A face recognition method based on deep neural network

Abstract

The application discloses a face recognition method based on a deep neural network, and relates to the field of face recognition, and comprises the following steps: acquiring and preprocessing face data; constructing a wearing object-skin coupling graph according to wearing object outlines, skin boundaries and cross-frame displacements; generating visual coupling residuals according to stable identity anchors and wearing object boundaries; inputting the visual coupling residuals, the stable identity anchors and a face sequence to be recognized into a residual coding layer to generate deformation features, standard identity states and wearing identity states; performing eigen decomposition, orthogonal reconstruction verification, coefficient stability verification and coupling consistency verification to generate candidate same-condition features; generating identity consistency amounts, deformation consistency amounts, conflict amounts and causal verification results; and outputting a recognized identity, an unknown person result and a reacquisition instruction. The application adopts dual-domain modeling and causal verification, realizes face recognition in a wearing state, and has the advantages of high stability and high reliability.

Description

Technical Field

[0001] This invention relates to the field of face recognition, and more particularly to a face recognition method based on deep neural networks. Background Technology

[0002] Currently, in the field of facial recognition, existing technologies typically extract global features of the entire face using convolutional neural networks and perform similarity matching with registered templates. For occlusions caused by masks, glasses, and other headwear, occlusion enhancement or local feature compensation are often used, which struggles to characterize the boundaries of the headwear, skin regions, and their cross-frame displacement relationships, leading to an overlay of identity features and local deformation features.

[0003] At the same time, existing technologies usually output recognition results directly based on single similarity or classification scores, lacking a verification mechanism for the source of deformation and the relationship between candidate identities. It is difficult to determine whether the feature changes are caused by the wearing object, and it is also difficult to distinguish between registered identities, unknown persons, and cases with insufficient collection quality, which can easily lead to misidentification or unreliable output. Summary of the Invention

[0004] One objective of this invention is to propose a face recognition method based on deep neural networks. This invention employs dual-domain modeling and causal verification to achieve face recognition in the wearing state, and has the advantages of strong stability and high reliability.

[0005] A face recognition method based on a deep neural network according to an embodiment of the present invention includes: Acquire and preprocess facial data to generate a sequence of faces to be identified and a registered facial template; Identify head and face objects and construct an object-skin coupling map based on their contours, skin boundaries, and cross-frame displacement. Stable identity anchors are selected based on the device-skin coupling map, and visual coupling residuals are generated based on the cross-frame relative position changes between the stable identity anchors and the device boundary. Visual coupling residuals, stable identity anchors, and the sequence of faces to be identified are input into the residual coding layer of a dual-domain Mamba-2 model to generate deformation features, normalized identity states, and wearing identity states. The deformation features are subjected to intrinsic decomposition to obtain orthogonal basis and deformation coefficients. Orthogonal reconstruction, coefficient stabilization and coupling consistency verification are performed in sequence. Candidate identities are retrieved from the registered face templates based on stable identity anchors. The candidate identity features are transformed in the feature space using the verified orthogonal basis and deformation coefficients to generate candidate same-condition features. Input the candidate same-condition features, standard identity status and wearing identity status into the dual-domain verification layer to generate identity consistency quantity, deformation consistency quantity and conflict quantity. Use the deformation coefficient as the intervention variable to perform zeroing, reversal and substitution interventions. Compare the changes of the three quantities before and after the intervention, and generate causal verification results according to the relationship between deformation consistency quantity and identity consistency quantity in the same direction and conflict quantity in the opposite direction. The causal verification results are input into the joint uncertainty decision layer to generate identity uncertainty and deformation uncertainty, and output the results of identity identification, unknown personnel and re-collection instructions.

[0006] Optionally, the generation of the face sequence to be identified and the registered face template includes: Continuous face images are acquired through a face capture device, and registered face images are read from the personnel registration database. Face region localization and quality screening are performed on the continuous face images and registered face images to obtain valid face frames and valid registered face images. Inter-frame association is performed on adjacent valid face frames, and face alignment, image correction and size unification are performed on each associated valid face frame and valid registered face image. The processed valid face frames are arranged according to the acquisition order to generate a face sequence to be identified. The face features of the processed valid registered face images are fused according to the registered identity to generate a registered face template.

[0007] Optionally, the construction of the wearable-skin coupling map includes: The device is used to identify the head and face accessories in each face frame of the face sequence to determine the accessory area and accessory type. Extract the outline of the worn object from the worn object area, extract the skin boundary from the junction of the worn object area and the visible skin area, and establish the boundary correspondence between the worn object outline and the skin boundary; Track the outline of the object and the skin boundary in adjacent face frames according to the boundary correspondence, and determine the cross-frame displacement of the object outline and the cross-frame displacement of the skin boundary respectively. Using the wearable area, wearable outline, skin boundary, and adjacent skin area as graph nodes, graph connections are established based on spatial adjacency relationships and cross-frame displacement correspondences to construct a wearable-skin coupling graph.

[0008] Optionally, the generation of the visual coupling residual includes: Candidate identity regions are screened from the wearer-skin coupling map and divided into local image blocks. Anchor point stabilization states are generated based on the cross-frame visibility ratio, continuous tracking length, texture changes, and positional fluctuations of each local image block. Local image blocks that directly overlap with the wearer boundary, have texture changes exceeding the stable range, or have interrupted positional continuity are removed. Determine representative locations and local identity features from the preserved local image patches, and combine the representative locations, local identity features, and corresponding cross-frame connections into stable identity anchors; Based on the positional changes of stable identity anchor points and the boundary of the worn object in adjacent face frames, anchor point displacement sequences and boundary displacement sequences are generated respectively. A relative displacement reference is generated based on the anchor point displacement sequence and the boundary displacement sequence. The relative displacement in adjacent face frames is compared with the relative displacement reference to obtain the cross-frame positional deviation and orientation deviation, and a visual coupling residual is generated.

[0009] Optionally, the generation of the deformation features, standardized identity status, and wearing identity status includes: Based on visual coupling residuals, stable identity anchors, and the sequence of faces to be identified, an anchor-aligned input group is generated. Extract face features, anchor position features, local identity features and visual coupling residuals from the anchor-aligned input group, map each feature to a token of the same dimension, and write it into time encoding, position encoding and data type encoding to generate a residual encoding sequence; The residual coding sequence is input into the residual coding layer of the coupled dual-domain Mamba-2 model, which includes a residual coding layer, a dual-domain verification layer, and a joint uncertainty decision layer. In the residual coding layer, a coupling gate is generated based on the visual coupling residual. The coupling gate is used to decompose the face feature response into deformation response, normalization domain response and wearing domain response. Deformation features are generated based on the deformation response corresponding to each effective anchor point. Perform temporal state scanning on the normative domain response and the wearing domain response respectively, and aggregate the valid temporal states to generate the normative identity state and the wearing identity state.

[0010] Optionally, the generation of the candidate conditional features includes: Perform symmetric matrix eigenvalue decomposition on the deformation features to form an orthogonal basis, and project the effective deformation response that forms the deformation features onto the orthogonal basis to generate deformation coefficients; The effective deformation response is reconstructed based on the orthogonal basis and deformation coefficients, and the reconstruction error is generated. Coefficient stability verification is performed on the deformation coefficients that pass the orthogonal reconstruction verification. Compare the changes in deformation coefficients that have passed the coefficient stability check with the changes in visual coupling residuals, perform coupling consistency check, and retain the orthogonal basis and deformation coefficients that have passed the orthogonal reconstruction check, coefficient stability check and coupling consistency check in sequence; Anchor point query features are generated based on stable identity anchor points. The anchor point query features are matched with the registered face templates for similarity. Candidate identities are filtered based on the matching results, and the corresponding candidate identity features are read. The candidate identity features are transformed into a feature space consistent with the orthogonal basis. Deformation components are generated based on the verified orthogonal basis and deformation coefficients. The deformation components are then superimposed on the candidate identity features to generate candidate features with the same conditions.

[0011] Optionally, the generation of the causality check result includes: The candidate features with the same conditions, the standardized identity status, and the wearing identity status are input into the dual-domain verification layer, transformed into a unified feature space, and a dual-domain verification feature group is generated. Based on the candidate identical condition features and the standard identity status and wearing identity status, generate identity consistency quantity, deformation consistency quantity and conflict quantity; Using the deformation coefficient as the intervention variable, zeroing intervention, reverse intervention, and alternative intervention were performed respectively to generate zeroing intervention features, reverse intervention features, and alternative intervention features; Based on the zero-intervention feature, reverse intervention feature, and alternative intervention feature, the identity consistency quantity, deformation consistency quantity, and conflict quantity corresponding to various interventions are regenerated, and the changes of the three quantities before and after the intervention are calculated. When the change in deformation consistency quantity is in the same direction as the change in identity consistency quantity and in the opposite direction to the change in conflict quantity, the corresponding intervention is identified as a causal consistency intervention, and a causal verification result is generated based on the number of causal consistency interventions and the magnitude of the changes in the three quantities.

[0012] Optionally, the generation of the identification result, the unknown person result, and the re-collection instruction includes: The causal verification results are input into the joint uncertainty decision layer. Candidate decision data are generated based on the causal verification score, the number of causal consistent interventions, changes in the three quantities, and the causal verification marker. Candidate identity distribution and identity uncertainty are generated based on the causal verification score. Based on the causal consistency status and the magnitude of change of the three quantities corresponding to various interventions, the intervention consistency ratio and the change discrete quantity are generated, and the deformation uncertainty is generated based on the intervention consistency ratio and the change discrete quantity. Based on the candidate identity with the highest causal verification score, identity uncertainty, deformation uncertainty, and corresponding causal verification marker, a joint determination is performed, and the results of identity recognition, unknown personnel identification, and re-collection instructions are output.

[0013] The beneficial effects of this invention are: This invention proposes a face recognition method based on deep neural networks. By identifying head and face accessories and combining the accessory contour, skin boundary, and their cross-frame displacement relationship to construct an accessory-skin coupling map, the face recognition process no longer relies solely on the static appearance features of the entire face. Furthermore, a coupled dual-domain Mamba-2 model is used to decompose the face feature response into deformation response, canonical domain response, and wearing domain response, generating canonical identity state and wearing identity state respectively, thereby reducing the aliasing between accessory information and true identity information. Through intrinsic decomposition, orthogonal reconstruction, coefficient stabilization, and coupling consistency verification of deformation features, stable and interpretable deformation components can be retained. Furthermore, conditional transformations are applied to candidate identity features within the feature space, ensuring more consistent comparison conditions between the registered face template and the face features in the current wearing state, improving the reliability of candidate identity retrieval and matching under occlusion and local deformation conditions.

[0014] This invention also verifies the causal consistency between candidate identities and the current deformation state by comparing the changes in the three quantities before and after intervention, avoiding direct output of recognition results based solely on a single similarity score. The joint uncertainty decision layer further combines causal verification scores, candidate identity distribution, intervention consistency ratio, and change discreteness to generate identity uncertainty and deformation uncertainty, enabling the system to simultaneously evaluate the degree of confusion between candidate identities and the stability of deformation evidence. Therefore, when a candidate identity is unique and the verification evidence is complete, the recognized identity is output; when all candidate identities lack consistent evidence, the result for an unknown person is output; and when there are candidate identity conflicts, insufficient intervention data, or incomplete deformation evidence, a re-collection instruction is output. This achieves differentiated processing of registered persons, unknown persons, and low-quality collection states, reducing misidentification and unreliable forced output, and improving the reliability, stability, and engineering applicability of face recognition results under complex wearing conditions. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a face recognition method based on deep neural networks proposed in this invention; Figure 2 This is a flowchart of the construction of the wearable-skin coupling interaction map and the generation of visual coupling residuals in a face recognition method based on deep neural networks proposed in this invention. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0017] refer to Figures 1-2 A face recognition method based on deep neural networks includes: Acquire and preprocess facial data to generate a sequence of faces to be identified and a registered facial template; Identify head and face objects and construct an object-skin coupling map based on their contours, skin boundaries, and cross-frame displacement. Stable identity anchors are selected based on the device-skin coupling map, and visual coupling residuals are generated based on the cross-frame relative position changes between the stable identity anchors and the device boundary. Visual coupling residuals, stable identity anchors, and the sequence of faces to be identified are input into the residual coding layer of a dual-domain Mamba-2 model to generate deformation features, normalized identity states, and wearing identity states. The deformation features are subjected to intrinsic decomposition to obtain orthogonal basis and deformation coefficients. Orthogonal reconstruction, coefficient stabilization and coupling consistency verification are performed in sequence. Candidate identities are retrieved from the registered face templates based on stable identity anchors. The candidate identity features are transformed in the feature space using the verified orthogonal basis and deformation coefficients to generate candidate same-condition features. Input the candidate same-condition features, standard identity status and wearing identity status into the dual-domain verification layer to generate identity consistency quantity, deformation consistency quantity and conflict quantity. Use the deformation coefficient as the intervention variable to perform zeroing, reversal and substitution interventions. Compare the changes of the three quantities before and after the intervention, and generate causal verification results according to the relationship between deformation consistency quantity and identity consistency quantity in the same direction and conflict quantity in the opposite direction. The causal verification results are input into the joint uncertainty decision layer to generate identity uncertainty and deformation uncertainty, and output the results of identity identification, unknown personnel and re-collection instructions.

[0018] In this embodiment, the generation of the face sequence to be identified and the registered face template includes: Continuous face images are acquired through a face capture device, and registered face images are read from the personnel registration database. Face region localization and quality screening are performed on the continuous face images and registered face images to obtain valid face frames and valid registered face images. The face recognition device includes a visible light camera and a data processing unit. The visible light camera has an output resolution of 1920×1080 pixels and a frame rate of 25 frames per second. After the person to be recognized enters the acquisition range of 0.8 meters to 1.5 meters from the camera, a face video of 2 to 3 seconds is continuously acquired. The face video is then decomposed into continuous face images. Each continuous face image is written with the acquisition time, access control channel number, and acquisition batch number. The personnel registration database stores the registered identity and the corresponding registered face image. The same registered identity includes a frontal face image and registered face images with different head postures. The original resolution of the registered face image is not less than 640×480 pixels. Both the continuous face image and the registered face image contain images with different head and face wearing states. For quality screening in preprocessing, the short side dimension of the face region is not less than 96 pixels, 8-bit brightness is used, the average brightness of the face region is limited to 40 to 220, the image clarity is determined according to the grayscale change of the face region, when the clarity evaluation value is lower than 80, the corresponding image is marked as a blurry image, and the corresponding face region is retained when the horizontal attitude deflection angle does not exceed 45 degrees and the vertical attitude deflection angle does not exceed 30 degrees. Inter-frame association is performed on adjacent valid face frames, and face alignment, image correction and size unification are performed on each associated valid face frame and valid registered face image. The processed valid face frames are arranged according to the acquisition order to generate a face sequence to be identified. The face features of the processed valid registered face images are fused according to the registered identity to generate a registered face template.

[0019] In this embodiment, the construction of the wearable-skin coupling map includes: The device is used to identify the head and face accessories in each face frame of the face sequence to determine the accessory area and accessory type. Each face frame in the face sequence to be identified is sequentially input into the head and face accessory segmentation network according to the acquisition time. The accessory category and face region category are determined, and the accessory segmentation result with the same size as each face frame is output. The head and face accessory segmentation network adopts the SegFormer-B2 semantic segmentation network, and the model output categories are set as background, visible skin, eye occlusion accessory, head and contour occlusion accessory, mouth, nose and lower part occlusion accessory, and cross-region occlusion accessory. Extract the outline of the worn object from the worn object area, extract the skin boundary from the junction of the worn object area and the visible skin area, and establish the boundary correspondence between the worn object outline and the skin boundary; The extraction of the object contour specifically includes: reading the object segmentation results corresponding to each face frame, performing morphological opening and closing operations on the object region, removing isolated pixels outside the object region and filling the segmentation gaps inside the object region, performing connected component analysis on the corrected object region, retaining connected regions with an area not less than 1% of the face region area, and determining the retained connected regions as the object region, using the Suzuki-Abe boundary tracking algorithm to perform boundary search along the outer edge of the object region, starting from the first boundary pixel in the upper left of the object region, sequentially searching for the next boundary pixel adjacent to the current boundary pixel in the eight-neighbor direction, until returning to the starting boundary pixel or reaching the end position of the open contour segment, and determining the boundary pixels connected in the search order as the object contour; The extraction of skin boundaries specifically includes: extracting visible skin regions based on the visible skin category in the wearable segmentation results, performing edge extraction using the Canny operator to obtain the edges of the visible skin regions, expanding the wearable region outward by 3 pixels, comparing its position with the edges of the visible skin regions, and determining the edges of the visible skin regions that are within the expansion range and continuously adjacent to the wearable contour as the skin boundaries; The generation of boundary correspondence specifically includes: selecting contour points of the wearable object at 3-pixel intervals along the contour of the wearable object; retaining at least the first endpoint, middle point, and last endpoint of contour segments with a length of less than 12 pixels; searching for skin boundaries along the normal direction of the corresponding contour, taking each contour point of the wearable object as the starting point; determining the skin boundary point that is closest to the search point with a search distance of no more than 8 pixels as the corresponding skin boundary point; when a skin boundary point corresponds to multiple contour points of the wearable object at the same time, retaining the correspondence between the contour point of the wearable object with the smallest image distance and the skin boundary point. Track the outline of the object and the skin boundary in adjacent face frames according to the boundary correspondence, and determine the cross-frame displacement of the object outline and the cross-frame displacement of the skin boundary respectively. The generation of cross-frame displacement of the wearable contour and skin boundary specifically includes: reading adjacent face frames according to the acquisition time; correcting the overall translation, rotation and scale changes of the face region based on the visible face key points in the adjacent face frames to ensure that the face regions in the adjacent face frames are under a unified position reference; determining the cross-frame correspondence between the wearable contour points and the skin boundary points using bidirectional optical flow tracking; verifying the tracking reliability and boundary pair consistency of the cross-frame correspondence; and generating the cross-frame displacement of the wearable contour and the cross-frame displacement of the skin boundary based on the position changes of the corresponding boundary points that have passed the verification. The process of boundary point tracking and tracking reliability verification specifically includes: using the pyramid Lucas-Kanade optical flow method to perform forward tracking on the wearable contour points and skin boundary points in adjacent face frames. The wearable contour points and skin boundary points in the previous face frame are used as the initial tracking positions. A search window with a side length of 16 pixels is set around the corresponding position in the next face frame. Based on the pixel grayscale changes and local texture changes within the search window, candidate corresponding points in the next face frame are determined. The same optical flow method is used to perform reverse tracking on the determined candidate corresponding points. The candidate corresponding points in the next face frame are used as the tracking starting point. Reverse corresponding points are determined in the previous face frame. The positional distance between the forward tracking starting point and the reverse corresponding point is calculated. When the positional distance is no more than 2 pixels, the corresponding tracking relationship is determined as a valid tracking relationship; otherwise, it is determined as an invalid tracking relationship. For wearable contour points and skin boundary points that have not formed a valid tracking relationship, nearest neighbor matching is performed on the same type of wearable contour and skin boundary in the next face frame. The candidate boundary point with the smallest positional distance between the candidate boundary points is selected to establish a supplementary tracking relationship. The specific steps of the boundary pair consistency verification process include: based on the boundary correspondence between the wearable contour and the skin boundary, forming boundary point pairs with corresponding wearable contour points and skin boundary points; generating wearable contour point displacement vectors and skin boundary point displacement vectors based on the position changes of the two boundary points in adjacent face frames; calculating the inner product of the two displacement vectors; normalizing the product of the lengths of the two displacement vectors; performing inverse cosine processing on the normalization result to obtain the angle between the two displacement vectors as the displacement direction difference; when the displacement direction difference of the boundary point pair is greater than 90 degrees, the corresponding tracking relationship is determined to be a direction conflict relationship; when the displacement amplitude of any boundary point is greater than 3 times the median displacement amplitude of each effective boundary point in the continuous boundary segment, the corresponding tracking relationship is determined to be an amplitude anomaly relationship; discarding the tracking results corresponding to the direction conflict relationship and the amplitude anomaly relationship, and retaining the tracking results that pass the boundary pair consistency verification. The generation of cross-frame displacement specifically includes: determining the horizontal displacement, vertical displacement, and displacement amplitude of the corresponding wearable contour points based on the position changes of the corresponding wearable contour points verified in adjacent face frames, and combining them according to the arrangement order of the wearable contour points in continuous wearable contour segments to generate the cross-frame displacement of the wearable contour; and determining the horizontal displacement, vertical displacement, and displacement amplitude of the skin boundary points based on the position changes of the corresponding skin boundary points verified in adjacent face frames, and combining them according to the arrangement order of the skin boundary points in continuous skin boundary segments to generate the cross-frame displacement of the skin boundary. Using the wearable area, wearable outline, skin boundary, and adjacent skin area as graph nodes, graph connections are established based on spatial adjacency relationships and cross-frame displacement correspondences to construct a wearable-skin coupling graph; The construction of the device-skin coupling graph specifically includes: generating graph nodes based on the device region, device outline, skin boundary and adjacent skin region; establishing intra-frame connections within the same face frame and cross-frame connections between adjacent face frames; writing the state of each region, boundary state and cross-frame displacement into the corresponding graph nodes and graph connections; and associating them according to the acquisition time of the face frames to construct the device-skin coupling graph. The generation of graph nodes specifically includes: setting each continuous wearable area as a wearable area node, setting each continuous contour segment in the wearable outline as a wearable outline node, setting the continuous skin boundary segment that forms a boundary relationship with the wearable outline as a skin boundary node, and setting the continuous visible skin area located outside the skin boundary and adjacent to the skin boundary as an adjacent skin area node. The generation of intra-frame connections specifically includes: in the same face frame, establishing a region affiliation connection between the wearable region node and the wearable contour node belonging to the wearable region, establishing a boundary correspondence connection between the wearable contour node that forms a boundary correspondence relationship and the skin boundary node, and establishing a spatial adjacency connection between the skin boundary node and the corresponding adjacent skin region node. The generation of cross-frame connections specifically includes: between adjacent face frames, determining the cross-frame correspondence of nodes of the same type based on the type of wearable device, region location, and boundary tracking results; connecting wearable device region nodes with the same type of wearable device, continuous region location, and corresponding contours that can be continuously tracked to generate cross-frame connections for wearable device regions; connecting wearable device contour nodes whose contour points can form an effective tracking relationship to generate cross-frame connections for wearable device contours; connecting skin boundary nodes whose skin boundary points can form an effective tracking relationship to generate cross-frame connections for skin boundaries; and connecting adjacent skin region nodes whose positions are continuous and whose visibility states remain consistent to generate cross-frame connections for skin regions.

[0020] In this embodiment, the generation of the visual coupling residual includes: Candidate identity regions are screened from the wearer-skin coupling map and divided into local image blocks. Anchor point stabilization states are generated based on the cross-frame visibility ratio, continuous tracking length, texture changes, and positional fluctuations of each local image block. Local image blocks that directly overlap with the wearer boundary, have texture changes exceeding the stable range, or have interrupted positional continuity are removed. The generation of candidate identity regions specifically includes: identifying adjacent skin regions that remain visible in the face sequence to be identified and have cross-frame connectivity as initial candidate regions; calculating the image distance between each initial candidate region and the boundary of the wearable device; comparing the image distance with the length of the short side of the corresponding face image to generate relative boundary distances; retaining the initial candidate regions whose relative boundary distances are within the distance range to generate candidate identity regions. The distance range is determined by the adjacent skin regions that can be continuously tracked in historical recognition records and subsequently identified as stable identity anchor points; statistically analyzing the relative boundary distances corresponding to adjacent skin regions; arranging them in ascending order of relative boundary distances; determining the relative boundary distance corresponding to the 5th percentile position as the lower limit of distance; and determining the relative boundary distance corresponding to the 95th percentile position as the upper limit of distance. The generation of local image patches and anchor point stable states specifically includes: dividing the candidate identity region into 8×8 pixel local windows, setting the movement step size of adjacent local windows to 4 pixels, counting the number of visible face frames in the local image patch, comparing the number of visible face frames with the total number of frames in the face sequence to be identified to generate the cross-frame visibility ratio, counting the number of face frames that form a continuous effective tracking relationship in the local image patch according to the acquisition time of the face frame to generate the continuous tracking length, when the cross-frame connection is interrupted, counting the continuous tracking length before and after the interruption, and retaining the tracking segment with the larger continuous tracking length, extracting local texture features for each local image patch using the existing local binary mode method, and extracting edge direction features using the directional gradient histogram method, comparing the cosine similarity of the local texture features and edge direction features of corresponding local image patches in adjacent face frames to obtain texture similarity and edge direction similarity, and taking the smaller value of the two similarities as... For cross-frame texture consistency, the median value of the cross-frame texture consistency corresponding to each effective adjacent face frame is used as the texture stability. When the texture stability reaches the texture stability threshold, the local image patch is marked as texture stable. The center position of each local image patch after the overall translation, rotation and scale change correction of the face is read, and the horizontal coordinate sequence and vertical coordinate sequence are obtained respectively. The median position of the horizontal coordinate sequence is calculated. The median value of the absolute deviation between each horizontal coordinate and the median position of the horizontal coordinate is determined as the horizontal position fluctuation. The median position of the vertical coordinate sequence is calculated. The median value of the absolute deviation between each vertical coordinate and the median position of the vertical coordinate is determined as the vertical position fluctuation. When the horizontal position fluctuation and the vertical position fluctuation are not greater than the position stability threshold, the local image patch is marked as position stable. The cross-frame visible ratio, continuous tracking length, texture stability, texture stability mark, horizontal position fluctuation, vertical position fluctuation and position stability mark are combined to form the anchor point stability state. The determination of texture stability threshold and position stability threshold specifically includes: taking local image patches in the effective face sequence in the historical recognition record as stable samples, calculating the texture stability, horizontal position fluctuation and vertical position fluctuation of each stable sample to form a texture stability sample set and a position fluctuation sample set, arranging them in ascending order of value and selecting the texture stability corresponding to the 5th percentile position as the texture stability threshold, taking the larger value between the horizontal position fluctuation and the vertical position fluctuation of each stable sample as the position fluctuation value of the stable sample, and arranging the position fluctuation sample set in ascending order of value and selecting the position fluctuation value corresponding to the 95th percentile position as the position stability threshold; The retention and removal of local image blocks specifically includes: reading the cross-frame visible ratio, continuous tracking length, texture stability and position stability of each local image block; retaining local image blocks with a cross-frame visible ratio not lower than the visible ratio threshold, a continuous tracking length not less than the tracking length threshold, and with texture stability and position stability markers; removing local image blocks that do not meet any of the above conditions; further performing boundary interference verification on the retained local image blocks; removing local image blocks that overlap with the wearable area or intersect with the wearable contour; and determining the remaining local image blocks as local image blocks that pass the stability test. The determination of the visible ratio threshold and the tracking length threshold specifically includes: statistically analyzing the cross-frame visible ratio of local image patches in historical effective face sequences; determining the low quantile value of the cross-frame visible ratio that can maintain stable identity features without significant occlusion changes as the visible ratio threshold; determining the tracking length threshold based on the acquisition frame rate of the face image and the minimum continuous stable duration of the local image patch; and rounding up the product of the acquisition frame rate and the minimum continuous stable duration to generate the tracking length threshold, wherein the minimum continuous stable duration is determined based on the shortest observation time required to distinguish between instantaneous boundary jitter and continuous local deformation in historical effective face sequences. Determine representative locations and local identity features from the preserved local image patches, and combine the representative locations, local identity features, and corresponding cross-frame connections into stable identity anchors; The generation of stable identity anchors specifically includes: using the geometric center of a local image patch as the initial representative position, detecting corner points within the 8×8 pixel area covered by the local image patch, determining the corner point with the largest response value as the representative position when a corner point is detected, and retaining the geometric center as the representative position when no corner point is detected. The corner point detection adopts the Shi-Tomasi corner point detection method, and the local texture features and edge direction features corresponding to the local image patch are sequentially concatenated. Vector normalization is then performed on the concatenation result to generate local identity features. The local image patch number, representative position, local identity features, anchor point stability state, and corresponding cross-frame connection are written into the same data record to generate stable identity anchors. Based on the positional changes of stable identity anchor points and the boundary of the worn object in adjacent face frames, anchor point displacement sequences and boundary displacement sequences are generated respectively. The determination of the representative position of the wearable boundary specifically includes: based on the wearable-skin coupling map, determining the corresponding wearable boundary within the same face frame for each stable identity anchor point; when a stable identity anchor point corresponds to multiple wearable contour nodes, selecting the wearable contour node corresponding to the shortest image distance between the representative position of the stable identity anchor point and each wearable contour node as the target wearable boundary; determining the contour point closest to the representative position of the stable identity anchor point on the target wearable boundary; and selecting continuous contour points along the two extension directions of the wearable contour with this contour point as the center; and combining the median value of the horizontal coordinate and the median value of the vertical coordinate of the contour points to form the representative position of the wearable boundary. The generation of the anchor point displacement sequence specifically includes: reading adjacent face frames sequentially according to the acquisition time, obtaining the representative positions of the same stable identity anchor point in the previous and next face frames, subtracting the abscissa of the representative position in the next face frame from the abscissa of the representative position in the previous face frame to generate the horizontal displacement of the anchor point, subtracting the ordinate of the representative position in the next face frame from the ordinate of the representative position in the previous face frame to generate the vertical displacement of the anchor point, combining the horizontal and vertical displacements of the anchor point into an anchor point displacement vector, and calculating the length and direction of the anchor point displacement vector. The data format of the anchor point displacement vector includes the anchor point number, adjacent face frame number, horizontal displacement of the anchor point, vertical displacement of the anchor point, displacement amplitude, and displacement direction. The anchor point displacement vectors corresponding to the same stable identity anchor point are arranged sequentially according to the acquisition time of the face frames to generate the anchor point displacement sequence. The generation of the boundary displacement sequence specifically includes: reading the representative position of the wearable boundary corresponding to the stable identity anchor point, and processing it according to the adjacent face frames with the same anchor point displacement; subtracting the abscissa of the representative position of the wearable boundary in the previous face frame from the abscissa of the representative position of the wearable boundary in the next face frame to generate the horizontal boundary displacement; subtracting the ordinate of the representative position of the wearable boundary in the previous face frame from the ordinate of the representative position of the wearable boundary in the next face frame to generate the vertical boundary displacement; combining the horizontal and vertical boundary displacements into a boundary displacement vector; and calculating the length and direction of the boundary displacement vector. The data format of the boundary displacement vector includes the corresponding anchor point number, adjacent face frame number, horizontal boundary displacement, vertical boundary displacement, displacement amplitude, and displacement direction. The boundary displacement vectors corresponding to the same stable identity anchor point are arranged sequentially according to the acquisition time of the face frames to generate the boundary displacement sequence. A relative displacement reference is generated based on the anchor point displacement sequence and the boundary displacement sequence. The relative displacement in adjacent face frames is compared with the relative displacement reference to obtain the cross-frame positional deviation and orientation deviation, and a visual coupling residual is generated. The generation of relative displacement specifically includes: for each set of displacement vectors with the same acquisition time in the anchor point displacement sequence and the boundary displacement sequence, subtracting the boundary horizontal displacement from the anchor point horizontal displacement to obtain the relative horizontal displacement, subtracting the boundary vertical displacement from the anchor point vertical displacement to obtain the relative vertical displacement, combining the relative horizontal displacement and the relative vertical displacement into a relative displacement vector, and removing relative displacement vectors with invalid tracking status, incomplete wearable boundary, or obviously abnormal displacement amplitude, determining the median value of the relative horizontal displacement in the effective relative displacement vector as the reference horizontal displacement, determining the median value of the relative vertical displacement as the reference vertical displacement, and combining the reference horizontal displacement and the reference vertical displacement into a relative displacement reference; The generation of cross-frame position deviation specifically includes: for each set of valid relative displacement vectors, subtracting the reference horizontal displacement from the relative horizontal displacement to obtain the horizontal position deviation, subtracting the reference vertical displacement from the relative vertical displacement to obtain the vertical position deviation, calculating the square of the horizontal position deviation and the square of the vertical position deviation, adding the two together and performing a square root operation to obtain the position deviation magnitude, and combining the horizontal position deviation, vertical position deviation and position deviation magnitude to obtain the cross-frame position deviation. The generation of directional deviation specifically includes: for each set of valid relative displacement vectors, combining the current relative horizontal displacement and the current relative vertical displacement into the current relative displacement vector, combining the reference horizontal displacement and the reference vertical displacement into the relative displacement reference vector, calculating the inner product of the current relative displacement vector and the relative displacement reference vector, calculating the lengths of the current relative displacement vector and the relative displacement reference vector respectively, dividing the inner product by the product of the lengths of the two vectors to generate a directional similarity value, using an interval limiting function to restrict the directional similarity value to a numerical range of -1 to 1, and performing an inverse cosine operation on the directional similarity value to obtain the angle between the current relative displacement vector and the relative displacement reference vector, which is taken as the directional deviation; The visual coupling residual vector is obtained by sequentially arranging and combining the horizontal position deviation, vertical position deviation, position deviation magnitude and direction deviation corresponding to each group of adjacent face frames.

[0021] In this embodiment, the generation of deformation features, standardized identity status, and wearing identity status includes: Based on visual coupling residuals, stable identity anchors, and the sequence of faces to be identified, an anchor-aligned input group is generated. Based on the anchor point number, the previous face frame number, and the next face frame number recorded in the visual coupling residual, each visual coupling residual is associated with the corresponding stable identity anchor point and the next face frame to form an anchor point alignment input item, including face frame number, anchor point number, anchor point representative position, local identity features, horizontal position deviation, vertical position deviation, position deviation magnitude, direction deviation, and valid state flag. Anchor point alignment input items belonging to the same face sequence to be identified are arranged in ascending order of face frame number, and arranged in the same face frame according to anchor point number to generate an anchor point alignment input group. When a stable identity anchor point has no valid cross-frame connection in the current face frame, the anchor point number and face frame number are retained, the corresponding input item is set as an invalid input item, and the valid state flag prevents it from participating in subsequent state aggregation. Extract face features, anchor position features, local identity features and visual coupling residuals from the anchor-aligned input group, map each feature to a token of the same dimension, and write it into time encoding, position encoding and data type encoding to generate a residual encoding sequence; The generation of facial features and anchor point location features specifically includes: inputting each facial frame in the sequence of faces to be identified into the convolutional feature extraction unit, performing convolution, normalization, and nonlinear mapping in sequence to generate a facial feature map; based on the representative position of the stable identity anchor point, converting the representative position from facial image coordinates to facial feature map coordinates, and using bilinear interpolation to extract the local facial features at the corresponding position from the facial feature map; dividing the horizontal and vertical coordinates of the representative position of the stable identity anchor point by the width and height of the facial image respectively to generate normalized anchor point coordinates; dividing the width and height of the local image block by the width and height of the facial image respectively to generate normalized local dimensions; concatenating the normalized anchor point coordinates, normalized local dimensions, and anchor point stable state to generate an anchor point location input vector; inputting the anchor point location input vector into a fully connected mapping layer and performing an affine transformation to obtain anchor point location features of fixed dimensions. The residual coding sequence is input into the residual coding layer of the coupled dual-domain Mamba-2 model, which includes a residual coding layer, a dual-domain verification layer, and a joint uncertainty decision layer. The training process of this model specifically includes: using a sequence of faces to be identified as a training sample, each training sample includes 16 to 32 frames of face images, with no more than 16 stable identity anchors retained per frame. All tokens are uniformly set to 256 dimensions. The training samples are divided into training, validation, and test sets according to the principle of mutual exclusion of registered identities, with a ratio of 8:1:1. For the same registered identity, registered face features with low occlusion and high quality screening results are selected as the standard identity reference. Face sequences with head and face accessories are used as worn identity samples, and the feature differences between the worn identity sample features and the standard identity reference are used as deformation references. For each worn identity sample, a real registered identity and different registered identities with high identity retrieval scores are configured, forming positive and negative candidate identities respectively. The corresponding frame-level deformation coefficients are zeroed, inverted, and replaced to form intervention training samples. Training labels include registered identity labels, candidate identity matching labels, deformation direction consistency labels, deformation amplitude consistency labels, conflict labels, and result type labels. Among these, candidate features corresponding to the real registered identity are marked as identity matching, while other registered identities are marked as identity mismatch. Real candidate samples generated using the original frame-level deformation coefficients are labeled as deformation-consistent. Intervention samples generated using zeroed, reversed, or replaced deformation coefficients are labeled as causally consistent or causally inconsistent based on the relationship between their three quantities. Samples with existing registered identities and unique candidate identities that pass causality verification are labeled as identity-identified samples. Samples with registered identities not written into the current registered face template are labeled as unknown person samples. Samples with insufficient stable identity anchors, fewer than two types of effective interventions, missing intervention data, or multiple candidate identities that pass causality verification simultaneously are labeled as re-collected samples. During training, the residual coding sequence is first input... The residual coding layer generates deformation features, normalized identity status, and wearing identity status. The deformation features are processed through fixed intrinsic decomposition, three-level verification, and candidate feature transformation to generate candidate identical features. The candidate identical features, normalized identity status, and wearing identity status are then input into the dual-domain verification layer to generate identity consistency quantity, deformation consistency quantity, and conflict quantity. The corresponding causal verification results are input into the joint uncertainty decision layer. The candidate identity ranking, orthogonal basis dimension selection and zeroing, reversal, and substitution processing adopt fixed operations and do not participate in gradient updates. Gradient propagation is preserved between the candidate identical features and the dual-domain verification layer. The model employs a phased training approach. The total training loss consists of the sum of identity classification loss, two-domain alignment loss, deformation fitting loss, response orthogonality loss, two-domain verification loss, and decision consistency loss. Identity classification loss uses additive angular-margin cross-entropy; two-domain alignment loss uses supervised contrastive loss to narrow the feature distance between different wearing states of the same identity and separate different identities; deformation fitting loss uses smoothing L1 loss and cosine direction loss to constrain the deformation response to maintain consistency with the deformation reference in amplitude and direction; response orthogonality loss uses the square of the cosine similarity between the deformation response and the normalized domain response; and two-domain verification loss uses... Binary cross-entropy is used to constrain identity consistency, deformation consistency, and conflict, respectively, and ranking loss is added. The decision consistency loss consists of candidate identity distribution cross-entropy, unknown personnel candidate score suppression term, and uncertainty ranking term of re-collected samples, and is adjusted according to the validation set results. The AdamW optimizer is used to update the model parameters. The initial learning rate is set to 0.0001, the weight decay coefficient is set to 0.01, and the batch size is set to 16. When the recognition accuracy improves by no more than 0.1 percentage points for 5 consecutive rounds, the model is considered to have reached convergence and training is stopped. The model parameters with the lowest total loss in the validation set are saved. The improvements to this model include: First, addressing the issue that the original Mamba-2 only performs state space scanning on a single sequence of features and has difficulty distinguishing between identity information and wearer deformation information, anchor alignment, multi-source token encoding, and visual coupling gating are added to the input to decompose the face response into deformation response, normalization domain response, and wearer domain response. Second, the original Mamba-2's single-path temporal scanning is changed to a normalization domain scanning branch and a wearer domain scanning branch with mutually independent parameters, generating normalized identity state and wearer identity state respectively, to reduce the interference of wearer occlusion and local skin deformation on stable identity features. Third, a dual-domain verification layer and a joint uncertainty decision layer are added after the state scanning, generating candidate identical features through orthogonal basis and deformation coefficients, and using zeroing, reversal, and substitution interventions to verify the changing relationship between identity consistency quantity, deformation consistency quantity, and conflict quantity, enabling the model to further distinguish between identifying identities, unknown persons, and situations requiring re-collection. In the residual coding layer, a coupling gate is generated based on the visual coupling residual. The coupling gate is used to decompose the face feature response into deformation response, normalization domain response and wearing domain response. Deformation features are generated based on the deformation response corresponding to each effective anchor point. The generation of the residual coding sequence specifically includes: inputting global face features, local face features, local identity features, and visual coupling residuals into fully connected mapping layers with independent parameters, converting them into global face tokens, local face tokens, anchor tokens, and residual tokens of the same dimension, and using anchor position features as position tokens. Each fully connected mapping layer is used to project input features of different dimensions and different data sources to a feature space of the same dimension. The face tokens, anchor tokens, position tokens, and residual tokens corresponding to the same anchor-aligned input items are arranged in order, and all valid anchor-aligned input items are connected sequentially according to the acquisition time of the face frames to generate the residual coding sequence. The generation of the coupled gating quantity specifically includes: concatenating the face token, anchor token, position token and residual token corresponding to the input item with the same anchor point, inputting the concatenated token into the fully connected layer and performing a nonlinear transformation using the GELU activation function to generate gating intermediate features, inputting the gating intermediate features into the fully connected layer again and using the Sigmoid function to map the gating response of each dimension to the numerical range of 0 to 1 to generate a coupled gating quantity with the same dimension as the face token; The generation of the three types of responses specifically includes: adding the face token corresponding to the local face features and the anchor token dimension by dimension, and performing layer normalization to generate the current identity response; concatenating the current identity response with the corresponding residual token to generate the deformation input vector; inputting the deformation input vector into a fully connected layer to convert it into deformation intermediate features; performing a nonlinear transformation on the deformation intermediate features using the GELU activation function; converting it into a candidate deformation response through a fully connected layer; performing layer normalization on the candidate deformation response; multiplying the candidate deformation response with the coupling gate variable dimension by dimension to generate the deformation response; subtracting the deformation response from the current identity response dimension by dimension, and performing layer normalization on the subtraction result to generate the normalized domain response; concatenating the current identity response and the deformation response and inputting them into the wearing domain mapping unit; and performing the same fully connected transformation, nonlinear activation, full connected transformation again, and layer normalization to generate the wearing domain response. The generation of deformation features specifically includes: arranging the deformation responses corresponding to all valid anchor points in the same face sequence to be identified according to the face frame number and anchor point number to form a deformation response matrix. The data size of this deformation response matrix is ​​the product of the number of valid deformation responses and the dimension of the deformation response. Subtracting the mean vector of all valid deformation responses from each deformation response yields a centered deformation response. Calculating the cross product of each centered deformation response with its transpose vector, summing all cross product results and dividing by the number of valid deformation responses yields a symmetric deformation correlation matrix. Calculating the average value of each diagonal element of the symmetric deformation correlation matrix, multiplying the average value by the numerical stability coefficient yields a numerical stability term, and adding the numerical stability term to each diagonal position of the symmetric deformation correlation matrix to generate deformation features. The numerical stability coefficient is determined based on the condition number, minimum eigenvalue, and eigenvalue decomposition convergence state of the symmetric deformation correlation matrix before and after adding the numerical stability term. The initial engineering value is set to 0.00001. Perform temporal state scanning on the normative domain response and the wearing domain response respectively, and aggregate the valid temporal states to generate the normative identity state and the wearing identity state; The generation of standardized identity states specifically includes: performing masked average pooling on the standardized domain responses within the same face frame to generate standardized domain anchor point aggregation features; concatenating the standardized domain anchor point aggregation features with the global face token to form a standardized domain frame input vector, and converting it into a standardized domain frame token through a fully connected layer; arranging the standardized domain frame tokens corresponding to each face frame according to the acquisition time to generate a standardized domain temporal sequence; for each standardized domain frame token, performing layer normalization first, and then inputting it into independent fully connected layers to generate state scan inputs and gated inputs of the same dimension; performing one-dimensional depthwise convolution and SiLU activation on the state scan inputs to generate state transition parameters and input parameters. The system calculates the number and output parameters, and performs selective state space scanning in conjunction with the canonical domain scanning state corresponding to the previous face frame to generate the current canonical domain scanning state. It then performs SiLU activation on the gated input to generate a canonical domain gate vector. The canonical domain gate vector is multiplied dimension by dimension with the current canonical domain scanning state, and after transformation by the output fully connected layer, it is added to the token of the current canonical domain frame to generate the canonical domain output state corresponding to the current face frame. After completing the state scanning of all face frames according to the acquisition time, the system performs masked average pooling on the effective canonical domain output states and concatenates them with the canonical domain output state corresponding to the last effective face frame. Finally, it is converted into a canonical identity state through a fully connected layer. The generation of the wearing identity state is the same as that of the standard identity state. The wearing domain response within the same face frame is subjected to masked average pooling, and then concatenated with the global face token and fully connected to generate a wearing domain frame token. The wearing domain frame tokens are arranged according to the acquisition time to generate a wearing domain temporal sequence. Layer normalization, dual-path fully connected mapping, selective state space scanning, gated fusion, output transformation and residual addition are performed in sequence to generate the wearing domain output state corresponding to each face frame. Masked average pooling is performed on the valid wearing domain output state. The pooling result is concatenated with the wearing domain output state corresponding to the last valid face frame and converted into the wearing identity state through a fully connected layer.

[0022] In this embodiment, the generation of candidate features with the same conditions includes: Perform symmetric matrix eigenvalue decomposition on the deformation features to form an orthogonal basis, and project the effective deformation response that forms the deformation features onto the orthogonal basis to generate deformation coefficients; The generation of orthogonal basis and deformation coefficients specifically includes: performing eigenvalue decomposition on the symmetric deformation correlation matrix to obtain multiple eigenvalues ​​and their corresponding eigenvectors. Since the symmetric deformation correlation matrix is ​​a real symmetric matrix, the eigenvectors are mutually orthogonal. Each eigenvector is normalized to a unit length. Each eigenvalue and its corresponding eigenvector form multiple eigen pairs. The eigen pairs are arranged in descending order of eigenvalue. Starting from the first eigen pair, the current eigenvalue is accumulated to the accumulated eigenvalues, and the ratio of the accumulated eigenvalues ​​to the sum of all eigenvalues ​​is calculated to generate the cumulative eigenvalue contribution rate corresponding to the current eigen pair. When the cumulative eigenvalue contribution rate reaches the contribution rate threshold for the first time, the number of eigenvectors from the first eigen pair to the current eigen pair is determined as the orthogonal basis dimension. The corresponding eigenvectors are arranged in the column direction to generate the orthogonal basis. For each effective deformation response, the mean deformation response is subtracted from the effective deformation response to generate the centered deformation response. The product of the transpose of the orthogonal basis and the centered deformation response is calculated to generate the corresponding deformation coefficient. The determination of the contribution rate threshold specifically includes: setting multiple candidate contribution rate thresholds, including 90%, 92%, 95%, 97%, and 99%; determining the corresponding orthogonal basis dimension based on each candidate contribution rate threshold; calculating the reconstruction error of the deformation response, the subsequent identity recognition accuracy, and the computational cost of a single feature transformation in historical valid samples; and selecting the candidate contribution rate threshold with the smaller orthogonal basis dimension from among the candidate contribution rate thresholds where the reconstruction error is not higher than the reconstruction error requirement and the identity recognition accuracy meets the accuracy requirement, as the contribution rate threshold. The effective deformation response is reconstructed based on the orthogonal basis and deformation coefficients, and the reconstruction error is generated. Coefficient stability verification is performed on the deformation coefficients that pass the orthogonal reconstruction verification. The orthogonal reconstruction verification specifically includes: multiplying the orthogonal basis with the deformation coefficients to generate reconstructed deformation components, and superimposing the mean of the deformation response onto the reconstructed deformation components to generate a reconstructed deformation response; calculating the Euclidean distance between the effective deformation response and the corresponding reconstructed deformation response; and dividing the Euclidean distance by the sum of the vector length of the effective deformation response and the numerical stability quantity to generate a normalized reconstruction error. The numerical stability quantity is used to avoid division by zero when the vector length of the effective deformation response is close to zero. The initial engineering value is set to 0.00001. The normalized reconstruction error corresponding to all effective deformation responses is calculated respectively. The median value of the normalized reconstruction error is used as the reconstruction error of the current orthogonal basis. When the reconstruction error is not greater than the reconstruction error threshold, the current orthogonal basis and the corresponding deformation coefficients are determined to have passed the orthogonal reconstruction verification. The coefficient stability verification specifically includes: arranging the deformation coefficients that have passed the orthogonal reconstruction verification according to the face frame number and anchor point number; when the same face frame corresponds to multiple valid stable identity anchor points, performing masked average pooling on the deformation coefficients within the same face frame to generate frame-level deformation coefficients; arranging the frame-level deformation coefficients corresponding to each face frame according to the acquisition time to generate a deformation coefficient sequence; for adjacent face frames, subtracting the frame-level deformation coefficient of the previous face frame from the frame-level deformation coefficient of the subsequent face frame, calculating the length of the difference vector, and dividing the length by the sum of the length of the frame-level deformation coefficient of the previous face frame and the numerical stability value to obtain the adjacent coefficient change rate, with the numerical stability value set to 0.00001; determining the median value of the adjacent coefficient change rates corresponding to all valid adjacent face frames as the coefficient fluctuation value; when the coefficient fluctuation value is not greater than the coefficient stability threshold, the corresponding deformation coefficient is determined to have passed the coefficient stability verification. The determination of the verification threshold specifically includes: reading face sequences from historical recognition records that have been manually verified as belonging to the same person, with clear status of the worn items and continuous and effective visual coupling residuals; calculating the reconstruction error, coefficient fluctuation, and coupling consistency of each historical face sequence to form a reconstruction error sample set, a coefficient fluctuation sample set, and a coupling consistency sample set; determining the 95th percentile of the reconstruction error sample set as the reconstruction error threshold; determining the 95th percentile of the coefficient fluctuation sample set as the coefficient stability threshold; and determining the 5th percentile of the coupling consistency sample set as the coupling consistency threshold. In the initial debugging stage of the model, the reconstruction error threshold can be set to 0.15, the coefficient stability threshold to 0.3, and the coupling consistency threshold to 0.5. Compare the changes in deformation coefficients that have passed the coefficient stability check with the changes in visual coupling residuals, perform coupling consistency check, and retain the orthogonal basis and deformation coefficients that have passed the orthogonal reconstruction check, coefficient stability check and coupling consistency check in sequence; The coupling consistency verification specifically includes: dividing the horizontal position deviation, vertical position deviation, and position deviation magnitude in the corresponding visual coupling residual by the length of the short side of the face image, dividing the directional deviation by 180 degrees, generating a four-dimensional normalized residual vector, calculating the Euclidean length of the four-dimensional normalized residual vector, generating the visual residual magnitude, calculating the Euclidean length of the frame-level deformation coefficient corresponding to the same face frame, generating the coefficient magnitude, matching the visual residual magnitude and coefficient magnitude corresponding to each face frame one-to-one according to the acquisition time, forming a visual residual magnitude sequence and a coefficient magnitude sequence, using the Spearman rank correlation coefficient calculation method, encoding the sequence values ​​in the visual residual magnitude sequence and the coefficient magnitude sequence according to their numerical values, calculating the correlation coefficient between the corresponding rank sequences as the coupling consistency quantity, and determining that the corresponding orthogonal basis and deformation coefficient pass the coupling consistency verification when the coupling consistency quantity is not less than the coupling consistency threshold; Anchor point query features are generated based on stable identity anchor points. The anchor point query features are matched with the registered face templates for similarity. Candidate identities are filtered based on the matching results, and the corresponding candidate identity features are read. The generation of anchor query features specifically includes: reading the local identity features and anchor stability states of each stable identity anchor; inputting each local identity feature into the same fully connected layer in sequence to convert it into an anchor query vector with the same feature dimension as the registered face template; performing unit length normalization on each anchor query vector; filtering effective anchor query vectors based on texture stability, cross-frame visibility ratio, and continuous tracking length in the anchor stability state; performing masked average pooling on all effective anchor query vectors; and performing unit length normalization again on the pooling result to generate anchor query features. The retrieval of candidate identities specifically includes: reading the facial features merged according to the registered identity in the registered face template, performing unit length normalization on each registered face feature, calculating the cosine similarity between the anchor query feature and each registered face feature, when the same registered identity includes multiple registered face features, selecting the three registered face features with the highest similarity, calculating their average similarity, generating the identity retrieval score for the corresponding registered identity, arranging each registered identity according to the identity retrieval score from high to low, selecting the top-ranked candidate registered identities as candidate identities, and reading the corresponding candidate identity features; The candidate identity features are transformed into a feature space consistent with the orthogonal basis. Deformation components are generated based on the verified orthogonal basis and deformation coefficients. The deformation components are superimposed on the candidate identity features to generate candidate features with the same conditions. The generation of candidate identical conditional features specifically includes: inputting candidate identity features into a fully connected layer, converting them into a feature space with the same dimension as the deformation response, and performing unit length normalization; for each valid face frame, reading the frame-level deformation coefficients that have passed three-level verification, calculating the product of the orthogonal basis and the frame-level deformation coefficients to generate the deformation component of the corresponding face frame; adding the deformation component to the candidate identity features corresponding to each candidate identity dimension by dimension, and performing layer normalization and unit length normalization on the addition result to generate candidate identical conditional features of the candidate identity under the current wearing object conditions; arranging the candidate identical conditional features of each face frame corresponding to the same candidate identity according to the acquisition time to generate a candidate identical conditional feature sequence; combining the candidate identical conditional feature sequences corresponding to all candidate identities according to the order of identity retrieval scores to generate candidate identical conditional features, and establishing associations with the candidate identity number, face frame number, orthogonal basis dimension, frame-level deformation coefficient, and identity retrieval score.

[0023] In this embodiment, the generation of the causality check result includes: The candidate features with the same conditions, the standardized identity status, and the wearing identity status are input into the dual-domain verification layer, transformed into a unified feature space, and a dual-domain verification feature group is generated. The generation of dual-domain verification feature groups specifically includes: according to the candidate identity number and face frame number, each candidate conditional feature is combined with the corresponding standard identity state and wearing identity state to form a dual-domain verification input item, which is then input into a fully connected layer with independent parameters, and converted into candidate verification features, standard domain verification features and wearing domain verification features with the same dimension. Then, layer normalization and unit length normalization are performed respectively, and combined into a dual-domain verification feature group. Based on the candidate identical condition features and the standard identity status and wearing identity status, generate identity consistency quantity, deformation consistency quantity and conflict quantity; The generation of the identity consistency quantity specifically includes: calculating the cosine similarity between the candidate verification feature and the normalized domain verification feature to obtain the normalized domain identity similarity; calculating the cosine similarity between the candidate verification feature and the wearing domain verification feature to obtain the wearing domain identity similarity; mapping the normalized domain identity similarity and the wearing domain identity similarity to the value range of 0 to 1 respectively; multiplying the two mapping results and performing a square root operation to obtain the identity consistency quantity. The generation of deformation consistency quantity specifically includes: subtracting the normalized domain verification features from the candidate verification features dimension by dimension to obtain the predicted deformation vector; subtracting the normalized domain verification features from the wearing domain verification features dimension by dimension to obtain the observed deformation vector; calculating the cosine similarity between the predicted deformation vector and the observed deformation vector; mapping the cosine similarity to the range of 0 to 1 to generate deformation direction consistency; calculating the Euclidean lengths of the predicted deformation vector and the observed deformation vector respectively; dividing the smaller of the two Euclidean lengths by the sum of the larger value and the numerical stability quantity to generate deformation amplitude consistency; averaging the deformation direction consistency and deformation amplitude consistency to generate deformation consistency quantity; and setting the numerical stability quantity to 0.000001. The generation of the conflict quantity specifically includes: calculating the absolute difference between the identity similarity in the normative domain and the identity similarity in the wearing domain to obtain the dual-domain identity difference quantity; subtracting the deformation consistency quantity from 1 to obtain the deformation deviation quantity; and averaging the dual-domain identity difference quantity and the deformation deviation quantity to obtain the conflict quantity. Using the deformation coefficient as the intervention variable, zeroing intervention, reverse intervention, and alternative intervention were performed respectively to generate zeroing intervention features, reverse intervention features, and alternative intervention features; The generation of zero-set intervention features specifically includes: reading the orthogonal basis and frame-level deformation coefficients associated with the candidate same-condition features; determining the frame-level deformation coefficients as intervention variables; setting all coefficients in each dimension of the frame-level deformation coefficients to 0 to generate zero-set deformation coefficients; multiplying the orthogonal basis with the zero-set deformation coefficients to obtain zero-set deformation components; adding the zero-set deformation components to the corresponding candidate identity features dimension by dimension; and performing layer normalization and unit length normalization according to the generation method of candidate same-condition features to obtain zero-set intervention features. The generation of reverse intervention features specifically includes: multiplying the coefficients of each dimension in the frame-level deformation coefficients by -1 to generate reverse deformation coefficients; multiplying the orthogonal basis with the reverse deformation coefficients to obtain reverse deformation components; adding the reverse deformation components with the corresponding candidate identity features dimension by dimension; and performing layer normalization and unit length normalization to obtain reverse intervention features. The generation of alternative intervention features specifically includes: reading the same frame-level deformation coefficient dimension as the current frame-level deformation coefficient dimension from other valid face frames of the same face sequence to be identified, using it as a candidate alternative coefficient; calculating the directional angle between each candidate alternative coefficient and the current frame-level deformation coefficient; selecting the candidate alternative coefficient with the largest directional angle; calculating the Euclidean length between the current frame-level deformation coefficient and the selected candidate alternative coefficient; scaling the candidate alternative coefficient by amplitude so that the scaled candidate alternative coefficient has the same Euclidean length as the current frame-level deformation coefficient; generating alternative deformation coefficients; multiplying the orthogonal basis with the alternative deformation coefficients to obtain alternative deformation components; adding the alternative deformation components to the corresponding candidate identity features dimension by dimension; and performing layer normalization and unit length normalization to generate alternative intervention features. Based on the zero-intervention feature, reverse intervention feature, and alternative intervention feature, the identity consistency quantity, deformation consistency quantity, and conflict quantity corresponding to various interventions are regenerated, and the changes of the three quantities before and after the intervention are calculated. The three changes after the intervention specifically include: according to the same feature conversion and calculation method as before the intervention, regenerating the zero-intervention intervention consistency quantity, zero-intervention deformation consistency quantity, and zero-intervention conflict quantity by combining the zero-intervention intervention feature with the standard identity state and the wearing identity state; regenerating the reverse intervention identity consistency quantity, reverse intervention deformation consistency quantity, and reverse intervention conflict quantity by combining the reverse intervention feature with the standard identity state and the wearing identity state; and regenerating the alternative intervention identity consistency quantity, alternative intervention deformation consistency quantity, and alternative intervention conflict quantity by combining the alternative intervention feature with the standard identity state and the wearing identity state. For each type of intervention, subtracting the post-intervention consistency quantity from the pre-intervention consistency quantity yields the corresponding consistency change quantity. When the change in deformation consistency quantity is in the same direction as the change in identity consistency quantity and in the opposite direction to the change in conflict quantity, the corresponding intervention is identified as a causal consistency intervention, and a causal verification result is generated based on the number of causal consistency interventions and the magnitude of the changes in the three quantities. For each causal consistency intervention, calculate the absolute values ​​of the changes in identity consistency, deformation consistency, and conflict, and calculate the arithmetic mean of the three absolute values ​​to generate the average change magnitude of the three quantities for the corresponding intervention. Count the number of causal consistency interventions among zeroing interventions, reverse interventions, and substitution interventions to generate the number of causal consistency interventions. When the number of causal consistency interventions is greater than 0, calculate the arithmetic mean of the average change magnitudes of the three quantities for each causal consistency intervention to generate the average change magnitude of the causal process. When the number of causal consistency interventions is 0, set the average change magnitude of the causal process to 0. Calculate the ratio between the number of causal consistency interventions and the total number of interventions to generate the causal consistency ratio. Multiply the causal consistency ratio by the average change magnitude of the causal process to obtain the causal verification score. When at least two of the three types of interventions are identified as causal consistency interventions, set the causal verification mark to pass. When fewer than two types of causal consistency interventions are identified, set the causal verification mark to fail. The candidate identity number, the change in identity consistency, change in deformation consistency and change in conflict corresponding to various interventions, the type of causal consistent intervention, the number of causal consistent interventions, the proportion of causal consistency, the average change in causality, the causal verification score and the causal verification mark are written into the same data record to generate the causal verification result.

[0024] In this embodiment, the generation of identity verification results, unknown personnel results, and re-collection instructions includes: The causal verification results are input into the joint uncertainty decision layer. Candidate decision data are generated based on the causal verification score, the number of causal consistent interventions, changes in the three quantities, and the causal verification marker. Candidate identity distribution and identity uncertainty are generated based on the causal verification score. For candidate decision data, the causal verification score, the number of causal consistency interventions, the changes in the three quantities, and the causal verification mark corresponding to the same candidate identity are written into the same decision record, and the decision records are arranged from high to low according to the causal verification score to generate candidate decision data. For the candidate identity distribution, the causal verification scores less than 0 are corrected to 0. The sum of the causal verification scores corresponding to all candidate identities is calculated. The causal verification score of each candidate identity is divided by the sum of the scores to obtain the normalized candidate score of the corresponding candidate identity. The candidate identity numbers and their corresponding normalized candidate scores are arranged from high to low according to the causal verification scores to generate the candidate identity distribution. For identity uncertainty, the Shannon entropy calculation method is used. The product of each normalized candidate score and its natural logarithm is calculated separately. All products are summed and the negative value is taken to generate the candidate distribution entropy. The candidate distribution entropy is divided by the natural logarithm of the number of candidate identities to obtain the normalized distribution entropy. The difference between the normalized candidate scores of the first and second ranked candidates in the candidate identity distribution is used to obtain the candidate score interval. When there is only 1 candidate identity, the normalized candidate score of the second ranked candidate is set to 0. The difference between 1 and the candidate score interval is determined as the score proximity. The arithmetic mean of the normalized distribution entropy and the score proximity is calculated to obtain the identity uncertainty. Based on the causal consistency status and the magnitude of change of the three quantities corresponding to various interventions, the intervention consistency ratio and the change discrete quantity are generated, and the deformation uncertainty is generated based on the intervention consistency ratio and the change discrete quantity. The generation of the intervention consistency ratio specifically includes: for the candidate identity with the highest causal verification score, reading the causal consistency status corresponding to the zero-intervention, reverse intervention, and alternative intervention respectively, dividing the number of intervention types determined as causal consistent interventions by the total number of effective interventions to obtain the intervention consistency ratio; The generation of the variable discrete quantity specifically includes: calculating the arithmetic mean of the average change amplitudes of the three quantities corresponding to each effective intervention, generating the average change amplitude of the intervention, calculating the difference between the average change amplitudes of the three quantities and the average change amplitude of the intervention, and performing square root calculation based on the square mean of each difference to obtain the standard deviation of change, dividing the standard deviation of change by the sum of the average change amplitude of the intervention and the numerically stable quantity to generate the relative variable discrete quantity, and using an interval limiting function to limit the relative variable discrete quantity to the numerical range of 0 to 1 to obtain the variable discrete quantity, with the initial engineering value of the numerically stable quantity set to 0.000001; For deformation uncertainty, subtract the intervention consistency ratio from 1 to obtain the intervention inconsistency quantity, and calculate the arithmetic mean of the intervention inconsistency quantity and the change discrete quantity to generate deformation uncertainty; Based on the candidate identity with the highest causal verification score, identity uncertainty, deformation uncertainty and corresponding causal verification mark, a joint judgment is performed, and the results of identity recognition, unknown persons and re-collection instructions are output. The joint judgment and result output specifically include: reading the causal verification mark, causal verification score, identity uncertainty, deformation uncertainty, and intervention data integrity status corresponding to each candidate identity, and arranging the candidate identities from high to low according to the causal verification score. When only the candidate identity at the top of the ranking passes the causal verification and the intervention data is complete, the candidate identity is output as the identification identity. When all candidate identities fail the causal verification and the intervention data is complete, the result of unknown personnel is output. When multiple candidate identities pass the causal verification, the top ranking is not unique, or the intervention data is missing, a re-collection instruction is output and the corresponding reason is recorded. Among them, the completeness of intervention data specifically means that the candidate decision data is complete and the number of effective interventions in zeroing intervention, reverse intervention, and alternative intervention is not less than two types.

[0025] Example 1: To verify the feasibility of this invention in practice, it was applied to the personnel access control area of ​​a comprehensive office building. Personnel passing through often wear masks, glasses, sunglasses, or hats. The existing equipment uses full-frame facial features to match the registration template, which is prone to mixing identity features with local skin changes, resulting in the generation of two similar candidate identities for the same person or increased feature fluctuations in adjacent frames.

[0026] Visible light cameras and edge data processing units are deployed at each access control entrance. The image resolution is 1920×1080 pixels, and the acquisition frame rate is 25 frames per second. Personnel are continuously captured for 2 to 3 seconds within a range of 0.8 to 1.5 meters from the camera. A total of 6240 verification samples were collected, including 4800 valid sequences from registered personnel, 960 sequences from unregistered personnel, and 480 sequences from blurred, occluded, or abnormal postures, covering mask, eyeglasses, sunglasses, hats, headphones, and combinations thereof. The system first locates the face region and performs quality screening, assembling the valid face frames into a sequence of faces to be identified according to the acquisition order. A head and face adornment segmentation network determines the adornment region and type, and the system extracts the adornment contour and skin boundary at the junction of the adornment and visible skin. Adjacent frames are tracked using bidirectional optical flow to trace contour points and skin boundary points, with a search window of 16 pixels. Tracking relationships are preserved when the forward and backward positional distance is no greater than 2 pixels. The system constructs an object-skin coupling graph by using the wearable area, wearable outline, skin boundary, and adjacent skin areas as graph nodes, and spatial adjacency relationships and cross-frame displacement correspondences as graph connections. The system divides candidate identity regions into 8×8 pixel local image blocks with a movement step of 4 pixels, and filters stable identity anchors based on cross-frame visibility ratio, continuous tracking length, texture stability, and positional fluctuation. The system calculates the cross-frame displacement of both stable identity anchors and wearable boundaries, and generates a visual coupling residual based on their relative displacement, thus recording the overall head movement and the local changes caused by the wearable effect separately.

[0027] A visually coupled residual model, a stable identity anchor, and a dual-domain Mamba-2 model are coupled in the input of the face sequence to be identified. The residual coding layer generates deformation response, normalized domain response, and wearing domain response. The normalized domain and wearing domain are scanned temporally. After intrinsic decomposition, orthogonal reconstruction, coefficient stabilization, and coupling consistency verification, the deformation features are transformed using orthogonal bases and deformation coefficients to generate candidate identity features with the same conditions. The dual-domain verification layer generates identity consistency, deformation consistency, and conflict quantities, and forms causal verification results through zeroing, reversal, and substitution interventions. The joint uncertainty decision layer outputs identification results, unknown person results, or re-acquisition instructions based on the candidate identity distribution and the discrete state of the intervention changes. To facilitate the evaluation of the implementation effect of this invention, all methods are compared using the same acquisition equipment, verification samples, and output caliber. The comparison values ​​are shown in Table 1.

[0028] Table 1. Comparison of Face Recognition Performance under Daily Wear Conditions (Same Standard)

[0029] As shown in Table 1, the registration identity recognition accuracy of the method of this invention is 96.5%, which is 3.3 percentage points higher than the ArcFace occlusion enhancement method and 7.1 percentage points higher than the FaceNet method. The detection rate of unknown persons reaches 95.1%, and the false acceptance rate is reduced to 1.2%. Compared with the ArcFace occlusion enhancement method, the false acceptance rate of this invention is reduced by 1.3 percentage points, a relative reduction of approximately 52.0%, indicating that causal verification and joint uncertainty decision-making can reduce the direct acceptance of highly similar candidate identities. The accuracy of re-collection and judgment reaches 92.1%, which is 8.8 percentage points higher than the ArcFace occlusion enhancement method, and can retain the uncertain state when there are insufficient anchor points, missing intervention data, or multiple candidate identities pass the verification at the same time.

[0030] The cross-frame feature fluctuation of the same person in this invention is 0.064, which is about 40.7% lower than that of the ArcFace occlusion enhancement method. This indicates that the wearer-skin coupling map and stable identity anchor points can reduce the interference of mask boundaries, glasses reflections, hat brim occlusion, and headphone bracket displacement on identity representation. The dual-domain Mamba-2 model forms the normalized identity state and the wearing identity state, respectively. The orthogonal basis and deformation coefficients transform the candidate identity features to the current wearing conditions, improving the matching consistency under daily occlusion conditions. The average single-sequence processing latency is 114.3 milliseconds, which is less than the 2 to 3 second acquisition cycle and can meet the engineering processing requirements of office building access control.

[0031] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A face recognition method based on deep neural networks, characterized in that, include: Acquire and preprocess facial data to generate a sequence of faces to be identified and a registered facial template; Identify head and face objects and construct an object-skin coupling map based on their contours, skin boundaries, and cross-frame displacement. Stable identity anchors are selected based on the device-skin coupling map, and visual coupling residuals are generated based on the cross-frame relative position changes between the stable identity anchors and the device boundary. Visual coupling residuals, stable identity anchors, and the sequence of faces to be identified are input into the residual coding layer of a dual-domain Mamba-2 model to generate deformation features, normalized identity states, and wearing identity states. The deformation features are subjected to intrinsic decomposition to obtain orthogonal basis and deformation coefficients. Orthogonal reconstruction, coefficient stabilization and coupling consistency verification are performed in sequence. Candidate identities are retrieved from the registered face templates based on stable identity anchors. The candidate identity features are transformed in the feature space using the verified orthogonal basis and deformation coefficients to generate candidate same-condition features. Input the candidate same-condition features, standard identity status and wearing identity status into the dual-domain verification layer to generate identity consistency quantity, deformation consistency quantity and conflict quantity. Use the deformation coefficient as the intervention variable to perform zeroing, reversal and substitution interventions. Compare the changes of the three quantities before and after the intervention, and generate causal verification results according to the relationship between deformation consistency quantity and identity consistency quantity in the same direction and conflict quantity in the opposite direction. The causal verification results are input into the joint uncertainty decision layer to generate identity uncertainty and deformation uncertainty, and output the results of identity identification, unknown personnel and re-collection instructions.

2. The face recognition method based on a deep neural network according to claim 1, characterized in that, The generation of the face sequence to be identified and the registered face template includes: Continuous face images are acquired through a face capture device, and registered face images are read from the personnel registration database. Face region localization and quality screening are performed on the continuous face images and registered face images to obtain valid face frames and valid registered face images. Inter-frame association is performed on adjacent valid face frames, and face alignment, image correction and size unification are performed on each associated valid face frame and valid registered face image. The processed valid face frames are arranged according to the acquisition order to generate a face sequence to be identified. The face features of the processed valid registered face images are fused according to the registered identity to generate a registered face template.

3. The face recognition method based on a deep neural network according to claim 1, characterized in that, The construction of the wearable-skin coupling map includes: The device is used to identify the head and face accessories in each face frame of the face sequence to determine the accessory area and accessory type. Extract the outline of the worn object from the worn object area, extract the skin boundary from the junction of the worn object area and the visible skin area, and establish the boundary correspondence between the worn object outline and the skin boundary; Track the outline of the object and the skin boundary in adjacent face frames according to the boundary correspondence, and determine the cross-frame displacement of the object outline and the cross-frame displacement of the skin boundary respectively. Using the wearable area, wearable outline, skin boundary, and adjacent skin area as graph nodes, graph connections are established based on spatial adjacency relationships and cross-frame displacement correspondences to construct a wearable-skin coupling graph.

4. The face recognition method based on a deep neural network according to claim 1, characterized in that, The generation of the visual coupling residual includes: Candidate identity regions are screened from the wearer-skin coupling map and divided into local image blocks. Anchor point stabilization states are generated based on the cross-frame visibility ratio, continuous tracking length, texture changes, and positional fluctuations of each local image block. Local image blocks that directly overlap with the wearer boundary, have texture changes exceeding the stable range, or have interrupted positional continuity are removed. Determine representative locations and local identity features from the preserved local image patches, and combine the representative locations, local identity features, and corresponding cross-frame connections into stable identity anchors; Based on the positional changes of stable identity anchor points and the boundary of the worn object in adjacent face frames, anchor point displacement sequences and boundary displacement sequences are generated respectively. A relative displacement reference is generated based on the anchor point displacement sequence and the boundary displacement sequence. The relative displacement in adjacent face frames is compared with the relative displacement reference to obtain the cross-frame positional deviation and orientation deviation, and a visual coupling residual is generated.

5. The face recognition method based on a deep neural network according to claim 1, characterized in that, The generation of the deformation features, standardized identity status, and wearing identity status includes: Based on visual coupling residuals, stable identity anchors, and the sequence of faces to be identified, an anchor-aligned input group is generated. Extract face features, anchor position features, local identity features and visual coupling residuals from the anchor-aligned input group, map each feature to a token of the same dimension, and write it into time encoding, position encoding and data type encoding to generate a residual encoding sequence; The residual coding sequence is input into the residual coding layer of the coupled dual-domain Mamba-2 model, which includes a residual coding layer, a dual-domain verification layer, and a joint uncertainty decision layer. In the residual coding layer, a coupling gate is generated based on the visual coupling residual. The coupling gate is used to decompose the face feature response into deformation response, normalization domain response and wearing domain response. Deformation features are generated based on the deformation response corresponding to each effective anchor point. Perform temporal state scanning on the normative domain response and the wearing domain response respectively, and aggregate the valid temporal states to generate the normative identity state and the wearing identity state.

6. The face recognition method based on a deep neural network according to claim 1, characterized in that, The generation of the candidate features with the same conditions includes: Perform symmetric matrix eigenvalue decomposition on the deformation features to form an orthogonal basis, and project the effective deformation response that forms the deformation features onto the orthogonal basis to generate deformation coefficients; The effective deformation response is reconstructed based on the orthogonal basis and deformation coefficients, and the reconstruction error is generated. Coefficient stability verification is performed on the deformation coefficients that pass the orthogonal reconstruction verification. Compare the changes in deformation coefficients that have passed the coefficient stability check with the changes in visual coupling residuals, perform coupling consistency check, and retain the orthogonal basis and deformation coefficients that have passed the orthogonal reconstruction check, coefficient stability check and coupling consistency check in sequence; Anchor point query features are generated based on stable identity anchor points. The anchor point query features are matched with the registered face templates for similarity. Candidate identities are filtered based on the matching results, and the corresponding candidate identity features are read. The candidate identity features are transformed into a feature space consistent with the orthogonal basis. Deformation components are generated based on the verified orthogonal basis and deformation coefficients. The deformation components are then superimposed on the candidate identity features to generate candidate features with the same conditions.

7. The face recognition method based on a deep neural network according to claim 1, characterized in that, The generation of the causality check result includes: The candidate features with the same conditions, the standardized identity status, and the wearing identity status are input into the dual-domain verification layer, transformed into a unified feature space, and a dual-domain verification feature group is generated. Based on the candidate identical condition features and the standard identity status and wearing identity status, generate identity consistency quantity, deformation consistency quantity and conflict quantity; Using the deformation coefficient as the intervention variable, zeroing intervention, reverse intervention, and alternative intervention were performed respectively to generate zeroing intervention features, reverse intervention features, and alternative intervention features; Based on the zero-intervention feature, reverse intervention feature, and alternative intervention feature, the identity consistency quantity, deformation consistency quantity, and conflict quantity corresponding to various interventions are regenerated, and the changes of the three quantities before and after the intervention are calculated. When the change in deformation consistency quantity is in the same direction as the change in identity consistency quantity and in the opposite direction to the change in conflict quantity, the corresponding intervention is identified as a causal consistency intervention, and a causal verification result is generated based on the number of causal consistency interventions and the magnitude of the changes in the three quantities.

8. The face recognition method based on a deep neural network according to claim 1, characterized in that, The generation of the identification result, unknown person result, and re-collection instruction includes: The causal verification results are input into the joint uncertainty decision layer. Candidate decision data are generated based on the causal verification score, the number of causal consistent interventions, changes in the three quantities, and the causal verification marker. Candidate identity distribution and identity uncertainty are generated based on the causal verification score. Based on the causal consistency status and the magnitude of change of the three quantities corresponding to various interventions, the intervention consistency ratio and the change discrete quantity are generated, and the deformation uncertainty is generated based on the intervention consistency ratio and the change discrete quantity. Based on the candidate identity with the highest causal verification score, identity uncertainty, deformation uncertainty, and corresponding causal verification marker, a joint determination is performed, and the results of identity recognition, unknown personnel identification, and re-collection instructions are output.

Images