Image recognition method for cross-domain access authentication

Abstract

The invention relates to the technical field of image recognition, and discloses an image recognition method for cross-domain access authentication, which comprises the following steps: acquiring a to-be-authenticated image uploaded by an image acquisition node; converting the to-be-authenticated image to a frequency domain, and extracting a plurality of response signals in multiple preset directions and scales; calculating a phase consistency measure by using the real part and imaginary part response components, and generating a phase consistency feature map; carrying out statistics on local phase distribution of the phase consistency feature map, determining a phase distribution entropy representing physical attributes of an access medium, and carrying out weight reduction suppression on the phase consistency feature map according to the phase distribution entropy; according to the method, phase interference introduced by an unnatural display medium is recognized and suppressed, in-vivo detection is achieved in identity verification, and the influence of heterogeneous imaging amplitude distortion on recognition precision is eliminated.

Description

TECHNICAL FIELD

[0001] The present application relates to an image recognition method for cross-domain access authentication, belonging to the technical field of image recognition. BACKGROUND

[0002] At present, in the distributed access authentication system, extracting features from the original image collected by the terminal to be authenticated and comparing the identity belong to the mainstream technical direction. The existing technology usually uses deep neural network to perform feature mapping in the image spatial domain, trying to convert the pixel distribution of different sensors to a unified feature space through nonlinear transformation. Due to the hardware heterogeneity of the access terminal, there are performance differences between the sensor sensitivity and the lens resolution, resulting in nonlinear amplitude distortion in the original image. This amplitude response deviation caused by hardware makes the spatial feature extraction process rely on a large amount of cross-domain alignment data. When the terminal to be authenticated is in a low-illumination environment or an optical resolution limited working condition, the spatial feature mapping function is easy to deviate, reducing the recognition accuracy. The industry increases the model depth or introduces domain adversarial network to fit the cross-domain feature distribution, which increases the computational load of the edge device, and fails to eliminate the interference of hardware optical gain on image structure semantics from the physical mechanism level.

[0003] In addition to the inherent performance differences of the front-end hardware, the existing access control method also shows logical limitations in processing the consistency of cross-domain signals. For example, the Chinese invention patent with authorization publication number CN114037457B discloses an industrial complex product terminal cross-domain access authentication method based on identity identification, which uses identification password technology and session key negotiation mechanism to optimize the registration process during terminal movement at the protocol level, effectively reducing the resource loss of inter-domain authentication. However, this kind of control logic often presets the standardization and purity of the front-end collected data, ignoring the physical gain interference in the imaging process. When the signal to be authenticated is affected by sensor thermal noise, directional motion blur or wide-angle lens geometric distortion, it is difficult to make up for the precision collapse of the original image in physical representation by simply relying on the handshake logic of the upper protocol. At the same time, the existing scheme generally lacks the ability to perceive the physical properties of the access medium, and cannot identify and block the injection of non-natural phase signals such as high-definition screen flipping within the authentication framework, resulting in insufficient security redundancy of the system when facing physical layer attacks in distributed heterogeneous environments. In view of the amplitude distortion and environmental adaptability bottleneck caused by distributed heterogeneous terminals, bypassing the unstable spatial pixel intensity distribution, extracting structure-invariant features decoupled from hardware gain characteristics from the frequency space has become a research direction to improve the reliability of cross-domain authentication.

[0004] Therefore, how to construct an image recognition mechanism that can decouple the terminal hardware optical characteristics and has environmental adaptability, and realize identity authentication in a heterogeneous terminal environment, has become a technical problem to be solved by the present application. SUMMARY

[0005] To solve the problems presented in the background art, the technical solutions of the present application are as follows: An image recognition method for cross-domain access authentication, comprising the following steps:

[0006] Step S1, acquiring an image to be authenticated uploaded by an image information collection node;

[0007] Step S2, transforming the image to be authenticated from a spatial domain to a frequency domain, and performing a filtering operation on the frequency domain image using a multi-scale log-Gaussian filter bank to extract complex frequency response signals of the image to be authenticated at multiple preset directions and multiple preset scales, the complex frequency response signals being composed of real part response components and imaginary part response components;

[0008] Step S3, calculating the sum of the local energies of the real part response components and the imaginary part response components as a local total energy, calculating the modulus of the sum of the response vectors of the real part response components and the imaginary part response components at each pixel position, determining the ratio of the modulus to the local total energy as a phase consistency measure of the image to be authenticated at each pixel position, and generating a phase consistency feature map according to the phase consistency measures of each pixel position;

[0009] Step S4, counting the phase angle distribution of the phase consistency feature map at each pixel position with respect to the preset scale, calculating a phase distribution entropy representing the physical properties of the access medium, performing a numerical subtraction operation on the phase distribution entropy and a preset medium discrimination threshold, and mapping the difference obtained by the operation as a feature suppression factor for each pixel position; performing a weight reduction operation on the feature amplitude of the corresponding pixel position in the phase consistency feature map using the feature suppression factor to obtain a suppressed feature map;

[0010] Step S5, extracting a set of local extreme points from the suppressed feature map as structure anchor points, constructing an identity representation vector according to the geometric distribution topological relationship of the structure anchor points, and performing a matching operation on the identity representation vector and a preset registration template, and outputting an authentication instruction for the image information collection node according to the matching operation result.

[0011] Preferably, the calculation of the phase consistency measure in step S3 further comprises: obtaining the component amplitude sum of the real part response components and the imaginary part response components at each preset scale, and accumulating the component amplitude sum at each preset scale to obtain the local total energy; the modulus of the sum of the local response vectors is obtained by calculating the square root of the sum of the squares of the sum of the real part response components and the sum of the imaginary part response components at each preset scale, so as to represent the edge structure of the image to be authenticated by calculating the phase resonance strength between each preset scale.

[0012] Preferably, in step S2, the filters included in the multi-scale log-Gaussian filter bank are distributed in a geometric progression in the log-frequency space, and the center frequencies of the filters overlap with each other to lock the invariant features of the image to be authenticated at different optical resolutions.

[0013] Preferably, the method further comprises step S6: obtaining a focal length parameter of the optical imaging module in the image information acquisition node; calculating a radial distance of each pixel point to the center of the image , and calculating a compensation phase term of each preset scale of the filter operator , which is expressed by the formula: , wherein and are preset edge distortion correction coefficients; and performing phase pre-compensation processing on the complex frequency response signal using the compensation phase term to correct the phase shift caused by the radial distortion of the optical imaging module.

[0014] Preferably, in step S4, the phase distribution entropy is obtained by calculating the phase angle variance of each pixel position at each preset scale, so as to quantify the phase arrangement consistency of the image to be authenticated in the frequency domain space.

[0015] Preferably, in step S4, performing the weight reduction operation using the feature suppression factor includes: when the phase distribution entropy is greater than a medium discrimination threshold, determining that the value of the feature suppression factor is less than 1; and multiplying the feature amplitude at the corresponding position by the feature suppression factor to weaken the periodic stripe features introduced by the non-natural display medium.

[0016] Preferably, in step S5, constructing the identity representation vector based on the geometric distribution topological relationship of the structural anchor points includes: extracting local phase information and coordinate information of each structural anchor point, taking the local phase information as a feature element, and encoding the identity representation vector according to the arrangement order of the coordinate information.

[0017] Preferably, before step S1, the method further comprises: identifying a device type label of the image information acquisition node, and calling a corresponding preset channel noise model according to the device type label to perform gray mean value equalization processing on the original image stream to be collected.

[0018] Preferably, step S2 specifically includes: mapping the image to be authenticated from the spatial domain to the frequency domain using fast Fourier transform; performing point-by-point multiplication operation on the mapped image signal and the log-Gaussian filter operator in the frequency domain; and performing inverse Fourier transform on the result of the multiplication operation to obtain real part response components and imaginary part response components corresponding to the direction and the scale.

[0019] Preferably, the method further comprises step S7: when the authentication instruction is authentication pass, generating an authorization message containing an encrypted access token, and delivering the authorization message to the image information acquisition node.

[0020] The beneficial effects of the present application compared to the prior art are:

[0021] 1. In image recognition, a feature alignment mechanism based on the orthogonality of physical signals is established to eliminate the amplitude response deviation between heterogeneous acquisition terminals. By extracting the phase consistency measure of the image in the frequency domain, the recognition process is freed from the dependence on the absolute value of the pixel amplitude. The phase resonance characteristics of the image structural features in the frequency space are utilized. In the calculation process, the common hardware-related optical calibration gain term contained in the numerator and denominator is offset by the normalization ratio operation, the amplitude response and the structural semantics are decoupled, the original invariant feature map independent of the terminal sensor light sensitivity difference and lens resolution fluctuation is obtained, and the cross-domain authentication process no longer depends on the pre-training of massive cross-domain alignment data, thereby eliminating the feature alignment failure caused by uneven terminal hardware performance in distributed access scenarios.

[0022] 2. Multi-dimensional feature reuse of intermediate variables in the frequency domain is realized to synergistically enhance the recognition stability in extreme environments. The complex response vector generated by the multi-scale log-Gaussian transform is used to simultaneously perform noise floor adaptive masking and asymmetric spectral entropy directional compensation. The local energy variance between adjacent scales is extracted to identify electronic noise and effective structure. In the fusion stage, weights are assigned according to the spectral entropy of different angular directions, so that the system can actively suppress the random phase disturbance caused by sensor thermal noise in near-zero illumination or complex environments with directional motion blur, bypass the damaged physical information channel, and reconstruct the identity representation vector only with the undisturbed directional components, thereby improving the feature purity of the authentication system in extreme dynamic working conditions without introducing additional denoising operators.

[0023] 3. An endogenous anti-counterfeiting barrier based on cross-scale phase locking characteristics is constructed to identify the physical properties of the access medium. The difference in information redundancy between real physical entities and artificially reproduced media in the multi-scale frequency domain space is utilized. The phase distribution entropy is calculated by statistically analyzing the local phase angle distribution of each pixel position at all observation scales. It identifies the non-natural phase mutations introduced by the sub-pixel arrangement of display devices or printing dots. When the phase distribution entropy of a specific area exceeds the preset medium discrimination threshold, the system automatically performs local saliency weighting on the feature map, blocks the information transmission of the counterfeit access source, and makes the recognition algorithm complete the liveness detection simultaneously while performing identity verification, solving the security risks brought by high-definition screen screen capture in a single processing framework. BRIEF DESCRIPTION OF DRAWINGS

[0024] Fig. 1 The flowchart of the cross-domain access authentication image recognition core algorithm of the present application;

[0025] Fig. 2 The hardware deployment architecture and data interaction principle diagram of the authentication system of the present application. DETAILED DESCRIPTION

[0026] The technical solutions in the embodiments of the present application will be described clearly and completely below. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making objective judgments are within the scope of protection of the present application. The following embodiments are intended to explain and illustrate the present application, rather than to limit the scope of protection of the present application.

[0027] The present application provides an image recognition method for cross-domain access authentication. By converting the image to be authenticated from a spatial domain to a frequency space, using the phase resonance characteristics of different frequency components at the visual structure, extracting the structural feature map decoupled from the terminal hardware optical gain, and cooperatively using the intermediate variables in the frequency domain to perform noise floor suppression, directional blur compensation, medium attribute recognition, and radial distortion calibration, finally outputting the identity authentication instruction based on the structural anchor point topological relationship. To solve the problem of nonlinear amplitude distortion of images collected by heterogeneous terminals, the system obtains the image information collected by the nodes uploaded by the image to be authenticated, maps the image to be authenticated from the spatial domain to the frequency domain by using the fast Fourier transform, and calls the preset multi-scale log-Gaussian filter set to perform frequency domain decomposition. The multi-scale log-Gaussian filter set is distributed in a geometric multiple in the log frequency space, and the center frequencies overlap with each other, so as to capture the complex frequency response signals under different optical resolutions. The complex frequency response signal is composed of a real part response component and an imaginary part response component , wherein is the filter scale, is the filter direction, in a specific engineering deployment scenario, the system receives an image to be authenticated with a resolution of 640x480, selects a filter set with 4 scales and 6 directions to perform point-by-point multiplication operation, and performs inverse Fourier transform on the operation result, thereby obtaining a complex response vector field reflecting the local frequency characteristics at each pixel position, thereby realizing the conversion of visual structure information from amplitude dependence to frequency distribution; due to the difference in photosensitivity of different sensors reflected on the pixel amplitude, the recognition features are shifted when crossing the domain, based on this, the system uses the phase consistency measure to construct the structure invariant feature map, specifically by calculating the sum of the local amplitudes of each pixel position coordinate as the local total energy, that is , and calculating the modulus value of the sum of the response vectors under each scale, determining the ratio of the modulus value to the local total energy as the phase consistency measure , the calculation formula is as follows: , wherein is the phase consistency measure of each pixel position, is the frequency weight factor, is the local amplitude, is a phase deviation function, is a noise threshold parameter determined based on a non-structural region energy distribution, is a constant for avoiding zero denominator, for example, taking a value of 0.0001, the calculation process offsets the hardware-related optical calibration gain item commonly contained in the numerator and the denominator through the normalization ratio operation, realizes the decoupling of the amplitude response and the structural semantics, and makes the generated phase consistency feature map stable to illumination fluctuations.

[0028] noise threshold parameter The calibration selects the non-structural reference region at the edge of the framing frame after obtaining the image to be authenticated, counts the component amplitude and statistical distribution characteristics of the pixel position in each preset scale in the reference region, calculates the component amplitude and mean and the standard deviation , determines the noise threshold parameter according to the formula , wherein is an energy threshold for removing false feature points, is the average value of the component amplitude in the reference region, is the component amplitude and the standard deviation, is a preset confidence adjustment coefficient, and the adjustment coefficient The numerical value is selected in the interval of 2.0 to 3.0, and the noise floor suppression benchmark is adjusted under different illumination conditions, so that the phase consistency measure calculation focuses on the physical structure edge; in order to solve the problem that random phase disturbance caused by sensor thermal noise in low-illumination environment causes false structure interference in the feature map, the system adopts an adaptive noise floor mask method, the system identifies effective features and electronic noise by monitoring the local energy variance of the frequency domain response between adjacent scales, specifically selects a 3x3 spatial window, and counts the dispersion degree of the energy values of each scale in the window, if the energy variance of a certain pixel point exceeds the preset noise floor distribution model threshold, and the average energy is lower than the preset energy threshold, it is determined that the point is dominated by noise, at this time the system generates a feature suppression factor, and performs weight shrinkage processing on the phase consistency measure of the position, and the feature amplitude of the corresponding position is multiplied by a decay coefficient between 0 and 0.5, thereby filtering out false feature points caused by low signal-to-noise ratio without losing the sharpness of the real edge, and ensuring the physical certainty of the identification benchmark in dark environment; in order to solve the problem of directional motion blur caused by user handheld terminal jitter, the system adopts a directionality compensation method based on spectral entropy, the system counts the total spectral energy of each pixel position in different directions and calculates the spectral entropy ​wherein the direction with higher entropy value corresponds to clear structural edge, and the direction with lower entropy value indicates the existence of frequency attenuation, i.e. blur, according to the size of the spectrum entropy of each direction, the system determines the weight distribution by using a weight formula, and the weight formula is as follows: wherein, is the fusion weight corresponding to the direction ; is the spectrum entropy of the direction, by adjusting the weight of the clear direction in feature fusion and suppressing the phase shift interference of the blurred direction, the system bypasses the damaged physical information channel, reconstructs the identity representation vector by using the direction component with lower interference degree, and improves the stability of the recognition algorithm in the dynamic interaction scene.

[0029] To prevent identity impersonation attacks using high-definition screens or printed materials, the system identifies the physical properties of the access medium through cross-scale phase locking characteristics. Since display devices introduce a specific periodic sampling pattern when reconstructing images, the phase distribution of the display devices in the multi-scale frequency domain space has phase discontinuity. The system extracts the local phase angle of each pixel position at different scales and calculates the phase distribution entropy thereof. If the phase distribution entropy of a specific region exceeds a preset medium discrimination threshold, which is 0.85 in this embodiment, it is determined that there is a risk of medium injection in the region, and the weight of the region in the phase consistency feature map is forced to be zero at this time, thereby realizing the living body detection function. The feature suppression factor of each pixel position is determined by calculating the difference between the phase distribution entropy and the medium discrimination threshold , which represents the strength of the non-natural property of the access medium. The feature suppression factor is determined by using the formula , wherein is the weight coefficient acting on the corresponding position of the phase consistency feature map, is the difference between the phase distribution entropy and the medium discrimination threshold value, is a preset suppression strength gain coefficient, and the gain coefficient has a value range of 1.2 to 3.5 according to the sub-pixel arrangement frequency of 2025 mainstream display terminals, so that the response amplitude of the region with high entropy value of the fake signal in the feature map is attenuated to below the background noise level, and the non-natural signal propagation path is cut off at the physical feature level.

[0030] To solve the problem of radial phase distortion at the edge of the image caused by the wide-angle lens of the distributed terminal, the system implements coordinate-driven nonlinear phase pre-collision, calculates the normalized radial distance of each pixel position relative to the center coordinate of the image , and determines the phase compensation offset at each scale according to a correction formula, and the correction formula is as follows: wherein, is the phase compensation offset; to normalize the radial distance; to filter the scale; and to pre-calibrate the edge distortion correction coefficient according to the lens focal length parameter, the system will compensate the phase offset of each pixel position as a phase shift operator acting on the complex frequency response signal, through the complex rotation factor and the response component of the filter operator performs point-by-point multiplication operation, so as to preset the phase delay compensation in the frequency domain decomposition stage, so that the frequency components of the distortion area are physically aligned in the spatial position after inverse transformation.

[0031] After obtaining the phase consistent feature map, the system extracts a set of local extreme points from the suppressed feature map, determines the set as a structure anchor point, converts the local phase information and coordinate information of each structure anchor point into an identity representation vector according to a preset encoding order by recording the local phase information and coordinate information of each structure anchor point, and constructs the geometric distribution topological relationship of the structure anchor point and the identity representation vector using a polar coordinate system cascade encoding method. By taking the geometric center of the image to be authenticated as the coordinate origin, the radial distance , polar axis angle and local phase value of each structure anchor point are extracted, and the polar axis angle is sequentially sorted from 0 to , and the local phase value corresponding to each point after sorting is extracted to perform cascade operation to form a fixed-length original feature sequence, and an identity representation vector is generated by using a normalization function to perform amplitude scaling , wherein is a matching operation feature vector, which eliminates the interference of image rotation or translation on feature expression consistency by determining the encoding order, so that the matching operation is anchored in the physical feature dimension. The vector and the pre-stored registration template in the server perform matching operation, and if the matching score exceeds the authentication threshold, an authorization message containing an encrypted access token is generated and sent to the image information collection node. The whole process is realized by a determined physical feature model, which reduces the computing load of the gateway side.

[0032] Embodiment one: in the night access authentication scene of the logistics park, due to the low environmental light intensity below 5 lux and the nonlinear difference of the amplitude response curve of different access terminals, the traditional spatial pixel feature will lose recognition due to the lack of contrast when compared across domains. At this time, the image recognition method obtains the image to be authenticated through the image information collection node, maps the image to be authenticated from the spatial domain to the frequency domain, calls the multi-scale logarithmic Gaussian filter group to extract the complex frequency response signal in 4 scales and 6 directions, and calculates the pixel position coordinates The sum of local amplitudes is taken as the local total energy. At the same time, the magnitude of the sum of response vectors at each scale is calculated, and the ratio of the magnitude to the local total energy is determined as the phase consistency measure. The calculation formula is as follows: ,in, A measure of phase consistency at each pixel location; Frequency weighting factor; This refers to the local amplitude. The phase deviation function is defined as a fixed calibration vector pre-stored in a static buffer in the calculation process. Its value originates from the dark current phase shift sampling of 4096 pixels of the sensor in a dark environment during the initial power-on of the system. The value range is calibrated between 0.005 rad and 0.025 rad, serving as a static denominator correction term for the phase consistency measure. The phase pre-adjustment process in step S6 performs secondary real-time fine-tuning for the dynamic distortion caused by the shift of the lens's physical center of gravity. The two processes work together to lock the physical phase coordinates through linear accumulation. This is the noise threshold parameter; The value is a constant of 0.0001. This calculation process cancels out the hardware-related optical calibration gain term through normalized ratio operation, thereby decoupling the amplitude response from the structural semantics.

[0033] Under conditions where sensor thermal noise and directional motion ambiguity are superimposed, image recognition methods monitor the local energy variance of the frequency domain response between adjacent scales. To identify electronic noise, a 3×3 spatial window is selected, and the dispersion of energy values ​​at each scale within this window is statistically analyzed. If the energy variance of a certain pixel is... If the value exceeds a preset threshold and its mean energy is below a preset threshold, then the point is determined to be dominated by noise. In this case, a feature suppression factor is generated to perform weighted shrinkage processing on the phase consistency measure, and the position of each pixel in different directions is statistically analyzed. Sum the spectral energy and calculate the spectral entropy of the signal in each direction. The weight allocation is determined based on the magnitude of the spectral entropy in each direction using a weighting formula, as follows: ,in, For weighting; To address the spectral entropy, by increasing the weight of the sharp direction and suppressing phase shift interference in the blurred direction, the image recognition method bypasses the damaged physical information channel. It reconstructs the identity representation vector using the less disturbed direction components. To address the radial phase distortion caused by wide-angle lenses, the image recognition method implements coordinate-driven nonlinear phase pre-counterbalancing, calculating the normalized radial distance of each pixel position relative to the image center coordinates. The phase compensation offset at each scale is determined according to the correction formula. ,in, is a phase compensation offset, is a normalized radial distance, is a filter scale, and is an edge distortion correction coefficient, a phase shift correction is performed on the complex response vector by using the phase compensation offset, the frequency components of the distortion area are re-aligned in the corresponding geometric position to eliminate the phase tailing phenomenon, and consistent recognition accuracy is ensured for the user within the global viewfinder frame; after obtaining the phase consistency feature map, the image recognition method extracts a structural anchor point from the feature map, records its local phase information and coordinate information, converts it into an identity representation vector according to a preset encoding order, and performs cosine similarity matching with a server pre-stored registration template; when the matching score exceeds an authentication threshold of 0.85, an authorization message containing an encrypted access token is generated and issued to an image information collection node, and the whole process is realized through a determined phase domain physical feature model, so that the identity authentication result exhibits inherent stability to light fluctuations, sensor noise and lens distortion.

[0034] Example two: in the night industrial logistics park access scene with illumination in the 1 lux to 10 lux gradient, due to the access terminal covering the fixed focus monitoring unit and the mobile intelligent terminal, the nonlinear difference of the imaging unit sensitivity causes the image signal to appear recognition offset when cross-domain comparison, in order to verify the effectiveness of the mechanism, the test is established on a distributed computing verification platform, which contains two image collection nodes with different dynamic range characteristics, in the test process, Gaussian white noise with a signal-to-noise ratio of 25 dB is injected into the original image data to simulate sensor thermal noise disturbance, the test sets the sampling resolution to 1280x720, the filter group scale n is set to 4 and the direction is set to 6, taking the 5 lux working condition as an example, the average contrast of the original image data in the spatial domain is less than 0.2 and is accompanied by pixel-level particle interference, at this time, the feature map output by the control group exhibits edge collapse phenomenon, and the average value of its recognition accuracy measurement is 62.4%; the sample group of the present application maps the original image containing 25 dB noise to the frequency domain by performing fast Fourier transform, the system calculates the local energy variance between adjacent scales, the measurement data shows that at the noise dominant point, the energy variance is between 1.28 and 1.56, while the energy variance at the structural edge is less than 0.35, based on this physical feature difference, the system uses an adaptive noise floor mask to suppress the phase weight of the noise area by 85%, in the suppressed complex response field, the spectral entropy of the clear edge direction is 4.82, while the directional spectral entropy affected by the directional motion blur is reduced to 2.15, by using the formula to determine the weight distribution, the fusion weight The peak signal-to-noise ratio of the processed phase consistency feature map under 5 lux illumination was increased to 0.72, from 18.4 dB in the original data to 28.6 dB, confirming that the adaptive noise floor mask provides a high signal-to-noise ratio input prerequisite for directional compensation.

[0035] For key parameter filter scale The value boundary is determined by setting multiple gradient comparison sample groups in the experiment. When the median value is 4, the system achieves a recognition accuracy of 91.2% at 1 lux illumination, while when... When the value is lowered to 2, the recognition accuracy drops to 76.5%. Increasing the threshold to 6 only improved the recognition accuracy to 91.8%, but resulted in a 142ms increase in system computation latency, indicating that the system... The system reaches a performance saturation point around 4. Further verification using a partially missing control group showed that after removing the adaptive noise floor mask, the false recognition rate in a 1-lux environment increased from 0.8% in the original sample to 12.4%. This performance inflection point confirms the causal relationship between phase resonance characteristics and noise suppression mechanisms under low illumination. To verify the correction effectiveness of radial phase distortion, the system addresses the edge phase trailing phenomenon caused by wide-angle lenses using a correction formula. Perform phase shift compensation at normalized radial distance For the viewfinder edge region with a value of 0.9, the measured value of the uncompensated complex response phase deviation is 0.82 rad. After offsetting with correction coefficients, the residual phase deviation is reduced to less than 0.05 rad. As a result, the topological consistency measure of the extracted structural anchor points in coordinate space is restored from 0.64 to 0.92. Finally, the consistency deviation of the matching score generated based on cosine similarity between heterogeneous terminals remains within 2.8%.

[0036] Example 3: This example combines Figs. 1-2 This describes an image recognition method for cross-domain access authentication, such as... Fig. 1As shown, the step S1 is performed to acquire the image information collected by the image information collection node, and the step S2 is performed to convert the to-be-authenticated image from the spatial domain to the frequency domain, perform filtering operation by using the multi-scale log-Gaussian filter set, extract the complex frequency response signal containing the real part and the imaginary part under the multiple preset directions and multiple preset scales, and then perform the step S3 to calculate the sum of the local energy of the real part and the imaginary part, that is, the local total energy, and the modulus value of the sum of the response vectors, determine the phase consistency measure of each pixel position by using the ratio of the modulus value to the local total energy, generate the phase consistency feature map, and then perform the step S4 to count the phase angle distribution of the phase consistency feature map with the scale change, calculate the phase distribution entropy representing the physical properties of the access medium, generate the feature suppression factor according to the difference between the phase distribution entropy and the medium discrimination threshold, perform the weight reduction operation on the feature amplitude to obtain the suppressed feature map, and finally perform the step S5 to extract the local extreme point set from the suppressed feature map as the structure anchor point, and construct the identity representation vector according to the geometric distribution topological relationship, perform the matching operation with the registration template, and output the authentication instruction according to the result.

[0037] As shown in the figure, Fig. 2 The image information collection node is deployed at the front end of the system, which covers device types such as access terminals, fixed-focus monitoring units or mobile intelligent terminals, and the internal hardware is integrated with an optical imaging module, a sensor, a wide-angle lens and a preset channel noise model, which is responsible for uploading the to-be-authenticated image to the gateway side and receiving the authorization message containing the encrypted access token, the data processing core of the system is the gateway side embedded device, which performs the identity authentication task, and the internal logic components include a multi-scale log-Gaussian filter set and an identity authentication module, wherein the identity authentication module specifically performs the operation logic of phase consistency measure calculation, phase distribution entropy calculation, live detection and identity representation vector construction, and the device is internally configured with a storage unit for storing edge distortion correction coefficients, filter center frequency and noise threshold parameters, the gateway side embedded device is connected with the server providing background data support through the matching operation interaction interface, and mainly responsible for maintaining the pre-stored registration template.

[0038] In the access authentication risk test environment including a high-definition display terminal with a resolution not less than 1920x1080 and a paper print medium, the image recognition method acquires the to-be-authenticated image through the image information collection node, in order to determine the noise suppression boundary in the phase domain decomposition stage, the system counts the local total energy of the non-structural component in the 32x32 pixel reference area at the edge of the image , calculates the energy mean value and the energy standard deviation of the non-structural component , and determines the noise threshold parameter according to the noise threshold formula , wherein Energy mean; Energy standard deviation; Gain adjustment coefficient, in the embodiment, it is 2.5, so that the phase feature extraction process is self-adapted according to the current imaging unit noise floor level; in order to identify the physical properties of the access medium, the image recognition method performs a phase distribution entropy extraction procedure, specifically, the local phase angle of each pixel position at different scales is mapped to a phase histogram containing 16 quantization intervals, and the frequency of occurrence of each quantization interval is counted The phase distribution entropy of each pixel region is calculated, wherein, is the phase distribution probability of the first quantization interval, when the phase distribution entropy in the continuous 8x8 pixel region exceeds the medium discrimination threshold of 0.85, it is determined that there is a risk of non-natural display medium injection in the region, at this time, the weight of the region in the phase consistency feature map is set to 0 through the feature suppression factor.

[0039] After extracting the set of structure anchor points, the system starts a 7x7 pixel sliding detection window in the suppressed feature map, when the phase consistency value of the center point of the window is higher than 1.25 times the average value of all pixels in the window, and the absolute value of the center point exceeds the hard determination threshold of 0.15, it is determined as a structure anchor point, in this way, the pseudo feature points caused by low contrast texture are filtered out, when reading the local phase information of each structure anchor point, the system uses an 8-direction hierarchical quantization method to map the phase angle of each feature point to a preset quantization interval index set, and cooperates with the normalized radius coordinate and angle coordinate in the polar coordinate system to form a feature tuple, each tuple is arranged in descending order of feature point degree to form a fixed-length identity representation vector, which is input into the matching module, the system calculates the Hamming distance between the identity representation vector and the pre-stored registration template of the server, if the Hamming distance is less than the preset authentication tolerance, an authorization message containing an encrypted access token is generated and sent to the image information collection node, this process realizes the closed loop of identity authentication result at the physical feature level through quantization parameter calibration and feature sequence uniqueness mapping.

[0040] Embodiment five: in a distributed access scene containing multiple specifications of imaging units, since the cutoff frequency distribution of the original image data shifts with the change of the photosensitive pixel size, the system performs parameter initialization of the multi-scale log-Gaussian filter group before deploying the image information collection node, by inputting a preset sinusoidal pulse texture signal to the image collection node, the system monitors the amplitude concentration of the complex response signal at each scale , according to the real part response component and the imaginary part response component The energy distribution law, select the filter center frequency of the structure edge response, the filter center frequency is written into the storage unit as the phase consistent performance energy extraction stage of the calibration reference.

[0041] When the system is facing electronic noise floor interference caused by sensor manufacturing tolerance, the system acquires the reference bias data of the image information acquisition node in the dark environment, the complex response distribution characteristics of the reference bias data in the frequency domain are counted, and the energy mean value in the non-structure background is calculated And the energy standard deviation And according to the safety level of the current application scene, the value of the gain adjustment coefficient The noise threshold parameter Is generated and injected into the gateway side identity authentication module, so that the system establishes a feature noise suppression baseline matched with the hardware properties of the image information acquisition node in the process of executing hash mapping and hamming distance calculation.

[0042] Embodiment six: in the distributed deployment scene involving multiple types of wide-angle imaging units, the system executes the calibration program of the radial phase distortion parameter, acquires the image of the checkerboard calibration target located at the preset object distance, calculates the normalized radial distance of each corner point coordinate relative to the image center coordinate , while extracting the local phase response value of each corner point in the multi-scale filter space, recording the cross-scale phase deviation observation value and constructing the observation matrix containing the normalized radial distance , filter scale And phase compensation offset The edge distortion correction coefficient And The numerical solution is determined by using the least square method to perform surface fitting operation on the observation matrix, and is written into the storage unit, and the calibration process provides operator input parameters for executing coordinate-driven nonlinear phase preflush.

[0043] In the initialization stage of the system identity authentication module, the determination of the authentication threshold and the medium discrimination threshold is based on the statistical procedure of the preset safety level, the system selects the verification sample set containing authorized identity and unauthorized identity for matching operation, and the false rejection rate distribution in different matching score intervals is counted. Select the score value with false rejection rate less than 0.01% to determine the threshold of identity authentication. For the sampling features generated by the display medium, the system tests the phase distribution entropy measurement value under different refresh rates, and determines the medium discrimination threshold as 1.2 times of the distribution mean value. This debugging procedure converts the statistical law into a logic judgment threshold, so that the image recognition method can establish a discrimination standard matched with the physical working condition when facing different imaging hardware and access environment.

[0044] It is apparent to those skilled in the art that the application is not limited to the details of the foregoing exemplary embodiments and that the present application can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application.

[0045] Finally, it should be noted that the above embodiments are merely used to illustrate, but not limit the present application, and although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical scheme of the present application can be modified or equivalently replaced without departing from the spirit and scope of the present application.

Claims

1. An image recognition method for cross-domain access authentication, characterized in that, Includes the following steps: Step S1: Obtain the image to be authenticated uploaded by the image information acquisition node; Step S2: Transform the image to be authenticated from the spatial domain to the frequency domain, and use a multi-scale logarithmic Gaussian filter bank to perform filtering operation on the frequency domain image to extract the complex frequency response signal of the image to be authenticated in multiple preset directions and multiple preset scales. The complex frequency response signal is composed of real response components and imaginary response components. Step S3: Calculate the sum of the local energies of the real response component and the imaginary response component as the local total energy, calculate the magnitude of the sum of the response vectors of the real response component and the imaginary response component at each pixel position, determine the ratio of the magnitude to the local total energy as the phase consistency measure of the image to be authenticated at each pixel position, and generate a phase consistency feature map based on the phase consistency measure of each pixel position. Step S4: Statistically analyze the phase angle distribution of the phase consistency feature map at each pixel position as the preset scale changes, and calculate the phase distribution entropy that characterizes the physical properties of the access medium; perform a numerical subtraction operation between the phase distribution entropy and the preset medium discrimination threshold, and map the difference obtained into a feature suppression factor for each pixel position. By using a feature suppression factor to perform a weighting operation on the feature amplitude at the corresponding pixel position in the phase consistency feature map, a suppressed feature map is obtained. Step S5: Extract the set of local extreme points from the suppressed feature map as structural anchor points, construct an identity representation vector based on the geometric distribution and topological relationship of the structural anchor points, perform a matching operation between the identity representation vector and the preset registration template, and output an authentication command for the image information acquisition node based on the matching operation result.

2. The image recognition method for cross-domain access authentication according to claim 1, characterized in that, Step S3, which calculates the phase consistency measure, also includes: obtaining the sum of the component amplitudes of the real and imaginary response components at each preset scale, and summing the component amplitudes at each preset scale to obtain the local total energy; the magnitude of the sum of local response vectors is obtained by calculating the square root of the sum of the squares of the real and imaginary response components at each preset scale, so as to characterize the edge structure of the image to be authenticated by calculating the phase resonance intensity between each preset scale.

3. The image recognition method for cross-domain access authentication according to claim 1, characterized in that, In step S2, the filters in the multi-scale logarithmic Gaussian filter bank are distributed in a geometric multiple in the logarithmic frequency space, and the center frequencies of each filter overlap with each other, so as to lock the invariant features of the image to be authenticated under different optical resolutions.

4. The image recognition method for cross-domain access authentication according to claim 1, characterized in that, The method also includes step S6: obtaining the focal length parameters of the optical imaging module in the image information acquisition node; based on the radial distance from the pixel to the image center... Calculate each preset scale Compensation phase term of the lower filter operator The formula is expressed as follows: ,in, and The preset edge distortion correction coefficients are used; the complex frequency response signal is pre-counted using a compensation phase term to correct the phase shift caused by the radial distortion of the optical imaging module.

5. The image recognition method for cross-domain access authentication according to claim 1, characterized in that, In step S4, the phase distribution entropy is obtained by calculating the phase angle variance of each pixel position at each preset scale, which is used to quantify the phase alignment consistency of the image to be authenticated in the frequency domain space.

6. The image recognition method for cross-domain access authentication according to claim 1, characterized in that, In step S4, the weighting operation using the feature suppression factor includes: when the phase distribution entropy is greater than the medium discrimination threshold, determining that the value of the feature suppression factor is less than 1; multiplying the feature amplitude at the corresponding position with the feature suppression factor to weaken the periodic stripe features introduced by the non-natural display medium.

7. The image recognition method for cross-domain access authentication according to claim 1, characterized in that, In step S5, constructing the identity representation vector based on the geometric distribution topology of the structural anchor points includes: extracting the local phase information and coordinate information of each structural anchor point, and encoding the local phase information as a feature element according to the arrangement order of the coordinate information into an identity representation vector.

8. The image recognition method for cross-domain access authentication according to claim 1, characterized in that, Before step S1, the method further includes: identifying the device type label of the image information acquisition node, and retrieving the corresponding preset channel noise model according to the device type label, and performing grayscale mean equalization processing on the raw image stream to be acquired.

9. The image recognition method for cross-domain access authentication according to claim 1, characterized in that, Step S2 specifically includes: mapping the image to be authenticated from the spatial domain to the frequency domain using a fast Fourier transform; performing a point-by-point product operation between the mapped image signal and the log-Gaussian filter operator in the frequency domain; and performing an inverse Fourier transform on the result of the product operation to obtain the real and imaginary response components for the corresponding direction and scale.

10. The image recognition method for cross-domain access authentication according to claim 1, characterized in that, The method also includes step S7, which generates an authorization message containing an encrypted access token when the authentication command indicates successful authentication, and sends the authorization message to the image information acquisition node.

Images