datasets for two modules. In the PF module dataset, object detection boxes are used to annotate face data. The annotation information includes: complete face and partial face information. The partial face information includes eyes, ears, nose and mouth. The label information is saved as a JSON file, and the label file name is consistent with the face data file name.

3. The blockchain light node face recognition method based on the PF-PR mode according to claim 1, characterized in that: The method for constructing the PR dataset in S2 is as follows: The PR module dataset saves the complete face data and partial face information from the unified face image data in the same folder, and labels the corresponding personnel identity information and saves it as a CSV file, with the label file name consistent with the data folder name; the two datasets are divided according to the ratio of 7:1:2 to construct the training set, validation set and test set.

4. The blockchain light node face recognition method based on the PF-PR mode according to claim 1, characterized in that: The method for constructing the recognition system in S5 is as follows: the trained PF model and PR model are combined. After the face data is collected at the edge, the face data is input into the PF model to complete the target recognition of face information and local information. Based on the target recognition result, the data is cropped and the cropped data is input into the PR model to complete the face recognition.

Images