A parking fee payment method based on parking lot system

Abstract

The present application discloses a parking fee payment method based on a parking lot system, which relates to the technical field of parking fee payment, and includes: collecting a user's facial image; performing contour segmentation by an image processing method to extract a facial area; adopting an image spectrum reconstruction method based on visible light to obtain spectral features of facial skin; constructing a multi-layer feedforward neural network to judge the authenticity of facial skin according to the extracted spectral features; when the facial skin judgment result is true, inputting the collected facial image into the constructed convolutional neural network model to extract a facial feature vector; matching the extracted facial feature vector with a registered user feature library, and if the match is successful, confirming the user identity; querying the associated user account information according to the confirmed user identity, obtaining the vehicle entry time, and calculating the parking fee; in view of the problem of low accuracy of user identity recognition when paying parking fees in the prior art, the present application improves the accuracy of user identity recognition.

Description

Technical Field

[0001] The present application relates to the technical field of parking fee payment, and in particular to a parking fee payment method based on a parking lot system. Background Art

[0002] With the continuous acceleration of urbanization, the demand for parking lots is increasing, and parking management systems are also facing the requirements of being more efficient and intelligent. The traditional parking payment system has the problem of low accuracy in user identity recognition, which leads to inconvenience in parking fee payment and reduced management efficiency. In response to this problem, people are in urgent need of a new payment method that can efficiently and accurately identify user identities.

[0003] Traditional parking payment systems mainly rely on license plate recognition or parking cards to identify users, but these methods have many limitations. For example, license plate recognition is affected by factors such as lighting and occlusion, and is prone to misidentification; parking cards are easily forgotten or lost, resulting in poor user experience. Therefore, parking payment methods based on facial recognition technology have become one of the research hotspots. However, traditional facial recognition technology often has low accuracy in complex environments and cannot meet the parking system's needs for efficient and accurate identity recognition.

[0004] In the related art, for example, Chinese patent document CN116311559A provides a parking lot deduction method and system based on facial payment. A parking lot deduction method based on facial payment includes S1: obtaining vehicle entry information; S2: obtaining vehicle exit information; S3: calculating parking fees; S4: obtaining the driver's facial image, and judging whether the driver is a parking lot member through a facial recognition algorithm. If the driver is a parking lot member, the parking fee is directly deducted from the driver's member account. The driver's identity is judged only by relying on a two-dimensional facial image, and the identification is only performed from the perspective of facial feature matching. Due to the limitations of the facial recognition algorithm, the recognition accuracy of the same person will decrease under changes in lighting, posture, expression, etc. In particular, for side face images that are not frontal faces, the recognition accuracy will drop significantly. Summary of the invention

[0005] 1. Technical issues to be solved

[0006] In response to the problem of low accuracy of user identity recognition when paying parking fees in the prior art, the present application provides a parking fee payment method based on a parking system. The driver's facial image is captured by a camera, and the driver's identity is automatically identified through facial positioning, spectral reconstruction, detection and identity matching, thereby improving the accuracy of user identity recognition.

[0007] 2. Technical solution

[0008] The purpose of this application is achieved through the following technical solutions.

[0009] The embodiment of the present specification provides a parking fee payment method based on a parking lot system, including: setting multiple cameras at the entrance and exit of the parking lot to collect user's facial images; performing contour segmentation on the collected facial images through an image processing method to extract the facial area; in the extracted facial area, using a visible light-based image spectrum reconstruction method to obtain the spectral features of the facial skin; constructing a multi-layer feedforward neural network to determine the authenticity of the facial skin based on the extracted spectral features; when the facial skin determination result is true, inputting the collected facial image into the constructed convolutional neural network model to extract the facial feature vector; matching the extracted facial feature vector with a registered user feature library, if the match is successful, confirming the user identity, the user identity includes registered users and unregistered users; according to the confirmed user identity, querying the associated user account information, obtaining the vehicle entry time, and calculating the parking fee; for registered users, directly deducting the calculated parking fee from the associated account through an interface; for unregistered users, generating a QR code of the parking fee through a mobile payment interface and outputting it to a display interface.

[0010] Among them, the image processing method refers to an algorithm or means for operating or converting image signals to achieve the purpose of improving image quality, enhancing the interesting area of ​​the image or extracting image-related information. In this application, filtering, sharpening and other algorithms are used to remove image noise and improve image quality. Canny edge detection and other algorithms are used to analyze image gradient features and locate facial contours. Distance transform and other algorithms are used to extend facial contour morphology to the entire facial area. Bilateral filtering and other algorithms are used to smooth the image while retaining facial structure information. Threshold segmentation, binarization and other algorithms are used to extract facial area masks. Image cropping is performed based on the mask to segment the facial area of ​​interest.

[0011] Among them, the facial area refers to the image block area in the facial image that contains the entire facial contour and the main facial tissue features. It includes facial structures such as the forehead, eyebrows, eyes, nose, cheeks, lips, etc., covering the main facial biometric features. In this application, extracting the facial area through image processing methods is a prerequisite for obtaining facial biometric features. Spectral analysis of the facial area can avoid background interference in feature extraction. The facial area image is cropped and sent to the convolutional neural network as input to extract facial features. The facial area contains complete facial biometric features, which is conducive to subsequent facial matching and identity recognition.

[0012] Among them, the image spectrum reconstruction method based on visible light refers to using the image information of the three visible light bands of red, green and blue to restore the spectral reflectance curve of the face in the entire visible light band through a spectral reconstruction algorithm. In this application, three filtered images of red, green and blue are obtained through three filter lenses. The light intensity of each pixel in the three bands is extracted. A conversion model between light intensity and spectral reflectance is established. The spectral reflectance of each pixel is obtained through the conversion model. The spectral reflectance curve of the full band is obtained using a spectral reconstruction algorithm.

[0013] Among them, spectral features refer to characteristic parameters with biometric significance extracted from the spectral reflectance curve of facial skin, such as absorption valleys, reflection peaks, etc. In this application, it contains optical property information of facial tissue. It can be used to distinguish between real faces and fake faces, providing a key basis for liveness detection.

[0014] The multi-layer feedforward neural network is an artificial neural network structure consisting of an input layer, an output layer, and multiple hidden layers. Each layer is forward-connected, and information is only transmitted forward without reverse connections. The input layer receives the extracted facial skin spectral features. The hidden layer performs nonlinear transformation on the input features layer by layer to extract abstract features. The output layer gives a binary classification result of the authenticity of the facial skin. Through the cascade of multiple hidden layers, complex facial features are extracted to achieve the discrimination of true and false faces. The network parameters are adjusted by training real face samples.

[0015] Among them, the facial feature vector is a series of vector values ​​that map the facial biometric features contained in the facial image to a high-dimensional digital vector space. This set of vectors, as a digital representation of the face, contains the unique biometric information of the face. In this application, the collected facial image is used as input and sent to the constructed convolutional neural network. The convolutional network extracts abstract facial features through operations such as convolution and pooling. The output results of the high-level network are sorted to form a facial feature vector. This feature vector is used as a digital expression of facial biometrics for subsequent face matching and recognition. By training the network parameters, a stable and accurate vectorized expression of facial features is obtained.

[0016] Specifically, the facial feature vector is converted into a standard vector of fixed length after extraction and regularization. The registered user feature library stores the standard facial feature vectors of all registered users. During matching, the facial feature vector to be identified is input, and the distance or similarity between it and each registered vector in the user library is calculated. Common distance calculation methods include Euclidean distance, cosine similarity, etc. The higher the similarity, the better the match. Compare the distance or similarity between the input feature vector and all registered vectors, and select the closest registered vector. If the similarity is greater than the preset threshold, the match is determined to be successful, and the input face is confirmed to be the identity of the corresponding registered user. If all similarities are lower than the threshold, it is determined to be an unregistered user identity. The registered user library needs to be trained with a large number of known identity face samples to obtain feature extraction and matching models. As the number of registered users increases, methods such as sorting retrieval can be used to accelerate the matching process.

[0017] Specifically, according to the user identity confirmed by the aforementioned facial recognition, the account information corresponding to the identity is queried in the user information database. The user information database will store the basic information and account of the registered user, including user ID, license plate number, mobile phone number, etc. Through account association, the entry time of the user's vehicle in the current parking lot is obtained. Calculate the parking time of the vehicle from the entry time to the current time. According to the parking lot's charging standards, calculate the parking fee to be charged according to the parking time. Generally, it is charged by the hour, and less than one hour is calculated as one hour. The charging standards can set different rates for different time periods, and discounts can be given according to membership status. Calculate the final payment fee = parking time * hourly rate. If there is a membership, the parking fee is calculated according to the membership rate. Finally, the parking fee payable for the confirmed identity is obtained. In this way, the user account can be directly associated through facial recognition to automatically calculate the parking fee.

[0018] Specifically, for registered users: the user identity has been confirmed and the parking fee has been calculated. The user's payment account information is searched in the user information database. Through the open interface of the payment system, the deduction API is called to enter the user account information and the parking fee amount. After the payment system verifies the information, the parking fee is directly deducted from the user's account. The deduction result is returned. If successful, the payment is completed; if failed, an error code is returned. The entire payment process is completed quickly without the user's presence. For unregistered users: If the user's identity cannot be confirmed, their payment account cannot be known. Call the payment system's payment collection code generation API and enter the parking fee amount to be paid. The payment system generates a payment collection QR code for the corresponding amount. The collection QR code is displayed on the interface of the detection terminal. The user scans the code through the mobile terminal and selects a payment channel to make payment. After the payment is completed, the parking fee payment for the unregistered user is completed.

[0019] Furthermore, extracting the facial region includes: performing Gaussian smoothing on the collected facial image; calculating the gradient size and direction of the facial image after smoothing to obtain the gradient amplitude and direction angle matrix of the facial image; performing non-maximum suppression on the gradient amplitude to determine the image boundary points; performing double threshold detection on the obtained image boundary points to obtain the facial boundary points to be connected; for the facial boundary points to be connected, connecting the boundary points according to the direction angle matrix through a threshold connection method to obtain the closed contour of the face; performing distance transformation on the extracted closed contour of the face to obtain the distance field of the facial region; performing bilateral filtering on the distance field to obtain the facial distance map; setting a threshold, binarizing the facial distance map to generate a binary mask of the facial region; and cropping the input facial image according to the non-zero area of ​​the binary mask to obtain the facial region.

[0020] Among them, Gaussian smoothing is a method of filtering an image using a Gaussian function as a convolution kernel. In this application, a two-dimensional Gaussian distribution is used as a convolution kernel. The collected original facial image is convolved with the Gaussian kernel. The center value of the Gaussian kernel is the largest, and the peripheral values ​​decrease according to the Gaussian distribution. This can smooth the noise points in the image while retaining detailed features such as the edge of the image. The noise introduced in the acquisition process is removed to improve the image quality of subsequent processing. The degree of smoothing can be adjusted by controlling the Gaussian kernel parameters. The smoothed image is more suitable for edge detection and feature extraction.

[0021] Specifically, for the facial image after Gaussian smoothing, the grayscale change of each pixel in the horizontal and vertical directions is calculated to obtain the gradient G of the point x and G y . Gradient G x It can be obtained by calculating the right pixel value of each pixel point minus the left pixel value. y It can be obtained by subtracting the pixel value above from the pixel value below. Calculate the gradient amplitude of each pixel: Calculate the gradient direction angle of each pixel: The gradient amplitudes of all pixels are integrated to form the gradient amplitude matrix Grad Mat. The gradient direction angles of all pixels are integrated to form the direction angle matrix Theta Mat. Grad Mat reflects the edge intensity distribution of the facial image. Theta Mat reflects the direction distribution of the edge. Through the above matrix operations, the overall gradient features of the facial image are obtained, which provides a basis for subsequent edge detection and feature extraction.

[0022] Among them, non-maximum suppression (NMS) is a method to eliminate non-edge maximum points in an image. It retains the points with the local maximum gradient amplitude at the boundary and suppresses non-maximum points. In this application, the calculated gradient amplitude matrix is ​​input. For each pixel point, compare it with the values ​​of the adjacent points on both sides in its gradient direction. If it is not a local maximum, suppress this point and set it to 0. Finally, the suppressed gradient map is obtained. Image boundary points: refers to the pixel points located on the edge contour of the image. These points contain the position information of the image boundary. For the suppressed gradient map, all non-zero value points are extracted. These non-zero points determine the boundary points of the facial image. The boundary points identify the contour of the facial area. Provide key boundary information for the subsequent generation of closed contours.

[0023] Among them, dual threshold detection is an image segmentation method based on dual thresholds. It uses two different thresholds to divide image pixels into three categories. In this application, the input is the gradient amplitude matrix after non-maximum suppression. Set two thresholds, a high threshold and a low threshold. If the amplitude is higher than the high threshold, it is marked as a strong boundary point. If the amplitude is lower than the low threshold, the point is discarded. If the amplitude is between the two thresholds, it is marked as a weak boundary point. Finally, a set of strong boundary points and a set of weak boundary points are obtained. These two types of points are the facial boundary points to be connected. Strong boundary points correspond to facial contours, and weak boundary points are used to connect contours. Through dual threshold detection, the boundary points corresponding to the facial contour are effectively extracted, providing a basis for the subsequent generation of closed contours. Specifically, in this application, dual threshold detection can be used: OTSU method: by statistically analyzing the image pixel histogram, the optimal threshold is automatically calculated. Then the high and low thresholds are determined in combination with the standard deviation calculation. Iterative method: set the initial threshold, calculate the average value of the two types of pixel clusters as the new threshold, and iterate until the threshold converges. Maximum inter-class variance method: traverse all possible dual threshold combinations, calculate the inter-class variance of two types of pixel clusters, and select the threshold corresponding to the largest one. Maximum entropy method: construct a maximum entropy model under dual threshold conditions, and determine the optimal threshold through an iterative optimization algorithm. Clustering method: use clustering algorithms such as K-means to divide pixel values ​​into two categories, and the cluster center is the threshold.

[0024] Among them, the threshold connection method refers to a method of connecting boundary points according to the gradient direction using a set threshold rule to generate a closed contour. In this application, the input is the boundary points to be connected after double threshold detection. For each boundary point, search for adjacent points in its gradient direction. If the gradient directions of adjacent points are similar (the difference in direction angles is less than a preset threshold), connect the two points. Continue to search and connect points that meet the threshold rules until a closed contour is generated. The direction angle matrix provides the gradient direction information of each point. The threshold rule is used to ensure that the gradient directions of the connected boundary points are consistent and belong to the same contour. Finally, the external closed contour of the facial area is obtained.

[0025] Among them, distance transform is an operation that maps the distance between the contour pixel on the image and the non-contour background. Distance field represents the distance value from each pixel on the image to the nearest contour pixel. In this application, the input is the extracted closed contour of the face. For each pixel in the contour, the distance from it to the contour pixel is calculated to obtain the distance field. The contour pixel distance is 0, and the greater the distance, the greater the value. The distance field reflects the spatial distance distribution from each pixel to the facial contour. By setting a threshold according to the distance field, the range of the facial area can be obtained. Distance transform expands the contour range and obtains the entire facial area. Through distance transform and distance field, the incomplete contour information can be effectively solved and the overall picture of the facial area can be determined.

[0026] Among them, bilateral filtering is a filtering method that can smooth an image while retaining edge details. It combines the advantages of spatial filtering and range filtering. In this application, the input is the facial distance field obtained by distance transformation. Bilateral filtering is applied to the distance field. The spatial distance of image pixels and their grayscale value differences are considered at the same time. The distance field is smoothed while retaining the edge of the facial contour. Noise and isolated points in the distance field are eliminated. The distance field image obtained by bilateral filtering is smoothed while retaining the contour edge. It accurately reflects the range and contour information of the facial area. The threshold is set according to the distance map to accurately segment the facial area. The facial area is extracted to provide a basis for subsequent facial feature analysis.

[0027] Specifically, a binary mask of the facial area is generated, and the input is the facial distance map obtained by bilateral filtering. The distribution of distance map values ​​is analyzed to determine a suitable threshold T. The distance map is traversed and a threshold judgment is performed: if the distance map pixel value is less than T, it is set to 1. If the distance map pixel value is greater than T, it is set to 0. After the above threshold judgment, a binary distance map is obtained. In the binary image, the area with a pixel value of 1 constitutes the facial area mask. The mask area is all the pixels within the facial contour. According to the actual application, multiple thresholds can also be set to generate multiple mask layers. The last mask layer contains the entire head and shoulder area. Through multi-level masks, facial features in different parts can be identified.

[0028] Specifically, cropping the facial region according to the binary mask includes: inputting the original facial image and the generated binary mask. The non-zero area in the mask represents the valid facial contour range. Calculating the enclosing rectangular frame of the non-zero area of ​​the mask. The rectangular frame determines the range of the facial region. On the original image, the image is cropped according to the position of the rectangular frame. The facial image area within the rectangular frame is retained. The invalid background area is cropped. The range of the rectangular frame can also be slightly enlarged to ensure that the complete face is included. If there are multiple levels of masks, local facial regions such as eyes and nose can be cropped in sequence. Finally, a clear facial region image is obtained, and the background part is removed. The facial region image is input into the subsequent algorithm for recognition and feature extraction.

[0029] Furthermore, the Chebyshev distance is used to perform distance transformation on the closed contour of the extracted face; wherein, the Chebyshev distance (Chebyshev, distance) is a method for measuring the distance between two points. It is defined as the maximum difference between the corresponding coordinates of two points. In the present application, the distance transformation is performed on the extracted closed contour of the face. The Chebyshev distance from each inner point of the contour to the contour pixel is calculated. The Chebyshev distance is defined as the maximum difference between the coordinates of two points on the x-axis and the y-axis. The Chebyshev distance field from each inner point to the contour is obtained. The Chebyshev distance field reflects the shortest Manhattan distance from each inner point to the contour. A facial region mask is generated based on the Chebyshev distance field. The Chebyshev distance transformation is efficient in calculation, and the distance value accurately reflects the spatial distribution. Finally, the Chebyshev distance map of each point in the contour is obtained. The use of Chebyshev distance is conducive to quickly and accurately obtaining facial region contour information.

[0030] Furthermore, the distance field is bilaterally filtered to obtain the facial distance map; the formula for bilateral filtering is: Among them, F(p) is the pixel value after filtering, f represents the spatial weight function, g represents the gray value similarity weight function, I(q) is the gray value of pixel q; p represents the coordinates of the central pixel in the image; q represents the coordinates of the adjacent pixels in the filter template area; I(p) represents the gray value of the central pixel p; I(q) represents the gray value of the adjacent pixel q.

[0031] Specifically, the input is the facial distance field image I obtained after distance transformation. The spatial weight function f(x) of bilateral filtering is defined as a Gaussian distribution function: where σ s represents the spatial standard deviation parameter of the Gaussian kernel; |pq| represents the Euclidean distance between the central pixel p and the adjacent pixel q. For each pixel p in the image I, the weighted average of the adjacent pixels q in its neighborhood S is:

[0032] k(p) is the normalization coefficient so that the sum of the weights is 1. The weight f(pq) is determined by the spatial distance, and the weight g(I(p)-I(q)) is determined by the grayscale difference. Calculate the filtering results of all pixels to obtain a smooth face distance map F. By adjusting the parameter σ s and σ r Controls the degree of filtering.

[0033] Furthermore, the spatial weight function f adopts a Gaussian function, where σ s represents the spatial standard deviation parameter of the Gaussian kernel; |pq| represents the Euclidean distance between the central pixel p and the adjacent pixel q.

[0034] Specifically, the spatial weight function f is in the form of a two-dimensional Gaussian function. For the central pixel p and the adjacent pixel q, their Euclidean distance d = |pq| is calculated. Taking d as the independent variable of the Gaussian function, the spatial weight is obtained: where σ s is the spatial standard deviation parameter of the Gaussian kernel. When d is 0, f(d) reaches a maximum value of 1. As d increases, f(d) decays rapidly according to the Gaussian distribution. By adjusting σ s The decay rate can be controlled to determine the scope of the spatial weight. s The larger the value, the larger the range of spatial weighting, and the smoother the corresponding filtering effect. According to the image content and filtering purpose, choose the appropriate σ s , smoothing while preserving the edge of the filter. Spatial Gaussian weights implement spatial filtering that assigns weights based on pixel distance.

[0035] Furthermore, the similarity weight function g adopts an exponential function; where σ r represents the standard deviation parameter of the gray value similarity kernel; |I(p)-I(q)| represents the absolute value of the gray value difference between the central pixel p and the adjacent pixel q.

[0036] Specifically, the similarity weight function g is in the form of an exponential function. For the central pixel p and the adjacent pixel q, the absolute value of the difference between their gray values ​​d = |I(p)-I(q)| is calculated. Taking d as the independent variable of the exponential function, the similarity weight is obtained: Where σ_r is the standard deviation parameter of the gray value similarity kernel. When d is 0, g(d) reaches a maximum value of 1. As d increases, g(d) decays exponentially. By adjusting σ r The decay rate can be controlled to determine the scope of the similarity weight. r The larger the similarity weight is, the larger the range of similarity weight is, and more edge details are retained. According to the image content and the requirement of retaining edges, choose the appropriate σ rThe similarity weights implement range filtering by assigning weights based on grayscale similarity.

[0037] Furthermore, a method for image spectrum reconstruction based on visible light includes: using three independent band filters of red, green and blue with non-overlapping wavelength ranges to filter and shoot the facial area of ​​the same user in sequence, and obtain three filtered images of red, green and blue respectively; performing image analysis on each facial pixel in the three filtered images, and extracting the light intensity value of each facial pixel in the three bands of red, green and blue; calculating the spectral reflectance value of each facial pixel in the three bands of red, green and blue according to the conversion relationship between the image pixel light intensity and the spectral reflectance in each filtered band; according to the spectral reflectance values ​​of the three bands of red, green and blue, using a spectral reconstruction algorithm, obtaining the spectral reflectance curve of the corresponding pixel in the full visible light band; extracting the spectral characteristics of the facial skin according to the obtained spectral reflectance curve, the spectral characteristics including the absorption valley and the reflection peak;

[0038] Among them, the band filter is an optical filter element that only allows light in a specific wavelength range to pass through. In this application, bandpass filters of three bands, red, green and blue, are selected. The red band filter only allows red light to pass through, and the cutoff wavelength is set to 620nm to 750nm. The green filter only allows green light to pass through, with a wavelength of 500nm to 570nm. The blue filter only allows blue light to pass through, with a wavelength of 430nm to 500nm. Use these three filters to filter and shoot the face of the same user respectively. Obtain a facial image R under red light illumination, a facial image G under green light illumination, and a facial image B under blue light illumination. The three filtered images reflect the performance of the face under different colors of light. Provide source data for the acquisition of multi-spectral facial images.

[0039] Among them, the light intensity value refers to the light intensity value of each pixel in the image in a specific band. In this application, the input is a facial image taken in three bands: red, green, and blue. For each facial pixel, its pixel value r is read in the red light image R, its pixel value g is read in the green light image G, and its pixel value b is read in the blue light image B. The pixel values ​​r, g, and b are the light intensity values ​​of the pixel in the red, green, and blue bands, respectively. The light intensity value reflects the light intensity response of the pixel under illumination in different bands. By extracting the light intensity value of each pixel, the spectral characteristics of the pixel in multiple bands are obtained. Finally, the spectral vector [r, g, b] of each facial pixel is constructed to express the spectral signature of the facial pixel. The spectral signature is used in facial recognition and verification.

[0040] Among them, the spectral reflectance value is a physical quantity that represents the surface reflection characteristics of each pixel in the image in a specific band. It has a corresponding relationship with the light intensity value of the pixel. In this application, for each facial pixel, its light intensity values ​​r, g, and b in the red, green, and blue bands have been extracted. According to the optical characteristic parameters of the imaging system, a mathematical conversion model between the light intensity value and the reflectance value in each band can be established. Substitute the light intensity values ​​r, g, and b into the model to calculate the spectral reflectance rr, rg, and rb of the pixel in the red, green, and blue bands. The reflectance value is an intrinsic parameter with a clear physical meaning and is not affected by shooting conditions. The reflectance is used to construct the spectral vector [rr, rg, rb] of the facial pixel as the spectral signature of the pixel. The reflectance spectral vector reflects the optical characteristics of the facial pixel and can be used for facial recognition and feature analysis.

[0041] Among them, the spectral reconstruction algorithm is a type of algorithm that reconstructs the spectral information of the pixel in the entire band from the spectral data of a limited number of bands. Commonly used methods include linear reconstruction method, non-negative matrix decomposition method, etc. Spectral reflectance curve: represents the spectral reflectance change curve of a pixel in the image in the entire band, which can be identified and analyzed using the spectral features on the curve. In this application, for each facial pixel, its spectral reflectance rr, rg, rb in the three bands of red, green and blue are input. Use the spectral reconstruction algorithm with rr, rg, rb as known sample points. Reconstruct the spectral reflectance curve R(λ) of the pixel in the full band range. R(λ) describes the continuous spectral characteristics of the pixel in the visible light and near-infrared light bands. Face recognition and analysis are performed using the shape of the spectral curve, characteristic wavelength, etc.

[0042] Among them, spectral features: refers to the spectral reflectance variation pattern of facial skin within a specific wavelength range. Absorption valley: a local minimum point on the spectral reflectance curve, in which light is strongly absorbed by the skin. Reflection peak: a local maximum point on the spectral reflectance curve, in which there is strong light reflection. In this application, the spectral reflectance curve R(λ) of the reconstructed facial skin area pixels is input. The shape of the R(λ) curve is analyzed to detect the absorption valley and reflection peak points. For example, the absorption valley near wavelengths of 420nm and 550nm. The reflection peak near a wavelength of 680nm. These features correspond to the specific spectral absorption and reflection characteristics of the skin. These spectral features are extracted to form a spectral feature vector of the skin pixel. Spectral features can be used for skin recognition, judging human attributes, etc.

[0043] Furthermore, the conversion relationship between the light intensity of the image pixel and the spectral reflectance in each filter band is obtained, including: constructing a multi-layer feedforward neural network, the neural network includes an input layer, an output layer and a hidden layer; the input layer includes 3 input nodes, which respectively receive the light intensity values ​​of the three bands of red, green and blue; the output layer includes 3 output nodes, which respectively represent the predicted spectral reflectance in the three bands of red, green and blue;

[0044] Specifically, a three-layer feedforward fully connected neural network is constructed, including an input layer, a hidden layer, and an output layer. The input layer contains three input nodes, which receive the light intensity values ​​r, g, and b of the three bands of red, green, and blue respectively. The hidden layer contains an appropriate number of neurons, such as 10-20, using activation functions such as ReLU. The output layer contains three output nodes, which predict the spectral reflectance values ​​rr, rg, and rb in the three bands of red, green, and blue respectively. The network connection weights are obtained through supervised learning training. Samples with matching light intensity and reflectance data are collected as training sets to train the network model. The model parameters are optimized so that the output results can accurately predict the spectral reflectance. The final light intensity to reflectance conversion network model is obtained. The newly input light intensity data is directly passed into the network input layer to obtain the reflectance prediction result.

[0045] Furthermore, the hidden layer uses the hyperbolic tangent function tanh as the activation function; wherein the hyperbolic tangent function tanh(x) is defined as: It is an S-shaped nonlinear activation function with a range of [-1, 1]. In this application, it is used as the activation function of the hidden layer of the neural network. A nonlinear transformation is performed on the linear weighted input of each neuron in the hidden layer. The introduction of nonlinear factors increases the ability of the network to express complex patterns. Compared with the sigmoid function, the gradient of the tanh function is easier to flow, which is conducive to model training. When the input value is close to 0, the tanh gradient is the largest, which helps to accelerate the gradient descent rate. The tanh function limits the output of the hidden layer neurons to the range of [-1, 1] to avoid saturation. Using tanh as an activation function can enhance the network's ability to fit the spectral reflectance conversion relationship and improve prediction accuracy.

[0046] Furthermore, the training objective function of the multi-layer feedforward neural network is set as the mean square error loss function of the difference between the predicted spectral value and the true spectral value. Among them, the matching light intensity data and spectral reflectance data are collected as the training set. The light intensity of the training set sample is input to the neural network to obtain the predicted spectral reflectance output. The mean square error between the predicted spectral value and the true spectral value is calculated as the loss function: Where N is the number of samples, y pred To predict the spectral reflectance, y trueis the true spectral reflectance. The network parameters W, b are updated by the gradient descent algorithm to minimize the loss function L. Repeat the process of inputting samples, forward propagation, calculating losses, backward propagation, and updating parameters. The network training goal is to minimize the mean square error between the predicted spectral value and the true value. When the loss function tends to be stable, the final optimized network model is obtained. This model can accurately predict the spectral reflectance of new samples.

[0047] 3. Beneficial effects

[0048] Compared with the prior art, the advantages of this application are:

[0049] (1) By adopting multiple technical means such as facial image acquisition, facial positioning, spectral reconstruction, detection and identity matching, this solution effectively improves the accuracy of user identity recognition. It overcomes the problems of traditional methods affected by factors such as lighting, posture, and occlusion, and ensures robustness in diverse environments;

[0050] (2) Gaussian smoothing, gradient calculation, contour segmentation and other technologies are used to make the extraction of facial areas more accurate, reduce errors and improve the stability of subsequent processing steps;

[0051] (3) Based on the visible light image spectrum reconstruction method, the spectral reflectance values ​​are calculated through filters in the red, green, and blue bands to obtain the spectral characteristics of the facial skin. This effectively improves the accuracy of the spectral characteristics of the facial skin and helps to more reliably authenticate user identities;

[0052] (4) Using the constructed neural network to detect and judge the authenticity of facial skin through comprehensive spectral features. This effectively prevents deception methods such as static images or models and improves the security of the system;

[0053] (5) The user's identity is quickly confirmed by matching the extracted facial features with the registered user feature library. For registered users, the parking fee can be directly deducted from the associated account, which improves the efficiency of parking fee payment. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] This specification will be further described in the form of exemplary embodiments, which will be described in detail by the accompanying drawings. These embodiments are not restrictive, and in these embodiments, the same number represents the same structure, wherein:

[0055] Figure 1 is an exemplary flow chart of a parking fee payment method based on a parking lot system according to some embodiments of this specification;

[0056] Figure 2 is an exemplary flow chart of extracting a face area according to some embodiments of this specification;

[0057] Figure 3 is an exemplary flow chart of extracting spectral features according to some embodiments of the present specification. DETAILED DESCRIPTION

[0058] The method and system provided in the embodiments of this specification are described in detail below with reference to the accompanying drawings.

[0059] Figure 1 This is an exemplary flow chart of a parking fee payment method based on a parking lot system according to some embodiments of the present specification, wherein multiple cameras are set at the entrance and exit of the parking lot to collect the user's facial image; the collected facial image is segmented by an image processing method to extract the facial area; within the extracted facial area, a visible light-based image spectrum reconstruction method is used to obtain the spectral features of the facial skin; a multi-layer feedforward neural network is constructed to determine the authenticity of the facial skin based on the extracted spectral features; when the facial skin determination result is true, the collected facial image is input into the constructed convolutional neural network model to extract the facial feature vector; the extracted facial feature vector is matched with the registered user feature library, and if the match is successful, the user identity is confirmed, and the user identity includes registered users and unregistered users; according to the confirmed user identity, the associated user account information is queried, the vehicle entry time is obtained, and the parking fee is calculated; for registered users, the calculated parking fee is directly deducted from the associated account through the interface; for unregistered users, a QR code of the parking fee is generated through the mobile payment interface and output to the display interface.

[0060] Cameras are set up at the entrance and exit of the parking lot to collect the driver's facial image. A high-definition camera is installed above the entrance and exit of the parking lot, and the position must ensure that the driver's facial image can be clearly captured. The camera is connected to the image acquisition card and transmitted to the computer through the PCIE interface. In the case of unattended operation, the camera can be triggered by the magnetic signal generated when the vehicle opens the door. After obtaining the front image of the vehicle and the driver, the face is located. The face area mask is generated by distance transformation and bilateral filtering. The face image area is extracted according to the mask and irrelevant background is removed. The face image is enhanced and cropped to output a clear and standardized front head portrait image. The head portrait image is passed as input to the subsequent face recognition system to complete the driver's identity authentication. After the verification is passed, the barrier is opened remotely to complete the entrance and exit management. The head portraits of the incoming vehicles are continuously collected to form a driver database, which can realize the automatic intelligent management of the entrance and exit.

[0061] Figure 2This is an exemplary flow chart for extracting facial regions according to some embodiments of this specification. The collected facial image is preprocessed, including smoothing, edge detection, etc., to generate a facial region mask. The collected original facial image is Gaussian filtered to smooth the image and remove noise: the collected original facial image I is input. According to the image content characteristics, the Gaussian filter window size is set, such as 5x5 pixels. The Gaussian kernel weight parameter σ is calculated to control the filter strength, which is generally 1-3. A two-dimensional Gaussian kernel G(x, y) is constructed. For each pixel point P(x, y) in the image I, Gaussian weighted average is performed in its neighborhood window: F(x, y) = ∑G(i, j)*I(x+i, y+j), and all pixels are repeatedly processed to obtain a smoothed image F. The Gaussian filter uses weighted average in the neighborhood to effectively suppress random noise. The filtering degree is adjusted by controlling σ to avoid blurring the image. The smoothed facial image is output to provide clear data for subsequent processing. According to the facial mask, the facial image is cropped to obtain a facial region image. For the face area image, three filter images are obtained using red, green and blue filter lenses.

[0062] Use the Sobel operator and other methods to detect the edge of the face contour: Input the smoothed face image I. Construct the Sobel filter kernels Sx and Sy in the horizontal and vertical directions. Perform Sx and Sy convolution on the image I respectively to obtain the horizontal and vertical gradient approximations Gx and Gy. Calculate the gradient magnitude image: Edge = sqrt(Gx^2+Gy^2). Perform threshold processing on the gradient magnitude image to obtain the binary edge image E. Select the appropriate threshold parameter to retain the key facial contour edges and remove false edges. The Sobel filter operator is actually a first-order derivative operation, which can effectively detect the strong edges of the image. The boundaries of facial features such as facial contours, eyebrows, and eyes appear as strong edges in the gradient image. Keep these strong edges and generate the facial contour edge image. The edge image is sent to the distance transform to calculate the facial contour distance field.

[0063] Perform distance transformation on the edge image and calculate the distance from each pixel to the edge: Input the obtained face edge binary image E. Initialize the distance image D, which is the same size as E, and all pixel values ​​are initialized to 0. Scan all edge pixel points P(x, y) in E, with coordinates (x, y). With P as the center, perform distance diffusion and calculate the Euclidean distance from the surrounding non-edge pixels to P. For each surrounding non-edge pixel point Q(x+dx, y+dy): D(x+dx, y+dy) = sqrt((x-x+dx)^2+(y-y+dy)^2), retain the minimum value between D(x+dx, y+dy) and the original value. Repeat steps 3-6 until the distance values ​​of all pixels in image D are calculated. Finally, the distance field image D of the face edge is obtained. The larger the value in D, the farther the point is from the edge.

[0064] Perform bilateral filtering on the distance field image to generate a smoothed face distance map: Input the already calculated facial edge distance field image D. Initialize the output distance map F with the same size as D. Define the spatial Gaussian kernel Gs and the gray value similarity kernel Gr of the bilateral filtering operator. For each pixel point P(x, y) in D, perform bilateral filtering within its neighborhood window: F(x, y) = ∑Gs(dist)*Gr(gray)*D(x + i, y + j), where dist is the spatial distance between P and its neighboring point; gray is the gray difference between P and its neighboring point. Gs assigns weights according to the spatial distance, and Gr assigns weights according to the gray similarity. Repeat the calculation for all pixel points to obtain the smoothed distance map F. The abrupt changes at the boundaries are retained, effectively smoothing the noise in the distance field. Output the smoothed and stable face distance map F after bilateral filtering.

[0065] Analyze the value distribution of the distance map to determine an appropriate distance threshold. Perform threshold processing on the distance map, setting points with a distance less than the threshold to 1 and points greater than the threshold to 0: Input the already obtained smoothed face distance map F. Calculate the histogram distribution of all pixel distance values in F. Analyze the histogram to find the first significant peak point Tp of the distance value distribution. Determine an appropriate distance threshold Ts slightly greater than Tp, for example, Ts = Tp + 5. Perform threshold segmentation on F:

[0066] If F(x, y) < Ts: Mask(x, y) = 1, otherwise: Mask(x, y) = 0. Set the pixel points with a distance less than Ts to 1 to generate the initial face mask Mask. Perform morphological processing on Mask to obtain a clear and complete final face region mask.

[0067] Extract the corresponding face image region according to the mask. After threshold segmentation, obtain the binary mask image of the region within the face contour. Perform hole filling and edge correction on the mask image to generate a clear and complete face region mask. The region with a mask pixel value of 1 represents the valid face region. Extract the corresponding face image region according to the mask to complete preprocessing and localization. Input the already obtained initial face mask image Mask. Perform morphological dilation and erosion operations on Mask to fill the holes. Analyze the Mask boundary to correct the missing and irregular edge regions. Finally, generate a clear and complete face region mask Final_Mask. In Final_Mask, the region with a pixel value of 1 represents the valid face region. For the originally acquired face image I, read its pixel position (x, y). If Final_Mask(x, y) == 1, then retain I(x, y). If Final_Mask(x, y) == 0, then discard I(x, y). Repeat the process for all pixels to extract the corresponding face region according to the mask. Output the localized face image to complete image preprocessing.

[0068] Figure 3 This is an exemplary flow chart for extracting spectral features according to some embodiments of this specification. The light intensity value of each pixel is extracted from three filtered images and input into a pre-trained neural network model. The filtered face images R, G, and B are taken for the red, green, and blue bands. The face area in R, G, and B is segmented and extracted. Each pixel point (x, y) in the three images is traversed. The light intensity r=R(x, y) is read from R. The light intensity g=G(x, y) is read from G. The light intensity b=B(x, y) is read from B. The filtered face images of the red, green, and blue bands are read. Select the pixel coordinates (x, y) for which the light intensity is to be extracted. In the red filtered image R, the light intensity value of the pixel is read:

[0069] r=R(x,y), in the green filter image G, read the light intensity value of the pixel: g=G(x,y). In the blue filter image B, read the light intensity value of the pixel: b=B(x,y). Combine the light intensity values ​​of the three bands into a 3-element vector: [r,g,b]. This vector represents the light intensity characteristics of the pixel in the red, green and blue bands. Repeat this process to extract the light intensity vectors of all pixels in the image. The light intensity vector reflects the optical color characteristics of the pixel. The light intensity vector can be used as the input of the neural network model to predict the spectral characteristics. Repeat to extract the light intensity values ​​of all facial pixels. Input the light intensity dataset into the pre-trained neural network model. The model contains 3 nodes in the input layer, which receive [r,g,b] respectively. The forward propagation obtains the predicted value of the spectral reflectance of the pixel. Finally, the spectral signature of all facial pixels is output.

[0070] The neural network model outputs the spectral reflectance values ​​of the three bands of red, green and blue corresponding to each pixel. Construct a neural network model, which includes an input layer, a hidden layer and an output layer. The input layer contains 3 nodes, which receive the light intensity vector [r, g, b] of the image pixel. The hidden layer adopts a fully connected structure and contains an appropriate number of neurons. The output layer of the neural network contains 3 fully connected nodes, which are responsible for predicting the spectral reflectance of the three bands of red, green and blue. The pixel light intensity vector [r, g, b] extracted by the input layer is forward propagated. The red band reflectance prediction node receives the weighted input from the previous layer and calculates the activation output rr. The green band reflectance prediction node receives the weighted input from the previous layer and calculates the activation output rg. The blue band reflectance prediction node receives the weighted input from the previous layer and calculates the activation output rb. The outputs of the three bands are combined into the predicted spectral reflectance vector [rr, rg, rb] of the pixel. rr represents the predicted reflectance value of the pixel in the red band. Similarly, rg and rb represent the predicted reflectance of the green and blue bands, respectively. This process is repeated for each input pixel to obtain the spectral reflectance map result.

[0071] The prediction result [rr, rg, rb] of a pixel by the neural network represents the predicted spectral reflectance of the pixel in the red, green and blue bands. Among them, rr is the predicted reflectance value of the pixel in the red band. rg is the predicted reflectance value of the pixel in the green band. rb is the predicted reflectance value of the pixel in the blue band. [rr, rg, rb] is used as the spectral reflectance feature vector of the pixel. Repeatedly using the neural network to predict each pixel in the image can obtain a [rr, rg, rb] vector. The spectral reflectance vectors of all pixels are integrated together. This constitutes the spectral reflectance map or spectral signature of the entire facial image. The spectral signature reflects the spectral reflectance characteristics of each position on the face. It can be used for further face analysis and recognition.

[0072] The spectral reflectance value of each pixel in each band is input into the spectral reconstruction algorithm to obtain the full-band spectrum. The neural network outputs the spectral reflectance [rr, rg, rb] of each pixel in the red, green and blue bands. Construct a mapping matrix M to describe the conversion relationship from RGB to the full band. For each pixel [rr, rg, rb], perform matrix multiplication: [r1, r2,......., rn] = M*[rr, rg, rb]. Among them, [r1, r2,......., rn] is the spectral reflectance of the full band of the pixel. n is the number of spectral bands, for example, 400-720nm, 10nm per band, n = 32. Repeat this process to predict the full-band spectra of all pixels. Stack the full-band spectra of each pixel. Finally, the hyper spectral cube data of the entire image is formed. Each pixel in the cube contains a spectral reflectance vector of length n. The reconstruction from RGB image to full-band spectral image is realized.

[0073] Extract spectral features from the reconstructed spectrum and input them into the detection model. Preprocess the reconstructed image cube, including correction and denoising. Select pixels or image regions of interest in the cube. Select discriminative bands from the spectral reflectance vector of the region. Extract the reflectance values ​​of these bands to form a spectral feature vector. For example, select the 700nm and 760nm bands to form [r700, r760]. Repeat this process to extract spectral features from all regions of interest. Construct a detection model with the same number of input layer nodes as the feature vector dimension. Input the extracted spectral feature vector as a sample into the detection model. The model outputs detection results, such as lesion recognition results. According to the output results of the model, complete the analysis and detection of the image area.

[0074] The detection model determines whether the spectral feature belongs to a real face and outputs the detection result. Collect spectral samples of real faces and spoofed faces (photos, videos, etc.). Extract spectral features that distinguish the two categories from the samples and build a training data set. Select an appropriate model structure, such as a support vector machine. Use the training data set to train the model and learn the spectral distinction rules between real and spoofed faces. Import the input spectral feature vector of the face to be detected into the model. The model determines whether the feature belongs to a real face based on the learned rules. If it belongs to the real person category, the model outputs a prediction result of 1. If it belongs to the spoofing category, the model outputs a prediction result of 0. Based on the prediction results output by the model, the detection of facial authenticity is completed. If the result is 1, it is determined that the face is real and the verification is passed; if the result is 0, it is an attack spoof and the verification fails.

[0075] When the detection result is true, a convolutional neural network is used to extract facial features. Obtain a facial image and perform preprocessing, including correction and normalization. Construct a convolutional neural network model, including convolutional layers, pooling layers, etc. The network input layer receives the preprocessed facial image. The convolutional layer extracts multi-level spatial features, including edge, texture and other information. The pooling layer aggregates features to enhance feature robustness. Flatten the convolutional pooling features and input them into the fully connected layer. The fully connected layer integrates multi-level features to form an overall description vector of the face. When the face authenticity detection result is 1, the convolutional network is enabled to extract facial features. If the detection result is 0, feature extraction is not used. Finally, the feature vector corresponding to the verified facial image is obtained.

[0076] The extracted facial features are matched with the registered user feature library to confirm the driver's identity. During registration, facial features are extracted for all authenticated users and a feature database is constructed. Each user corresponds to a facial feature vector, which stores the feature ID and feature value. When the verification detection is a real face, the feature vector of the face is further extracted. The feature vector of each user is traversed in the registration library. The distance or similarity between the feature vector to be verified and each registered feature vector is calculated. The registered feature with the smallest distance or the largest similarity is selected. If the similarity exceeds the preset threshold, the match is successful, and it is confirmed that the detected face belongs to the corresponding registered user. Read the user ID corresponding to the registered feature to complete the identity confirmation. If all do not match, an unregistered illegal user is detected. The driver's identity is verified through a one-to-one feature comparison.

[0077] According to the confirmed driver identity, query the account information and complete the deduction. The identity authentication module confirms the detected driver identity ID. According to the identity ID, search for the corresponding account information in the user account database. Retrieve the basic information of the user from the account table, including name, account balance, etc. Calculate the fees to be deducted based on the mileage data and the unit mileage rate of the vehicle model. Check whether the account balance is sufficient for the deduction. If the balance is sufficient, deduct the calculated fee from the account balance. Update the latest balance and consumption record of the user in the account table. If the balance is insufficient, output a prompt that the driver's balance is insufficient, please recharge. After completing the deduction transaction, send a consumption notification to the user. Complete the automatic deduction of the corresponding driver by searching and updating the account information.

[0078] The invention of the present application and its implementation methods are described schematically above. The description is not restrictive. Without departing from the spirit or basic features of the present application, the present application can be implemented in other specific forms. What is shown in the accompanying drawings is only one of the implementation methods of the invention of the present application. The actual structure is not limited to this, and any figure mark in the claims should not limit the claims involved. Therefore, if ordinary technicians in this field are inspired by it, without departing from the purpose of the present invention, they can creatively design structural methods and embodiments similar to the technical solution, which should all belong to the scope of protection of the present application. In addition, the word "including" does not exclude other elements or steps, and the word "one" before the element does not exclude the inclusion of "multiple" elements. The multiple elements stated in the product claim can also be implemented by one element through software or hardware. The words first, second, etc. are used to indicate names, and do not indicate any specific order.

Claims

1. A parking fee payment method based on a parking lot system, comprising: Multiple cameras are set up at the entrances and exits of the parking lot to collect users' facial images; For the collected facial images, contour segmentation is performed through image processing methods to extract the facial area; In the extracted facial area, a visible light-based image spectrum reconstruction method is used to obtain the spectrum characteristics of the facial skin; Construct a multi-layer feed-forward neural network to identify the authenticity of facial skin based on the extracted spectral features; When the facial skin identification result is true, the collected facial image is input into the constructed convolutional neural network model to extract the facial feature vector; The extracted facial feature vector is matched with the registered user feature library. If the match is successful, the user identity is confirmed. The user identity includes registered users and unregistered users. Based on the confirmed user identity, query the associated user account information, obtain the vehicle entry time, and calculate the parking fee; For registered users, the calculated parking fee is deducted directly from the associated account through the interface; For unregistered users, a QR code for parking fees is generated through the mobile payment interface and output to the display interface; Extract the face area, including: Perform Gaussian smoothing on the collected facial images; Calculate the gradient size and direction of the smoothed facial image to obtain the gradient amplitude and direction angle matrix of the facial image; Perform non-maximum suppression on the gradient amplitude to determine the image boundary points; Perform double threshold detection on the obtained image boundary points to obtain the face boundary points to be connected; For the facial boundary points to be connected, the threshold connection method is used to connect the boundary points according to the direction angle matrix to obtain the closed contour of the face; Perform distance transformation on the extracted closed contour of the face to obtain the distance field of the face area; Perform bilateral filtering on the distance field to obtain the facial distance map; Set the threshold, binarize the face distance map, and generate a binary mask of the face area; According to the non-zero area of ​​the binary mask, the input face image is cropped to obtain the face area; The Chebyshev distance is used to transform the closed contour of the extracted face; Perform bilateral filtering on the distance field to obtain the facial distance map; The formula for bilateral filtering is: Among them, F(p) is the pixel value after filtering, f represents the spatial weight function, g represents the gray value similarity weight function, I(q) is the gray value of pixel q; p represents the coordinates of the central pixel of the image; q represents the coordinates of the adjacent pixels in the filter template area; I(p) represents the gray value of the central pixel p; I(q) represents the gray value of the adjacent pixel q; k(p) is the normalization coefficient so that the sum of the weights is 1; The spatial weight function f adopts a Gaussian function: where σ s represents the spatial standard deviation parameter of the Gaussian kernel; |pq| represents the Euclidean distance between the central pixel p and the adjacent pixel q; The similarity weight function g adopts an exponential function; where σ r represents the standard deviation parameter of the gray value similarity kernel; |I(p)-I(q)| represents the absolute value of the gray value difference between the central pixel p and the adjacent pixel q.

2. The parking fee payment method based on the parking lot system according to claim 1, characterized in that: The image spectrum reconstruction method based on visible light includes: Using three independent band filters with non-overlapping wavelength ranges, red, green and blue, the facial area of ​​the same user is filtered and photographed in sequence to obtain three filtered images, red, green and blue, respectively; Perform image analysis on each facial pixel in the three filtered images and extract the light intensity value of each facial pixel in the three bands of red, green and blue; According to the conversion relationship between the image pixel light intensity and the spectral reflectance in each filter band, the spectral reflectance value of each facial pixel in the three bands of red, green and blue is calculated; According to the spectral reflectance values ​​of the three bands of red, green and blue, a spectral reconstruction algorithm is used to obtain the spectral reflectance curve of the corresponding pixel point in the full visible light band; According to the obtained spectral reflectance curve, the spectral features of the facial skin are extracted, and the spectral features include absorption valleys and reflection peaks.

3. The parking fee payment method based on the parking lot system according to claim 2, characterized in that: Obtain the conversion relationship between the image pixel light intensity and spectral reflectance in each filter band, including: Construct a multi-layer feedforward neural network, which includes an input layer, an output layer, and a hidden layer; The input layer contains three input nodes, which receive the light intensity values ​​of the red, green and blue bands respectively; The output layer contains three output nodes, which represent the predicted spectral reflectance in the red, green and blue bands respectively.

4. The parking fee payment method based on the parking lot system according to claim 3 is characterized in that: The hidden layer uses the hyperbolic tangent function tanh as the activation function.

5. The parking fee payment method based on the parking lot system according to claim 4 is characterized in that: The training objective function of the multilayer feedforward neural network is set as the mean square error loss function of the difference between the predicted spectral value and the true spectral value.

Images