An intelligent home face recognition method based on a graph neural network

Abstract

The application discloses a kind of intelligent house face recognition methods based on graph neural network, comprising the following steps: S1, cutting face segment to be identified;S2, constitute face structure diagram;S3, face structure diagram is input to improved DGCNN;S4, initial face dynamic graph is constructed by dynamic graph construction layer;S5, EdgeConv side convolution layer executes edge-by-edge convolution operation, forms edge-by-edge convolution response, and merges and receives benchmark response;S6, call side residual recursive correction mechanism, and the feature of candidate neighborhood edge state table is updated to the center structure point by incorporating residual error into center structure point;S7, dynamic graph update layer reduces candidate neighborhood edge, and forms recursive correction face dynamic graph;S8, global aggregation is executed to recursive correction face dynamic graph, and intelligent house face recognition result is formed.The application reduces shielding and false detection interference by side residual recursive correction face dynamic graph, and improves intelligent house face recognition stability.

Description

Technical Field

[0001] This invention relates to the field of smart home identity recognition technology, and in particular to a smart home face recognition method based on graph neural networks. Background Technology

[0002] With the widespread adoption of smart locks, indoor monitoring devices, and home security terminals, facial recognition is gradually becoming an important method for identity verification in smart homes. Existing smart home facial recognition methods mostly rely on single-frame image texture features or ordinary deep learning models for identity matching. While they achieve good results under conditions of frontal faces, stable lighting, and no occlusion, in real-world home scenarios, the captured images are often affected by low light at night, backlighting at doorways, rapid passage, side profiles, mask or hat occlusion, and camera installation angle deviations. This leads to missing facial areas, key point offsets, and unstable contour connections. Ordinary convolutional models primarily extract image textures, making it difficult to explicitly express the spatial relationships between the eyes, nose, mouth, and facial contours. Although ordinary DGCNN can construct dynamic maps based on facial structure points, its EdgeConv edge convolutional layers typically aggregate abnormal neighborhood edge responses directly into the central structure point features. During dynamic map updates, it is easy to repeatedly introduce deviated connections caused by occlusion or false detections, resulting in unstable identity features, misidentification of strangers, and false door lock triggering.

[0003] Therefore, how to provide a smart home facial recognition method based on graph neural networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a smart home face recognition method based on graph neural networks. This invention constructs a facial structure graph by using facial structural points and initial neighborhood relationships, shifting the smart home face recognition process from simple texture matching to modeling facial spatial relationships. Even in low-light, side-view, and partially occluded scenarios, it can still preserve stable structural information between the eyes, nose, mouth, and contour regions. Simultaneously, an edge residual recursive correction mechanism is set between the EdgeConv edge convolutional layer and the dynamic graph update layer of the improved DGCNN, utilizing a stable set of connecting edges. The baseline response is merged, and the difference between the edge-by-edge convolution response and the baseline response is divided into edge response residuals. The baseline residual and the deviation residual are further decomposed, so that the baseline residual strengthens the update feature of the central structure point, and the deviation residual is written into the candidate neighborhood edge state table and drives the reduction of abnormal candidate neighborhood edges. Thus, the recursive correction of the facial dynamic map can retain continuous and stable facial baseline edges layer by layer, suppress abnormal connections caused by occlusion, false detection and pose shift, reduce the risk of stranger misidentification and door lock false triggering, and improve the stability and reliability of face recognition results in the complex acquisition environment of smart homes.

[0005] A smart home face recognition method based on graph neural networks according to an embodiment of the present invention includes the following steps: S1. The smart home edge node collects facial image streams within the target home area and crops out the facial segments to be identified; S2. Locate facial structure points in the face segment to be identified, organize the initial neighborhood relationship according to the spatial connection relationship between the facial structure points, and form a facial structure map. S3. Input the facial structure map into the improved DGCNN. The improved DGCNN includes a dynamic graph construction layer, an EdgeConv edge convolutional layer and a dynamic graph update layer. An edge residual recursive correction mechanism is set between the EdgeConv edge convolutional layer and the dynamic graph update layer, and a candidate neighborhood edge state table is set in the dynamic graph update layer. S4. The dynamic graph construction layer screens the neighborhood structure points corresponding to the central structure point according to the initial neighborhood relationship, and includes the candidate neighborhood edges that maintain the facial space connection into the stable connection edge set to construct the initial facial dynamic graph. S5, EdgeConv edge convolutional layers perform edge-by-edge convolution operations along the initial facial dynamic map to form edge-by-edge convolutional responses, and merge the baseline responses based on the stable set of receiving edges; S6, EdgeConv edge convolutional layer calls the edge residual recursive correction mechanism, divides the difference between the edge convolution response and the baseline response into edge response residual, decomposes the baseline residual and deviation residual, incorporates the baseline residual into the center structure point update feature, and writes the deviation residual into the candidate neighborhood edge state table. S7. The dynamic graph update layer reduces candidate neighborhood edges that deviate from the residual hit according to the candidate neighborhood edge state table, and retains candidate neighborhood edges that continuously hit the residual to form a recursive corrected face dynamic graph. S8. Perform global aggregation on the recursive modified facial dynamic map to form facial identity features, and map the facial identity features to the smart home face recognition result.

[0006] Optionally, S1 specifically includes: S11. The smart home edge node performs frame-by-frame face region localization on the face image stream according to the acquisition time sequence, and classifies face regions with overlapping contours and continuous positions in adjacent frames into the same face candidate segment. S12. Perform boundary clipping and facial integrity checks on the face region within the face candidate segment, and remove candidate frames with truncated contours, faces off-center from the acquisition area, and broken consecutive frames. S13. Extract the retained candidate frames into face segments to be identified according to the acquisition time sequence.

[0007] Optionally, S2 specifically includes: S21. Locate the boundaries of key facial regions and facial contours in the face segment to be identified, and merge the intersection points, local turning points and center points of the regions into facial structure points. S22. Configure the structure point position record according to the spatial position of the facial structure points in the face segment to be identified, and connect adjacent facial structure points in the same facial region as local neighborhood edges. S23. Connect adjacent facial structural points across facial regions as receiving neighboring edges, connect corresponding facial structural points on the left and right as symmetrical neighboring edges, and remove connecting edges that deviate from the facial spatial order. S24. The facial structure graph is constructed by combining facial structure points, local neighborhood edges, receiving neighborhood edges, and symmetrical neighborhood edges, and the local neighborhood edges, receiving neighborhood edges, and symmetrical neighborhood edges are used as the initial neighborhood relations.

[0008] Optionally, S4 specifically includes: S41. The dynamic graph construction layer sets the facial structure points in the facial structure graph as the center structure points one by one according to the spatial arrangement order, and extracts the connection edges directly connected to the current center structure point from the initial neighborhood relationship to form the candidate neighborhood edge set corresponding to the current center structure point. S42. Locate the neighborhood structure points along the candidate neighborhood edge set one by one, and perform edge-end attribution verification on the facial structure points at both ends of the candidate neighborhood edge. Retain the candidate neighborhood edges whose edges still belong to the adjacent facial region, symmetrical facial region, or contour-bound region. S43. Perform a connection direction check on the retained candidate neighborhood edges. Push the candidate neighborhood edges whose connection direction is consistent with the spatial connection direction in the initial neighborhood relationship into the stable connection edge set. Retain the candidate neighborhood edges whose connection direction has shifted but have not deviated from the facial structure graph as ordinary candidate edges. S44. Rewrite the candidate neighborhood edges in the stable receiving edge set as the dynamic graph neighborhood edges corresponding to the central structure point, and label the facial structure points at both ends of the dynamic graph neighborhood edges as the central structure point and the neighborhood structure point, respectively. S45. Arrange the ordinary candidate edges after the stable receiving edge set to form the candidate neighborhood supplementary edge set corresponding to the current central structure point; S46. Based on the central structural point, encapsulate the set of neighborhood edges and candidate neighborhood supplementary edges of the dynamic graph into local dynamic graph units, and connect each local dynamic graph unit in series according to the spatial connection relationship between facial structural points to construct the initial facial dynamic graph.

[0009] Optionally, S5 specifically includes: S51, EdgeConv convolutional layer receives the initial facial dynamic map, expands the dynamic map neighborhood edges in each local dynamic map unit one by one, and pairs the center structure point and neighborhood structure point at both ends of the dynamic map neighborhood edge as edge end pairs. S52. Perform edge feature splitting on each edge pair, place the central structure point feature at the central end, and place the neighboring structure point feature at the neighboring end. S53. Perform a dimension-wise subtraction between the neighborhood end features and the center end features to obtain the neighborhood offset features, and then place the neighborhood offset features and the center end features in the same position to form the edge convolution input of the corresponding dynamic graph neighborhood edge. S54. Perform edge-by-edge mapping on the edge convolution input, mapping the edge convolution input corresponding to each neighborhood edge of the dynamic graph to an edge-by-edge convolution response, and retaining the association relationship between the edge convolution response and the neighborhood edge of the dynamic graph after mapping. S55. Compare the edge-by-edge convolutional responses belonging to the same central structural point dimension by dimension, and fill the maximum response value in each dimension back into the feature position of the central structural point to form the updated feature of the central structural point. S56. Extract the corresponding edge-by-edge convolutional response from the stable receiving edge set, and perform dimension-by-dimensional accumulation and dimension-by-dimensional averaging on the extracted edge-by-edge convolutional response according to the center structure point to form the receiving benchmark response corresponding to the current center structure point. S57. The edge-by-edge convolutional response, the central structure point update feature, and the supporting baseline response are split and retained at the output of the EdgeConv edge convolutional layer.

[0010] Optionally, S6 specifically includes: S61, EdgeConv edge convolutional layer calls the edge residual recursive correction mechanism, and according to the belonging relationship between the edge convolutional response and the candidate neighborhood edge, the edge convolutional response is back-pasted to the corresponding candidate neighborhood edge. S62. Perform dimension-wise subtraction between the edge-by-edge convolutional response and the baseline response under the same central structural point to form the edge response residuals corresponding to each candidate neighborhood edge. S63. According to the spatial connection direction of the candidate neighborhood edge in the initial neighborhood relationship, perform direction verification on the edge response residual and classify the residual component that continues along the spatial connection direction into the connection residual. S64. Residual components that deviate from the spatial bearing direction, cross non-bearing face regions, or coincide with the reduction states in the candidate neighborhood edge state table are classified as deviation residuals. S65. The inherited residuals are incorporated into the updated features of the central structure points according to their central structure point affiliation, and the affiliation relationship between the inherited residuals and the corresponding candidate neighborhood edges is preserved. S66. Remove the deviation residual from the center structure point update feature and write it into the candidate neighborhood edge state table according to the candidate neighborhood edge position. S67. The candidate neighborhood edge state table performs inter-layer accumulation on the deviation residuals of the same candidate neighborhood edge, performs inter-layer retention on the receiving residuals of the same candidate neighborhood edge, and rewrites the candidate neighborhood edges with continuously accumulated deviation residuals to the reduction state, and rewrites the candidate neighborhood edges with continuously retained receiving residuals to the retention state.

[0011] Optionally, S7 specifically includes: S71. The dynamic graph update layer reads the candidate neighborhood edge state table, assigns candidate neighborhood edges written to the reduction state to the reduction edge set, assigns candidate neighborhood edges written to the retention state to the retention edge set, and assigns candidate neighborhood edges not written to the reduction state or retention state to the edge set to be checked. S72. Perform edge connection pruning on the cut edge set, remove the candidate neighborhood edges in the cut edge set from the candidate neighborhood range constructed by the next layer dynamic graph, and simultaneously mask the deviation residuals corresponding to the cut edge set. S73. Perform edge connection solidification on the retained edge set, and write the candidate neighborhood edges and the structural points at both ends in the retained edge set into the retained neighborhood range constructed by the next layer of dynamic graph; S74. Perform inter-layer state comparison on the set of edges to be checked. Transfer the edges to be checked that deviate from the residuals written in the current graph convolution stage to the set of edges to be reduced. Transfer the edges to be checked that accept the residuals written in the convolution stages of adjacent graphs to the set of edges to be retained. S75. Rearrange the candidate neighborhood edges in the retained edge set according to the affiliation of the central structural point, and perform alignment between the rearranged candidate neighborhood edges and the unpruned initial neighborhood relationships. S76. After the candidate neighborhood edges are aligned, they are repackaged into local dynamic graph units, and each local dynamic graph unit is connected in series according to the spatial connection relationship between facial structure points to form a recursive corrected facial dynamic graph.

[0012] Optionally, S8 specifically includes: S81. Receive the recursive correction of the facial dynamic map, include the updated features of the center structure points associated with the edge set in the aggregation range, and exclude the deviation residuals associated with the reduced edge set from the aggregation range. S82. Based on the spatial arrangement of the central structural points in the facial structure map, the update features of the central structural points included in the aggregation range are sequentially aggregated to form a sequence of facial structural point update features. S83. Perform a dimension-wise comparison on the updated feature sequence of facial structure points, extract the maximum response of each dimension, and perform a dimension-wise accumulation and averaging on the updated feature sequence of facial structure points to extract the average response of each dimension. S84. The maximum response and average response of each dimension are placed side by side to form a global facial structure response; S85. Encapsulate the global facial structure response with the candidate neighborhood edge attribution relationship in the retained edge set to form facial identity features; S86. Match the facial identity features with the identity records in the smart home identity database one by one to obtain the identity matching response between the face to be identified and each identity record. S87. Sort the numerical values ​​of each identity matching response and write the sorting results into the smart home face recognition results.

[0013] The beneficial effects of this invention are: 1. This invention organizes the facial fragments to be identified into a facial structure map and constructs initial neighborhood relationships according to the spatial connections between facial structural points. This allows the facial recognition process to no longer rely solely on single-frame texture features, but incorporates the structural connections between the eyes, nose, mouth, and facial contours into the recognition chain. In low-light, side-view, fast-moving, and partially occluded scenarios, facial textures are easily lost or blurred, but the stable connections between facial structural points can still provide a structural basis for identity recognition, thereby improving the stability of facial feature extraction in smart home scenarios.

[0014] 2. This invention improves DGCNN by setting up an edge residual recursive correction mechanism. It stabilizes the set of receiving edges and merges them into a baseline response. The edge response residual is then formed by differentiating the edge-by-edge convolutional response from the baseline response, and further decomposed into receiving residuals and deviation residuals. The receiving residual is incorporated into the central structure point update features, while the deviation residual is written into the candidate neighborhood edge state table. This prevents abnormal neighborhood edges formed by occlusion points, false detection points, and side profile offset points from continuously polluting the central structure point update features, thereby improving the anti-interference capability of facial identity features.

[0015] 3. This invention records deviation and continuation residuals across layers using a candidate neighborhood edge state table. In the dynamic graph update layer, it reduces candidate neighborhood edges that match deviation residuals while retaining those that continuously match continuation residuals. This allows the recursive correction of the facial dynamic graph to converge layer by layer towards a stable facial spatial continuation relationship. Compared to ordinary DGCNN, which is prone to repeatedly introducing anomalous edges during each layer's reconstruction, this invention reduces the risks of false stranger identification, occlusion-related false identification, and false door lock triggering, thus improving the reliability of smart home facial recognition results. Attached Figure Description

[0016] 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 a flowchart of a smart home face recognition method based on graph neural networks proposed in this invention; Figure 2This is a schematic diagram of the improved DGCNN model structure of a smart home face recognition method based on graph neural networks proposed in this invention. Figure 3 This is a schematic diagram illustrating the edge residual recursive correction mechanism of a smart home face recognition method based on graph neural networks proposed in this invention. Detailed Implementation

[0017] 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.

[0018] refer to Figures 1-3 A smart home face recognition method based on graph neural networks includes the following steps: S1. The smart home edge node collects facial image streams within the target home area and crops out the facial segments to be identified; S2. Locate facial structure points in the face segment to be identified, organize the initial neighborhood relationship according to the spatial connection relationship between the facial structure points, and form a facial structure map. S3. Input the facial structure map into the improved DGCNN. The improved DGCNN includes a dynamic graph construction layer, an EdgeConv edge convolutional layer and a dynamic graph update layer. An edge residual recursive correction mechanism is set between the EdgeConv edge convolutional layer and the dynamic graph update layer, and a candidate neighborhood edge state table is set in the dynamic graph update layer. S4. The dynamic graph construction layer screens the neighborhood structure points corresponding to the central structure point according to the initial neighborhood relationship, and includes the candidate neighborhood edges that maintain the facial space connection into the stable connection edge set to construct the initial facial dynamic graph. S5, EdgeConv edge convolutional layers perform edge-by-edge convolution operations along the initial facial dynamic map to form edge-by-edge convolutional responses, and merge the baseline responses based on the stable set of receiving edges; S6, EdgeConv edge convolutional layer calls the edge residual recursive correction mechanism, divides the difference between the edge convolution response and the baseline response into edge response residual, decomposes the baseline residual and deviation residual, incorporates the baseline residual into the center structure point update feature, and writes the deviation residual into the candidate neighborhood edge state table. S7. The dynamic graph update layer reduces candidate neighborhood edges that deviate from the residual hit according to the candidate neighborhood edge state table, and retains candidate neighborhood edges that continuously hit the residual to form a recursive corrected face dynamic graph. S8. Perform global aggregation on the recursive modified facial dynamic map to form facial identity features, and map the facial identity features to the smart home face recognition result.

[0019] In its implementation, this improved DGCNN builds upon the dynamic graph construction and EdgeConv edge convolution of ordinary DGCNN by introducing an edge residual recursive correction mechanism and a candidate neighborhood edge state table. Ordinary DGCNN typically reconstructs neighborhood relationships based on point features at each graph convolution stage and aggregates neighborhood edge responses into center point features. This invention, after obtaining the edge-by-edge convolution response at the EdgeConv edge convolution layer, retains the attribution relationship between the edge-by-edge convolution response and candidate neighborhood edges, and uses a stable set of receiving edges to merge and obtain the receiving baseline response. The edge residual recursive correction mechanism differs between the edge-by-edge convolution response and the receiving baseline response to obtain the edge response residual. Then, based on the spatial receiving direction of the candidate neighborhood edges in the initial neighborhood relationship, it splits the receiving residual and the deviation residual. The candidate neighborhood edge state table is used to record the receiving residual and deviation residual states of candidate neighborhood edges in the convolution stage of adjacent graphs. The dynamic graph update layer reduces candidate neighborhood edges that continuously hit the deviation residual according to the state table, and retains candidate neighborhood edges that continuously hit the receiving residual, thereby forming a recursive corrected face dynamic graph.

[0020] In this embodiment, S1 specifically includes: S11. The smart home edge node performs frame-by-frame face region localization on the face image stream according to the acquisition time sequence, and classifies face regions with overlapping contours and continuous positions in adjacent frames into the same face candidate segment. S12. Perform boundary clipping and facial integrity checks on the face region within the face candidate segment, and remove candidate frames with truncated contours, faces off-center from the acquisition area, and broken consecutive frames. S13. Extract the retained candidate frames into face segments to be identified according to the acquisition time sequence.

[0021] In the specific implementation process, the smart home edge node first acquires a continuous stream of facial images from cameras within the target home area. This image stream can originate from doorway cameras, indoor security cameras, or smart lock cameras. To avoid single-frame images being directly processed due to low light, head turning, or motion blur, this step does not directly select a single image as the recognition object. Instead, it performs frame-by-frame facial region localization on the image stream according to the acquisition sequence. Each located facial region in each frame is configured with a corresponding facial contour range and frame sequence position, facilitating continuous verification of facial regions in adjacent frames.

[0022] After frame-by-frame localization, the face regions in the current frame are compared with the face regions in the previous frame or several adjacent frames in terms of contour overlap. If the contour ranges of two face regions overlap, and the direction of movement of the face center position in adjacent frames is continuous and the magnitude of movement does not change abruptly, then the two face regions are grouped into the same face candidate segment. If the position of the face region in adjacent frames suddenly changes, the contour overlap is interrupted, or multiple discontinuous face regions appear at the same time, then different face candidate segments are created separately to avoid erroneously merging face regions of different people or different time periods.

[0023] For the generated candidate face segments, boundary cropping and facial integrity checks are also required. Boundary cropping checks primarily inspect whether the face region extends beyond the captured image boundary. If the face contour is truncated by the image edge, or if areas such as the forehead, jawline, or cheeks are significantly missing, the corresponding candidate frame will not be retained. Facial integrity checks primarily inspect whether the main facial area is within the recognizable range. If the face deviates from the captured area, the continuity of the facial contour is disrupted, or there are breaks in the face region between adjacent frames, the corresponding candidate frame will be removed. Candidate frames with slight localized missing areas due to brief head turning but still continuous facial contours can be retained within the segment for further filtering using facial structure point localization and dynamic image correction.

[0024] After candidate frame removal, the remaining candidate frames are cropped into face segments to be identified according to the acquisition sequence. During cropping, the facial contour range in the retained candidate frames is used as a base, with a small margin added outwards to ensure that the areas required for locating facial structure points such as the eyes, nose, mouth, and facial contours are fully included in the cropping range. The margin can be determined proportionally based on the width or height of the face region; for example, 5% to 10% of the face region width can be used as the horizontal expansion range, and 5% to 10% of the face region height as the vertical expansion range. When the expanded area exceeds the image boundary, it is truncated according to the image boundary.

[0025] In this embodiment, S2 specifically includes: S21. Locate the boundaries of key facial regions and facial contours in the face segment to be identified, and merge the intersection points, local turning points and center points of the regions into facial structure points. S22. Configure the structure point position record according to the spatial position of the facial structure points in the face segment to be identified, and connect adjacent facial structure points in the same facial region as local neighborhood edges. S23. Connect adjacent facial structural points across facial regions as receiving neighboring edges, connect corresponding facial structural points on the left and right as symmetrical neighboring edges, and remove connecting edges that deviate from the facial spatial order. S24. The facial structure graph is constructed by combining facial structure points, local neighborhood edges, receiving neighborhood edges, and symmetrical neighborhood edges, and the local neighborhood edges, receiving neighborhood edges, and symmetrical neighborhood edges are used as the initial neighborhood relations.

[0026] In the specific implementation process, when locating facial structural points for the face segment to be identified, the key regions of the face in the retained candidate frames are first segmented to determine the boundary ranges of the eye region, nose region, mouth region, eyebrow region, and facial contour region. The boundaries of the key facial regions can be jointly determined by the facial key point detection model, edge detection results, or facial region segmentation results. When there is slight jitter in the position of the same key region in consecutive frames, it is corrected by the overlapping position of the region boundaries and the trend of the center position change in adjacent frames to avoid the facial structural points jumping due to single-frame positioning deviation.

[0027] After determining the boundaries of key facial regions and facial contours, the intersections of these boundaries, local turning points, and regional center points are grouped into facial structural points. For example, facial structural points include locations that reflect the spatial relationships of the face, such as the corners of the eyes, corners of the mouth, jawline turning points, and the center of the nose.

[0028] After merging facial structure points, the structure point location records are configured according to their spatial positions within the face segment to be identified. Each structure point location record includes at least its planar position within the current face segment, its associated facial region, and its order relative to the facial contour. For the position of the same facial structure point in consecutive frames, consistency can be checked according to the acquisition sequence. When a structure point is missing in a single frame but consistently appears in adjacent frames, its structural affiliation within the segment is retained, and the corresponding structure point is not directly deleted, ensuring that the facial structure map remains essentially complete within the time segment.

[0029] When establishing initial neighborhood relationships, adjacent facial structural points within the same facial region are first linked as local neighborhood edges. For example, local neighborhood edges are established between adjacent structural points within the same eye region, between adjacent structural points within the same mouth contour region, and between adjacent structural points within the same face contour segment. Local neighborhood edges mainly preserve the continuous structure within the local region, enabling the dynamic graph construction layer to expand neighborhood filtering along the local facial structure.

[0030] For facial structural points that are adjacent across facial regions, points with spatially connected relationships are linked as connecting neighborhood edges. For example, adjacent structural points between the eye and nose regions, the nose and mouth regions, and the cheek and jawline contours can form connecting neighborhood edges. Connecting neighborhood edges emphasize the structural connections between key facial regions, ensuring that the facial structure diagram not only includes connections within a single region but also preserves the spatial connection chains from the eyes, nose, and mouth to the contour regions. For facial structural points that correspond to each other on the left and right sides, corresponding points are linked as symmetrical neighborhood edges. For example, symmetrical neighborhood edges are established between the left and right corners of the eyes, the left and right eyebrow transition points, the left and right cheek structural points, and the left and right jawline contour points to preserve the correspondence between the structures on both sides of the face.

[0031] After connecting local, adjacent, and symmetrical neighborhood edges, the spatial order of all connected edges is checked against facial contours. If a connected edge crosses multiple non-adjacent facial regions, its connection direction significantly conflicts with the facial contour direction, or the structural points at both ends of the connection do not have a stable spatial correspondence in consecutive frames, then that connected edge is discarded. This process avoids erroneous connections deviating from the true facial structure caused by local occlusion, side-face pose, or false detection of key points.

[0032] Finally, the preserved facial structure points are used as graph nodes, and local neighborhood edges, connecting neighborhood edges, and symmetrical neighborhood edges are used as graph edges to construct the facial structure graph. Local neighborhood edges, connecting neighborhood edges, and symmetrical neighborhood edges are used together as initial neighborhood relations to enter the improved DGCNN, enabling the dynamic graph construction layer to screen the neighborhood structure points corresponding to the central structure point under the facial spatial connecting constraints.

[0033] In this embodiment, S4 specifically includes: S41. The dynamic graph construction layer sets the facial structure points in the facial structure graph as the center structure points one by one according to the spatial arrangement order, and extracts the connection edges directly connected to the current center structure point from the initial neighborhood relationship to form the candidate neighborhood edge set corresponding to the current center structure point. S42. Locate the neighborhood structure points along the candidate neighborhood edge set one by one, and perform edge-end attribution verification on the facial structure points at both ends of the candidate neighborhood edge. Retain the candidate neighborhood edges whose edges still belong to the adjacent facial region, symmetrical facial region, or contour-bound region. S43. Perform a connection direction check on the retained candidate neighborhood edges. Push the candidate neighborhood edges whose connection direction is consistent with the spatial connection direction in the initial neighborhood relationship into the stable connection edge set. Retain the candidate neighborhood edges whose connection direction has shifted but have not deviated from the facial structure graph as ordinary candidate edges. S44. Rewrite the candidate neighborhood edges in the stable receiving edge set as the dynamic graph neighborhood edges corresponding to the central structure point, and label the facial structure points at both ends of the dynamic graph neighborhood edges as the central structure point and the neighborhood structure point, respectively. S45. Arrange the ordinary candidate edges after the stable receiving edge set to form the candidate neighborhood supplementary edge set corresponding to the current central structure point; S46. Based on the central structural point, encapsulate the set of neighborhood edges and candidate neighborhood supplementary edges of the dynamic graph into local dynamic graph units, and connect each local dynamic graph unit in series according to the spatial connection relationship between facial structural points to construct the initial facial dynamic graph.

[0034] In its implementation, the dynamic graph construction layer does not directly select neighboring points freely among all facial structural points using the K-nearest neighbor method in ordinary DGCNN. Instead, it first reads the already formed facial structural map and initial neighborhood relationships. Ordinary DGCNN typically treats the input point set as an unordered set and re-searches for neighbors for each point based on coordinate distance or feature distance. This step improves the dynamic graph construction method by using the existing spatial connections between facial structural points as constraints for dynamic graph construction. This ensures that the neighborhood edges of the dynamic graph preferentially come from the actual facial structural order, rather than relying entirely on free nearest neighbor search.

[0035] The dynamic graph construction layer first assigns each facial structural point in the facial structure map as a central structural point according to its spatial arrangement. This spatial arrangement can follow a top-to-bottom, mid-axis-to-side, or local-to-contour pattern, as long as the relative connections between the facial structural points are maintained. For each central structural point, the dynamic graph construction layer extracts connecting edges directly attached to it from the initial neighborhood relationships, forming a candidate neighborhood edge set. This candidate neighborhood edge set is limited to the initial neighborhood relationships comprised of local, connecting, and symmetrical neighborhood edges. This avoids the common DGCNN approach of creating incorrect neighborhood edges between points with similar feature distances but not structurally adjacent facial features.

[0036] Subsequently, neighborhood structural points are located edge by edge along the candidate neighborhood edge set, and edge-end attribution checks are performed on the facial structural points at both ends of the candidate neighborhood edges. Edge-end attribution checks mainly determine whether the two ends of the candidate neighborhood edge still belong to adjacent facial regions, symmetrical facial regions, or contour-connecting regions. For example, adjacent structural points within the same eye region can be retained as local connections, structural points that are consecutive between the eye and nose can be retained as cross-regional connections, and corresponding structural points on the left and right cheeks or the left and right corners of the eyes can be retained as symmetrical connections. If a candidate neighborhood edge crosses multiple non-connecting facial regions, such as crossing directly from the corner of the mouth to the distal corner of the eye, or crossing from a single contour point to a non-corresponding local point, then the candidate neighborhood edge is discarded and not included in the initial facial animation.

[0037] After completing the edge attribution verification, the remaining candidate neighborhood edges undergo a connection direction verification. This verification compares whether the connection direction of the candidate neighborhood edges aligns with the spatial connection direction in the initial neighborhood relationship. If a candidate neighborhood edge continues along the natural arrangement of the facial structure, such as along the eye boundary, bridge of the nose, mouth contour, or jaw contour, it is pushed into the stable connection edge set. If the connection direction of a candidate neighborhood edge shifts, but the structural points at both ends remain within the facial structure map and do not cross non-connecting facial regions, it is retained as a regular candidate edge. This process ensures that the stable connection edge set provides the source of the baseline response, while regular candidate edges serve as supplementary neighborhoods in the initial facial dynamic map construction. This preserves stable connections in the facial structure without directly excluding edges with slight pose changes.

[0038] In one embodiment, the spatial acceptance direction is determined by the edge type in the initial neighborhood relationship. The spatial acceptance direction of local neighborhood edges is determined according to the arrangement order of structural points within the same facial region; the spatial acceptance direction of accepting neighborhood edges is determined according to the connection order from the structural point of the previous region to the structural point of the next region; the spatial acceptance direction of symmetrical neighborhood edges is determined according to the mirror correspondence between corresponding structural points on the left and right sides. If both ends of a candidate neighborhood edge are still within the connection range defined by the above edge type, and the connection direction does not cross non-accepting facial regions, then the candidate neighborhood edge is determined to maintain facial spatial acceptance. If the connection direction of a candidate neighborhood edge is inconsistent with the above edge type, but the structural points at both ends are still within the facial structure diagram, then the candidate neighborhood edge is retained as a normal candidate edge and does not enter the stable accepting edge set.

[0039] After forming a stable set of receiving edges, the dynamic graph construction layer rewrites the candidate neighborhood edges in the stable receiving edge set as dynamic graph neighborhood edges corresponding to the central structural point, and labels the facial structural points at both ends of the dynamic graph neighborhood edges as the central structural point and the neighborhood structural point, respectively. This rewriting action transforms the candidate neighborhood edges from static connections in the facial structural graph into dynamic graph neighborhood relationships that the EdgeConv edge convolutional layer can unfold edge by edge. Ordinary candidate edges are then arranged after the stable receiving edge set, forming a supplementary set of candidate neighborhood edges corresponding to the current central structural point. This sorting method allows the EdgeConv edge convolutional layer to distinguish between stable receiving edges and ordinary candidate edges when performing edge-by-edge convolution, and the receiving baseline response can also be preferentially merged from the stable receiving edge set.

[0040] Finally, based on the attribution of the central structural point, the neighborhood edges and candidate supplementary neighborhood edges of the dynamic graph are encapsulated into local dynamic graph units. Each local dynamic graph unit corresponds to a central structural point and contains the central structural point, the neighborhood structural points directly connected to it, a stable set of connecting edges, and a set of candidate supplementary neighborhood edges. These local dynamic graph units are then concatenated according to the spatial connection relationships between facial structural points to construct the initial facial dynamic graph. The resulting initial facial dynamic graph retains the basic structure of DGCNN's layer-by-layer dynamic graph learning while introducing facial structural connection constraints in the first-layer dynamic graph construction stage, providing a stable graph structure foundation for the EdgeConv edge convolutional layer to generate edge-by-edge convolutional responses and connection baseline responses. The stable set of connecting edges provides the source of the connection baseline response, while ordinary candidate edges provide reserved space for neighborhood supplementation under pose changes.

[0041] In this embodiment, S5 specifically includes: S51, EdgeConv convolutional layer receives the initial facial dynamic map, expands the dynamic map neighborhood edges in each local dynamic map unit one by one, and pairs the center structure point and neighborhood structure point at both ends of the dynamic map neighborhood edge as edge end pairs. S52. Perform edge feature splitting on each edge pair, place the central structure point feature at the central end, and place the neighboring structure point feature at the neighboring end. S53. Perform a dimension-wise subtraction between the neighborhood end features and the center end features to obtain the neighborhood offset features, and then place the neighborhood offset features and the center end features in the same position to form the edge convolution input of the corresponding dynamic graph neighborhood edge. S54. Perform edge-by-edge mapping on the edge convolution input, mapping the edge convolution input corresponding to each neighborhood edge of the dynamic graph to an edge-by-edge convolution response, and retaining the association relationship between the edge convolution response and the neighborhood edge of the dynamic graph after mapping. S55. Compare the edge-by-edge convolutional responses belonging to the same central structural point dimension by dimension, and fill the maximum response value in each dimension back into the feature position of the central structural point to form the updated feature of the central structural point. S56. Extract the corresponding edge-by-edge convolutional response from the stable receiving edge set, and perform dimension-by-dimensional accumulation and dimension-by-dimensional averaging on the extracted edge-by-edge convolutional response according to the center structure point to form the receiving benchmark response corresponding to the current center structure point. S57. The edge-by-edge convolutional response, the central structure point update feature, and the supporting baseline response are split and retained at the output of the EdgeConv edge convolutional layer.

[0042] In the specific implementation process, after receiving the initial facial animation, the EdgeConv convolutional layer first unfolds the graph structure according to the local animation unit. Each local animation unit corresponds to a central structure point. The animation neighborhood edges within the unit are extracted one by one, and the central structure points and neighborhood structure points at both ends of the animation neighborhood edge are paired as edge pairs.

[0043] For each edge pair, the EdgeConv edge convolutional layer performs edge feature splitting, placing the features of the central structural point at the central end and the features of the neighboring structural points at the neighboring ends. Then, a dimension-wise subtraction is performed between the neighboring end features and the central end features to obtain the neighborhood offset features. These neighborhood offset features reflect the local structural changes of the neighboring structural points relative to the central structural point. Subsequently, the neighborhood offset features and the central end features are placed side-by-side to form the edge convolution input for the corresponding dynamic graph neighborhood edge.

[0044] After constructing the edge convolution input, the EdgeConv edge convolutional layer performs edge-by-edge mapping on the edge convolution input corresponding to each neighborhood edge of the dynamic graph. In one embodiment, the central feature dimension is 64-dimensional, and the neighborhood offset feature dimension is 64-dimensional. These two are arranged side-by-side to form a 128-dimensional edge convolution input. During edge-by-edge mapping, the 128 feature values ​​are read sequentially according to the dimensional order of the edge convolution input. Each feature value is multiplied by the corresponding mapping parameter within the EdgeConv edge convolutional layer, and the 128 multiplication results are accumulated. The accumulated result is then superimposed with the corresponding bias term to obtain the mapping value of the neighborhood edge of the dynamic graph in one response dimension. Mapping is performed on each of the 64 response dimensions using the same method to form a 64-dimensional edge response vector. Nonlinear compression is performed on the values ​​of each dimension in the edge response vector, and the compressed values ​​are limited to the range of -3 to 3, forming the edge-by-edge convolution response. After the edge-by-edge convolutional response is formed, the EdgeConv edge convolutional layer simultaneously retains the attribution relationship between the edge-by-edge convolutional response and the corresponding dynamic graph neighborhood edge. This attribution relationship corresponds to the central structure point, neighborhood structure point and local dynamic graph unit, which serves as the edge-level source for back-pasting the edge-by-edge convolutional response to the corresponding candidate neighborhood edge during the edge residual recursive correction process.

[0045] When forming the updated features of the central structure point, the EdgeConv edge convolutional layer uses the central structure point as the aggregation unit. It compares the edge-by-edge convolutional responses belonging to the same central structure point dimension by dimension, and fills the maximum response value in each dimension back into the feature position of the central structure point to form the updated features of the central structure point. For example, when the same central structure point is connected to multiple neighborhood edges of the dynamic graph, each edge will generate an edge-by-edge convolutional response; after comparing these responses one by one in the same dimension, the maximum response value is taken as the update result for that dimension.

[0046] To generate the baseline response, the EdgeConv edge convolutional layer extracts the corresponding edge-by-edge convolutional responses from the set of stable receiving edges. The set of stable receiving edges has already undergone spatial receiving direction verification in the dynamic graph construction layer; therefore, the edge-by-edge convolutional responses within it can serve as a source of stable neighborhood responses for the current central structural point. The extracted edge-by-edge convolutional responses are then accumulated dimension-wise according to the central structural point's affiliation, and then averaged dimension-wise according to the number of stable receiving edges to obtain the baseline response corresponding to the current central structural point. The baseline response characterizes the neighborhood response level of the central structural point under stable face spatial receiving relationships, providing a reference for edge response residual calculation.

[0047] In one implementation, if a central structural point corresponds to three stable receiving edges, and each stable receiving edge forms a 64-dimensional edge-wise convolutional response, then the three edge-wise convolutional responses are first accumulated dimension-wise from the 1st to the 64th dimension, and then the accumulated result of each dimension is divided by 3 to obtain the 64-dimensional receiving baseline response corresponding to the central structural point. If a central structural point corresponds to only one stable receiving edge, then the edge-wise convolutional response of that stable receiving edge is used as the receiving baseline response; if a central structural point does not form a stable receiving edge in the current graph convolution stage, then the edge-wise convolutional responses that still retain edge affiliation are selected from the ordinary candidate edges corresponding to the central structural point and temporarily averaged to obtain a temporary receiving baseline response, and the default receiving baseline record of the central structural point is retained in the candidate neighborhood edge state table.

[0048] Ultimately, the EdgeConv edge convolutional layer generates three types of results in this step: edge-wise convolutional response, central structure point update features, and the inherited baseline response. The edge-wise convolutional response is used for differencing with the inherited baseline response, the central structure point update features are used for inheriting residuals and global aggregation, and the inherited baseline response is used for calculating the edge response residuals.

[0049] In this embodiment, S6 specifically includes: S61, EdgeConv edge convolutional layer calls the edge residual recursive correction mechanism, and according to the belonging relationship between the edge convolutional response and the candidate neighborhood edge, the edge convolutional response is back-pasted to the corresponding candidate neighborhood edge. S62. Perform dimension-wise subtraction between the edge-by-edge convolutional response and the baseline response under the same central structural point to form the edge response residuals corresponding to each candidate neighborhood edge. S63. According to the spatial connection direction of the candidate neighborhood edge in the initial neighborhood relationship, perform direction verification on the edge response residual and classify the residual component that continues along the spatial connection direction into the connection residual. S64. Residual components that deviate from the spatial bearing direction, cross non-bearing face regions, or coincide with the reduction states in the candidate neighborhood edge state table are classified as deviation residuals. S65. The inherited residuals are incorporated into the updated features of the central structure points according to their central structure point affiliation, and the affiliation relationship between the inherited residuals and the corresponding candidate neighborhood edges is preserved. S66. Remove the deviation residual from the center structure point update feature and write it into the candidate neighborhood edge state table according to the candidate neighborhood edge position. S67. The candidate neighborhood edge state table performs inter-layer accumulation on the deviation residuals of the same candidate neighborhood edge, performs inter-layer retention on the receiving residuals of the same candidate neighborhood edge, and rewrites the candidate neighborhood edges with continuously accumulated deviation residuals to the reduction state, and rewrites the candidate neighborhood edges with continuously retained receiving residuals to the retention state.

[0050] In its implementation, after the EdgeConv edge convolutional layer completes the separation and preservation of the edge-by-edge convolutional response, the center structure point update feature, and the baseline response, it calls the edge residual recursive correction mechanism to process the edge-by-edge convolutional response again. In ordinary DGCNN, during EdgeConv processing, the edge convolutional responses of each neighboring edge are usually directly aggregated into the center point feature, and the edge-level responses are no longer retained separately after aggregation; when reconstructing the image in the next convolutional stage, the neighborhood relationships are mainly recalculated based on the updated point features. When a face image has side profiles, partial occlusion, low-light blur, or false keypoint detections, abnormal candidate neighboring edges may generate strong responses in EdgeConv. These abnormal responses, after being directly pushed into the center structure point update feature, will continue to affect neighborhood selection in the next convolutional stage, causing the dynamic graph structure to deviate layer by layer from the true facial connection relationships. Instead of following the single path of "directly aggregating edge responses into point features" in ordinary DGCNN, this invention decomposes the residual of the baseline response obtained by merging the edge-by-edge convolution response with the stable receiving edge set, and writes the deviation residual into the candidate neighborhood edge state table after separating it from the updated features of the central structure point. Thus, an edge-level recursive correction path is formed between the graph convolution stages.

[0051] Specifically, after the EdgeConv edge convolutional layer invokes the edge residual recursive correction mechanism, it first re-attaches each edge-by-edge convolutional response to its corresponding candidate neighbor edge according to the attribution relationship between the edge-by-edge convolutional response and the candidate neighbor edge. This re-attachment process utilizes the attribution relationship to reposition the edge-by-edge convolutional response to a specific central structural point, neighboring structural point, and candidate neighbor edge. For example, when a central structural point is connected to 5 candidate neighbor edges, the EdgeConv edge convolutional layer will generate 5 edge-by-edge convolutional responses. The edge residual recursive correction mechanism re-attaches each of the 5 edge-by-edge convolutional responses to its corresponding candidate neighbor edge, ensuring that each candidate neighbor edge has an independent response record.

[0052] Subsequently, the edge-by-edge convolutional response and the reference response at the same central structural point are subtracted dimension-by-dimensionally to form the edge response residuals corresponding to each candidate neighbor edge. In one embodiment, both the edge-by-edge convolutional response and the reference response are 64-dimensional vectors, so the subtraction is performed item by item from the 1st to the 64th dimension to obtain the 64-dimensional edge response residuals. If the edge-by-edge convolutional response of a candidate neighbor edge is the response vector of the current edge, and the reference response is the average response vector obtained by merging the stable receiving edges at the same central structural point, then the edge response residuals reflect the degree of offset of the candidate neighbor edge relative to the stable face receiving relationship.

[0053] After obtaining the edge response residuals, a direction check is performed on the edge response residuals according to the spatial connection direction of the candidate neighbor edges in the initial neighborhood relation. During direction check, the connection direction of the candidate neighbor edges in the initial neighborhood relation is first read, such as the sequential connection direction within a local region, the connection direction across facial regions, or the symmetrical connection direction corresponding to the left and right sides; then, it is checked whether the component in the edge response residual consistent with this connection direction remains continuous. For residual components that continue along the spatial connection direction, this residual component is classified into the connection residual. The connection residual indicates that although the candidate neighbor edge differs from the connection reference response, the difference still falls within the connection direction allowed by the facial structure, and can serve as an effective supplement to the updated features of the central structural point.

[0054] In one embodiment, the direction check includes edge direction check and residual direction check. Edge direction check is used to confirm whether the structural points at both ends of a candidate neighborhood edge are still connected along the edge type in the initial neighborhood relationship; residual direction check is used to confirm whether the edge response residual continues along the original connection direction of the candidate neighborhood edge. For local neighborhood edges, if the response enhancement position corresponding to the edge response residual still falls within the same face region, the residual component is classified as a receiving residual; for receiving neighborhood edges, if the response enhancement position corresponding to the edge response residual still falls between adjacent face regions, the residual component is classified as a receiving residual; for symmetrical neighborhood edges, if the response enhancement position corresponding to the edge response residual still maintains a left-right correspondence, the residual component is classified as a receiving residual. If the response enhancement position corresponding to the residual component crosses a non-receiving face region, points to a reduced candidate neighborhood edge, or deviates from the original edge attribution, the residual component is classified as a deviation residual.

[0055] For residual components that deviate from the spatial bearing direction, cross non-bearing facial regions, or overlap with the reduction state in the candidate neighbor edge state table, these residual components are classified as deviation residuals. For example, if a candidate neighbor edge originally corresponds to a local neighbor edge within the eye region, but the edge response residual shows that the edge response is closer to the connection change across the eye and non-adjacent contour regions, then this residual component will no longer be included in the bearing residual and will be classified as a deviation residual. Similarly, if a candidate neighbor edge has already been rewritten to a reduction state in the candidate neighbor edge state table during the previous convolution stage, and this candidate neighbor edge again generates a residual that deviates from the bearing direction in this stage, then this residual component will continue to be classified as a deviation residual and treated as an inter-layer accumulation object.

[0056] After decomposing the receiving and deviation residuals, the receiving residuals are incorporated into the updated features of the central structural points according to their assignment. During incorporation, the assignment position of the updated features of the central structural points is not changed; instead, the effective components of the receiving residuals that continue along the spatial receiving direction are superimposed at the corresponding central structural point feature positions. Simultaneously, the assignment relationship between the receiving residuals and their corresponding candidate neighborhood edges is preserved. This process can retain the contribution of stable facial structural edges to identity features and avoid directly weakening effective neighborhood edges due to slight pose changes or local expression changes.

[0057] For deviation residuals, they are extracted from the updated features of the central structural point and written into the candidate neighbor edge state table according to the candidate neighbor edge position. The candidate neighbor edge state table records candidate neighbor edges, performs inter-layer accumulation on deviation residuals of the same candidate neighbor edge, and performs inter-layer retention on the receiving residuals of the same candidate neighbor edge. In one embodiment, if the same candidate neighbor edge is written with deviation residuals in two consecutive graph convolution stages, the candidate neighbor edge state table rewrites the candidate neighbor edge to a reduced state; if the same candidate neighbor edge is retained with receiving residuals in two consecutive graph convolution stages, the candidate neighbor edge state table rewrites the candidate neighbor edge to a retained state. After the state rewriting result enters the dynamic graph update layer, the candidate neighbor edge corresponding to the reduced state is removed from the candidate neighbor range constructed by the next layer of dynamic graph, and the candidate neighbor edge corresponding to the retained state enters the retained neighbor range constructed by the next layer of dynamic graph.

[0058] In one embodiment, the candidate neighborhood edge state table records each candidate neighborhood edge row by row according to its candidate neighborhood edge number. Each row corresponds to one candidate neighborhood edge and records the center structure point, neighborhood structure point, current graph convolution stage state, cumulative number of deviation residuals, consecutive number of receiving residuals, and connection state of the candidate neighborhood edge. The current graph convolution stage state includes three categories: receiving hit, deviation hit, and miss. The connection state includes normal state, reduction state, and retention state. If the same candidate neighborhood edge writes deviation residuals in two consecutive graph convolution stages, the cumulative number of deviation residuals is incremented and the connection state is changed to reduction state. If the same candidate neighborhood edge writes receiving residuals in two consecutive graph convolution stages, the consecutive number of receiving residuals is incremented and the connection state is changed to retention state. If the candidate neighborhood edge does not write receiving residuals or deviation residuals in the current graph convolution stage, the connection state of the previous graph convolution stage is retained, and the check continues in the next graph convolution stage.

[0059] Through the above processing, the improvement of the edge residual recursive correction mechanism compared with ordinary DGCNN is that: ordinary DGCNN aggregates edge responses into point features and then reconstructs the graph, while this invention retains the source of each edge response before and after aggregation, and uses the receiving baseline response as a stable reference to decompose each edge response into receiving residual and deviation residual; receiving residual continues to strengthen the update features of the central structure point, and deviation residual no longer pollutes the update features of the central structure point, but is written into the candidate neighborhood edge state table and drives the reduction of candidate neighborhood edges in the next graph convolution stage.

[0060] In this embodiment, S7 specifically includes: S71. The dynamic graph update layer reads the candidate neighborhood edge state table, assigns candidate neighborhood edges written to the reduction state to the reduction edge set, assigns candidate neighborhood edges written to the retention state to the retention edge set, and assigns candidate neighborhood edges not written to the reduction state or retention state to the edge set to be checked. S72. Perform edge connection pruning on the cut edge set, remove the candidate neighborhood edges in the cut edge set from the candidate neighborhood range constructed by the next layer dynamic graph, and simultaneously mask the deviation residuals corresponding to the cut edge set. S73. Perform edge connection solidification on the retained edge set, and write the candidate neighborhood edges and the structural points at both ends in the retained edge set into the retained neighborhood range constructed by the next layer of dynamic graph; S74. Perform inter-layer state comparison on the set of edges to be checked. Transfer the edges to be checked that deviate from the residuals written in the current graph convolution stage to the set of edges to be reduced. Transfer the edges to be checked that accept the residuals written in the convolution stages of adjacent graphs to the set of edges to be retained. S75. Rearrange the candidate neighborhood edges in the retained edge set according to the affiliation of the central structural point, and perform alignment between the rearranged candidate neighborhood edges and the unpruned initial neighborhood relationships. S76. After the candidate neighborhood edges are aligned, they are repackaged into local dynamic graph units, and each local dynamic graph unit is connected in series according to the spatial connection relationship between facial structure points to form a recursive corrected facial dynamic graph.

[0061] In the specific implementation, after receiving the candidate neighborhood edge state table, the dynamic graph update layer no longer searches for neighborhoods again based solely on the updated structural point features of the current layer, as is the case with ordinary DGCNN. This step uses the candidate neighborhood edge state table as the edge state constraint for the dynamic graph update layer. Before constructing the next dynamic graph layer, the candidate neighborhood edges are first reduced, retained, and checked to ensure that the recursive correction of the facial dynamic graph does not repeatedly introduce abnormal edges that have already been hit by the deviated residual.

[0062] The dynamic graph update layer first reads the candidate neighborhood edge state table. This table records the pruning state, retention state, deviation residual record, and continuation residual record written in the previous graph convolution stage or adjacent graph convolution stages. The dynamic graph update layer checks each candidate neighborhood edge: candidate neighborhood edges with pruning states are added to the pruning edge set; candidate neighborhood edges with retention states are added to the retention edge set; candidate neighborhood edges without pruning or retention states are added to the edge set to be checked. If the same candidate neighborhood edge has both deviation residual and continuation residual records, the state of the latest graph convolution stage is read first; if deviation residuals accumulate continuously in two adjacent graph convolution stages, the candidate neighborhood edge is preferentially added to the pruning edge set to prevent abnormal edges from continuing into the next layer's edge convolution calculation.

[0063] For the edge reduction set, the dynamic graph update layer performs edge connection pruning. During the construction of the next dynamic graph layer, the dynamic graph construction layer prioritizes reading candidate neighborhood edges within the retained neighborhood range and arranges them at the front of the candidate edge sequence corresponding to the central structure point. For candidate neighborhood edges in the reduced edge set, they are not repeatedly pulled from the initial neighborhood relations. For candidate neighborhood edges in the edge set to be checked, their ordinary candidate states are retained, and edge response residual decomposition continues after edge convolution processing in the next layer's EdgeConv layer. Thus, the candidate neighborhood edge state table not only exists as a record structure but also directly changes the candidate neighborhood range of the next dynamic graph construction. Edge connection pruning directly removes candidate neighborhood edges from the candidate neighborhood range of the next dynamic graph construction. During removal, the dynamic graph update layer simultaneously locates the central structure point and neighborhood structure point at both ends of the candidate neighborhood edge in the reduced edge set, prunes the corresponding edge connection from the candidate edge list of the next local dynamic graph unit, and simultaneously masks the deviation residual corresponding to the reduced edge set. After this processing, candidate neighborhood edges that are continuously hit by the deviated residuals will not participate in the construction of the next dynamic graph layer again, nor will they generate new edge-by-edge convolutional responses in the next EdgeConv edge convolutional layer.

[0064] For the retained edge set, the dynamic graph update layer performs edge connection solidification. Edge connection solidification refers to writing the candidate neighborhood edges and their endpoint structural points from the retained edge set into the retained neighborhood range constructed in the next dynamic graph layer. Candidate neighborhood edges in the retained neighborhood range are preferentially arranged at the front of the candidate edge sequence during the construction of the next dynamic graph layer and continue to serve as stable facial spatial connections in the EdgeConv edge convolutional layer. Unlike ordinary DGCNN where each layer freely reconstructs the neighborhood, this step retains candidate neighborhood edges that continuously hit the connecting residuals, ensuring that the connections between stable facial structures such as the eyes, nose, and contours persist throughout the multi-layer graph convolutional stages, providing a stable structural source for facial identity features.

[0065] For the set of edges to be checked, the dynamic graph update layer performs inter-layer state comparison. This comparison doesn't re-evaluate the distance to candidate neighboring edges, but rather checks the residual writing status of each candidate neighboring edge in the current graph convolution stage and adjacent graph convolution stages. When an edge to be checked writes a deviation residual in the current graph convolution stage, the dynamic graph update layer moves that edge to the reduction edge set; when an edge to be checked continuously writes a receiving residual in adjacent graph convolution stages, the dynamic graph update layer moves that edge to the retention edge set; when an edge has neither continuous receiving records nor a current deviation record, the dynamic graph update layer temporarily retains the edge's normal candidate state and continues to accept edge residual recursive correction in the next graph convolution stage. Through this process, the reduction or retention of candidate neighboring edges is determined recursively by the inter-layer residual state.

[0066] After processing the edges to be cut, retained, and checked, the dynamic graph update layer rearranges the candidate neighborhood edges in the retained edge set according to the affiliation of the central structural point. Specifically, retained edges corresponding to the same central structural point are first merged into the same local neighborhood, and then the retained edges are arranged according to the spatial connection relationship between the facial structural points. Subsequently, the rearranged retained edges are aligned with the initial neighborhood relationships that have not been pruned. During alignment, if a candidate edge in the initial neighborhood relationship has been pruned by the edge cut set, the candidate edge will not be included in the alignment result; if a candidate edge in the initial neighborhood relationship remains in a normal candidate state, the candidate edge is arranged after the retained edges as a supplementary edge for the construction of the next layer of dynamic graph.

[0067] Finally, the dynamic graph update layer repackages the candidate neighborhood edges after in-situ docking into local dynamic graph units. Each local dynamic graph unit still corresponds to a central structural point, and retains the fixed stable connecting edges and unpruned ordinary candidate edges within the unit. These local dynamic graph units are then concatenated according to the spatial connection relationships between facial structural points to form a recursively corrected facial dynamic graph. This recursively corrected facial dynamic graph serves as the input graph structure for the next graph convolution stage, continuing into the dynamic graph construction layer and the EdgeConv edge convolution layer for the next round of edge convolution processing.

[0068] Through the above processing, candidate neighborhood edges that deviate from the facial spatial continuity relationship are recursively reduced, while candidate neighborhood edges that maintain continuous continuity relationship are recursively retained. The recursively corrected facial dynamic map can converge towards a stable facial structure layer by layer.

[0069] In this embodiment, S8 specifically includes: S81. Receive the recursive correction of the facial dynamic map, include the updated features of the center structure points associated with the edge set in the aggregation range, and exclude the deviation residuals associated with the reduced edge set from the aggregation range. S82. Based on the spatial arrangement of the central structural points in the facial structure map, the update features of the central structural points included in the aggregation range are sequentially aggregated to form a sequence of facial structural point update features. S83. Perform a dimension-wise comparison on the updated feature sequence of facial structure points, extract the maximum response of each dimension, and perform a dimension-wise accumulation and averaging on the updated feature sequence of facial structure points to extract the average response of each dimension. S84. The maximum response and average response of each dimension are placed side by side to form a global facial structure response; S85. Encapsulate the global facial structure response with the candidate neighborhood edge attribution relationship in the retained edge set to form facial identity features; S86. Match the facial identity features with the identity records in the smart home identity database one by one to obtain the identity matching response between the face to be identified and each identity record. S87. Sort the numerical values ​​of each identity matching response and write the sorting results into the smart home face recognition results.

[0070] In the specific implementation process, after the recursive correction of the facial dynamic map is completed, global aggregation no longer directly performs pooling on the updated features of all central structural points and the convolutional responses of all edges. This step determines the aggregation range based on the recursive correction of the facial dynamic map and the candidate neighborhood edge state table. Only the updated features of the central structural points associated with the edge set are included in the identity feature aggregation range, and the deviation residuals associated with the reduced edge set are excluded from the aggregation range. In this way, the recursive correction result of the dynamic map is implemented in the final facial identity feature.

[0071] Specifically, the process first receives the recursively corrected facial animation and reads the set of retained edges and the set of cut edges already formed in the animation update layer. For the set of retained edges, the center structural points and neighboring structural points at both ends of the retained edges are located, and the updated features of the corresponding center structural points are included in the aggregation range. For the set of cut edges, the deviation residuals and candidate neighboring edge positions corresponding to the cut edges are located, and the deviation residuals are excluded from the aggregation range. If a certain center structural point is associated with both retained edges and cut edges, the updated features of the center structural point after the residuals are incorporated are retained, and the deviation residuals caused by the cut edges are excluded to avoid breaking the facial structure chain due to whole-point deletion.

[0072] After determining the aggregation range, the updated features of the central structural points included in the aggregation range are sequentially aggregated according to their spatial arrangement in the facial structure map. During sequential aggregation, the updated features of the central structural points are first organized according to the order of the facial central axis region, the left and right symmetrical regions, and the contour-connecting regions. Within the same facial region, the updated features of the central structural points are arranged according to the order of their local neighbor edges. For central structural points spanning multiple facial regions, the updated features are arranged according to the connection direction of their neighbor edges. Through this process, the aggregation object is no longer an unordered set of point features, but rather a sequence of updated facial structural points that maintains consistency with the spatial connection relationships of the face.

[0073] After generating the facial structural point update feature sequence, a dimension-by-dimensional comparison is performed on the update features of each central structural point within the sequence, extracting the maximum response in each dimension. Taking a 64-dimensional vector as an example, the values ​​of the update features of all central structural points included in the aggregation range are compared sequentially in dimensions 1 to 64, and the maximum value in each dimension is taken to obtain the 64-dimensional maximum response vector. The maximum response vector retains the most significant identity-distinguishing features in the stable facial transition regions, such as the strong structural responses in the eye-nose transition region or the contour transition region.

[0074] Simultaneously, the updated feature sequences of facial structural points are accumulated and averaged dimension-wise to extract the average response for each dimension. Taking a 64-dimensional vector as an example, the updated feature values ​​of all central structural points included in the aggregation range are accumulated along the same dimension, and then divided by the number of central structural points included in the aggregation range to obtain the 64-dimensional average response vector. The average response vector preserves the overall facial structure distribution, reducing the excessive influence of a single strong response point on identity features. Then, the 64-dimensional maximum response vector and the 64-dimensional average response vector are arranged side-by-side in the same dimension order to form a 128-dimensional global facial structural response. This global facial structural response simultaneously contains both local salient structure and overall stable structure information.

[0075] After generating the global facial structure response, the global facial structure response is co-encapsulated with the candidate neighborhood edge attribution relationships in the retained edge set to form facial identity features. During co-encapsulation, the global facial structure response is jointly bound with the central structure point attribution, neighborhood structure point attribution, and facial spatial connection relationship corresponding to the retained edge set. This ensures that the facial identity features not only include the aggregated response value but also retain the structural source of the response from stable facial connection edges.

[0076] During identity mapping, facial identity features are matched item by item with identity records in the smart home identity database. Each identity record in the smart home identity database can store the registered facial identity features of family members, corresponding identity identifiers, and historical recognition update features. During matching, the response difference or similarity response between the facial identity features of the face to be identified and each identity record is calculated item by item to obtain the identity matching response between the face to be identified and each identity record. In one embodiment, the 128-dimensional facial identity features can be subtracted dimension by dimension from the 128-dimensional registration features in the identity records, and the absolute values ​​of the differences in each dimension can be accumulated and then the accumulated result can be converted into an identity matching response; alternatively, the two sets of features can be multiplied dimension by dimension and accumulated to obtain an identity matching response in the form of similarity. The specific matching method used can be determined based on the training method of the smart home identity database.

[0077] After obtaining the responses to each identity match, the responses are numerically sorted, and the sorting results are written into the smart home face recognition results. The sorting results may include the identity record with the highest matching degree, the matching order of each candidate identity record, and the corresponding response value. If the smart home system is configured with a stranger recognition strategy, the comparison process between the first ranked response and the identity confirmation threshold can be further explained in other parts of the instruction manual. In this step, the identity output results should at least include the sorting relationship between the face to be recognized and each identity record in the identity database, providing an identity recognition basis for smart door lock control, family member arrival records, or stranger alarms.

[0078] In one embodiment, the smart home facial recognition result includes an identity ranking result and an identity confirmation result. If the difference between the identity matching response of the first ranked person and the identity matching response of the second ranked person meets the identity differentiation requirement, the identity record corresponding to the first ranked person is written into the identity confirmation result; if the identity matching response of the first ranked person does not meet the identity confirmation requirement, or the difference between the responses of the first ranked person and the second ranked person is insufficient, the face to be identified is written into the unconfirmed identity result. The unconfirmed identity result can proceed to a stranger alert, manual review, or secondary data collection process. This process can avoid forcibly outputting the wrong family member identity in low-light, occluded, or extreme profile scenarios.

[0079] Through the above processing, the global aggregation process no longer performs indiscriminate pooling on all DGCNN output features. Instead, it limits the aggregation range based on recursively correcting the facial dynamic map, incorporating the updated features of the central structural points associated with the retained edge set into the facial identity features, and excluding the deviation residuals associated with the reduced edge set from the aggregation range. This processing extends the reduced and retained states generated by the recursive correction mechanism of edge residuals to the final identity feature construction stage, reducing the interference of occluded edges, misconnected edges, and pose offset edges on the identity mapping results, thereby improving the stability of face recognition results in complex smart home acquisition environments.

[0080] Example 1: To verify the feasibility of this invention in practice, it was applied to a smart door lock and indoor monitoring linkage scenario in a residential community. The test area included three data acquisition locations: the entrance door, the foyer, and the living room entrance. Each location was equipped with a 2-megapixel camera with a resolution of 1920×1080 and a frame rate of 25fps. The smart home edge node used an 8-core ARM processor and 4GB of RAM. The test subjects included 48 family members from 12 households, and an additional 36 unregistered individuals were collected as stranger samples. For each family member, six types of scene clips were collected: frontal face, side face, low light, wearing a hat, wearing a mask, and quickly passing by. Approximately 60 video clips were collected for each scene type, resulting in 8640 registered samples and 7200 test samples, including 2400 unobstructed test clips, 1800 partially obstructed test clips, 1200 side face test clips, 1200 low light test clips, and 600 stranger test clips.

[0081] During testing, the smart home edge node first extracts the face segment to be recognized from the face image stream and locates facial structural points. A facial structural map is then constructed based on the spatial relationships between these points. The number of facial structural points is controlled to within 68, and 5 to 9 consecutive frames are selected from each face segment to be recognized for structural point stability verification. The improved DGCNN dynamic graph construction layer first screens candidate neighborhood edges from the initial neighborhood relations and includes candidate neighborhood edges that maintain facial spatial continuity into a stable continuity edge set. The EdgeConv edge convolutional layer performs edge-by-edge convolution operations on the initial facial dynamic graph to form an edge-by-edge convolution response, which is then merged into the baseline response by the stable continuity edge set. The edge residual recursive correction mechanism differs between the edge-by-edge convolution response and the baseline response, decomposing the continuity residual and the deviation residual. The continuity residual is incorporated into the center structure point to update features, while the deviation residual is written into the candidate neighborhood edge state table. The dynamic graph update layer reduces the candidate neighborhood edges that are hit by the deviation residual based on the candidate neighborhood edge state table, retaining the candidate neighborhood edges that continuously hit the continuity residual, forming a recursively corrected facial dynamic graph. Finally, global aggregation is performed on the recursively corrected facial dynamic graph to obtain facial identity features, which are then matched item by item with the identity records in the smart home identity database to output the smart home face recognition result.

[0082] To verify the effectiveness of the method of this invention, it was compared with traditional CNN face recognition methods and ordinary DGCNN face recognition methods. Traditional CNN face recognition methods directly perform convolutional feature extraction and identity classification on the cropped face image; ordinary DGCNN face recognition methods input facial structure points into the DGCNN, but do not set a stable set of supporting edges, a supporting baseline response, an edge residual recursive correction mechanism, or a candidate neighborhood edge state table. The three methods used the same test samples, and the recognition accuracy, partial occlusion scene recognition accuracy, stranger false recognition rate, average recognition time per segment, and number of false door lock triggers were statistically analyzed. The results are shown in Table 1.

[0083] Table 1. Comparison of recognition performance of different face recognition methods in smart home scenarios.

[0084] Among them, the overall recognition accuracy rate represents the proportion of correct identity recognition in all test segments; the partial occlusion scene recognition accuracy rate represents the proportion of correct identity recognition in scenarios where the face is covered by masks, hat brims, hands, etc.; the side face scene recognition accuracy rate represents the proportion of correct identity recognition when the face is turned at a large angle; the stranger false recognition rate represents the proportion of non-registered persons being incorrectly identified as family members; the average recognition time per segment represents the average processing time from the input of the face segment to be recognized to the output recognition result; and the number of door lock false triggers represents the number of times the door lock is opened or the unlocking command is triggered due to incorrect identity confirmation during the test. The method of this invention reduces the false identity confirmation caused by occlusion and false detection by reducing the candidate neighborhood edges that deviate from the residual hit. Therefore, the number of door lock false triggers is lower than that of traditional CNN methods and ordinary DGCNN methods.

[0085] As shown in Table 1, the method of this invention outperforms traditional CNN face recognition methods and ordinary DGCNN face recognition methods in terms of overall recognition accuracy, accuracy in partial occlusion scene recognition, and accuracy in profile scene recognition. Traditional CNN face recognition methods rely directly on image texture features. When encountering masks, hat brims, low lighting, and profile scenes, the lack of local texture leads to unstable identity features. Although ordinary DGCNN introduces graph relationships between facial structure points, its dynamic graph update still easily re-includes occluded points, false detection points, or pose shift points into the neighborhood, and abnormal edge responses continue to propagate in multi-layer graph convolutions.

[0086] This invention's method stabilizes the set of receiving edges and merges them into a baseline response. Then, it utilizes a recursive correction mechanism based on edge residuals to decompose receiving residuals and deviation residuals. Deviation residuals are written into a candidate neighbor edge state table, and candidate neighbor edges that match deviation residuals are reduced during the dynamic graph update phase. This process can suppress abnormal connections caused by occlusion, side profiles, and false detections of key points layer by layer during the graph structure recursion process, while retaining candidate neighbor edges that continuously match receiving residuals. This makes the recursively corrected facial dynamic graph more closely resemble the actual spatial relationships of the face. While slightly increasing the time consumption, this invention reduces the false recognition rate of strangers and the number of false door lock triggers, meeting the edge recognition requirements of smart homes.

[0087] 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 smart home face recognition method based on graph neural networks, characterized in that, Includes the following steps: S1. The smart home edge node collects facial image streams within the target home area and crops out the facial segments to be identified; S2. Locate facial structure points in the face segment to be identified, organize the initial neighborhood relationship according to the spatial connection relationship between the facial structure points, and form a facial structure map. S3. Input the facial structure map into the improved DGCNN. The improved DGCNN includes a dynamic graph construction layer, an EdgeConv edge convolutional layer and a dynamic graph update layer. An edge residual recursive correction mechanism is set between the EdgeConv edge convolutional layer and the dynamic graph update layer, and a candidate neighborhood edge state table is set in the dynamic graph update layer. S4. The dynamic graph construction layer screens the neighborhood structure points corresponding to the central structure point according to the initial neighborhood relationship, and includes the candidate neighborhood edges that maintain the facial space connection into the stable connection edge set to construct the initial facial dynamic graph. S5, EdgeConv edge convolutional layers perform edge-by-edge convolution operations along the initial facial dynamic map to form edge-by-edge convolutional responses, and merge the baseline responses based on the stable set of receiving edges; S6, EdgeConv edge convolutional layer calls the edge residual recursive correction mechanism, divides the difference between the edge convolution response and the baseline response into edge response residual, decomposes the baseline residual and deviation residual, incorporates the baseline residual into the center structure point update feature, and writes the deviation residual into the candidate neighborhood edge state table. S7. The dynamic graph update layer reduces candidate neighborhood edges that deviate from the residual hit according to the candidate neighborhood edge state table, and retains candidate neighborhood edges that continuously hit the residual to form a recursive corrected face dynamic graph. S8. Perform global aggregation on the recursive modified facial dynamic map to form facial identity features, and map the facial identity features to the smart home face recognition result.

2. The smart home face recognition method based on graph neural networks according to claim 1, characterized in that, S1 specifically includes: S11. The smart home edge node performs frame-by-frame face region localization on the face image stream according to the acquisition time sequence, and classifies face regions with overlapping contours and continuous positions in adjacent frames into the same face candidate segment. S12. Perform boundary clipping and facial integrity checks on the face regions within the face candidate segments, and remove candidate frames with truncated contours, faces that deviate from the acquisition area, and broken consecutive frames. S13. Extract the retained candidate frames into face segments to be identified according to the acquisition time sequence.

3. The smart home face recognition method based on graph neural networks according to claim 1, characterized in that, S2 specifically includes: S21. Locate the boundaries of key facial regions and facial contours in the face segment to be identified, and merge the intersection points, local turning points and center points of the regions into facial structure points. S22. Configure the structure point position record according to the spatial position of the facial structure points in the face segment to be identified, and connect adjacent facial structure points in the same facial region as local neighborhood edges. S23. Connect adjacent facial structural points across facial regions as receiving neighboring edges, connect corresponding facial structural points on the left and right as symmetrical neighboring edges, and remove connecting edges that deviate from the facial spatial order. S24. The facial structure graph is constructed by combining facial structure points, local neighborhood edges, receiving neighborhood edges, and symmetrical neighborhood edges, and the local neighborhood edges, receiving neighborhood edges, and symmetrical neighborhood edges are used as the initial neighborhood relations.

4. The smart home face recognition method based on graph neural networks according to claim 1, characterized in that, S4 specifically includes: S41. The dynamic graph construction layer sets the facial structure points in the facial structure graph as the center structure points one by one according to the spatial arrangement order, and extracts the connection edges directly connected to the current center structure point from the initial neighborhood relationship to form the candidate neighborhood edge set corresponding to the current center structure point. S42. Locate the neighborhood structure points along the candidate neighborhood edge set one by one, and perform edge-end attribution verification on the facial structure points at both ends of the candidate neighborhood edge. Retain the candidate neighborhood edges whose edges still belong to the adjacent facial region, symmetrical facial region, or contour-bound region. S43. Perform a connection direction check on the retained candidate neighborhood edges. Push the candidate neighborhood edges whose connection direction is consistent with the spatial connection direction in the initial neighborhood relationship into the stable connection edge set. Retain the candidate neighborhood edges whose connection direction has shifted but have not deviated from the facial structure graph as ordinary candidate edges. S44. Rewrite the candidate neighborhood edges in the stable receiving edge set as the dynamic graph neighborhood edges corresponding to the central structure point, and label the facial structure points at both ends of the dynamic graph neighborhood edges as the central structure point and the neighborhood structure point, respectively. S45. Arrange the ordinary candidate edges after the stable receiving edge set to form the candidate neighborhood supplementary edge set corresponding to the current central structure point; S46. Based on the central structural point, encapsulate the set of neighborhood edges and candidate neighborhood supplementary edges of the dynamic graph into local dynamic graph units, and connect each local dynamic graph unit in series according to the spatial connection relationship between facial structural points to construct the initial facial dynamic graph.

5. The smart home face recognition method based on graph neural networks according to claim 1, characterized in that, S5 specifically includes: S51, EdgeConv convolutional layer receives the initial facial dynamic map, expands the dynamic map neighborhood edges in each local dynamic map unit one by one, and pairs the center structure point and neighborhood structure point at both ends of the dynamic map neighborhood edge as edge end pairs. S52. Perform edge feature splitting on each edge pair, place the central structure point feature at the central end, and place the neighboring structure point feature at the neighboring end. S53. Perform a dimension-wise subtraction between the neighborhood end features and the center end features to obtain the neighborhood offset features, and then place the neighborhood offset features and the center end features in the same position to form the edge convolution input of the corresponding dynamic graph neighborhood edge. S54. Perform edge-by-edge mapping on the edge convolution input, mapping the edge convolution input corresponding to each neighborhood edge of the dynamic graph to an edge-by-edge convolution response, and retaining the association relationship between the edge convolution response and the neighborhood edge of the dynamic graph after mapping. S55. Compare the edge-by-edge convolutional responses belonging to the same central structural point dimension by dimension, and fill the maximum response value in each dimension back into the feature position of the central structural point to form the updated feature of the central structural point. S56. Extract the corresponding edge-by-edge convolutional response from the stable receiving edge set, and perform dimension-by-dimensional accumulation and dimension-by-dimensional averaging on the extracted edge-by-edge convolutional response according to the center structure point to form the receiving benchmark response corresponding to the current center structure point. S57. The edge-by-edge convolutional response, the central structure point update feature, and the supporting baseline response are split and retained at the output of the EdgeConv edge convolutional layer.

6. The smart home face recognition method based on graph neural networks according to claim 1, characterized in that, S6 specifically includes: S61, EdgeConv edge convolutional layer calls the edge residual recursive correction mechanism, and according to the belonging relationship between the edge convolutional response and the candidate neighborhood edge, the edge convolutional response is back-pasted to the corresponding candidate neighborhood edge. S62. Perform dimension-wise subtraction between the edge-by-edge convolutional response and the baseline response under the same central structural point to form the edge response residuals corresponding to each candidate neighborhood edge. S63. According to the spatial connection direction of the candidate neighborhood edge in the initial neighborhood relationship, perform direction verification on the edge response residual and classify the residual component that continues along the spatial connection direction into the connection residual. S64. Residual components that deviate from the spatial bearing direction, cross non-bearing face regions, or coincide with the reduction states in the candidate neighborhood edge state table are classified as deviation residuals. S65. The inherited residuals are incorporated into the updated features of the central structure points according to their central structure point affiliation, and the affiliation relationship between the inherited residuals and the corresponding candidate neighborhood edges is preserved. S66. Remove the deviation residual from the center structure point update feature and write it into the candidate neighborhood edge state table according to the candidate neighborhood edge position. S67. The candidate neighborhood edge state table performs inter-layer accumulation on the deviation residuals of the same candidate neighborhood edge, performs inter-layer retention on the receiving residuals of the same candidate neighborhood edge, and rewrites the candidate neighborhood edges with continuously accumulated deviation residuals to the reduction state, and rewrites the candidate neighborhood edges with continuously retained receiving residuals to the retention state.

7. The smart home face recognition method based on graph neural networks according to claim 1, characterized in that, Specifically, S7 includes: S71. The dynamic graph update layer reads the candidate neighborhood edge state table, assigns candidate neighborhood edges written to the reduction state to the reduction edge set, assigns candidate neighborhood edges written to the retention state to the retention edge set, and assigns candidate neighborhood edges not written to the reduction state or retention state to the edge set to be checked. S72. Perform edge connection pruning on the cut edge set, remove the candidate neighborhood edges in the cut edge set from the candidate neighborhood range constructed by the next layer dynamic graph, and simultaneously mask the deviation residuals corresponding to the cut edge set. S73. Perform edge connection solidification on the retained edge set, and write the candidate neighborhood edges and the structural points at both ends in the retained edge set into the retained neighborhood range constructed by the next layer of dynamic graph; S74. Perform inter-layer state comparison on the set of edges to be checked. Transfer the edges to be checked that deviate from the residuals written in the current graph convolution stage to the set of edges to be reduced. Transfer the edges to be checked that accept the residuals written in the convolution stages of adjacent graphs to the set of edges to be retained. S75. Rearrange the candidate neighborhood edges in the retained edge set according to the affiliation of the central structural point, and perform alignment between the rearranged candidate neighborhood edges and the unpruned initial neighborhood relationships. S76. After the candidate neighborhood edges are aligned, they are repackaged into local dynamic graph units, and each local dynamic graph unit is connected in series according to the spatial connection relationship between facial structure points to form a recursive corrected facial dynamic graph.

8. The smart home face recognition method based on graph neural networks according to claim 1, characterized in that, S8 specifically includes: S81. Receive the recursive correction of the facial dynamic map, include the updated features of the center structure points associated with the edge set in the aggregation range, and exclude the deviation residuals associated with the reduced edge set from the aggregation range. S82. Based on the spatial arrangement of the central structural points in the facial structure map, the update features of the central structural points included in the aggregation range are sequentially aggregated to form a sequence of facial structural point update features. S83. Perform a dimension-wise comparison on the updated feature sequence of facial structure points, extract the maximum response of each dimension, and perform a dimension-wise accumulation and averaging on the updated feature sequence of facial structure points to extract the average response of each dimension. S84. The maximum response and average response of each dimension are placed side by side to form a global facial structure response; S85. Encapsulate the global facial structure response with the candidate neighborhood edge attribution relationship in the retained edge set to form facial identity features. S86. Match the facial identity features with the identity records in the smart home identity database one by one to obtain the identity matching response between the face to be identified and each identity record; S87. Sort the numerical values ​​of each identity matching response and write the sorting results into the smart home face recognition results.

Images