Deep learning-based face recognition methods

Abstract

This invention discloses a deep learning-based face recognition method, belonging to the field of face recognition technology, including the following steps: At the moment of high-speed occlusion release of a person's face, a photon accumulation profile is established based on the principle of temporal inversion, and an inverse phase compensation curve is constructed under the photon accumulation profile to predict and simulate the trajectory of photons that have not completely dissipated within the exposure cycle; under the constraint of the photon accumulation profile, a cross-frame energy flow map is generated, and abnormal energy and normal energy are differentially projected in the cross-frame energy flow map. This invention, by performing multi-dimensional compensation and decoupling of photon trajectories and afterimage signals in high-speed occlusion scenarios, ensures the stability of facial features in continuous frames, and achieves layer-by-layer stripping and correction within the deep feature mapping space. Combined with dynamic masking to suppress afterimage pixels, it ensures the integrity and continuity of recognition features, effectively improving the accuracy and robustness of face recognition and meeting the security requirements of highly sensitive application scenarios.

Description

Technical Field

[0001] This invention relates to the field of facial recognition technology, and more specifically to a facial recognition method based on deep learning. Background Technology

[0002] Deep learning-based facial recognition technology originated from the deep integration of artificial intelligence and computer vision. By constructing and training deep neural networks, it can automatically extract multi-level feature information from a large number of facial images, forming a high-dimensional representation of the face. Unlike traditional methods that rely on handcrafted features, the end-to-end training method of deep learning allows the model to maintain high recognition accuracy under different lighting, poses, expressions, and even occlusion conditions. This technology has extremely wide applications, from identity verification, security monitoring, and financial payments to smart city construction and human-computer interaction, all of which place higher demands on the real-time performance and accuracy of facial recognition. With the help of deep learning methods such as convolutional neural networks (CNNs), generative adversarial networks (GANs), and attention mechanisms, researchers and industry have achieved organic integration of face detection, feature extraction, and similarity measurement, making facial recognition a crucial cornerstone supporting trusted identity authentication and intelligent services in the digital society.

[0003] The existing technology has the following shortcomings:

[0004] In dynamic scenes, when a person's face is occluded at high speed for an extremely short period and then released instantaneously, the camera imaging chain often cannot complete the image update of complete dissipation within a single exposure cycle, resulting in a short-term ghosting effect at the sensor output. This ghosting signal, when entering the deep learning recognition network, superimposes with the real facial features, forming a mixed and unstable feature encoding, causing a severe shift in key feature dimensions of the network. This shift not only disrupts the continuity of deep features during face recognition but also causes the system to reject legitimate users during the authentication phase, leading to a significant decrease in the reliability of the identity verification process. In highly sensitive scenarios involving financial payments and security control, this can easily lead to serious consequences.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a deep learning-based face recognition method to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based face recognition method, comprising the following steps:

[0008] S1. At the moment when the face of the person is subjected to high-speed occlusion and release, a photon accumulation profile is established based on the time-series inversion principle, and an inverse phase compensation curve is constructed under the photon accumulation profile to predict and simulate the trajectory of photons that have not completely dissipated within the exposure cycle, so as to provide a time reference for the stable recognition of subsequent facial features.

[0009] S2. Under the constraint of the photon accumulation profile, a cross-frame energy flow map is generated, and the abnormal energy and normal energy are differentially projected in the cross-frame energy flow map to form a refined energy layer. Boundary conditions are established for afterimage detection in the energy layer.

[0010] S3. With the support of energy layering, a force field traction mechanism is introduced to apply reverse phase traction to the high-energy anomalous layer in energy layering, so as to dynamically decouple the afterimage region from the real face region, shrink the afterimage boundary and provide conditions for feature reconstruction.

[0011] S4. Based on the shrinkage of the afterimage boundary, construct a temporal replay link, perform counterfactual playback on the key frames before and after the occlusion, compare the evolution trajectory of the real face features with the afterimage signal frame by frame, identify the offset pattern of the afterimage on the time axis, and provide anchor points for feature correction.

[0012] S5. After the offset mode of the afterimage is locked, a multi-scale projection stripping strategy is established to strip the afterimage features from the depth feature mapping space step by step. During the step-by-step stripping process, the real face features are corrected layer by layer and written back to the main space to prevent the collapse of the depth features during the face recognition process.

[0013] S6. After completing the layer-by-layer correction, a residual prediction matrix is ​​constructed between consecutive exposure cycles to detect abnormal brightness residuals caused by high-speed occlusion in real time. The residuals are then mapped to dynamic masks and injected into the deep learning feature extraction network to automatically weaken afterimage pixels and maintain the continuity and stability of face recognition features.

[0014] Preferably, the steps of establishing a photon accumulation profile and constructing an inverse phase compensation curve at the instant when the face experiences high-speed occlusion release include:

[0015] The photon stream received by the sensor is collected in a time-division manner at the moment the face is occluded and released. The single exposure cycle is divided into multiple continuous time segments, and the photon incident intensity, arrival time and spatial distribution of each time segment are recorded one by one. The photon accumulation profile is obtained based on the inversion method.

[0016] After the photon accumulation profile is established, a time series analysis is performed on the energy change trend of the profile to identify the sharp rise segment and the slow decay segment. An energy decrease curve in the opposite direction is constructed in the sharp rise segment, and a gradually enhanced reverse phase compensation curve is applied in the slow decay segment to form an inverse phase compensation curve.

[0017] After the inverse phase compensation curve is constructed, it is superimposed point by point with the photon accumulation profile to obtain the compensated and corrected energy change curve. Based on this curve, the trajectory of the photon that has not completely dissipated is extended and predicted. The prediction result is mapped to the recognition time axis to provide a stable reference for subsequent face feature recognition.

[0018] Preferably, the steps of generating a cross-frame energy flow map and forming a refined energy layer under the constraint of the photon accumulation profile include:

[0019] Under the time constraint of the photon accumulation profile, the spatial distribution data of photon energy of consecutive adjacent exposure frames are collected and arranged in chronological order to generate a cross-frame energy flow map to record the propagation and dissipation trajectory of photons between multiple frames.

[0020] Based on the cross-frame energy flow map, the dissipation law of each energy trajectory is compared with the photon accumulation profile. Trajectories that deviate from the law are projected as abnormal energy layers, and trajectories that conform to the law are projected as normal energy layers, thus realizing differentiated projection of abnormal energy and normal energy.

[0021] After differential projection is completed, the abnormal energy layer and normal energy layer are further divided into multiple energy sub-layers, and a refined energy layering is formed by referring to the migration path of the cross-frame energy flow map.

[0022] With the support of refined energy stratification, the ghosting detection boundary is set in the boundary area between the abnormal energy sub-layer and the normal energy sub-layer according to the degree of deviation of the abnormal energy trajectory and the energy deviation value, so as to delineate the range of ghosting signal existence.

[0023] Preferably, when setting the image retention detection boundary with the support of refined energy stratification, the region where the energy decay rate is lower than the dissipation law of the photon accumulation profile is identified as the image retention region, and the boundary is locked at the outer edge of the abnormal energy sublayer to improve the accuracy of image retention detection.

[0024] Preferably, the steps of introducing a force field traction mechanism with the support of energy stratification include:

[0025] In the refined energy stratification results, high-energy anomalous layers are identified. Their energy intensity is compared point by point with the standard dissipation curve of the photon accumulation profile. When the energy intensity continuously exceeds the standard range and the decay rate is significantly lagging, the energy layer is identified as a high-energy anomalous layer and marked in combination with the anomalous trajectory information of the differential projection.

[0026] After the high-energy anomalous layer is identified, the phase shift direction and magnitude are determined based on the constraints of the photon accumulation profile and the inverse phase compensation curve. A force field vector is introduced in this direction to apply reverse phase pulling, and the pulling intensity is monitored in real time to avoid acting on the normal energy layer.

[0027] Under the reverse phase traction, the afterimage region and the real face region are gradually decoupled dynamically. By dynamically adjusting the range of the traction vector, the overlap between the two is reduced, so that the afterimage region and the real region form a relatively independent distribution.

[0028] After dynamic decoupling is completed, the afterimage boundary is gradually shrunk, and the shrunk afterimage boundary is compared with the cross-frame energy flow map to ensure that the afterimage signal is limited to a controllable range, thereby providing stable conditions for subsequent feature reconstruction.

[0029] Preferably, the steps of constructing a temporal replay link and performing counterfactual playback based on the shrinkage of the afterimage boundary include:

[0030] Under the premise of shrinking afterimage boundaries, key frames containing complete facial features before occlusion and key frames with mixed afterimage features after occlusion disappear from continuous video frames are extracted, and a temporal replay link is established in chronological order to preserve energy distribution and photon trajectory features.

[0031] Based on the temporal replay link, the key frame before occlusion is used as the reference trajectory for the evolution of real face features, and the key frame after occlusion is replayed frame by frame in a counterfactual manner. The evolution difference between the overall contour and local details is compared, and the abnormal parts concentrated at the afterimage detection boundary are identified as afterimage offset signals.

[0032] After the ghosting offset signal is identified, its offset direction, amplitude and duration on the time axis are extracted, and the offset pattern is mapped to the corresponding position of the time-series replay link to set anchor points, so that subsequent feature correction can use the anchor points to correct ghosting interference in advance when processing new input frames.

[0033] Preferably, the steps for establishing a multi-scale projection stripping strategy after the afterimage's offset mode is locked include:

[0034] Under the premise that the afterimage offset mode is locked, a multi-scale projection structure is established, the feature mapping space is divided into edge feature layer, texture feature layer and semantic feature layer, and the afterimage offset mode is mapped to each scale layer to locate the afterimage distribution.

[0035] With the support of a multi-scale projection structure, the afterimage features are stripped off step by step in the edge layer, texture layer and semantic layer, while maintaining the continuity of the real features to avoid over-stripping.

[0036] While peeling away the image layer by layer, the real features affected by the afterimage are corrected layer by layer, and the corrected edge information, texture distribution and semantic contours are written back to the main space in turn to accumulate stable feature representations.

[0037] After multi-scale stripping and writing back are completed, consistency detection is performed on the main space, the corrected features of each layer are compared with the original features, and when residual images are found, a second stripping and correction is performed to ensure that the final features remain stable and consistent.

[0038] Preferably, the steps of constructing the residual prediction matrix and injecting the dynamic mask after completing the layer-by-layer correction include:

[0039] After the layer-by-layer correction is completed, the adjacent exposure cycles are regarded as time windows, the corrected facial feature photon energy distribution is recorded, and the previous exposure cycle is used as a reference to compare the brightness difference with the current exposure cycle point by point. The residual prediction matrix is ​​accumulated in time order.

[0040] Under the constraint of the residual prediction matrix, the brightness difference is compared with the reference photon distribution. When the difference exceeds the threshold and is consistent with the trajectory of the afterimage shift mode, it is determined to be an abnormal brightness residual. Real-time detection is performed by combining the persistence and position shift of the residual.

[0041] After completing the abnormal brightness residual detection, a dynamic mask is generated based on the residual prediction matrix. The abnormal region is assigned a lower weight and works synchronously with the input image during the feature extraction process to automatically weaken afterimage pixels and maintain the continuity and stability of face recognition features.

[0042] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0043] This invention establishes a photon accumulation profile at the moment of high-speed blocking release and combines it with an inverse phase compensation curve. With the support of cross-frame energy flow maps and force field traction mechanisms, it dynamically decouples the afterimage signal from the real face features in both time and space dimensions, allowing the afterimage boundary to shrink. This effectively suppresses the superposition interference of afterimages on real features, maintains the stability of face features in consecutive frames, and makes the features received by the recognition network at the input end clearer and purer, significantly improving the accuracy and reliability of the recognition results.

[0044] This invention constructs an offset pattern for identifying afterimages in a temporal replay link, and based on this, establishes a multi-scale projection stripping strategy and a residual prediction matrix. This achieves layer-by-layer stripping of afterimage features and layer-by-layer correction and rewriting of true features, ensuring the consistency and integrity of the deep feature mapping space during recognition. This process not only avoids feature collapse but also automatically weakens afterimage pixels during the deep learning feature extraction stage through dynamic masks. This allows face recognition to maintain the continuity and stability of feature expression even in dynamic and complex scenes, thus meeting the security and robustness requirements of highly sensitive applications. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0046] Figure 1 This is a flowchart of the deep learning-based face recognition method of the present invention. Detailed Implementation

[0047] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0048] This invention provides, for example Figure 1 The deep learning-based face recognition method shown includes the following steps:

[0049] S1. At the moment when the face of the person is subjected to high-speed occlusion and release, a photon accumulation profile is established based on the time-series inversion principle, and an inverse phase compensation curve is constructed under the photon accumulation profile to predict and simulate the trajectory of photons that have not completely dissipated within the exposure cycle, so as to provide a time reference for the stable recognition of subsequent facial features.

[0050] The specific implementation process of this step is as follows:

[0051] At the instant the facial occlusion is released, the photon stream received by the sensor is acquired in a time-division manner, and a photon accumulation profile is established based on the inversion principle. Specifically, the sensor's single exposure cycle is artificially divided into multiple consecutive time segments. For each time segment, the intensity of incident photons, their arrival time, and their spatial distribution on the photosensitive surface are acquired. Because the release of the occlusion causes a sudden increase in the number of photons in a short period and results in an uneven distribution at different locations, it is necessary to record the photon distribution data for each time segment individually. After completing the multi-segment acquisition, the photon distributions are reversed in chronological order using an inversion method, thus obtaining the cumulative change trend of photon energy throughout the entire exposure cycle. This trend curve is defined as the photon accumulation profile. The photon accumulation profile not only shows the energy superposition of photons in different time segments but also reveals the delayed energy release phenomenon caused by the occlusion release. For example, after the masking is released, some photons continue to enter the photosensitive surface in a low-intensity form during the middle and later stages of the exposure cycle due to reflection and refraction effects. These delayed photon signals will appear as a tail feature with slowly decreasing energy on the profile, thus providing an intuitive basis for subsequent compensation and prediction.

[0052] It should be noted that:

[0053] The sensor here primarily refers to the image acquisition device in the camera imaging chain, typically including an image sensor array (such as a CMOS image sensor or a CCD image sensor). The sensor's function is to capture and convert light signals reflected or transmitted from the subject's face under external lighting conditions during the exposure cycle. In other words, the sensor is the key hardware that converts light energy entering the lens into electrical signals. It not only records the intensity information of the light signals but also precisely divides the arrival sequence of the light signals in time. In this embodiment, the sensor's role is to perform time-division acquisition of the incoming photon stream according to preset time segments, allowing the sudden photon changes generated at the moment of release from obstruction to be characterized in detail, rather than being averaged over the entire exposure cycle. Through this time-division acquisition, the sensor can provide raw data support for subsequent time-series inversion and the establishment of photon accumulation profiles.

[0054] In this context, a photon refers to a single quantum particle of light emitted from the external environment (especially a light source) and entering the photosensitive surface of the sensor after being reflected, scattered, or refracted by the subject's face. A photon is the smallest unit of light energy, and each photon carries energy and phase information. During high-speed blocking and release of the blocking mechanism, the distribution of photons is significantly disturbed: when the blocking object is present, the incident amount of photon flow is significantly reduced; when the blocking object is suddenly removed, a large number of photons rush in within a very short time, and due to different paths, some photons arrive at the photosensitive surface only in the middle and later stages of the exposure cycle, creating a delayed effect. These delayed photons are the source of image retention. In this embodiment, photons are not only considered as energy carriers but also used to construct a cumulative profile in the time dimension. By recording the intensity changes of photons at different time segments, the entire process of image retention generation and dissipation can be revealed, thus providing a reliable physical basis for inverse phase compensation and trajectory prediction.

[0055] After establishing the photon accumulation profile, the energy distribution information within this profile is further utilized to construct an inverse phase compensation curve. Specifically, the energy change trend contained in the photon accumulation profile is first analyzed over time to identify the sharp rise segment and the subsequent slow decay segment that occur at the moment of release from blocking. In the sharp rise segment, photon energy exhibits a sudden, large increase, while in the slow decay segment, it exhibits a prolonged residual effect. Compensation strategies are formulated for these two different energy change characteristics. For the sharp rise segment, the time position and magnitude of the energy increase are first identified, and an energy decrease curve in the opposite direction is constructed at that position to cancel out the rise segment over time. For the slow decay segment, the energy deviation is obtained by comparing the baseline photon distribution before release from blocking with the current decay curve, and a gradually increasing inverse phase compensation curve is applied to the time series, gradually weakening the residual photon signal. The inverse phase compensation curve is constructed by referencing the slope changes and energy inflection point positions in the photon accumulation profile. This ensures that the compensation effect is precisely applied to the segment where the afterimage signal appears, without affecting the authenticity of the photon distribution of normal facial features. In this way, the inverse phase compensation curve can play a corrective role in the energy evolution process, accelerating the recovery speed of the accumulation profile after occlusion release, thus laying the foundation for subsequent trajectory prediction.

[0056] After constructing the inverse phase compensation curve, the compensated profile is used to predict and simulate the trajectory of incompletely dissipated photons, providing a stable time reference for subsequent feature recognition. The specific steps are as follows: First, the inverse phase compensation curve is superimposed point-by-point onto the photon accumulation profile to obtain a compensated energy change curve, which accurately reflects the actual attenuation process of photons throughout the exposure cycle. After obtaining this curve, the portion with residual energy is extended for prediction. The prediction process uses a frame-by-frame time window approach, using the collected photon energy values ​​as input and combining them with the energy change trend provided by the compensation curve to estimate the possible energy level and distribution location of photons within future time windows. For example, if there is still a slow energy decrease at the tail of the photon profile, the prediction can determine the specific attenuation rate and direction of the afterimage signal in the subsequent time windows. The prediction results not only numerically represent the duration of the afterimage signal but also spatially pinpoint the area where the afterimage may appear. Subsequently, these prediction results are mapped onto the timeline of deep learning face recognition, enabling the recognition process to pre-mark and avoid afterimage signals. In this way, when the recognition network extracts facial features, it can use these prediction anchors to avoid the superposition of afterimage signals and real features, thereby ensuring the stability and continuity of feature encoding.

[0057] Through the above steps, this implementation method realizes a complete link from time-division photon acquisition and inversion, the establishment of photon accumulation profiles, to the construction of inverse phase compensation curves, and finally to the prediction and simulation of undissipated photon trajectories. By introducing the concepts of photon accumulation profiles and inverse phase compensation, the image retention problem is solved in the time domain of the exposure link, thereby avoiding contamination of the deep learning network at the input end.

[0058] S2. Under the constraint of the photon accumulation profile, a cross-frame energy flow map is generated, and the abnormal energy and normal energy are differentially projected in the cross-frame energy flow map to form a refined energy layer. Boundary conditions are established for afterimage detection in the energy layer.

[0059] The specific implementation process of this step is as follows:

[0060] A cross-frame energy flow map is generated under the time constraint of a photon accumulation profile. The specific process is as follows: Spatial distribution data of photon energy is collected for each of several consecutive adjacent exposure frames and arranged chronologically. Using the photon accumulation profile obtained in the previous step as a reference, the changes in photon energy in each frame are mapped one-to-one with the accumulation profile to obtain the energy evolution trajectory in both time and space. To ensure the integrity of the trajectory, when generating the cross-frame energy flow map, it is necessary to record not only the spatial diffusion of photons within the same frame but also the energy migration relationship between frames. For example, if photons in a certain region exhibit high energy in the previous frame and gradually weaken in the next frame, this is reflected as an energy migration path in the cross-frame energy flow map. In this way, the cross-frame energy flow map can comprehensively describe the propagation and dissipation process of photons across multiple frames and provide basic data for subsequent energy differentiation projection.

[0061] Based on the cross-frame energy flow map, anomalous and normal energies are projected differentially. Specifically, firstly, the energy trajectories in the cross-frame energy flow map are compared with the energy dissipation pattern revealed in the photon accumulation profile. If the energy intensity or decay rate of a trajectory significantly deviates from the trend of the accumulation profile, it is determined to be an anomalous energy trajectory; conversely, if the energy change of a trajectory is consistent with the accumulation profile, it is determined to be a normal energy trajectory. Next, the anomalous energy trajectories are projected onto an independent energy space layer, where their distribution range and decay path are individually marked. Simultaneously, the normal energy trajectories are projected onto another energy space layer, maintaining their energy distribution corresponding to real facial features. Through this differential projection, anomalous and normal energies are effectively separated spatially, laying the foundation for subsequent refined layering.

[0062] After differentiating the projections of anomalous and normal energies, a more refined energy stratification is further developed. Specifically, the anomalous and normal energy layers are refined, dividing energy trajectories of different intensity levels into multiple energy sub-layers. For example, in the anomalous energy layer, it can be divided into high-intensity, medium-intensity, and low-intensity anomalous sub-layers according to energy intensity from high to low; in the normal energy layer, it can be divided into stable feature sub-layers and transitional feature sub-layers based on the stability of photon distribution. This refined division allows for the differentiation of afterimage signals from true signals at a finer granularity, thus avoiding confusion between the two in overlapping energy regions. Simultaneously, during the formation of the refined energy stratification, the dynamic migration path of the cross-frame energy flow map is referenced to record the connection relationships between different energy sub-layers, ensuring that energy in the stratified structure possesses both static differentiation and dynamic evolution characteristics. This refined stratification structure provides higher resolution and more precise localization accuracy for afterimage detection.

[0063] Finally, with the support of refined energy stratification, clear boundary conditions are established for ghosting detection. Specifically, a ghosting detection boundary line is defined between each energy sub-layer, especially at the boundary between abnormal and normal energy sub-layers. This boundary line is based on the degree of deviation of the abnormal energy trajectory and the energy deviation value in the cumulative profile. For example, when the energy attenuation rate of a certain region across frames is significantly lower than that of the normal energy trajectory, this region is identified as a ghosting region, and its boundary line is locked at the outer edge of the abnormal energy sub-layer. Conversely, when the energy trajectory of a certain region perfectly conforms to the dissipation law of the cumulative profile, it is classified as a normal region and excluded from the ghosting detection boundary. In this way, the ghosting detection boundary can accurately delineate the range of ghosting signals, allowing the recognition algorithm to specifically correct ghosting regions in the subsequent face feature extraction process without affecting the feature extraction of normal regions. Ultimately, the cross-frame energy flow map, differential projection, refined stratification, and the establishment of boundary conditions form a complete processing chain, giving ghosting detection high accuracy and high robustness.

[0064] Through the execution of the above steps, this implementation achieves a complete process under the constraint of photon accumulation profile, from cross-frame capture of energy evolution trajectories, to the distinction between abnormal and normal energies, to refined energy stratification, and the establishment of afterimage detection boundaries. This process can not only effectively distinguish afterimage signals from real face signals at the physical level, but also simultaneously establish clear detection boundaries in both spatial and temporal dimensions, thereby providing highly stable input data for subsequent face recognition.

[0065] S3. With the support of energy layering, a force field traction mechanism is introduced to apply reverse phase traction to the high-energy anomalous layer in energy layering, so as to dynamically decouple the afterimage region from the real face region, shrink the afterimage boundary and provide conditions for feature reconstruction.

[0066] The specific implementation process of this step is as follows:

[0067] High-energy anomalous layers are identified in the energy stratification results. Specifically, in a refined energy stratification structure, high-energy anomalous layers typically manifest as regions with energy intensities significantly higher than normal energy layers, exhibiting abnormal delays or uneven diffusion trends in cross-frame energy flow maps. To ensure accurate identification, the energy intensity of the anomalous layer is compared point-by-point with the standard dissipation curve in the photon accumulation profile. When the intensity of a certain energy layer consistently exceeds the range of the standard dissipation curve, and its decay rate significantly lags behind that of the normal energy layer, that energy layer is identified as a high-energy anomalous layer. In this process, the anomalous trajectory information obtained during differential projection is combined with spatial location and temporal delay features to obtain a complete high-energy anomalous layer labeling result. The purpose of this identification step is to provide a precise target for subsequent traction operations, avoiding interference with normal energy layers.

[0068] After the high-energy anomalous layer is identified, a reverse phase pull is applied. Specifically, the phase shift direction and amplitude of the high-energy anomalous layer are first determined based on the constraints of the photon accumulation profile and the inverse phase compensation curve. Then, a force field vector is introduced in this direction, pulling the energy trajectory in the opposite direction to the afterimage diffusion. The essence of this pulling process is to counteract the spatial diffusion trend of the afterimage energy through an artificially set phase reversal, preventing the afterimage signal from extending infinitely and instead causing it to gradually converge towards a predetermined contraction direction. During the pulling process, to avoid accidental damage to the real face area, the intensity of the pulling vector is monitored in real time. When the pulling intensity exceeds the stability threshold of the normal energy layer, the pulling force is automatically reduced, ensuring that the pulling only applies to the high-energy anomalous layer. In this way, reverse phase pull can effectively pull the afterimage signal back from the energy level and create a contraction trend.

[0069] Under the reverse phase traction, dynamic decoupling between the afterimage region and the real face region is gradually achieved. Specifically, as the high-energy anomalous layer is pulled in the opposite direction to the real energy layer, the overlap between it and the normal energy layer gradually decreases. To further enhance the decoupling effect, the range of the traction vector is dynamically adjusted during the decoupling process. For example, when there is still a large overlap between the boundary of the afterimage region and the real region, the traction vector covers a wider area to ensure that the afterimage region separates outward as a whole; while when the overlap area between the afterimage region and the real region gradually decreases, the traction range is reduced, concentrating on the local areas where adhesion still exists. This dynamic adjustment method ensures that the afterimage signal does not damage the real face features during the separation process. Ultimately, the afterimage region and the real face region will form a relatively independent distribution in the spatial dimension, thereby achieving dynamic decoupling.

[0070] Based on the dynamic decoupling of the afterimage region from the real region, the afterimage boundary is gradually shrunk, providing stable conditions for subsequent feature reconstruction. The specific process is as follows: First, the boundary of the separated afterimage region is located, identifying the outer contour line of the afterimage signal. Then, under the continuous action of reverse phase traction, this contour line gradually shrinks inward until the afterimage energy is compressed into a smaller spatial range. Simultaneously, the shrunk afterimage boundary needs to be compared with the cross-frame energy flow map to ensure that no residual parts of the afterimage are missed during the shrinkage process. In this way, the afterimage signal is effectively confined to a controllable range, preventing it from continuing to spread into the real region. After the afterimage boundary shrinkage is completed, the previously disturbed real face region regains a stable energy distribution, thus providing a reliable prerequisite for subsequent feature reconstruction steps based on the temporal replay link.

[0071] Through the above steps, this implementation method, with the support of energy stratification, achieves a complete process including the identification of high-energy anomaly layers, the application of reverse phase traction, the dynamic decoupling of the afterimage from the real region, and the contraction of the afterimage boundary. This process not only effectively suppresses the spread of the afterimage signal but also physically compresses and constrains the afterimage region, allowing the real facial features to be released and restored in the spatial dimension.

[0072] S4. Based on the shrinkage of the afterimage boundary, construct a temporal replay link, perform counterfactual playback on the key frames before and after the occlusion, compare the evolution trajectory of the real face features with the afterimage signal frame by frame, identify the offset pattern of the afterimage on the time axis, and provide anchor points for feature correction.

[0073] The specific implementation process of this step is as follows:

[0074] Under the premise of image retention boundary contraction, keyframes before and after the occlusion occur are extracted, and a temporal replay link is established. Specifically, during the process of a person's face undergoing high-speed occlusion and then releasing, several keyframes are selected from consecutive video frames, one before the occlusion occurs and the other after it disappears. The selection principle for keyframes is: keyframes before occlusion should contain relatively complete real facial features, while keyframes after occlusion should contain facial features that may be mixed with image retention signals. In this way, a set of frame sequences with before-and-after contrast can be obtained on the timeline. Next, a temporal replay link is constructed on this set of keyframe sequences, that is, these keyframes are arranged continuously in chronological order, preserving the energy distribution characteristics and photon trajectory characteristics of each frame during the arrangement process. Thus, the temporal replay link not only reflects the continuity before and after occlusion in the temporal dimension but also preserves the mixing of image retention signals and real signals in the spatial and energy dimensions, providing a basis for subsequent counterfactual playback.

[0075] Based on the temporal replay link, counterfactual playback is performed on keyframes before and after occlusion. Specifically, the keyframes before occlusion are first used as the reference trajectory for the evolution of real facial features. This trajectory includes the continuous process of facial features changing over time under normal conditions. Then, the keyframes containing afterimages after occlusion are replayed according to the same temporal sequence and compared frame-by-frame with the real trajectory before occlusion. During the comparison, if the facial feature distribution of a frame after occlusion differs significantly from the trajectory before occlusion, and this difference is concentrated near the afterimage detection boundary, then this difference is determined to originate from an afterimage signal. To further improve the accuracy of the determination, counterfactual playback not only compares the overall shape of the feature contours but also compares the detailed evolution process of local features, such as the movement of the eye contours and subtle changes in the corners of the mouth. When these detailed evolutions are inconsistent with the trend of the real trajectory, they can be marked as afterimage offset signals. In this way, counterfactual playback can effectively distinguish the evolution trajectory of real facial features from the interference trajectory of afterimage signals, providing data support for the recognition of offset patterns.

[0076] After identifying the difference between the afterimage signal and the true trajectory through counterfactual playback, the offset pattern of the afterimage on the time axis is further extracted, and anchor points are provided for feature correction based on this offset pattern. The specific process is as follows: First, the evolution trajectories of keyframes before and after occlusion are compared to extract the offset direction, magnitude, and duration of the afterimage signal relative to the true trajectory on the time axis. For example, a certain afterimage signal may have a fixed-direction time delay relative to the true trajectory, or it may exhibit a gradually drifting pattern across multiple consecutive frames. By statistically analyzing these differences, a complete afterimage offset pattern can be formed. Next, the offset pattern is mapped to the corresponding position in the temporal replay link, and anchor points are set at these positions. The role of the anchor points is to provide reference coordinates for subsequent feature correction. Specifically, when the deep learning recognition process processes new input frames, the system can use these anchor points to determine the possible temporal location and spatial range of the afterimage signal, and actively correct these areas during the recognition process. In this way, feature correction is not only targeted but also intervenes early in the temporal dimension, avoiding continuous contamination of deep features by the afterimage signal. In this way, the recognition of afterimage offset patterns and the establishment of anchor points realize a closed loop from detection to correction, which greatly improves the robustness of the face recognition process in dynamic scenes.

[0077] Through the above steps, this implementation method achieves a complete process from keyframe extraction and temporal replay, to frame-by-frame comparison in counterfactual playback, and then to the identification of ghosting offset patterns and the establishment of anchor points. This process further enhances the dynamic analysis capability of ghosting signals based on ghosting boundary contraction, effectively locking the ghosting not only spatially but also temporally. Furthermore, by utilizing the concepts of temporal replay links and counterfactual playback, multiple keyframes are chained together for overall analysis, thereby obtaining more complete and detailed ghosting offset features.

[0078] S5. After the offset mode of the afterimage is locked, a multi-scale projection stripping strategy is established to strip the afterimage features from the depth feature mapping space step by step. During the step-by-step stripping process, the real face features are corrected layer by layer and written back to the main space to prevent the collapse of the depth features during the face recognition process.

[0079] The specific implementation process of this step is as follows:

[0080] With the afterimage migration pattern locked, a multi-scale projection structure is established and the stripping strategy is initialized. Specifically, based on the hierarchical structure of the deep feature mapping space, the entire feature space is divided into multiple scale layers, each corresponding to a different level of feature representation, such as low-level edge features, mid-level texture features, and high-level semantic features. When establishing the multi-scale projection structure, the afterimage migration pattern is mapped to each scale layer to determine the distribution of the afterimage signal at different scales. Since afterimages often manifest as energy blurring and edge drift in low-level feature spaces, while they may manifest as semantic distortion and positional shift in high-level feature spaces, localization must be performed simultaneously at multiple scales. In this way, a complete spatial distribution reference can be provided for subsequent step-by-step stripping, ensuring the full coverage and targeted nature of the stripping process.

[0081] Supported by a multi-scale projection structure, afterimage features are progressively stripped. Specifically, at the lowest scale layer, blurry areas of the afterimage signal on edge features are first stripped by subtracting the energy trajectory corresponding to the afterimage offset pattern from the feature mapping space, resulting in clearer edge features. After low-level stripping, the process moves to the mid-level feature space to process abnormal signals from the texture layer caused by the afterimage. This stripping operation not only removes local texture distortions caused by the afterimage but also maintains the continuity of the real face texture, avoiding damage to real features due to over-stripping. Finally, in the high-level semantic feature space, semantic misalignments caused by the afterimage are stripped, restoring the overall shape and contour of the face to a normal state. The entire progressive stripping process relies on the afterimage offset pattern corrected in the previous step at each scale layer, ensuring that the afterimage signal can be completely removed across multiple scales.

[0082] During the step-by-step stripping process, the real facial features are corrected layer by layer and written back to the main space. Specifically, after stripping the afterimage signal at each scale layer, the real features affected by the afterimage at that scale layer are corrected. For example, in the low-level edge feature layer, the correction process includes redrawing the boundary lines blurred by the afterimage and writing the corrected edge information back to the main space, allowing subsequent layers to use more accurate edge data as a basis. In the mid-level texture feature layer, the correction process includes restoring the texture continuity disrupted by the afterimage, such as smooth transitions in skin texture or natural distribution of local facial shadows, and writing these corrected texture features back to the main space. In the high-level semantic feature layer, the correction process includes repairing the overall facial contour shift caused by the afterimage, such as misalignment of the bridge of the nose or distortion of the mouth contour, and similarly writing the repaired semantic features back to the main space. Through this step-by-step correction and writing back, the main space can gradually accumulate a more stable and complete facial feature representation, thereby avoiding the collapse of depth features caused by the accumulation of afterimage signals.

[0083] After completing multi-scale, step-by-step stripping and layer-by-layer correction and write-back, consistency detection and steady-state confirmation are performed on the entire main space. Specifically, the features written back at each scale layer are compared one by one with the original, unstripped features to determine if there are any new offsets or discontinuous regions. If residual ghosting signals are found at a certain scale layer, the process is repeated at that scale layer for secondary stripping and correction until the ghosting signals are completely eliminated. After consistency detection, the facial features in the main space maintain high consistency at the edge, texture, and semantic levels, and coincide with the previously established feature correction anchor points in the time dimension, thus forming a complete and stable feature representation. Finally, the facial features processed by the multi-scale projection stripping strategy are no longer affected by ghosting, and the deep learning recognition process can proceed normally with the support of real features, effectively avoiding recognition failures caused by feature collapse.

[0084] Through the implementation of the above steps, this embodiment, based on the locking of the afterimage offset mode, forms a complete closed-loop process by establishing a multi-scale projection structure, progressively stripping afterimage features, correcting real features layer by layer and writing them back to the main space, and finally performing consistency detection. This embodiment is the first to propose a strategy of progressively stripping afterimages in a multi-scale space, and simultaneously correcting and writing back real features at each level. This not only fundamentally eliminates the multi-level interference of afterimages on the feature space, but also enables positive information transfer and accumulation between feature levels, ensuring the stability and integrity of the depth features on which the final recognition depends.

[0085] S6. After completing the layer-by-layer correction, a residual prediction matrix is ​​constructed between consecutive exposure cycles to detect abnormal brightness residuals caused by high-speed occlusion in real time. The residuals are then mapped to dynamic masks and injected into the deep learning feature extraction network to automatically weaken afterimage pixels and maintain the continuity and stability of face recognition features.

[0086] The specific implementation process of this step is as follows:

[0087] After layer-by-layer correction, a residual prediction matrix covering consecutive exposure cycles is constructed. Specifically, multiple adjacent exposure cycles are treated as a single temporal window, and the photon energy distribution curve of the corrected real face features is recorded within this window. Since most of the ghosting signal has been eliminated through step-by-step stripping and correction in the previous stage, the face features in the main space are now largely stable. However, brightness residuals due to high-speed occlusion may still exist between consecutive exposure cycles. To detect these residuals, a prediction matrix reflecting brightness differences during different exposure cycles is first established. During construction, the correction result of the real face features in the previous exposure cycle is used as a reference, and its point-by-point comparison with the feature distribution of the current exposure cycle is performed to calculate the brightness difference. Subsequently, these differences are arranged chronologically and accumulated in the matrix to form a residual prediction matrix that dynamically displays the brightness change trend. This matrix not only reveals the residual intensity of the ghosting signal in the temporal dimension but also reflects the diffusion range of the ghosting in the spatial dimension, providing the initial basis for subsequent mask generation.

[0088] Under the constraints of the residual prediction matrix, abnormal brightness residuals caused by high-speed occlusion are detected in real time. Specifically, the brightness difference in the residual prediction matrix is ​​compared with the reference photon distribution before occlusion. When the brightness difference in a certain area exceeds a set threshold and is consistent with the trajectory of the afterimage shift pattern, it is determined that there is an abnormal brightness residual in that area. During the detection process, not only the magnitude of the residual is considered, but also its temporal persistence and spatial positional shift are taken into account. If a residual signal appears only in a single exposure cycle and has a small amplitude, it may belong to normal illumination fluctuations; however, if the residual signal persists in multiple consecutive exposure cycles and coincides with the previously identified afterimage boundary in spatial location, it can be determined that the signal is an afterimage residual caused by high-speed occlusion. In this way, real-time detection can dynamically track the afterimage signal in multiple consecutive exposure cycles and accurately distinguish it from normal illumination changes, ensuring the reliability and accuracy of the detection results.

[0089] After detecting abnormal brightness residuals, the residuals are mapped to a dynamic mask and injected into a deep learning feature extraction network to automatically weaken afterimage pixels and maintain the stability of facial recognition features. The specific process is as follows: First, based on the abnormal regions marked in the residual prediction matrix, a dynamic mask corresponding to the input image is generated. In this mask, afterimage regions are assigned lower weights, while normal regions retain their original weights. Subsequently, during deep learning feature extraction, the dynamic mask is synchronized with the input image, allowing the network to automatically reduce the contribution of afterimage region pixels when calculating feature mappings, thus preventing afterimage signals from dominating feature encoding. Simultaneously, the dynamic mask is continuously updated between consecutive exposure cycles to adapt to changes in afterimage signals over time. In this way, when new afterimage residuals appear, the mask can adjust instantly, ensuring that the recognition process always focuses on the stable parts of the real facial features. Finally, through the combined effect of the residual prediction matrix and the dynamic mask, the continuity of facial recognition features over time is guaranteed, and the deep learning network can maintain high accuracy and robustness even in high-speed dynamic scenes.

[0090] Through the above steps, this implementation method achieves further suppression of ghosting signals in the temporal dimension based on the completion of layer-by-layer correction. First, a residual prediction matrix is ​​constructed to capture brightness differences between exposure cycles. Second, the distribution range of abnormal residuals is locked through real-time detection. Finally, ghosting pixels are weakened during the depth feature extraction stage using a dynamic mask, preventing the recognition process from being interfered with by ghosting. Furthermore, this implementation method creatively combines temporal residual prediction with the depth feature extraction process, forming a closed-loop control mechanism from detection to suppression, thereby significantly improving the accuracy and stability of face recognition in dynamic scenes.

[0091] This invention establishes a photon accumulation profile at the moment of high-speed blocking release and combines it with an inverse phase compensation curve. With the support of cross-frame energy flow maps and force field traction mechanisms, it dynamically decouples the afterimage signal from the real face features in both time and space dimensions, allowing the afterimage boundary to shrink. This effectively suppresses the superposition interference of afterimages on real features, maintains the stability of face features in consecutive frames, and makes the features received by the recognition network at the input end clearer and purer, significantly improving the accuracy and reliability of the recognition results.

[0092] This invention constructs an offset pattern for identifying afterimages in a temporal replay link, and based on this, establishes a multi-scale projection stripping strategy and a residual prediction matrix. This achieves layer-by-layer stripping of afterimage features and layer-by-layer correction and rewriting of true features, ensuring the consistency and integrity of the deep feature mapping space during recognition. This process not only avoids feature collapse but also automatically weakens afterimage pixels during the deep learning feature extraction stage through dynamic masks. This allows face recognition to maintain the continuity and stability of feature expression even in dynamic and complex scenes, thus meeting the security and robustness requirements of highly sensitive applications.

[0093] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A face recognition method based on deep learning, characterized in that, Includes the following steps: At the moment when the face of a person experiences high-speed occlusion and release, a photon accumulation profile is established based on the principle of time-series inversion, and an inverse phase compensation curve is constructed under the photon accumulation profile to predict and simulate the trajectory of photons that have not completely dissipated within the exposure cycle. Under the constraint of the photon accumulation profile, a cross-frame energy flow map is generated, and the abnormal energy and normal energy are differentially projected in the cross-frame energy flow map to form a refined energy layer. Boundary conditions are established for image retention detection in the energy layer. With the support of energy stratification, a force field traction mechanism is introduced to apply reverse phase traction to the high-energy anomaly layer in the energy stratification, so as to dynamically decouple the afterimage region from the real face region and shrink the afterimage boundary. Based on the shrinkage of the afterimage boundary, a temporal replay link is constructed to perform counterfactual playback of key frames before and after occlusion. The evolution trajectory of real facial features is compared with the afterimage signal frame by frame to identify the offset pattern of the afterimage on the time axis and provide anchor points for feature correction. After the offset mode of the afterimage is locked, a multi-scale projection stripping strategy is established to strip the afterimage features from the depth feature mapping space step by step, and the real face features are corrected layer by layer and written back to the main space during the step-by-step stripping process. After completing the layer-by-layer correction, a residual prediction matrix is ​​constructed between consecutive exposure cycles to detect abnormal brightness residuals caused by high-speed occlusion in real time. The residuals are then mapped to dynamic masks and injected into a deep learning feature extraction network to automatically weaken afterimage pixels.

2. The deep learning-based face recognition method according to claim 1, characterized in that, The steps for establishing a photon accumulation profile and constructing an inverse phase compensation curve at the instant when a person's face experiences high-speed occlusion and release include: The photon stream received by the sensor is collected in a time-division manner at the moment the face is occluded and released. The single exposure cycle is divided into multiple continuous time segments, and the photon incident intensity, arrival time and spatial distribution of each time segment are recorded one by one. The photon accumulation profile is obtained based on the inversion method. After the photon accumulation profile is established, a time-series analysis is performed on the energy change trend of the profile to identify the sharp rise segment and the slow decay segment. An energy decrease curve in the opposite direction is constructed in the sharp rise segment, and a gradually enhanced reverse phase compensation curve is applied in the slow decay segment to form an inverse phase compensation curve. After the inverse phase compensation curve is constructed, it is superimposed point by point with the photon accumulation profile to obtain the compensated and corrected energy change curve. Based on this energy change curve, the trajectory of the photon that has not completely dissipated is extended and predicted, and the prediction result is mapped to the recognition time axis.

3. The deep learning-based face recognition method according to claim 1, characterized in that, The steps for generating cross-frame energy flow maps and forming refined energy stratification under the constraint of photon accumulation profiles include: Under the time constraint of the photon accumulation profile, the spatial distribution data of photon energy of consecutive adjacent exposure frames are collected and arranged in chronological order to generate a cross-frame energy flow map to record the propagation and dissipation trajectory of photons between multiple frames. Based on the cross-frame energy flow map, the dissipation law of each energy trajectory is compared with the photon accumulation profile. Trajectories that deviate from the law are projected as abnormal energy layers, and trajectories that conform to the law are projected as normal energy layers, thus realizing differentiated projection of abnormal energy and normal energy. After differential projection is completed, the abnormal energy layer and normal energy layer are further divided into multiple energy sub-layers, and a refined energy layering is formed by referring to the migration path of the cross-frame energy flow map. With the support of refined energy stratification, the residual image detection boundary is set in the boundary area between the abnormal energy sub-layer and the normal energy sub-layer according to the degree of deviation of the abnormal energy trajectory and the energy deviation value, thus defining the range of residual image signal existence.

4. The deep learning-based face recognition method according to claim 3, characterized in that, When setting the afterimage detection boundary with the support of refined energy stratification, the region where the energy decay rate is lower than the dissipation law of the photon accumulation profile is identified as the afterimage region, and the boundary is locked at the outer edge of the abnormal energy sublayer.

5. The deep learning-based face recognition method according to claim 3, characterized in that, The steps for introducing a force field traction mechanism with the support of energy stratification include: In the refined energy stratification results, high-energy anomalous layers are identified. Their energy intensity is compared point by point with the standard dissipation curve of the photon accumulation profile. When the energy intensity continuously exceeds the standard range and the decay rate is significantly lagging, the energy layer is identified as a high-energy anomalous layer and marked in combination with the anomalous trajectory information of the differential projection. After the high-energy anomalous layer is identified, the phase shift direction and magnitude are determined based on the constraints of the photon accumulation profile and the inverse phase compensation curve. A force field vector is introduced in this direction to apply reverse phase pulling, and the pulling intensity is monitored in real time to avoid acting on the normal energy layer. Under the reverse phase traction, the afterimage region and the real face region are gradually decoupled dynamically. By dynamically adjusting the range of the traction vector, the overlap between the two is reduced, so that the afterimage region and the real region form a relatively independent distribution. After dynamic decoupling is completed, the afterimage boundary is gradually shrunk, and the shrunk afterimage boundary is compared with the cross-frame energy flow map to ensure that the afterimage signal is limited to a controllable range.

6. The deep learning-based face recognition method according to claim 5, characterized in that, The steps for constructing a temporal replay link and performing counterfactual playback based on afterimage boundary shrinkage include: Under the premise of shrinking afterimage boundaries, key frames containing complete facial features before occlusion and key frames with mixed afterimage features after occlusion disappear from continuous video frames are extracted, and a temporal replay link is established in chronological order to preserve energy distribution and photon trajectory features. Based on the temporal replay link, the key frame before occlusion is used as the reference trajectory for the evolution of real face features, and the key frame after occlusion is replayed frame by frame in a counterfactual manner. The evolution difference between the overall contour and local details is compared, and the abnormal parts concentrated at the afterimage detection boundary are identified as afterimage offset signals. After the ghosting offset signal is identified, its offset direction, amplitude and duration on the time axis are extracted, and the offset pattern is mapped to the corresponding position of the time-series replay link to set anchor points, so that subsequent feature correction can use the anchor points to correct ghosting interference in advance when processing new input frames.

7. The deep learning-based face recognition method according to claim 1, characterized in that, The steps for establishing a multi-scale projection stripping strategy after the afterimage offset mode is locked include: Under the premise that the afterimage offset mode is locked, a multi-scale projection structure is established, the feature mapping space is divided into edge feature layer, texture feature layer and semantic feature layer, and the afterimage offset mode is mapped to each scale layer to locate the afterimage distribution. With the support of a multi-scale projection structure, the afterimage features are sequentially stripped away in the edge layer, texture layer and semantic layer. While peeling away the image layer by layer, the real features affected by the afterimage are corrected layer by layer, and the corrected edge information, texture distribution and semantic contours are written back to the main space in turn to accumulate stable feature representations. After multi-scale stripping and writing back are completed, consistency detection is performed on the main space, the corrected features of each layer are compared with the original features, and when residual images are found, a second stripping and correction is performed to ensure that the final features remain consistent.

8. The deep learning-based face recognition method according to claim 7, characterized in that, The steps for constructing the residual prediction matrix and injecting the dynamic mask after completing the layer-by-layer correction include: After the layer-by-layer correction is completed, the adjacent exposure cycles are regarded as time windows, the corrected facial feature photon energy distribution is recorded, and the previous exposure cycle is used as a reference to compare the brightness difference with the current exposure cycle point by point. The residual prediction matrix is ​​accumulated in time order. Under the constraint of the residual prediction matrix, the brightness difference is compared with the reference photon distribution. When the difference exceeds the threshold and is consistent with the trajectory of the afterimage shift mode, it is determined to be an abnormal brightness residual. Real-time detection is performed by combining the persistence and position shift of the residual. After completing the abnormal brightness residual detection, a dynamic mask is generated based on the residual prediction matrix, which assigns low weight to the abnormal region and works synchronously with the input image during the feature extraction process to automatically weaken the afterimage pixels.

Images