Process blocking method and system based on personnel image recognition

Abstract

The application belongs to the field of personnel image recognition, and particularly discloses a process blocking method and system based on personnel image recognition, which comprises image acquisition, image preprocessing, personnel recognition and decision judgment. The scheme constructs an image enhancement model comprising a resolution improvement block, a denoising enhancement block and an image fusion unit, effectively improves the image resolution through a super-resolution network and residual learning, uses a stacked denoising autoencoder to specifically process low-light noise, improves the light conditions of the image, and improves the overall quality of the image; a morphological mask method and an angle point detection algorithm are used to perform multi-color space conversion on the fused image, generate an initial skin region mask, fully utilize the advantages of different color spaces, narrow the detection range through special morphological mask processing, and combine bilinear interpolation sub-pixel accuracy positioning to effectively solve the problem of inaccurate face detection and key point positioning in a complex environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of personnel image recognition, and particularly relates to a process blocking method and system based on personnel image recognition. BACKGROUND

[0002] The process blocking based on face image recognition is a method of using image recognition technology to extract features of faces in collected images, comparing the features with personnel features stored in a database in advance, and triggering a process blocking action if a matching result meets a pre-set blocking rule. However, in the field of personnel image recognition, initial images are often affected by low resolution and low light noise, which leads to a decrease in subsequent recognition accuracy. Traditional methods are difficult to simultaneously consider resolution improvement and noise removal when processing such images, and the fusion effect is not good, which seriously affects the accuracy of face detection and feature extraction. Meanwhile, in the face detection process, complex environmental factors such as different light conditions, various skin color differences and noise in images can easily lead to false detection and missed detection of skin regions, and further affect the accurate detection of facial key points, resulting in problems such as pseudo-corner points, which seriously affects the accuracy of personnel recognition. SUMMARY

[0003] In view of the above problems, the present application provides a process blocking method and system based on personnel image recognition, which can effectively solve the technical problems that in the field of personnel image recognition, initial images are often affected by low resolution and low light noise, which leads to a decrease in subsequent recognition accuracy, and traditional methods are difficult to simultaneously consider resolution improvement and noise removal when processing such images, and the fusion effect is not good, which seriously affects the accuracy of face detection and feature extraction. The present application constructs an image enhancement model including a resolution improvement block, a denoising enhancement block and an image fusion unit, effectively improves the image resolution through super-resolution network and residual learning, makes the image details clearer, and improves the overall quality of the image by using the stacked denoising autoencoder to specifically process the low light noise and improve the light conditions of the image. In view of the technical problems that in the face detection process, complex environmental factors such as different light conditions, various skin color differences and noise in images can easily lead to false detection and missed detection of skin regions, and further affect the accurate detection of facial key points, resulting in problems such as pseudo-corner points, which seriously affects the accuracy of personnel recognition, the present application adopts a morphological mask method and a corner point detection algorithm to preliminarily detect the skin region, performs multi-color space conversion on the fused image to generate an initial skin region mask, fully utilizes the advantages of different color spaces, effectively improves the accuracy of skin region detection, narrows the detection range through special morphological mask processing, and combines bilinear interpolation sub-pixel accuracy positioning to effectively solve the problem of inaccurate face detection and key point positioning in complex environments.

[0004] The technical scheme adopted by the present application is as follows: The present application provides a process blocking method based on personnel image recognition, which comprises the following steps:

[0005] Step S1: image acquisition, deploy a high-definition camera to acquire an initial personnel image;

[0006] Step S2: image preprocessing, denoising and enhancing the initial personnel image based on a convolutional neural network and a stacked denoising autoencoder, processing low-resolution and low-light noise through two deep learning blocks respectively, and obtaining a fused image by using an image fusion technology;

[0007] Step S3: personnel recognition, using a morphological mask method and an angle point detection algorithm to intelligently detect the face in the fused image and output face key point coordinates;

[0008] Step S4: decision judgment, constructing an authorized personnel face feature database, extracting a feature vector of the face key point coordinates using a pre-trained ArcFace model, comparing the feature vector with feature data in the authorized personnel face feature database, setting a similarity threshold to determine whether it is an authorized personnel, if it is an authorized personnel, sending an allow signal, if it is not an authorized personnel, triggering a blocking process.

[0009] Further, in step S2, the image preprocessing specifically comprises the following steps:

[0010] Step S21: constructing an image enhancement model, the image enhancement model comprising a resolution improvement block, a denoising enhancement block and an image fusion unit, the resolution improvement block being used to improve the resolution of the initial personnel image to obtain a high-resolution image, the denoising enhancement block being used to improve the illumination effect and noise of the initial personnel image to obtain a denoised image, and the image fusion unit being used to fuse the high-resolution image and the denoised image to obtain a fused image;

[0011] Step S22: training the resolution improvement block, establishing and initializing a super-resolution network as the resolution improvement block, the super-resolution network taking ReLU as an activation function, introducing residual learning in the super-resolution network for training to avoid the gradient vanishing problem in deep network training, and the formula used being as follows:

[0012] ;

[0013] In the formula, is a high-resolution image output by the super-resolution network, is the initial personnel image, is a learning residual;

[0014] In the training process, the gradient range is set, the gradient clipping technology is adopted to prevent gradient explosion, and the loss function of the resolution enhancement block is defined as the Euclidean distance between the high-resolution image and the initial person image.

[0015] Step S23: training the denoising enhancement block, synthesizing dark noise processing by adding Gaussian noise to the initial person image to obtain a darkened image, and the formula used is as follows:

[0016] ;

[0017] In the formula, is the darkened image, is the maximum pixel intensity, is a gamma adjustment factor, when is less than 1, the initial person image is brightened, when is greater than 1, the initial person image is darkened, is Gaussian noise, , is the noise variance, is a random value with a value range of (0, 1);

[0018] A stacked denoising autoencoder is established and initialized as the denoising enhancement block, the stacked denoising autoencoder includes three autoencoders, the darkened image is segmented into blocks and input into the stacked denoising autoencoder for denoising to obtain a denoised image.

[0019] Step S24: image fusion, using the Symlet wavelet-based image fusion technology to fuse the high-resolution image and the denoised image, respectively performing discrete wavelet transform on the high-resolution image and the denoised image to obtain high-resolution wavelet coefficients and denoised wavelet coefficients, comparing the absolute values of the high-resolution wavelet coefficients and the denoised wavelet coefficients at each pixel position, selecting the larger wavelet coefficient as the fused wavelet coefficient, and performing inverse discrete wavelet transform on the fused wavelet coefficient to obtain a fused image.

[0020] Further, in step S3, the person recognition specifically includes the following steps:

[0021] Step S31: initial inspection of the skin area, performing multi-color space conversion on the fused image to convert it into an RGB space image, an HSV space image and an NTSC space image, for the RGB space image, establishing a skin color model using a skin color clustering algorithm to screen a suspected skin area, for the HSV space image, extracting the skin color range in the hue, saturation and brightness components through threshold segmentation, for the NTSC space image, filtering non-skin color areas using the I and Q components of the YIQ color model, and using a six-sigma statistical method to fuse the detection results of the RGB space image, the HSV space image and the NTSC space image to generate an initial skin area mask.

[0022] Step S32: Special morphological masking: Draw a rectangular ROI based on the initial skin region mask, generate a binary mask, and crop the fused image according to the boundary coordinates of the rectangular ROI to obtain a facial focus image;

[0023] Step S33: Facial corner detection. A preset corner threshold is set, and the facial corner response value is calculated. If the facial corner response value is greater than the corner threshold, it is marked as a candidate corner. The facial corner response function is defined as the minimum intensity change value of a pixel in the facial focused image across all possible directions. The formula used is as follows:

[0024] ;

[0025] In the formula, It is the facial corner response function. It is the center pixel of the facial focus image. and These are the grayscale values ​​of two adjacent pixels. It is the grayscale value of the center pixel in the facial focus image;

[0026] Step S34: Multi-grid layered processing. A pyramid multi-grid structure is used to decompose the facial focus image into layers of different resolutions. Candidate corner points are selected at the low-resolution layers, and the process is refined layer by layer to the original resolution layer. Bilinear interpolation is used to perform sub-pixel precision localization of the candidate corner points to obtain optimized corner point response values. The formula used is as follows:

[0027] ;

[0028] In the formula, It optimizes the corner response value. and These are the corner response values ​​of two adjacent grid layers;

[0029] Step S35: Key point localization. Combine candidate corner points with facial prior knowledge to establish a facial key point model. Verify the spatial relationship of key points using the Delaunay triangulation algorithm, eliminate false corner points, and output accurate facial key point coordinates.

[0030] The present invention provides a process blocking system based on personnel image recognition, comprising an image acquisition module, an image preprocessing module, a personnel recognition module, and a decision judgment module;

[0031] The image acquisition module is equipped with a high-definition camera to capture initial images of people.

[0032] The image preprocessing module is based on a convolutional neural network and a stacked denoising autoencoder to denoise and enhance the initial personnel image, processes low resolution and low light noise through two deep learning blocks respectively, and obtains a fused image by using an image fusion technology;

[0033] The personnel recognition module uses a morphological mask method and an angle point detection algorithm to intelligently detect the face in the fused image and outputs face key point coordinates.

[0034] The decision judgment module constructs an authorized personnel face feature database, extracts a feature vector of the face key point coordinates by using a pre-trained ArcFace model, compares the feature vector with feature data in the authorized personnel face feature database, sets a similarity threshold to judge whether the personnel is authorized or not, sends an allowing signal if the personnel is recognized as authorized, and triggers a blocking process if the personnel is not recognized as authorized.

[0035] The above scheme has the following beneficial effects:

[0036] (1) In the field of personnel image recognition, the initial image is often affected by low resolution and low light noise, which leads to a decrease in subsequent recognition accuracy. Traditional methods are difficult to simultaneously consider resolution improvement and noise removal, and the fusion effect is not good, which seriously affects the accuracy of face detection and feature extraction. The present scheme constructs an image enhancement model including a resolution improvement block, a denoising enhancement block and an image fusion unit, effectively improves the image resolution by using a super-resolution network and residual learning, makes the image details clearer, and improves the image quality as a whole by using a stacked denoising autoencoder to specifically process low light noise.

[0037] (2) In the face detection process, complex environmental factors such as different light conditions, various skin color differences and image noise can easily lead to false detection and missed detection of the skin area, thereby affecting the accurate detection of the face key point, resulting in pseudo-angle point and other problems, and the key point positioning accuracy is not high, which seriously affects the accuracy of personnel recognition. The present scheme uses a morphological mask method and an angle point detection algorithm to preliminarily detect the skin area, converts the fused image into multiple color spaces to generate an initial skin area mask, fully utilizes the advantages of different color spaces, effectively improves the accuracy of skin area detection, narrows the detection range by using a special morphological mask processing, and effectively solves the problem of inaccurate face detection and key point positioning in complex environments by using a bilinear interpolation sub-pixel accuracy positioning. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 A flowchart of a process blocking method based on personnel image recognition is provided.

[0039] Figure 2 A structural connection diagram of a process blocking system based on personnel image recognition provided by the present application.

[0040] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, and are used to explain the present application together with embodiments of the present application, and do not constitute a limitation on the present application. DETAILED DESCRIPTION

[0041] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0042] In the description of the present application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0043] Embodiment one: refer to Figure 1 The present embodiment provides a process blocking method and system based on personnel image recognition. The process blocking method based on personnel image recognition comprises the following steps:

[0044] Step S1: image acquisition, deploy a high-definition camera to acquire initial personnel images;

[0045] Step S2: image preprocessing, denoising and enhancing the initial personnel images based on a convolutional neural network and a stacked denoising autoencoder, processing low-resolution and low-light noise through two deep learning blocks, and obtaining a fused image using image fusion technology;

[0046] Step S3: personnel recognition, using a morphological mask method and a corner detection algorithm to intelligently detect the face in the fused image, and outputting face key point coordinates;

[0047] Step S4: decision judgment, constructing an authorized personnel face feature database, extracting a feature vector of the face key point coordinates using a pre-trained ArcFace model, comparing the feature vector with feature data in the authorized personnel face feature database, setting a similarity threshold to determine whether it is an authorized personnel, sending an allow signal if it is an authorized personnel, and triggering a blocking process if it is not an authorized personnel.

[0048] Embodiment two: refer to Figure 1 This embodiment is based on the above embodiment, in step S2, the image preprocessing specifically includes the following steps:

[0049] Step S21: constructing an image enhancement model, the image enhancement model includes a resolution enhancement block, a denoising enhancement block and an image fusion unit, the resolution enhancement block is used to enhance the resolution of the initial personnel image, to obtain a high-resolution image, the denoising enhancement block is used to improve the illumination effect and noise of the initial personnel image, to obtain a denoised image, the image fusion unit is used to fuse the high-resolution image and the denoised image, to obtain a fused image;

[0050] Step S22: training the resolution enhancement block, establishing and initializing a super-resolution network as the resolution enhancement block, the super-resolution network takes ReLU as the activation function, residual learning is introduced in the super-resolution network for training, to avoid the gradient vanishing problem in deep network training, the formula is as follows:

[0051] ;

[0052] In the formula, is the high-resolution image output by the super-resolution network, is the initial personnel image, is the learning residual;

[0053] In the training process, the gradient range is set, the gradient clipping technology is adopted to prevent gradient explosion, and the loss function of the resolution enhancement block is defined as the Euclidean distance between the high-resolution image and the initial personnel image;

[0054] Step S23: training the denoising enhancement block, adding Gaussian noise to the initial personnel image to perform synthetic dark noise processing to obtain a darkened image, the formula is as follows:

[0055] ;

[0056] In the formula, is the darkened image, is the maximum pixel intensity, is a gamma adjustment factor, when is less than 1, the initial personnel image becomes brighter, when is greater than 1, the initial personnel image becomes darker, is Gaussian noise, , is the noise variance, is a random value with a value range of (0, 1);

[0057] A stacked denoising autoencoder is established and initialized as a denoising enhancement block, the stacked denoising autoencoder comprises three autoencoders, the darkened image is segmented into blocks and input into the stacked denoising autoencoder for denoising to obtain a denoised image;

[0058] Step S24: image fusion, a high-resolution image and a denoised image are fused by using a Symlet wavelet-based image fusion technology, the high-resolution image and the denoised image are respectively subjected to discrete wavelet transform to obtain high-resolution wavelet coefficients and denoised wavelet coefficients, absolute values of the high-resolution wavelet coefficients and the denoised wavelet coefficients at each pixel position are compared, and a larger wavelet coefficient is selected as a fused wavelet coefficient, and the fused wavelet coefficient is subjected to inverse discrete wavelet transform to obtain a fused image.

[0059] Embodiment three: refer to Figure 1 This embodiment is based on the above-mentioned embodiment, and the personnel identification in step S3 specifically comprises the following steps:

[0060] Step S31: initial detection of a skin region, multi-color space conversion is performed on the fused image to respectively convert the fused image into an RGB space image, an HSV space image and an NTSC space image, for the RGB space image, a skin color model is established by using a skin color clustering algorithm to screen a suspected skin region, for the HSV space image, a skin color range in a hue, saturation and brightness component is extracted by threshold segmentation, for the NTSC space image, non-skin color regions are filtered by using I and Q components of a YIQ color model, and a six-sigma statistical method is used to fuse detection results of the RGB space image, the HSV space image and the NTSC space image to generate an initial skin region mask;

[0061] Step S32: special morphological mask processing, a rectangular ROI is drawn based on the initial skin region mask to generate a binary mask, and the fused image is cropped according to boundary coordinates of the rectangular ROI to obtain a face focusing image;

[0062] Step S33: face corner point detection, a preset corner point threshold is set, a face corner point response value is calculated, if the face corner point response value is greater than the corner point threshold, a candidate corner point is marked, and a face corner point response function is defined as a minimum intensity change value of a pixel point in the face focusing image in all possible directions, and the formula is as follows:

[0063] ;

[0064] In the formula, is the face corner point response function, is a center pixel point of the face focusing image, and are gray values of two adjacent pixels, is a gray value of the center pixel point of the face focusing image;

[0065] Step S34: multi-grid hierarchical processing, using a pyramid multi-grid structure, the face focus image is decomposed into different resolution levels, candidate corner points are screened in low resolution levels, and are refined layer by layer to the original resolution layer, and the bilinear interpolation is combined to sub-pixel accuracy positioning of the candidate corner points, to obtain the optimized corner point response value, the formula is as follows:

[0066] ;

[0067] In the formula, is the optimized corner point response value, and are the corner point response values of two adjacent grid layers.

[0068] Step S35: key point positioning, combining the candidate corner points with the face priori knowledge, establishing a face key point model, verifying the spatial relationship of the key points through the Delaunay triangulation algorithm, excluding false corner points, and outputting accurate face key point coordinates.

[0069] Embodiment four: refer to Figure 2 , this embodiment is based on the above-mentioned embodiment, the present application provides a process blocking system based on personnel image recognition, including image acquisition module, image preprocessing module, personnel identification module and decision-making module;

[0070] The image acquisition module is deployed with a high-definition camera to collect initial personnel images.

[0071] The image preprocessing module denoises and enhances the initial personnel images based on convolutional neural network and stacked denoising autoencoder, processes low resolution and low light noise through two deep learning blocks respectively, and acquires a fusion image by using image fusion technology.

[0072] The personnel identification module uses a morphological mask method and an angle point detection algorithm to intelligently detect the face in the fusion image and output face key point coordinates.

[0073] The decision-making module constructs an authorized personnel face feature database, extracts a feature vector of the face key point coordinates by using a pre-trained ArcFace model, compares the feature vector with feature data in the authorized personnel face feature database, sets a similarity threshold to judge whether it is an authorized personnel, if it is an authorized personnel, sends an allow signal, if it is not an authorized personnel, triggers a blocking process.

[0074] Embodiment five: this embodiment is based on the above-mentioned embodiments, and in the second embodiment, a super-resolution network is established and initialized as a resolution enhancement block, the super-resolution network adopts a 25-layer convolutional neural network, takes ReLU as an activation function, contains 64 3*3 convolution kernels, and in the training process, the initial learning rate is set to 0.1 and gradually reduced to 0.0001;

[0075] The darkened image is divided into blocks, the size of the blocks is 17*17 pixels, and in the training of the stacked denoising autoencoder, the learning rate of the first two layers of the denoising autoencoder is set to 0.1 and the learning rate of the last layer is set to 0.01.

[0076] Embodiment six: this embodiment is based on the above-mentioned embodiments, and in the third embodiment, the candidate corner points are combined with the face priori knowledge to establish a face key point model, the face priori knowledge includes the symmetry of the two eyes, the proportion of the distance between the eyes and eyebrows, and the relative position of the mouth and nose, and the face key points in the face key point model include the left eye corner, the right eye corner, the left eyebrow corner, the right eyebrow corner, the left lip corner, the right lip corner, the left nostril and the right nostril.

[0077] Embodiment seven: this embodiment is based on the above-mentioned embodiments, and in actual application, the living body detection is added, the blink detection is integrated, and the photo or video attack is prevented.

[0078] It should be noted that, in this paper, relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment.

[0079] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

[0080] The above describes the present application and its embodiments, which are not restrictive, and the embodiments shown in the drawings are only one of the embodiments of the present application, and the actual structure is not limited thereto. In general, if a person skilled in the art is inspired thereby, without departing from the purpose of the present application, without creative design, similar structure and embodiments to the technical solution can be designed and belong to the protection scope of the present application.

Claims

1. A process blocking method based on personnel image recognition, characterized in that, Includes the following steps: Step S1: Image acquisition, deploy high-definition cameras to acquire initial images of personnel; Step S2: Image preprocessing. The initial personnel image is denoised and enhanced based on a convolutional neural network and a stacked denoising autoencoder. Low-resolution and low-light noise are processed by two deep learning blocks respectively, and a fused image is obtained by using image fusion technology. Step S3: Person recognition. A morphological masking method and corner detection algorithm are used to intelligently detect faces in the fused image, outputting the coordinates of facial key points. This includes the following steps: Step S31: Initial skin region detection, perform multi-color space conversion on the fused image to generate an initial skin region mask; Step S32: Special morphological masking: Draw a rectangular ROI based on the initial skin region mask, generate a binary mask, and crop the fused image according to the boundary coordinates of the rectangular ROI to obtain a facial focus image; Step S33: Facial corner detection, preset corner threshold, calculate facial corner response value, if the facial corner response value is greater than the corner threshold, mark it as a candidate corner, and define the facial corner response function as the minimum intensity change value of a pixel in all possible directions in the facial focused image; Step S34: Multi-grid layered processing. A pyramid multi-grid structure is adopted to decompose the facial focus image into layers of different resolutions. Candidate corner points are selected at the low-resolution layers and refined layer by layer to the original resolution layer. Bilinear interpolation is used to perform sub-pixel precision positioning of the candidate corner points to obtain optimized corner point response values. Step S35: Key point localization. Combine candidate corner points with facial prior knowledge to establish a facial key point model. Verify the spatial relationship of key points using the Delaunay triangulation algorithm and output accurate facial key point coordinates. Step S4: Decision-making and judgment. Construct an authorized personnel facial feature database, use a pre-trained ArcFace model to extract feature vectors of facial key point coordinates, compare them with feature data in the authorized personnel facial feature database, set a similarity threshold to determine whether it is an authorized person. If it is identified as an authorized person, send an allow signal; if it is not identified as an authorized person, trigger the blocking process.

2. The process blocking method based on personnel image recognition according to claim 1, characterized in that, In step S2, the image preprocessing specifically includes the following steps: Step S21: Construct an image enhancement model, which includes a resolution enhancement block, a denoising enhancement block, and an image fusion unit. The resolution enhancement block is used to obtain a high-resolution image, the denoising enhancement block is used to obtain a denoised image, and the image fusion unit is used to obtain a fused image. Step S22: Train the resolution upscaling block. Establish and initialize a super-resolution network as the resolution upscaling block. The super-resolution network uses ReLU as the activation function and introduces residual learning for training. During training, a gradient range is set, and gradient clipping is used to prevent gradient explosion. The loss function of the resolution upscaling block is defined as the Euclidean distance between the high-resolution image and the initial person image. Step S23: Train the denoising enhancement block, add Gaussian noise to the initial personnel image to synthesize dark noise, and obtain a darkened image; A stacked denoising autoencoder is established and initialized as a denoising enhancement block. The stacked denoising autoencoder contains three autoencoders. The darkened image is segmented into blocks and input into the stacked denoising autoencoder for denoising to obtain a denoised image. Step S24: Image fusion. Using Symlet wavelet-based image fusion technology, the high-resolution image and the denoised image are fused. Discrete wavelet transform is performed on the high-resolution image and the denoised image respectively to obtain high-resolution wavelet coefficients and denoised wavelet coefficients. The absolute values ​​of the high-resolution wavelet coefficients and denoised wavelet coefficients at each pixel position are compared, and the larger wavelet coefficient is selected as the fusion wavelet coefficient. Inverse discrete wavelet transform is performed on the fusion wavelet coefficients to obtain the fused image.

3. A process blocking system based on personnel image recognition according to claim 1, used to implement a process blocking method based on personnel image recognition according to any one of claims 1-2, characterized in that, The process blocking system based on personnel image recognition includes an image acquisition module, an image preprocessing module, a personnel recognition module, and a decision-making module. The image acquisition module is equipped with a high-definition camera to capture initial images of people. The image preprocessing module denoises and enhances the initial personnel image based on a convolutional neural network and a stacked denoising autoencoder. It processes low-resolution and low-light noise through two deep learning blocks respectively, and uses image fusion technology to obtain a fused image. The personnel recognition module uses a morphological masking method and a corner detection algorithm to intelligently detect faces in the fused image and outputs the coordinates of facial key points. The decision-making module constructs a facial feature database of authorized personnel, uses a pre-trained ArcFace model to extract feature vectors of facial key point coordinates, compares them with feature data in the facial feature database of authorized personnel, sets a similarity threshold to determine whether the person is an authorized person, sends an permission signal if the person is identified as an authorized person, and triggers a blocking process if the person is not identified as an authorized person.

Images