Face access control recognition system and method based on image processing and storage medium

Abstract

The application discloses a face access control recognition system and method based on image processing and a storage medium, belongs to the technical field of access control, and solves the problem that the existing method cannot effectively capture local features in an image and lacks the sensing ability for image details when calculating face subgraphs by using Mobilefacenet to obtain face characteristic values, the method comprises the following steps: acquiring at least one set of access control collection data, pre-processing the access control collection data, executing a face recognition model, judging whether the pre-processing set is a living target, if the pre-processing set is a living target, the face recognition model performs recognition detection on the pre-processing set, and outputs a detection result; the face recognition model is constructed by combining distributed displacement convolution and a Segnet network, can more effectively capture local features in an image, thereby enhancing the sensing ability of the model for image details, and can also combine deep features and shallow features through a skip connection, thereby enhancing the adaptive ability of the model to complex backgrounds and different illumination conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of access control recognition, and in particular relates to a face access control recognition system, method and storage medium based on image processing. Background Art

[0002] With the increasing development of modern information technology, biometric recognition technology has been widely used in access control systems, with facial recognition being a common method. Facial recognition technology offers advantages such as speed, intuitiveness, and non-contact. Traditional facial recognition algorithms, however, are primarily based on model matching and feature matching. Both algorithms can be affected by various factors, such as lighting, posture, and facial expression, and their accuracy and real-time performance need to be improved.

[0003] Chinese patent CN109903433B discloses a facial recognition-based access control system and access control method. The system includes: a camera, configured to send a facial image to a microcontroller unit after obtaining a first session key with an intelligent gateway; a microcontroller unit, configured to calculate facial feature values ​​based on the received facial image; a secure element, configured to send a facial recognition command to a secure element; the secure element, configured to compare the facial feature values ​​in the received facial recognition command with feature values ​​in a facial database, obtain a facial recognition result based on the comparison result, and return the facial recognition result to the microcontroller unit; the microcontroller unit, further configured to receive the facial recognition result and send an access control operation command to an intelligent access control system; and the intelligent access control system, configured to decrypt the received access control operation command using the second session key after obtaining a second session key and perform access control operations based on the obtained facial recognition result. However, existing methods use MobileFaceNet to calculate facial subgraphs, but the facial feature values ​​obtained cannot effectively capture local features in the image and lack the ability to perceive image details. To address these issues, we propose a facial access control recognition system, method, and storage medium based on image processing. Summary of the Invention

[0004] The purpose of the present invention is to address the shortcomings of the existing technology and provide a face access control recognition system, method and storage medium based on image processing, thereby solving the problem that when the existing method uses Mobilefacenet to calculate the face subgraph, the facial feature values ​​obtained cannot effectively capture the local features in the image and the perception ability of image details is insufficient.

[0005] The present invention is implemented as follows: a face access control recognition method based on image processing, the face access control recognition method based on image processing specifically includes:

[0006] Obtain at least one set of access control collected data, pre-process the access control collected data, and obtain a pre-processed set;

[0007] The face recognition model of the improved Segnet network is constructed by combining distributed shift convolution. The standard image set in the access control database is retrieved and divided into training set, test set and validation set. The face recognition model is iteratively trained using the training set and the converged face recognition model is output.

[0008] Load the preprocessing set, use the preprocessing set as input, execute the face recognition model, and determine whether the preprocessing set is a living target. If the preprocessing set is a living target, the face recognition model will perform recognition detection on the preprocessing set and output the detection result;

[0009] If the pre-processed set is not a living target object, a re-verification instruction is triggered to collect real-time face data in real time, pre-process the real-time face data, and determine whether the real-time face data is a living target object based on the face recognition model. If the real-time face data is a living target object, the real-time face data is replaced with the associated pre-processed set, and the face recognition model performs recognition detection on the pre-processed set and outputs the detection result;

[0010] Obtain the test results, determine whether the test results exceed the preset result threshold, trigger the access control opening and closing instructions based on the preset result threshold, and turn on / off the access control system.

[0011] Preferably, the method for preprocessing access control data includes:

[0012] Load access control data and extract facial feature images from the data;

[0013] Extract the feature points of nose, eyebrows, eyes, mouth and ears in the facial feature image, and calculate the comprehensive grayscale value of the feature points in the facial feature image based on the Eigenfaces algorithm;

[0014] Integrate the facial feature image and the comprehensive grayscale values ​​of the feature points corresponding to the facial feature image, perform equalization processing on the facial feature image based on the cumulative distribution function, and determine the pixel distribution range of the facial feature image;

[0015] Obtaining the pixel distribution range of the facial feature image, and determining whether the pixel distribution range of the facial feature image meets the preset range threshold based on a preset range threshold. If the range meets the preset range threshold, retaining the facial feature image after equalization processing;

[0016] If it does not meet the preset range threshold, the facial feature image after equalization processing is deleted;

[0017] At least one group of facial feature images that meet a preset range threshold is integrated, and the facial feature images are set as a preprocessing set.

[0018] Preferably, the method of iteratively training a face recognition model using a training set specifically includes:

[0019] The Segnet network is used as the initial model of the face recognition model, wherein the initial model includes an encoder and a decoder. The encoder includes five groups of convolutional modules, and each group of the convolutional modules includes three groups of convolutional layers and one group of BN layers. The decoder includes three groups of decoding modules, and each group of the decoding modules includes three groups of upsampling layers, a convolutional layer, and one group of BN layers.

[0020] The initial model introduces the feature classifier of FaceLivenessNet, uses the feature discriminator as the liveness discriminator, and places the liveness discriminator in front of the encoder, connecting the liveness discriminator to the encoder.

[0021] Freeze the convolutional layers in the decoding module, replace them with distributed shifted convolutional structures, and replace the batch normalization layers in the decoding module with cascaded Adaboost modules;

[0022] Initialize and set the recognition node control indicators and hyperparameters of the feature classifier, encoder, and decoder in the initial model. Introduce the attention mechanism to comprehensively define the pixel-level classification layer of the Segnet network and adjust the position of the recognition node.

[0023] Conduct ablation tests on the feature classifier, encoder, and decoder to complete the construction of the initial model;

[0024] Obtain a training set, use the training set as input, set the number of training rounds, iteratively train the initial model, use forward propagation to calculate the loss value, update the model hyperparameters based on the backpropagation algorithm, and after each training, use the validation set to verify the performance of the initial model until the initial model converges;

[0025] Obtain a test set, use the test set as input, execute the converged initial model, and obtain the model test results;

[0026] Based on the preset model accuracy, the model test result is verified to see whether it exceeds the preset model accuracy. If it exceeds the preset model accuracy, the trained face recognition model is output.

[0027] Preferably, the method of introducing the attention mechanism to comprehensively define the pixel-level classification layer of the Segnet network specifically includes:

[0028] Obtain the decoder decoding feature map, and introduce the SimAM attention mechanism into the pixel-level classification layer to infer the 3D attention weights of the facial feature points nose, eyebrows, eyes, mouth, and ears in the decoded feature map;

[0029] The energy function of the neurons in the pixel-level classification layer is defined based on the 3D attention weight of the feature points, and the mean and variance of the energy function are calculated through the energy function of the neurons;

[0030] Combining the mean and variance of the energy function, the improved vector machine is used to establish the feature recognition space of the pixel-level classification layer neurons, and the difference vector of facial feature recognition in the space is calculated;

[0031] Obtain vector calculation results and use them to comprehensively define the pixel-level classification layer of the Segnet network.

[0032] Preferably, the method of executing the face recognition model to determine whether the preprocessed set is a living target object specifically includes:

[0033] Get the preprocessing set, and perform hybrid enhancement on the preprocessing set based on Mosaic+Mixup to obtain the image enhancement set;

[0034] The image enhancement set is loaded, and the feature discriminator extracts interactive feedback of the image enhancement set based on the FaceLivenessNet algorithm. The interactive feedback includes pixel motion patterns, thermal imaging information of facial feature points, and RDB image information.

[0035] Obtain interactive feedback for the image enhancement set and calculate the significance level of the interactive feedback based on principal component analysis;

[0036]

[0037] in, represents the significance level of interactive feedback, Respectively represent the pixel motion mode, facial feature point thermal imaging information, and RDB image information weight coefficients in the interactive feedback, is the number of samples in the image enhancement set, Respectively represent pixel motion mode, facial feature point thermal imaging information, and RDB image information input values;

[0038] A liveness significance threshold is preset to determine whether the significance level of the interactive feedback exceeds the liveness significance threshold. If it exceeds the liveness significance threshold, the preprocessed set is determined to be a live target.

[0039] Preferably, the method for the face recognition model to perform recognition detection on the preprocessing set specifically includes:

[0040] The encoder obtains the preprocessing set, and the convolution module in the encoder performs a first-level convolution on the preprocessing set to obtain a first-level convolution result. The encoding convolution result extracts the edges, lines, and angles of the facial features;

[0041] Obtain the first-level convolution result. The convolution module constructs a facial feature map based on the first-level convolution result, supplements the facial feature map with details based on the preset number of zero-padding layers, and outputs the facial feature map with the supplemented details.

[0042] Load the facial feature map, and the convolution module performs secondary convolution on the facial feature map to increase the void receptive field of the facial feature map and output the secondary convolution result;

[0043] Obtain the secondary convolution results, and the decoding module performs feature fusion on the secondary convolution results to obtain a feature fusion set;

[0044] Load the feature fusion set and build the initial decoding feature graph based on the feature fusion set;

[0045] The cascaded Adaboost module is used to process the initial decoding feature map, and the deviation of the identified initial decoding feature map is corrected for secondary processing, and the decoding feature map and map confidence are output;

[0046] The detection result value of the decoded feature graph is calculated based on the graph confidence, and the detection result is output.

[0047] On the other hand, the present invention also provides a face access control recognition system based on image processing, the face access control recognition system based on image processing specifically includes:

[0048] A data acquisition module is used to obtain at least one set of access control collected data, pre-process the access control collected data, and obtain a pre-processed set;

[0049] The model building module combines distributed shift convolution to build an improved Segnet network face recognition model, retrieves the standard image set from the access control database, divides the standard image set into training set, test set and validation set, uses the training set to iteratively train the face recognition model, and outputs a converged face recognition model;

[0050] The recognition and detection module loads the preprocessing set, uses the preprocessing set as input, executes the face recognition model, and determines whether the preprocessing set is a living target. If the preprocessing set is a living target, the face recognition model performs recognition detection on the preprocessing set and outputs the detection result;

[0051] The secondary verification module is used to obtain the judgment result of whether the pre-processed set is a living target object. If the pre-processed set is not a living target object, a re-verification instruction is triggered to collect real-time face data in real time, pre-process the real-time face data, and judge whether the real-time face data is a living target object based on the face recognition model. If the real-time face data is a living target object, the real-time face data is replaced with the associated pre-processed set. The face recognition model recognizes and detects the pre-processed set and outputs the detection result;

[0052] The access control module is used to obtain the detection results, determine whether the detection results exceed the preset result threshold, trigger the access control opening and closing instructions based on the preset result threshold, and open / close the access control system.

[0053] Preferably, the data acquisition module specifically includes:

[0054] A feature extraction unit is used to load access control data and extract facial feature images from the access control data;

[0055] Comprehensive grayscale calculation unit, used to extract the feature points of nose, eyebrows, eyes, mouth, and ears in the facial feature image, and calculate the comprehensive grayscale value of the feature points in the facial feature image based on the Eigenfaces algorithm;

[0056] An equalization processing unit is used to integrate the facial feature image and the comprehensive grayscale values ​​of the feature points corresponding to the facial feature image, perform equalization processing on the facial feature image based on the cumulative distribution function, and determine the pixel distribution range of the facial feature image;

[0057] The preprocessing set output unit is used to obtain the pixel distribution range of the facial feature image, and judge whether the pixel distribution range of the facial feature image meets the preset range threshold based on the preset range threshold. If it meets the preset range threshold, the facial feature image after equalization processing is retained; if it does not meet the preset range threshold, the facial feature image after equalization processing is deleted, and at least one group of facial feature images that meet the preset range threshold are integrated, and the facial feature images are set as the preprocessing set.

[0058] The present invention also provides a computer-readable storage medium, which stores computer program instructions. When the computer program instructions are executed, the image processing-based face access control recognition method is implemented.

[0059] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0060] In an embodiment of the present invention, a face access control recognition method based on image processing is provided. By extracting the feature points of the nose, eyebrows, eyes, mouth, and ears in the face feature image, the comprehensive grayscale value of the feature points is calculated, thereby facilitating the face recognition model to quickly detect and identify the access control collected data. At the same time, the face recognition model is constructed by combining distributed displacement convolution and Segnet network, which can more effectively capture local features in the image, thereby enhancing the model's perception of image details, and can also combine deep features with shallow features through jump connections, thereby enhancing the model's adaptability to complex backgrounds and different lighting conditions, overcoming the problem that the existing method uses Mobilefacenet to calculate the face subgraph, and cannot effectively capture the local features in the image when obtaining the face feature value, and the insufficient perception of image details.

[0061] In an embodiment of the present invention, when recognizing an image, the comprehensive grayscale values ​​of feature points in a facial feature image are calculated based on the Eigenfaces algorithm, and the facial feature image is equalized based on a cumulative distribution function, so that pixels originally concentrated in certain grayscale ranges are more evenly distributed within the grayscale range of the entire facial feature image. In this way, the contrast of the image can be significantly enhanced, and the details of the facial features can be highlighted. In addition, preprocessing of access control collected data using the Eigenfaces algorithm and the cumulative distribution function can also improve the performance of a facial recognition model, ensure that the facial recognition model is more likely to determine a suitable threshold for segmentation, and thus improve the accuracy and reliability of facial recognition model detection.

[0062] In an embodiment of the present invention, a face recognition model and a training method for the model are provided. The face recognition model is based on a Segnet network and distributed shift convolution, and introduces a feature classifier of FaceLivenessNet, so as to realize the simultaneous detection of living target objects and face recognition detection. The attention mechanism is introduced to comprehensively define the pixel-level classification layer of the Segnet network, thereby combining deep features with shallow features, enhancing the model's adaptability to complex backgrounds and different lighting conditions, and the distributed shift convolution and Segnet network can be combined with the attention mechanism to form an integrated learning system, further improving the accuracy and robustness of face recognition.

[0063] In an embodiment of the present invention, multimodal fusion of living object detection is achieved by combining image enhancement-focused pixel motion patterns, thermal imaging information of facial feature points, and RDB image information. By utilizing the advantages of various modalities in different environments and compensating for each other's shortcomings, the impact of abnormalities or errors in individual modal data on the entire detection system is reduced, thereby improving the robustness of the face recognition model. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 The figure is a flowchart of the face access control recognition method based on image processing provided by the present invention.

[0065] Figure 2 The figure shows a flow chart of the implementation of the method for preprocessing access control data.

[0066] Figure 3 A schematic diagram of the implementation process of the method for iteratively training a face recognition model using a training set is shown.

[0067] Figure 4 A schematic diagram of the implementation process of the method for comprehensively defining the pixel-level classification layer of the Segnet network by introducing the attention mechanism is shown.

[0068] Figure 5A schematic diagram of the implementation flow of the method for executing the face recognition model and determining whether the preprocessing set is a living target object is shown.

[0069] Figure 6 A schematic diagram of the implementation flow of the face recognition model for the recognition and detection method of the preprocessing set is shown.

[0070] Figure 7 It is a structural diagram of the face access control recognition system based on image processing provided by the present invention. DETAILED DESCRIPTION

[0071] Unless otherwise defined, all technical and scientific terms used herein have the same meanings as commonly understood by those skilled in the art to which this application belongs. The terms used in the specification of the application are only for the purpose of describing specific embodiments and are not intended to limit this application. The terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of this application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0072] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0073] The existing method uses Mobilefacenet to calculate the face subgraph, and when obtaining the facial feature value, it cannot effectively capture the local features in the image and has insufficient perception of image details. To address the above problems, we propose a face access control recognition system, method and storage medium based on image processing. In short, when the method is executed, at least one set of access control collection data is first obtained, the access control collection data is preprocessed, and at the same time, a face recognition model of the improved Segnet network is constructed in combination with distributed displacement convolution. Then, the preprocessing set is loaded, and the face recognition model is executed with the preprocessing set as input to determine whether the preprocessing set is a living target. If the preprocessing set is a living target, the face recognition model performs recognition detection on the preprocessing set and outputs the detection result. In an embodiment of the present invention, a face access control recognition method based on image processing is provided. By extracting the feature points of the nose, eyebrows, eyes, mouth, and ears in the face feature image, the comprehensive grayscale value of the feature points is calculated, thereby facilitating the face recognition model to quickly detect and identify the access control collected data. At the same time, the face recognition model is constructed by combining distributed displacement convolution and Segnet network, which can more effectively capture local features in the image, thereby enhancing the model's perception of image details, and can also combine deep features with shallow features through jump connections, thereby enhancing the model's adaptability to complex backgrounds and different lighting conditions, overcoming the problem that the existing method uses Mobilefacenet to calculate the face subgraph, and cannot effectively capture the local features in the image when obtaining the face feature value, and the insufficient perception of image details.

[0074] The embodiment of the present invention provides a face access control recognition method based on image processing, Figure 1 The figure shows a flow chart of the face access control recognition method based on image processing. The face access control recognition method based on image processing specifically includes:

[0075] Step S10, obtaining at least one set of access control collected data, preprocessing the access control collected data to obtain a preprocessed set;

[0076] It should be noted that the access control data is obtained through sensors built into the access control system. The access control data includes but is not limited to images, videos, sounds, and fingerprint data, and the sensors include but are not limited to infrared cameras, RGB cameras, optical fingerprint sensors, and sound collection sensors.

[0077] The access control system is composed of front-end hardware equipment (camera, display, access control lock, fingerprint reader), and the access control system can be vertical or wall-mounted.

[0078] Step S20: constructing an improved Segnet network face recognition model in combination with distributed shift convolution, retrieving a standard image set from the access control database, dividing the standard image set into a training set, a test set, and a validation set, iteratively training the face recognition model using the training set, and outputting a converged face recognition model;

[0079] It should be noted that the data volume ratio of the samples in the training set, test set, and validation set is 3:1:1, and the access control database is stored in a face access control recognition system based on image processing, and at least one set of access control data directories is set up in the system. Each folder in the directory is named after the name, and the image parameter specifications in the access control data directory are consistent. When a picture with the same name is put into the database, the original directory data will be automatically overwritten.

[0080] Step S30, loading the pre-processed set, using the pre-processed set as input, executing the face recognition model, and determining whether the pre-processed set is a living target object;

[0081] Step S40: If the pre-processed set is a living target object, the face recognition model performs recognition detection on the pre-processed set and outputs the detection result;

[0082] Step S50: If the pre-processed set is not a living target object, a re-verification instruction is triggered to collect real-time face data in real time, pre-process the real-time face data, and determine whether the real-time face data is a living target object based on the face recognition model. If the real-time face data is a living target object, the real-time face data is replaced with the associated pre-processed set, and the face recognition model performs recognition detection on the pre-processed set, outputs the detection result, and executes step S60;

[0083] It should be noted that the real-time facial data collected in real time is triggered when it is determined that the first facial recognition is not a living target, and the secondary collection instruction is triggered. The access control system re-enters the collected real-time facial data through voice prompts, and the preprocessing method of the real-time facial data is the same as the method of collecting access control data.

[0084] If the real-time facial data is a living target, the real-time facial data detection result is fed back, the access control is not opened, and the face recognition task is terminated;

[0085] Step S60: Obtain the detection result, determine whether the detection result exceeds a preset result threshold, trigger the access control opening and closing instruction based on the preset result threshold, and turn on / off the access control system.

[0086] In this embodiment, the preset result threshold can be set to 0.9-1. When the detection result is greater than or equal to the result threshold, the access control system unlocks the access lock.

[0087] In an embodiment of the present invention, a face access control recognition method based on image processing is provided. By extracting the feature points of the nose, eyebrows, eyes, mouth, and ears in the face feature image, the comprehensive grayscale value of the feature points is calculated, thereby facilitating the face recognition model to quickly detect and identify the access control collected data. At the same time, the face recognition model is constructed by combining distributed displacement convolution and Segnet network, which can more effectively capture local features in the image, thereby enhancing the model's perception of image details, and can also combine deep features with shallow features through jump connections, thereby enhancing the model's adaptability to complex backgrounds and different lighting conditions, overcoming the problem that the existing method uses Mobilefacenet to calculate the face subgraph, and cannot effectively capture the local features in the image when obtaining the face feature value, and the insufficient perception of image details.

[0088] The embodiment of the present invention provides a method for preprocessing access control collected data. Figure 2 The figure shows a flow chart of the method for preprocessing access control data. The method for preprocessing access control data specifically includes:

[0089] Step S101, loading access control data, and extracting facial feature images from the access control data;

[0090] Step S102, extracting the nose, eyebrows, eyes, mouth, and ear feature points in the facial feature image, and calculating the comprehensive grayscale values ​​of the feature points in the facial feature image based on the Eigenfaces algorithm;

[0091] Step S103, integrating the facial feature image and the comprehensive grayscale values ​​of the feature points corresponding to the facial feature image, performing equalization processing on the facial feature image based on the cumulative distribution function, and determining the pixel distribution range of the facial feature image;

[0092] It should be noted that when performing equalization processing on the facial feature image based on the cumulative distribution function, the histogram of the facial feature image is first calculated. For color images, the three channels (red, green, and blue) can be processed separately. Then, based on the cumulative distribution function, the sum of the occurrence probabilities of all pixels under a certain grayscale value is calculated. Then, the equalization formula is used to map the pixel values ​​of the original image, and finally, the images are recombined to achieve equalization processing of the facial feature image.

[0093] Step S104, obtaining the pixel distribution range of the facial feature image, and determining whether the pixel distribution range of the facial feature image meets the preset range threshold based on the preset range threshold;

[0094] Step S105: if the face meets the preset range threshold, retain the equalized face feature image;

[0095] Step S106: if the image does not meet the preset range threshold, the facial feature image after equalization processing is deleted;

[0096] Step S107 , integrating at least one group of facial feature images that meet a preset range threshold, and setting the facial feature images as a pre-processing set.

[0097] In an embodiment of the present invention, when recognizing an image, the comprehensive grayscale values ​​of feature points in a facial feature image are calculated based on the Eigenfaces algorithm, and the facial feature image is equalized based on a cumulative distribution function, so that pixels originally concentrated in certain grayscale ranges are more evenly distributed within the grayscale range of the entire facial feature image. In this way, the contrast of the image can be significantly enhanced, and the details of the facial features can be highlighted. In addition, preprocessing of access control collected data using the Eigenfaces algorithm and the cumulative distribution function can also improve the performance of a facial recognition model, ensure that the facial recognition model is more likely to determine a suitable threshold for segmentation, and thus improve the accuracy and reliability of facial recognition model detection.

[0098] The embodiment of the present invention provides a method for iteratively training a face recognition model using a training set. Figure 3 The figure shows a flow chart of implementing the method for iteratively training a face recognition model using a training set. The method for iteratively training a face recognition model using a training set specifically includes:

[0099] Step S201, using a Segnet network as the initial model of a face recognition model, wherein the initial model includes an encoder and a decoder, the encoder includes five groups of convolutional modules, and each group of the convolutional modules includes three groups of convolutional layers and one group of BN layers, and the decoder includes three groups of decoding modules, and each group of the decoding modules includes three groups of upsampling layers, a convolutional layer, and one group of BN layers;

[0100] It should be noted that the convolution kernel size of the three groups of convolution layers in the convolution module is 3*3, and the convolution layer contains 128 neurons. The activation function of the convolution module can be a RELU activation function.

[0101] Step S202: Introduce the feature classifier of FaceLivenessNet into the initial model, use the feature discriminator as the liveness discriminator, place the liveness discriminator in front of the encoder, and connect the liveness discriminator to the encoder;

[0102] Step S203: freeze the convolutional layer in the decoding module, replace the convolutional layer with a distributed shifted convolutional structure, and replace the batch normalization layer in the decoding module with a cascaded Adaboost module;

[0103] Step S204: Initialize and set the recognition node control indicators and hyperparameters of the feature classifier, encoder, and decoder in the initial model, introduce the attention mechanism to comprehensively define the Segnet network pixel-level classification layer, and adjust the recognition node position;

[0104] Step S205: Perform ablation tests on the feature classifier, encoder, and decoder to complete the construction of the initial model;

[0105] In this example, during ablation testing, only one component at a time—the feature classifier, encoder, or decoder—is modified. This allows for accurate assessment of each component's impact on model performance. The training process and results of each test are recorded for subsequent analysis and comparison. This example uses loss and runtime as evaluation metrics to ensure the accuracy and reliability of the test results.

[0106] Step S206, obtain a training set, use the training set as input, set training rounds, iteratively train the initial model, use forward propagation to calculate the loss value, update the model hyperparameters based on the backpropagation algorithm, and after each training, use the validation set to verify the performance of the initial model until the initial model converges; in this embodiment, the training rounds can be 200-1000 times.

[0107] Step S207: Obtain a test set, use the test set as input, execute the converged initial model, and obtain the model test results;

[0108] Step S208, verifying whether the model test result exceeds the preset model accuracy based on the preset model accuracy;

[0109] Step S209: If the accuracy exceeds the preset model accuracy, the trained face recognition model is output.

[0110] If the model accuracy does not exceed the preset value, the process returns to step S206 and continues to iteratively train the model.

[0111] In this embodiment, the preset model accuracy may be 0.9-0.98.

[0112] In an embodiment of the present invention, a face recognition model and a training method for the model are provided. The face recognition model is based on a Segnet network and distributed shift convolution, and introduces a feature classifier of FaceLivenessNet, so as to realize the simultaneous detection of living target objects and face recognition detection. The attention mechanism is introduced to comprehensively define the pixel-level classification layer of the Segnet network, thereby combining deep features with shallow features, enhancing the model's adaptability to complex backgrounds and different lighting conditions, and the distributed shift convolution and Segnet network can be combined with the attention mechanism to form an integrated learning system, further improving the accuracy and robustness of face recognition.

[0113] The embodiment of the present invention provides a method for introducing an attention mechanism to comprehensively define the pixel-level classification layer of the Segnet network. Figure 4 The following figure shows a schematic diagram of the implementation process of the method for introducing the attention mechanism to comprehensively define the pixel-level classification layer of the Segnet network. The method for introducing the attention mechanism to comprehensively define the pixel-level classification layer of the Segnet network specifically includes:

[0114] Step S301: Obtain the decoded feature map of the decoder, and introduce the SimAM attention mechanism into the pixel-level classification layer to infer the 3D attention weights of the facial feature points nose, eyebrows, eyes, mouth, and ears in the decoded feature map;

[0115] It should be noted that the attention weights determined by the attention mechanism can provide guidance for 3D modeling, enabling the model to pay more attention to the accurate position and shape of key feature points when reconstructing the face, thereby generating a more realistic and detailed 3D face model.

[0116] Step S302: defining the energy function of the neurons in the pixel-level classification layer based on the 3D attention weights of the feature points, and calculating the mean and variance of the energy function through the energy function of the neurons;

[0117] Step S303, combining the mean and variance of the energy function, using the improved vector machine to establish a feature recognition space for neurons in the pixel-level classification layer, and calculating the difference vector of facial feature recognition in the space;

[0118] Step S304: Obtain vector calculation results, and use the vector calculation results to comprehensively define the Segnet network pixel-level classification layer.

[0119] It should be noted that complex background scenes may be encountered in 3D face modeling. Vector calculation results can help the SegNet network better distinguish between foreground faces and backgrounds, improving the accuracy of pixel-level classification, thereby providing a cleaner data foundation for subsequent 3D modeling. At the same time, vector calculation can overcome the limitations of single features to a certain extent, comprehensively considering multiple factors, making classification results more stable and reliable.

[0120] The embodiment of the present invention provides a method for executing a face recognition model to determine whether a pre-processed set is a living target object. Figure 5 The flowchart of the method for executing the face recognition model and determining whether the preprocessed set is a living target is shown. The method for executing the face recognition model and determining whether the preprocessed set is a living target specifically includes:

[0121] Step S401, obtaining a pre-processing set, and performing hybrid enhancement on the pre-processing set based on Mosaic+Mixup to obtain an image enhancement set;

[0122] Step S402: Loading the image enhancement set, and extracting interactive feedback of the image enhancement set by the feature discriminator based on the FaceLivenessNet algorithm, wherein the interactive feedback includes pixel motion patterns, thermal imaging information of facial feature points, and RDB image information;

[0123] It should be noted that interactive feedback refers to establishing an interactive index with the image enhancement set. The image enhancement set is based on the pixel motion pattern, facial feature point thermal imaging information, and RDB image information fed back by the interactive index. The pixel motion pattern is obtained by analyzing the video sequence using the optical flow method, while the facial feature point thermal imaging information is obtained through an infrared camera, and the RDB image information can be collected through an RDB camera.

[0124] Step S403: obtaining interactive feedback of the image enhancement set, and calculating the significance level of the interactive feedback based on principal component analysis;

[0125]

[0126] in, represents the significance level of interactive feedback, Respectively represent the pixel motion mode, facial feature point thermal imaging information, and RDB image information weight coefficients in the interactive feedback, is the number of samples in the image enhancement set, Respectively represent pixel motion mode, facial feature point thermal imaging information, and RDB image information input values;

[0127] Step S404: Preset a living body significance threshold, and determine whether the significance level of the interactive feedback exceeds the living body significance threshold. If it exceeds the living body significance threshold, determine that the pre-processed set is a living body target.

[0128] In this embodiment, the living body significance threshold can be set to 0.8-0.85.

[0129] In an embodiment of the present invention, multimodal fusion of living object detection is achieved by combining image enhancement-focused pixel motion patterns, thermal imaging information of facial feature points, and RDB image information. By utilizing the advantages of various modalities in different environments and compensating for each other's shortcomings, the impact of abnormalities or errors in individual modal data on the entire detection system is reduced, thereby improving the robustness of the face recognition model.

[0130] The embodiment of the present invention provides a method for a face recognition model to perform recognition detection on a preprocessing set. Figure 6 The figure shows a flow chart of the face recognition model's recognition and detection method for the preprocessing set. The face recognition model's recognition and detection method for the preprocessing set specifically includes:

[0131] Step S501: The encoder obtains a preprocessing set, and the convolution module in the encoder performs a first-level convolution process on the preprocessing set to obtain a first-level convolution result, wherein the encoding convolution result extracts edges, lines, and angles of facial features;

[0132] Step S502: Obtain the first-level convolution result. The convolution module constructs a facial feature map based on the first-level convolution result, supplements the facial feature map with details based on a preset number of zero-padding layers, and outputs the facial feature map with the completed details supplemented.

[0133] Step S503: load the facial feature map, and the convolution module performs secondary convolution on the facial feature map to increase the hole receptive field of the facial feature map and output the secondary convolution result.

[0134] Step S504: obtaining the secondary convolution result, and the decoding module performs feature fusion on the secondary convolution result to obtain a feature fusion set;

[0135] Step S505: Load the feature fusion set, and construct an initial decoding feature graph based on the feature fusion set;

[0136] Step S506: Using a cascaded Adaboost module to process the initial decoding feature graph, performing secondary processing to correct the deviation of the identified initial decoding feature graph, and outputting the decoding feature graph and the graph confidence;

[0137] Step S507: Calculate the detection result value of the decoded feature graph based on the graph confidence, and output the detection result.

[0138] It should be noted that, in this embodiment, the use of cascaded Adaboost modules to process the initial decoding feature graph can correct the initial decoding feature graph, thereby avoiding recognition errors to the greatest extent and ensuring recognition accuracy and stability.

[0139] The embodiment of the present invention provides a face access control recognition system based on image processing, Figure 7 The structure diagram of the face access control recognition system based on image processing is shown. The face access control recognition system based on image processing specifically includes:

[0140] The data acquisition module 100 is used to obtain at least one set of access control data, pre-process the access control data, and obtain a pre-processed set;

[0141] The model building module 200 constructs a face recognition model of the improved Segnet network by combining distributed shift convolution, retrieves a standard image set from the access control database, divides the standard image set into a training set, a test set, and a validation set, iteratively trains the face recognition model using the training set, and outputs a converged face recognition model;

[0142] The recognition and detection module 300 loads the pre-processed set, uses the pre-processed set as input, executes the face recognition model, and determines whether the pre-processed set is a living target object. If the pre-processed set is a living target object, the face recognition model performs recognition detection on the pre-processed set and outputs the detection result;

[0143] Secondary verification module 400, for obtaining a determination result on whether the pre-processed set is a live target object. If the pre-processed set is not a live target object, a re-verification instruction is triggered to collect real-time facial data in real time, pre-process the real-time facial data, and determine whether the real-time facial data is a live target object based on a face recognition model. If the real-time facial data is a live target object, the real-time facial data is replaced with the associated pre-processed set. The face recognition model performs recognition detection on the pre-processed set and outputs the detection result.

[0144] The access control module 500 is used to obtain the detection result, determine whether the detection result exceeds a preset result threshold, trigger the access control opening and closing instruction based on the preset result threshold, and open / close the access control system.

[0145] It should be noted that it is understandable that the face access control recognition system based on image processing provided by the embodiment of the present invention corresponds to the above-mentioned face access control recognition method based on image processing. The explanation, examples, beneficial effects, etc. of the relevant contents can refer to the corresponding contents of the face access control recognition method based on image processing, and will not be repeated here. The data acquisition module 100, model construction module 200, recognition and detection module 300, secondary verification module 400, access control module 500 and the access control database are connected by Bluetooth, DTU or communication interaction.

[0146] In this embodiment, the data acquisition module 100 specifically includes:

[0147] The feature extraction unit 110 is used to load the access control data and extract the facial feature image from the access control data;

[0148] The comprehensive grayscale calculation unit 120 is used to extract the feature points of the nose, eyebrows, eyes, mouth, and ears in the facial feature image, and calculate the comprehensive grayscale values ​​of the feature points in the facial feature image based on the Eigenfaces algorithm;

[0149] An equalization processing unit 130 is used to integrate the facial feature image and the comprehensive grayscale values ​​of the feature points corresponding to the facial feature image, perform equalization processing on the facial feature image based on a cumulative distribution function, and determine the pixel distribution range of the facial feature image;

[0150] The pre-processing set output unit 140 is used to obtain the pixel distribution range of the facial feature image, and determine whether the pixel distribution range of the facial feature image meets the preset range threshold based on the preset range threshold. If it meets the preset range threshold, the facial feature image after equalization processing is retained; if it does not meet the preset range threshold, the facial feature image after equalization processing is deleted, and at least one group of facial feature images that meet the preset range threshold are integrated, and the facial feature images are set as the pre-processing set.

[0151] On the other hand, an embodiment of the present invention further provides a computer-readable storage medium storing computer program instructions, which can be executed by a processor. When the computer program instructions are executed, the method of any of the above embodiments is implemented.

[0152] Finally, it should be noted that the computer-readable storage medium (e.g., memory) herein may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. By way of example and not limitation, non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM), which may act as external cache memory. By way of example and not limitation, RAM may be available in various forms, such as synchronous RAM (DRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), and direct Rambus RAM (DRRAM). The memory devices of the disclosed aspects are intended to include, but are not limited to, these and other suitable types of memory.

[0153] In summary, the present invention provides a face access control recognition system, method and storage medium based on image processing. In an embodiment of the present invention, a face access control recognition method based on image processing is provided. By extracting the feature points of the nose, eyebrows, eyes, mouth and ears in the face feature image, the comprehensive grayscale value of the feature points is calculated, which facilitates the face recognition model to quickly detect and identify the access control collected data. At the same time, the face recognition model is constructed by combining distributed displacement convolution and Segnet network, which can more effectively capture local features in the image, thereby enhancing the model's perception of image details, and can also combine deep features with shallow features through jump connections, thereby enhancing the model's adaptability to complex backgrounds and different lighting conditions, overcoming the problem that the existing method uses Mobilefacenet to calculate the face subgraph, and cannot effectively capture local features in the image when obtaining the face feature value, and the perception of image details is insufficient.

[0154] It should be noted that for the aforementioned embodiments, for simplicity of description, they are all expressed as a series of action combinations. However, those skilled in the art should be aware that the present invention is not limited by the order of the actions described, because according to the present invention, certain steps may be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present invention.

[0155] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the scope of protection of the invention. Obviously, the embodiments described are only some embodiments of the present invention, rather than all embodiments. Based on these embodiments, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in this field can still combine, add, delete or make other adjustments to the features in the various embodiments of the present invention according to the circumstances without conflict, without making creative work, so as to obtain different other technical solutions that do not deviate from the concept of the present invention in essence, and these technical solutions also fall within the scope of protection of the present invention.

Claims

1. The face access control recognition method based on image processing is characterized by: The face access control recognition method based on image processing specifically includes: Obtain at least one set of access control collected data, pre-process the access control collected data, and obtain a pre-processed set; The face recognition model of the improved Segnet network is constructed by combining distributed shift convolution. The standard image set in the access control database is retrieved and divided into training set, test set and validation set. The face recognition model is iteratively trained using the training set and the converged face recognition model is output. Load the preprocessing set, use the preprocessing set as input, execute the face recognition model, and determine whether the preprocessing set is a living target. If the preprocessing set is a living target, the face recognition model will perform recognition detection on the preprocessing set and output the detection result; If the pre-processed set is not a living target object, a re-verification instruction is triggered to collect real-time face data in real time, pre-process the real-time face data, and determine whether the real-time face data is a living target object based on the face recognition model. If the real-time face data is a living target object, the real-time face data is replaced with the associated pre-processed set, and the face recognition model performs recognition detection on the pre-processed set and outputs the detection result; Obtain the test results, determine whether the test results exceed the preset result threshold, trigger the access control opening and closing instructions based on the preset result threshold, and open / close the access control system; The method of executing the face recognition model to determine whether the preprocessed set is a living target object specifically includes: Get the preprocessing set, and perform hybrid enhancement on the preprocessing set based on Mosaic+Mixup to obtain the image enhancement set; The image enhancement set is loaded, and the feature discriminator extracts interactive feedback of the image enhancement set based on the FaceLivenessNet algorithm. The interactive feedback includes pixel motion patterns, thermal imaging information of facial feature points, and RDB image information. Obtain interactive feedback for the image enhancement set and calculate the significance level of the interactive feedback based on principal component analysis; in, represents the significance level of interactive feedback, Respectively represent the pixel motion mode, facial feature point thermal imaging information, and RDB image information weight coefficients in the interactive feedback, is the number of samples in the image enhancement set, Respectively represent pixel motion mode, facial feature point thermal imaging information, and RDB image information input values; A liveness significance threshold is preset to determine whether the significance level of the interactive feedback exceeds the liveness significance threshold. If it exceeds the liveness significance threshold, the preprocessed set is determined to be a live target.

2. The face access control recognition method based on image processing according to claim 1, characterized in that: The method for preprocessing access control collected data specifically includes: Load access control data and extract facial feature images from the data; Extract the feature points of nose, eyebrows, eyes, mouth and ears in the facial feature image, and calculate the comprehensive grayscale value of the feature points in the facial feature image based on the Eigenfaces algorithm; Integrate the facial feature image and the comprehensive grayscale values ​​of the feature points corresponding to the facial feature image, perform equalization processing on the facial feature image based on the cumulative distribution function, and determine the pixel distribution range of the facial feature image; Obtaining the pixel distribution range of the facial feature image, and determining whether the pixel distribution range of the facial feature image meets the preset range threshold based on a preset range threshold. If the range meets the preset range threshold, retaining the facial feature image after equalization processing; If it does not meet the preset range threshold, the facial feature image after equalization processing is deleted; At least one group of facial feature images that meet a preset range threshold is integrated, and the facial feature images are set as a preprocessing set.

3. The face access control recognition method based on image processing according to claim 1, characterized in that: The method of iteratively training a face recognition model using a training set specifically includes: The Segnet network is used as the initial model of the face recognition model, wherein the initial model includes an encoder and a decoder. The encoder includes five groups of convolutional modules, and each group of the convolutional modules includes three groups of convolutional layers and one group of BN layers. The decoder includes three groups of decoding modules, and each group of the decoding modules includes three groups of upsampling layers, a convolutional layer, and one group of BN layers. The initial model introduces the feature classifier of FaceLivenessNet, uses the feature discriminator as the liveness discriminator, and places the liveness discriminator in front of the encoder, connecting the liveness discriminator to the encoder. Freeze the convolutional layers in the decoding module, replace them with distributed shifted convolutional structures, and replace the batch normalization layers in the decoding module with cascaded Adaboost modules; Initialize and set the recognition node control indicators and hyperparameters of the feature classifier, encoder, and decoder in the initial model. Introduce the attention mechanism to comprehensively define the pixel-level classification layer of the Segnet network and adjust the position of the recognition node. Conduct ablation tests on the feature classifier, encoder, and decoder to complete the construction of the initial model; Obtain a training set, use the training set as input, set the number of training rounds, iteratively train the initial model, use forward propagation to calculate the loss value, update the model hyperparameters based on the backpropagation algorithm, and after each training, use the validation set to verify the performance of the initial model until the initial model converges; Obtain a test set, use the test set as input, execute the converged initial model, and obtain the model test results; Based on the preset model accuracy, the model test result is verified to see whether it exceeds the preset model accuracy. If it exceeds the preset model accuracy, the trained face recognition model is output.

4. The face access control recognition method based on image processing according to claim 3, characterized in that: The method of introducing the attention mechanism to comprehensively define the pixel-level classification layer of the Segnet network specifically includes: Obtain the decoder decoding feature map, and introduce the SimAM attention mechanism into the pixel-level classification layer to infer the 3D attention weights of the facial feature points nose, eyebrows, eyes, mouth, and ears in the decoded feature map; The energy function of the neurons in the pixel-level classification layer is defined based on the 3D attention weight of the feature points, and the mean and variance of the energy function are calculated through the energy function of the neurons; Combining the mean and variance of the energy function, the improved vector machine is used to establish the feature recognition space of the pixel-level classification layer neurons, and the difference vector of facial feature recognition in the space is calculated; Obtain vector calculation results and use them to comprehensively define the pixel-level classification layer of the Segnet network.

5. The face access control recognition method based on image processing according to claim 4, characterized in that: The method for the face recognition model to perform recognition detection on the preprocessing set specifically includes: The encoder obtains the preprocessing set, and the convolution module in the encoder performs a first-level convolution on the preprocessing set to obtain a first-level convolution result. The encoding convolution result extracts the edges, lines, and angles of the facial features; Obtain the first-level convolution result. The convolution module constructs a facial feature map based on the first-level convolution result, supplements the facial feature map with details based on the preset number of zero-padding layers, and outputs the facial feature map with the supplemented details. The facial feature map is loaded, and the convolution module performs secondary convolution on the facial feature map to increase the void receptive field of the facial feature map and output the secondary convolution result.

6. The face access control recognition method based on image processing according to claim 5, characterized in that: The method for the face recognition model to perform recognition detection on the preprocessing set specifically includes: Obtain the secondary convolution results, and the decoding module performs feature fusion on the secondary convolution results to obtain a feature fusion set; Load the feature fusion set and build the initial decoding feature graph based on the feature fusion set; The cascaded Adaboost module is used to process the initial decoding feature map, and the deviation of the identified initial decoding feature map is corrected for secondary processing, and the decoding feature map and map confidence are output; The detection result value of the decoded feature graph is calculated based on the graph confidence, and the detection result is output.

7. A facial access control recognition system based on image processing, for implementing the facial access control recognition method based on image processing according to any one of claims 1 to 6, characterized in that: The face access control recognition system based on image processing specifically includes: A data acquisition module is used to obtain at least one set of access control collected data, pre-process the access control collected data, and obtain a pre-processed set; The model building module combines distributed shift convolution to build an improved Segnet network face recognition model, retrieves the standard image set from the access control database, divides the standard image set into training set, test set and validation set, uses the training set to iteratively train the face recognition model, and outputs a converged face recognition model; The recognition and detection module loads the preprocessing set, uses the preprocessing set as input, executes the face recognition model, and determines whether the preprocessing set is a living target. If the preprocessing set is a living target, the face recognition model performs recognition detection on the preprocessing set and outputs the detection result; The secondary verification module is used to obtain the judgment result of whether the pre-processed set is a living target object. If the pre-processed set is not a living target object, a re-verification instruction is triggered to collect real-time face data in real time, pre-process the real-time face data, and judge whether the real-time face data is a living target object based on the face recognition model. If the real-time face data is a living target object, the real-time face data is replaced with the associated pre-processed set. The face recognition model recognizes and detects the pre-processed set and outputs the detection result; The access control module is used to obtain the detection results, determine whether the detection results exceed the preset result threshold, trigger the access control opening and closing instructions based on the preset result threshold, and open / close the access control system.

8. The face access control recognition system based on image processing according to claim 7, characterized in that: The data acquisition module specifically includes: A feature extraction unit is used to load access control data and extract facial feature images from the access control data; Comprehensive grayscale calculation unit, used to extract the feature points of nose, eyebrows, eyes, mouth, and ears in the facial feature image, and calculate the comprehensive grayscale value of the feature points in the facial feature image based on the Eigenfaces algorithm; An equalization processing unit is used to integrate the facial feature image and the comprehensive grayscale values ​​of the feature points corresponding to the facial feature image, perform equalization processing on the facial feature image based on the cumulative distribution function, and determine the pixel distribution range of the facial feature image; The preprocessing set output unit is used to obtain the pixel distribution range of the facial feature image, and judge whether the pixel distribution range of the facial feature image meets the preset range threshold based on the preset range threshold. If it meets the preset range threshold, the facial feature image after equalization processing is retained; if it does not meet the preset range threshold, the facial feature image after equalization processing is deleted, and at least one group of facial feature images that meet the preset range threshold are integrated, and the facial feature images are set as the preprocessing set.

9. A computer-readable storage medium, characterized in that Computer program instructions are stored, and when the computer program instructions are executed, the face access control recognition method based on image processing according to any one of claims 1 to 6 is implemented.

Images