Operating personnel identification method and system based on deep learning
Through deep learning technology, the occlusion image feature points of the operators are identified and aligned, and the EfficientNet network is trained to solve the problem of low accuracy of occlusion image recognition, and the identification of safety personnel at the construction site is realized, reducing safety risks.
Patent Information
- Application Number
- CN202411097382.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-12
- Publication Date
- 2024-11-22
AI Technical Summary
In the prior art, when identifying face blocking images wearing masks, sunglasses, hats and other accessories, it is difficult to accurately identify face feature points, resulting in large errors in front face image, affecting the accuracy of face recognition, especially in the recognition of worker personnel. There are security risks in the scenario.
Using deep learning-based job recognition method, by collecting face image data and frontal image data of various facial occlusion situations, detecting and extracting feature points, and aligning feature points to the corresponding position using a specific face alignment algorithm, training improvement The EfficientNet network generates a face recognition model and inputs the construction site image data to achieve identity recognition.
It improves the accuracy and robustness of operator identification, ensures that only relevant operators can enter the construction site, effectively prevent unauthorized personnel from entering, thereby reducing safety risks.
Smart Images

Figure CN119007266A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of image recognition technology, and more specifically to a method and system for identifying operators based on deep learning. Background Art
[0002] Face recognition technology is a biometric technology that identifies people based on their facial features and has been widely used in various fields of life. In the process of face recognition, in order to improve recognition accuracy, it is usually necessary to convert the face image into a frontal face image before performing identity recognition. At present, the existing technology usually converts the face image to be identified into a frontal face image based on the facial feature point features of the face image to be identified. However, when identifying the face occluded image of a person wearing accessories such as a mask, sunglasses, or a hat, it is difficult to accurately identify the facial feature point features because the facial area in the face occlusion image is partially blocked, resulting in a large error in the converted frontal face image, which seriously affects the accuracy of subsequent face recognition and is in urgent need of improvement.
[0003] The application of deep learning in the field of face recognition has made significant progress and has become a powerful technical means. As an important application scenario of face recognition technology, operator identification has also received more and more attention.
[0004] Therefore, how to provide a worker identification method and system based on deep learning to ensure that only relevant workers can enter the construction site by performing facial recognition on construction workers is a technical problem that technical personnel in this field urgently need to solve. Summary of the invention
[0005] In view of this, the present invention provides a worker identification method and system based on deep learning, which performs facial recognition on construction workers to ensure that only relevant workers can enter the construction site, thereby effectively preventing unauthorized personnel from entering the construction area and reducing safety risks.
[0006] In order to achieve the above object, the present invention adopts the following technical solution: a method for identifying operators based on deep learning, comprising:
[0007] Collecting facial image data of operators in various facial occlusion situations and frontal image data of operators, and segmenting and classifying the facial image data and frontal image data to obtain classified data;
[0008] Detect and extract feature points in categorical data;
[0009] Using a specific face alignment algorithm, aligning feature points to corresponding positions based on the frontal image data to obtain a sample data set;
[0010] Taking the sample data set as input, training the improved deep learning EfficientNet network to obtain a face recognition model;
[0011] The image data of the construction site is input into the face recognition model to obtain the identity recognition result of the operator.
[0012] Preferably, the position coordinates of the feature points are expressed as:
[0013] Q=[(x1,y1),(x2,y2),…,(x m ,y m )];
[0014] Where m is the number of feature points, x m is the horizontal coordinate of the mth feature point, y m is the ordinate of the mth feature point.
[0015] Preferably, the improved deep learning EfficientNet network is trained, including:
[0016] Dividing the sample data set into a training set, a validation set and a test set;
[0017] The improved deep learning EfficientNet network is trained according to the training set to obtain a trained deep learning network, wherein the training of the improved deep learning EfficientNet network is based on adjusting the balance between network depth, width and resolution.
[0018] Preferably, the improved deep learning EfficientNet network includes 1 depthwise separable convolutional layer and 5 MBConv layers;
[0019] The deep separable convolution layer in the improved EfficientNet network structure separates spatial feature learning from channel feature learning, and each channel is independent of each other; there is an SK module in each MBConv. The SK module is a soft attention mechanism that is directly embedded in the model for training, allowing the network to adaptively adjust the size of the receptive field according to multiple scales of the input information;
[0020] Optimizing the trained deep learning network according to the verification set to obtain a face recognition model;
[0021] The test set is input into the face recognition model for recognition, and classified using a softmax classifier to obtain a face recognition result.
[0022] Preferably, the network depth, width and resolution in the deep learning network are adjusted according to the training set to obtain adjusted network depth-related parameters, width-related parameters and resolution-related parameters;
[0023] A trained deep learning network is obtained according to the network depth-related parameters, the width-related parameters and the resolution-related parameters.
[0024] Preferably, by using a specific face alignment algorithm, the feature points are aligned to corresponding positions based on the frontal image data to obtain a sample data set, including:
[0025] Based on a feature extraction algorithm or a pre-trained face alignment network, facial feature points are extracted from the acquired face occluded image, and then based on the extracted facial feature points, an affine transformation is performed on the acquired face occluded image to achieve an alignment operation on the face occluded image and obtain an aligned face occluded image.
[0026] Preferably, the depth data of the face image is converted into 3D voxelized data and input into a V2V-PoseNet network to obtain feature point data of the face image; the V2V-PoseNet network is a V2V-PoseNet network that has undergone pruning processing.
[0027] Preferably, the 3D voxelization data is performed according to the following steps: converting the depth data of the face image into a 3D volume form, reprojecting the points into the 3D space, and discretizing the continuous space, and setting the voxel value of each discrete space according to the voxel space position and the target object;
[0028] The 3D voxelized data is used as the input of the V2V-PoseNet network, the likelihood of each key point belonging to each voxel is calculated, the position corresponding to the highest likelihood of each key point is identified, and it is converted into real-world coordinates as head feature point data.
[0029] Preferably, a worker identification system based on deep learning comprises:
[0030] A data collection module, used to collect face image data of operators in various facial occlusion situations and frontal image data of operators, and to segment and classify the face image data and frontal image data to obtain classified data;
[0031] Feature point detection module, used to detect and extract feature points in classification data;
[0032] A face alignment module, used to align feature points to corresponding positions based on the frontal image data by using a specific face alignment algorithm to obtain a sample data set;
[0033] A model acquisition module, used to take the sample data set as input, train the improved deep learning EfficientNet network, and obtain a face recognition model;
[0034] The result generation module is used to input the image data of the construction site into the face recognition model to obtain the identity recognition result of the operator.
[0035] It can be known from the above technical solutions that compared with the prior art, the present invention discloses a method and system for identifying workers based on deep learning, including: collecting face image data containing various facial occlusion situations of workers and frontal image data of workers, and cutting and classifying the face image data and frontal image data to obtain classified data; detecting and extracting feature points in the classified data; aligning the feature points to corresponding positions based on the frontal image data through a specific face alignment algorithm to obtain a sample data set; using the sample data set as input to train the improved deep learning EfficientNet network to obtain a face recognition model; inputting the image data of the construction site into the face recognition model to obtain the identity recognition result of the worker. The present invention ensures that only relevant workers can enter the construction site by performing face recognition on construction workers, thereby effectively preventing unauthorized personnel from entering the construction area and reducing safety risks. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.
[0037] Figure 1 A flowchart of a method for identifying workers based on deep learning is provided in an embodiment of the present invention.
[0038] Figure 2 A schematic diagram of the structure of a worker identification system based on deep learning provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0039] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0040] The embodiment of the present invention discloses a worker identification method and system based on deep learning. By performing facial recognition on construction workers, it is ensured that only relevant workers can enter the construction site, thereby effectively preventing unauthorized persons from entering the construction area and reducing safety risks.
[0041] The embodiment of the present invention discloses a method for identifying operators based on deep learning, such as Figure 1 As shown, including:
[0042] Collect facial image data of operators with various facial occlusion conditions and frontal image data of operators, including those wearing masks, sunglasses, hats, etc. Use data enhancement technology to expand the data set and increase data diversity; and cut and classify the facial image data and frontal image data to obtain classified data;
[0043] Detect and extract feature points in the classified data; the feature points include key feature points such as eyes, nose and mouth, eyebrows, highest point of the face, lowest point of the face, etc.
[0044] Using a specific face alignment algorithm, aligning feature points to corresponding positions based on the frontal image data to obtain a sample data set;
[0045] The sample data set is used as input to train the improved deep learning EfficientNet network to obtain a face recognition model, so as to retain useful face information and reduce the impact of occlusion, thereby improving recognition accuracy.
[0046] The image data of the construction site is input into the face recognition model to obtain the identity recognition result of the operator.
[0047] The embodiments of the present invention improve the accuracy and robustness of operator identification through a large amount of data and deep learning.
[0048] Specifically, it also includes: comparing the obtained facial recognition results with the facial database of the workers at the construction site. If the facial recognition results are not matched in the facial database, an alarm message is generated to remind the management personnel; if the facial recognition results are matched in the facial database, their work information is recorded.
[0049] Specifically, the position coordinates of the feature points are expressed as:
[0050] Q=[(x1,y1),(x2,y2),…,(x m ,y m )];
[0051] Where m is the number of feature points, x m is the horizontal coordinate of the mth feature point, ym is the ordinate of the mth feature point.
[0052] Specifically, the improved deep learning EfficientNet network is trained, including:
[0053] Dividing the sample data set into a training set, a validation set and a test set;
[0054] The improved deep learning EfficientNet network is trained according to the training set to obtain a trained deep learning network, wherein the training of the improved deep learning EfficientNet network is based on adjusting the balance between network depth, width and resolution.
[0055] Specifically, the improved deep learning EfficientNet network includes 1 depthwise separable convolutional layer and 5 MBConv layers;
[0056] The deep separable convolution layer in the improved EfficientNet network structure separates spatial feature learning from channel feature learning, and each channel is independent of each other; there is an SK module in each MBConv. The SK module is a soft attention mechanism that is directly embedded in the model for training, so that the network can adaptively adjust the size of the receptive field according to multiple scales of the input information; that is, when extracting features, the network can strengthen the extraction of effective features and compress invalid features;
[0057] The trained deep learning network is optimized according to the verification set to obtain a face recognition model; the number of network iterations and the learning rate of the trained deep learning network are repeatedly adjusted according to the verification set to obtain an optimal deep learning network, i.e., a face recognition model.
[0058] The test set is input into the face recognition model for recognition, and classified using a softmax classifier to obtain a face recognition result.
[0059] The EfficientNet network used in the present invention is an expansion of the depth, width and resolution of the deep learning network under the maximum balance program, which greatly enriches the extracted features. The original model has a complex structure and is not suitable for a small sample set. Therefore, the original model is simplified in the recognition model of the present invention. At the same time, the SE module is improved to a SK module with better recognition effect, which strengthens the extraction of effective features and suppresses invalid features. The method of the present invention can improve the recognition accuracy of operators.
[0060] Specifically, the network depth, width and resolution in the deep learning network are adjusted according to the training set to obtain adjusted network depth-related parameters, width-related parameters and resolution-related parameters;
[0061] A trained deep learning network is obtained according to the network depth-related parameters, the width-related parameters and the resolution-related parameters.
[0062] Specifically, it also includes: adding a non-local module to the EfficientNet network, which directly captures long-range dependencies by calculating the interaction between any two unknowns, rather than being limited to adjacent points, and maintains more information through eigenvalue weighting operations. Inserting a non-local module has different degrees of improvement on the effects of visual tasks such as image classification, target detection, and target segmentation. The specific algorithm flow of the non-local module is shown in the following formula:
[0063]
[0064]
[0065]
[0066] Among them, y i represents the feature vector of the jth spatial position of a feature map of the deep learning network after non-local processing, x i With x j represents the feature vectors of the i-th and j-th spatial positions on the corresponding original feature map, where i, j∈(1, HW). f(·) is a function that calculates the similarity between any two points, where θ(·) and They are all convolution operations, C(x) represents the normalization factor, and e is a mathematical constant. It is a mapping function that maps a point into a vector, which can be regarded as calculating the feature of a point. The final feature y i The number of channels is C g .
[0067] Using the high-performance deep learning network EfficientNet as the basis, the improved network can well meet the application requirements of different scenarios after fine-tuning.
[0068] In the network structure after improving the output of the deep learning network, the output of each layer is used as the input of the next layer;
[0069] The non-local module is inserted into the last three high-order semantic layers, that is, the three non-local modules are placed directly after the 3rd, 4th, and 5th MBConv layers, and the calculation of f(·) is converted into matrix multiplication and convolution operations. The non-local module operates on the feature layer and can be easily inserted into any network. The embodiment preferably inserts the high-order semantic layer to reduce the amount of calculation.
[0070] Since the original EfficientNet network structure is very complex and has many convolution layers, the vehicle model data set samples used in the present invention are not very large, and the original network is prone to overfitting. The present invention simplifies the original network based on EfficientNet, simplifies the 16 blocks in the original network into 5 blocks, and unifies the convolution size in all blocks to 3×3. The simplified EfficientNet can optimize recognition accuracy and efficiency, and find better performance from the balance between network depth, width and resolution.
[0071] The SE (Squeeze-and-Excitation Module) module of the original EfficientNet network is replaced by the more effective SK (Selective Kernel Module) module. The SK module is a better soft attention mechanism than the SE module. The network channels are increased, and it focuses on the important features of the data set while suppressing other useless features and enhancing the properties, which can improve the recognition accuracy.
[0072] The specific algorithm of the EfficientNet network is as follows:
[0073]
[0074] In formula (1), N represents the classification network, X represents the input, and f i represents the basic network layer, i represents the stage, L i represents f i The number of repetitions of the structure in the i-th stage.
[0075] The following formula is to seek the optimal parameters required to maximize recognition accuracy. The parameters to be optimized are network depth (d), network width (w) and resolution (r). The goal of network scaling is to maximize the recognition accuracy of the network while satisfying the constraints on network parameters and computational complexity.
[0076]
[0077]
[0078] Memory(N)≤target_memory;
[0079] FLOPS(N)≤target_flops;
[0080] In the formula, Memory represents memory storage, and FLOPS represents the number of floating-point operations per second.
[0081] In order to find the appropriate three-dimensional parameters, the Φ parameter is introduced, and the three parameters to be optimized are all represented by the Φ index, as shown in formula (6):
[0082] depth:d=α φ
[0083] width:w=β φ
[0084] resolution:r=γ φ ;
[0085] stα·β 2 γ 2 ≈2
[0086] α≥1,β≥1,γ≥1
[0087] Through experiments, we found that the overall effect is best when α is 1.2, β is 1.1, and γ is 1.15.
[0088] In the process of designing the entire network model, the main issues to be addressed are how to improve the accuracy of the network, how to ensure that features are not lost while deepening the number of network layers, and how to extract effective features more efficiently in a deep network. In general, in order to extract more features, the network is deepened and widened, but when the network is deepened and widened, it is easy to cause overfitting and the loss of features. The general large-scale network structure often has the disadvantages of many parameters and easy overfitting. The present invention simplifies the EfficientNet deep learning network based on adjusting the balance between network depth, width and resolution, and improves a soft attention mechanism SK module with better effect. The design network structure contains 5 block structures, which effectively solves the above problems. The network of the present invention balances the three aspects of deepening the depth of the network, widening the width and increasing the resolution, and trains a recognition network that can be increased in three aspects compared to other networks, and has higher recognition accuracy than other networks.
[0089] The improved EfficientNet network structure includes 1 depthwise separable convolution and 5 MBConv. The depthwise separable convolution layer in the EfficientNet network structure separates spatial feature learning from channel feature learning, and each channel is independent of each other. By separating the convolution layer, fewer parameters are required and the amount of calculation is also smaller, so a smaller and faster model can be obtained.
[0090] There is an SK module in each MBConv. The SK module is a softattention mechanism that can be directly embedded in the model for training. It can enable the network to adaptively adjust the size of the receptive field according to the multiple scales of the input information, that is, to enable the network to enhance the extraction of effective features and compress invalid features when extracting features. The SK module is mainly divided into three operations: Split, Fuse, and Select. The specific steps are as follows:
[0091] a) The Split operation refers to the convolution operation of Kernel3×3 and Kernel5×5 on the input X to obtain the output and
[0092] b) Fuse operation is the same as SE module operation. gp is the global average pooling operation, F fc It is a two-layer fully connected layer that first reduces the dimension and then increases the dimension. It should be noted that the two output matrices a and b, where matrix b is a redundant matrix, b = 1-a in the case of two branches. Output
[0093] c) The Select operation uses two weight matrices a and b to and Perform weighted operations to find the final vector V.
[0094] After inputting the image, the specific parameters of each stage are as follows: In the first convolution layer, the convolution kernel size is 3×3, and the image size is 128×128. The convolution kernel size of depthwise separable convolution is 3×3, the input feature channels are 16, the output feature channels are 32, and the image size is 64×64. MBConv1 has 32 input convolution kernels, a convolution kernel size of 3×3, 16 output convolution kernels, and slides with a step size of 2, and the image size is 64×64. MBConv2 has 16 input convolution kernels, a convolution kernel size of 3×3, 32 output convolution kernels, and slides with a step size of 2, and the image size is 32×32. MBConv3 has 32 input convolution kernels, a convolution kernel size of 3×3, 64 output convolution kernels, and slides with a step size of 2, and the image size is 16×16. MBConv4 has 40 input convolution kernels, a convolution kernel size of 3×3, 80 output convolution kernels, sliding with a step size of 2, and an image size of 8×8. MBConv5 has 80 input convolution kernels, a convolution kernel size of 3×3, 120 output convolution kernels, sliding with a step size of 2, and an image size of 4×4.
[0095] Specifically, by using a specific face alignment algorithm, the feature points are aligned to corresponding positions based on the frontal image data to obtain a sample data set, including:
[0096] After obtaining the face occluded image, an alignment operation can be performed on the face occluded image based on the features of the facial feature points. It should be noted that the operation of aligning the face occluded image is also the process of converting the face occluded image into a frontal face. Specifically, facial feature points can be extracted from the acquired face occluded image based on a feature extraction algorithm or a pre-trained face alignment network. For example, 78 feature points of the eyes, nose, mouth, eyebrows and other areas can be extracted. Then, based on the extracted facial feature points, an affine transformation is performed on the acquired face occluded image to achieve the alignment operation of the face occluded image and obtain the aligned face occluded image.
[0097] An affine transformation is geometrically defined as an affine transformation or affine mapping between two vector spaces consisting of a non-singular linear transformation (a transformation performed by a linear function) followed by a translation. In the finite-dimensional case, each affine transformation can be given by a matrix A and a vector b, which can be written as A with an additional column b. An affine transformation corresponds to the multiplication of a matrix and a vector, while the composition of affine transformations corresponds to ordinary matrix multiplication, with an extra row added to the bottom of the matrix, which is all zeros except for a 1 on the right, and a 1 added to the bottom of the column vector.
[0098] It should be noted that the facial image to be recognized in the embodiment of the present invention is a facial image whose facial area is blocked by accessories such as masks, glasses, and hats. Since the facial area is blocked, it is difficult to accurately identify the facial feature points in the facial occluded image in this step, resulting in a large error in the facial occluded image obtained after feature alignment based on facial associated feature points. It is not an accurate frontal facial occluded image, and the following operations need to be continued to further perform spatial transformation on the aligned facial occluded image to obtain a standard frontal facial occluded image.
[0099] Specifically, the depth data of the face image is converted into 3D voxel data, and input into a V2V-PoseNet network to obtain feature point data of the face image; the V2V-PoseNet network is a pruned V2V-PoseNet network;
[0100] The depth data of the face image is obtained by the following steps: obtaining a depth image and an RGB image;
[0101] Input the RGB image into the YOLOv3 network to obtain the head bounding box;
[0102] The depth data is aligned with the RGB image, and the depth data is cropped according to the coordinates of the head bounding box to separate the head area and the background area to obtain the depth data of the face image.
[0103] Specifically, the 3D voxelization data is performed according to the following steps: the depth data of the face image is converted into a 3D volume form, the points are re-projected into the 3D space, and the continuous space is discretized, and the voxel value of each discrete space is set according to the voxel space position and the target object;
[0104] The 3D voxelized data is used as the input of the V2V-PoseNet network, the likelihood of each key point belonging to each voxel is calculated, the position corresponding to the highest likelihood of each key point is identified, and it is converted into real-world coordinates as head feature point data.
[0105] The embodiment of the present invention uses a deep learning target detection method based on the RGB image to extract the head bounding box, and separates the head depth image based on the bounding box and the depth image; secondly, the depth information is converted into a voxel representation, and the position of the head feature points is detected using a three-dimensional convolutional network. The embodiment of the present invention combines cutting-edge deep learning models to avoid the introduction of artificially defined features. The model has strong generalization and expression capabilities and good scalability. The existing models used in the system are pruned and optimized according to the needs of the task, which improves the speed of the model without affecting the accuracy.
[0106] The embodiment of the present invention also discloses a worker identification system based on deep learning, such as Figure 2 As shown, including:
[0107] A data collection module, used to collect face image data of operators in various facial occlusion situations and frontal image data of operators, and to segment and classify the face image data and frontal image data to obtain classified data;
[0108] Feature point detection module, used to detect and extract feature points in classification data;
[0109] A face alignment module, used to align feature points to corresponding positions based on the frontal image data by using a specific face alignment algorithm to obtain a sample data set;
[0110] A model acquisition module, used to take the sample data set as input, train the improved deep learning EfficientNet network, and obtain a face recognition model;
[0111] The result generation module is used to input the image data of the construction site into the face recognition model to obtain the identity recognition result of the operator.
[0112] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.
[0113] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for identifying workers based on deep learning, characterized in that: include: Collecting facial image data of operators in various facial occlusion situations and frontal image data of operators, and segmenting and classifying the facial image data and frontal image data to obtain classified data; Detect and extract feature points in categorical data; Using a specific face alignment algorithm, aligning feature points to corresponding positions based on the frontal image data to obtain a sample data set; Taking the sample data set as input, training the improved deep learning EfficientNet network to obtain a face recognition model; The image data of the construction site is input into the face recognition model to obtain the identity recognition result of the operator.
2. The method for identifying workers based on deep learning according to claim 1, characterized in that: The position coordinates of the feature points are expressed as: Q=[(x1,y1),(x2,y2),…,(x m ,y m )]; Where m is the number of feature points, x m is the horizontal coordinate of the mth feature point, y m is the ordinate of the mth feature point.
3. The method for identifying workers based on deep learning according to claim 1, characterized in that: Training the improved deep learning EfficientNet network includes: Dividing the sample data set into a training set, a validation set and a test set; The improved deep learning EfficientNet network is trained according to the training set to obtain a trained deep learning network, wherein the training of the improved deep learning EfficientNet network is based on adjusting the balance between network depth, width and resolution.
4. The method for identifying workers based on deep learning according to claim 3, characterized in that: The improved deep learning EfficientNet network includes 1 depthwise separable convolution layer and 5 MBConv layers; The deep separable convolution layer in the improved EfficientNet network structure separates spatial feature learning from channel feature learning, and each channel is independent of each other; there is an SK module in each MBConv. The SK module is a softattention mechanism that is directly embedded in the model for training, allowing the network to adaptively adjust the size of the receptive field according to multiple scales of the input information; Optimizing the trained deep learning network according to the verification set to obtain a face recognition model; The test set is input into the face recognition model for recognition, and classified using a softmax classifier to obtain a face recognition result.
5. The method for identifying workers based on deep learning according to claim 3, characterized in that: Adjust the network depth, width and resolution in the deep learning network according to the training set to obtain adjusted network depth-related parameters, width-related parameters and resolution-related parameters; A trained deep learning network is obtained according to the network depth-related parameters, the width-related parameters and the resolution-related parameters.
6. The method for identifying workers based on deep learning according to claim 1, characterized in that: By using a specific face alignment algorithm, the feature points are aligned to corresponding positions based on the frontal image data to obtain a sample data set, including: Based on a feature extraction algorithm or a pre-trained face alignment network, facial feature points are extracted from the acquired face occluded image, and then based on the extracted facial feature points, an affine transformation is performed on the acquired face occluded image to achieve an alignment operation on the face occluded image and obtain an aligned face occluded image.
7. The method for identifying workers based on deep learning according to claim 2, characterized in that: The depth data of the face image is converted into 3D voxel data, and input into a V2V-PoseNet network to obtain feature point data of the face image; the V2V-PoseNet network is a V2V-PoseNet network that has undergone pruning processing.
8. The method for identifying workers based on deep learning according to claim 7, characterized in that: The 3D voxelization data is performed in the following steps: the depth data of the face image is converted into a 3D volume form, the points are re-projected into the 3D space, and the continuous space is discretized, and the voxel value of each discrete space is set according to the voxel space position and the target object; The 3D voxelized data is used as the input of the V2V-PoseNet network, the likelihood of each key point belonging to each voxel is calculated, the position corresponding to the highest likelihood of each key point is identified, and it is converted into real-world coordinates as head feature point data.
9. A worker identification system based on deep learning, applying a worker identification method based on deep learning according to any one of claims 1 to 8, characterized in that: include: A data collection module, used to collect face image data of operators in various facial occlusion situations and frontal image data of operators, and to segment and classify the face image data and frontal image data to obtain classified data; Feature point detection module, used to detect and extract feature points in classification data; A face alignment module, used to align feature points to corresponding positions based on the frontal image data by using a specific face alignment algorithm to obtain a sample data set; A model acquisition module, used to take the sample data set as input, train the improved deep learning EfficientNet network, and obtain a face recognition model; The result generation module is used to input the image data of the construction site into the face recognition model to obtain the identity recognition result of the operator.
