Automatic recognition system for decor style matching

Abstract

This invention relates to the field of image recognition and discloses an automatic recognition system for decorative style matching, comprising: an image acquisition unit acquiring a two-dimensional real-scene image; a semantic anchor extraction unit extracting a first set of bounding boxes labeled with decorative style and a second set of bounding boxes labeled with style neutrality; a cross-semantic space interference screening unit determining initial weights based on the overlap ratio of bounding boxes; a local pixel mask verification unit correcting node weights using the actual intersecting pixel area; a spatial topology repair unit completing topological breaks caused by occlusion using the shortest connected path and constructing a spatial semantic topology graph; and a graph matching unit determining style matching similarity through graph edit distance. This invention effectively eliminates coordinate offset noise caused by clutter occlusion by constructing a heterogeneous semantic interference verification mechanism, ensuring the logical coherence of the spatial topology structure, and improving the system's robustness in complex real-scene environments without increasing the sensing hardware.

Description

Technical Field

[0001] This invention belongs to the field of image recognition technology, and in particular relates to an automatic recognition system for matching decorative styles. Background Technology

[0002] Current image recognition schemes typically use convolutional neural networks to extract global feature vectors from images and determine style consistency by calculating the feature space distance between the image to be matched and the standard sample. Such schemes treat all pixels in the image region as homogeneous semantic input sources, and their effectiveness is based on the ideal assumption that the target features occupy the main receptive field of the image and the background is relatively pure.

[0003] However, when the above solutions are applied to complex indoor real-world scenarios, the image content contains a large number of everyday semantic entities unrelated to the target style attributes, such as people, electronic devices, and randomly placed clutter. These neutral entities not only disrupt the visual coherence between style anchor points on the two-dimensional projection plane, but also cause random shifts in the center coordinates of the bounding boxes output by the target detection logic. Since the existing spatial topology construction mechanism assumes the absolute validity of all detection coordinates, this coordinate pollution caused by outlier pixels will directly translate into incorrect topological edge weights, resulting in geometric distortion of the final generated image feature manifold and reducing the matching accuracy of the recognition system. In-depth analysis reveals that such improvements still remain at the level of statistical probability weight allocation, without establishing a semantic conflict verification mechanism based on the geometric essence of physical occlusion. Simple feature enhancement cannot identify false centers caused by occlusion inside the bounding box. Secondly, the static graph network structure lacks self-healing ability when faced with the complete absence of key anchor points, causing the matching algorithm to deviate due to topological breaks.

[0004] Therefore, the technical problem to be solved by this invention is how to construct cross-semantic interference refraction logic at the two-dimensional pixel level and reconstruct the structural integrity of the spatial topology graph by combining the shortest connected path of the pixel trajectory to eliminate the recognition manifold distortion caused by foreground clutter occlusion. Summary of the Invention

[0005] This invention proposes an automatic recognition system for matching decorative styles, the system comprising: Image acquisition unit, used to acquire two-dimensional real-scene image data; The semantic anchor extraction unit is used to extract a first set of bounding boxes with decorative style semantic labels and a second set of bounding boxes with style-neutral semantic labels from two-dimensional real-scene image data. The cross-semantic space interference screening unit is used to calculate the screening ratio and determine the corresponding initial attenuation weight in a two-dimensional pixel coordinate system based on the overlap area of ​​each first bounding box and each second bounding box, as well as the total area of ​​the first bounding boxes. The local pixel mask verification unit is used to determine the intersecting pixel area of ​​decorative style semantic tags and style neutral semantic tags in the overlapping pixel area when the initial screening ratio exceeds the preset interference threshold. The overlapping area is replaced with the intersecting pixel area, and the initial attenuation weight is updated to obtain the real node weight. The spatial topology repair unit is used to construct a spatial semantic topology graph with the center coordinates of the first bounding box as the graph node. When the weight of the real node is 0, the corresponding graph node is removed. In the two-dimensional real scene image data, the shortest connected path length between adjacent node pairs in the graph node that does not pass through the pixels inside the second bounding box is determined. The reciprocal of the shortest connected path length is determined as the edge feature weight of the virtual bridge edge. The graph matching unit is used to calculate the graph edit distance between the spatial semantic topology graph and the standard prototype topology graph, and outputs the image style matching similarity.

[0006] Preferably, the semantic anchor extraction unit is used to input two-dimensional real-scene image data into a deep convolutional neural network to extract global texture features and local semantic features; use the candidate region proposal module to locate the bounding box containing the decorative component in the local semantic features; label the bounding box through the classification branch to establish a first bounding box set and a second bounding box set respectively, and determine the vertex coordinate sequence of each bounding box in the two-dimensional pixel coordinate system.

[0007] Preferably, the local pixel mask verification unit includes the following process: Step S31: Identify the intersecting pixel region where the first bounding box and the second bounding box geometrically overlap in the two-dimensional pixel coordinate system; Step S32: Extract the first pixel mask corresponding to the first bounding box set and the second pixel mask corresponding to the second bounding box set within the intersecting pixel region; Step S33: Calculate the intersection-union ratio (IUU) of the first pixel mask and the second pixel mask; Step S34: When the IUU is lower than a preset ratio threshold, determine that the intersecting pixel region is in a non-physical occlusion state and retain the first bounding box.

[0008] Preferably, the spatial topology repair unit is used to: obtain the center pixel coordinates and category probability values ​​of each first bounding box, and map the category probability values ​​to the node weights of graph nodes; calculate the Euclidean distance between any two graph nodes and generate a connection vector; associate the connection vector with its angular offset in a preset reference coordinate system to generate edge features with rotation invariance.

[0009] Preferably, the graph matching unit is used to map the spatial semantic topological graph to a high-dimensional feature space to generate a feature vector to be identified; retrieve standard topological features from a preset library to obtain the corresponding standard feature vector; and calculate the geodesic distance between the feature vector to be identified and the standard feature vector.

[0010] Preferably, the image matching unit is also used to introduce a deviation correction factor and to perform numerical compensation on the geodesic distance according to the initial screening ratio; when there are cluttered interference signals in the two-dimensional real scene image data, the weight ratio of the overlapping area in the matching process is reduced by the deviation correction factor.

[0011] Preferably, when the semantic anchor extraction unit outputs the second bounding box set, it is also used to perform pixel-level annotation of light sources and shadows in the indoor environment, and generate a style interference distribution map based on the annotation results.

[0012] Preferably, the system further includes: a memory for storing a preset style database and a standard prototype topology map; and a processor connected to the memory for implementing the functions of an image acquisition unit, a semantic anchor extraction unit, a cross-semantic space interference screening unit, a local pixel mask verification unit, a spatial topology repair unit, and a graph matching unit by calling program instructions stored in the memory.

[0013] Compared with existing technologies, the automatic recognition system for matching decoration styles of the present invention has the following advantages: 1. In the automatic identification of decorative style matching, the coupling of anchor point semantic extraction and spatial interference verification is used to accurately remove the false occlusion noise caused by the overlapping of irregular object projections. Since the overlap of rectangular bounding boxes is not equivalent to physical occlusion at the pixel level, this scheme triggers pixel-level semantic segmentation verification of the local overlapping area when the initial screening ratio exceeds the threshold, and corrects the initial attenuation weight using the total number of real intersecting pixels. This mechanism ensures that the feature weight attenuation is based on the physical fact of pixel overlap, avoids the erroneous suppression of effective decorative features in complex visual environments, and maintains the authenticity of the underlying feature resolution.

[0014] 2. The constructed cross-semantic space interference discounting mechanism achieves dynamic control of node weights in the spatial semantic topology graph by quantitatively measuring the degree of interference of style-neutral entities on the first bounding box. When the occlusion of clutter in the indoor real-world image causes random shifts in the center coordinates of the decorative anchor point bounding box, the system pulls down the weights of the interfered nodes through the feature attenuation coefficient, reducing the pollution of the global topology structure by the coordinate shift. This process counteracts the structural noise introduced by visual distortion from the underlying logic, prevents the final output recognition manifold from undergoing geometric distortion in multi-dimensional space, and ensures the robustness of the matching judgment logic.

[0015] 3. Utilizing a topological manifold adaptive repair mechanism, the system calculates the shortest connected path around interfering entities in a two-dimensional pixel coordinate system, enabling logical self-healing of the spatial topology under extreme occlusion conditions. Even if key decorative components are completely occluded, resulting in the removal of graph nodes, the system maintains the coherence of global spatial semantic relationships by establishing virtual bridging edges between adjacent nodes and mapping the inverse of the path length to edge feature weights. This technical approach eliminates the convergence deviation of the recognition model caused by node loss, solving the problem of inaccurate judgments in traditional recognition schemes when faced with topological breaks. Attached Figure Description

[0016] Figure 1 This is a block diagram showing the processing logic and functional modules of the decorative style matching system of the present invention; Figure 2 This is an interaction use case diagram of the user and decoration style recognition business process of the system of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0018] It should be noted that all directional and positional terms used in this invention, such as: up, down, left, right, front, back, vertical, horizontal, inner, outer, top, bottom, transverse, longitudinal, center, etc., are only used to explain the relative positional relationship and connection between components in a specific state (as shown in the accompanying drawings). They are only for the convenience of describing this invention and do not require that this invention be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention. In addition, the descriptions of "first," "second," etc., in this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated.

[0019] In the description of this invention, unless otherwise explicitly specified and limited, the terms installation, connection, and linking should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections; they can refer to direct connections or indirect connections through an intermediate medium; they can refer to the internal connection of two components. For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances.

[0020] In the description of this specification, references to the terms "an embodiment," "some embodiments," "illustrative embodiments," "examples," "specific examples," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example, and the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0021] An automatic recognition system for matching interior design styles, the system comprising: Image acquisition unit, used to acquire two-dimensional real-scene image data; The semantic anchor extraction unit is used to extract a first set of bounding boxes with decorative style semantic labels and a second set of bounding boxes with style-neutral semantic labels from two-dimensional real-scene image data. The cross-semantic space interference screening unit is used to calculate the screening ratio and determine the corresponding initial attenuation weight in a two-dimensional pixel coordinate system based on the overlap area of ​​each first bounding box and each second bounding box, as well as the total area of ​​the first bounding boxes. The local pixel mask verification unit is used to determine the intersecting pixel area of ​​decorative style semantic tags and style neutral semantic tags in the overlapping pixel area when the initial screening ratio exceeds the preset interference threshold. The overlapping area is replaced with the intersecting pixel area, and the initial attenuation weight is updated to obtain the real node weight. The spatial topology repair unit is used to construct a spatial semantic topology graph with the center coordinates of the first bounding box as the graph node. When the weight of the real node is 0, the corresponding graph node is removed. In the two-dimensional real scene image data, the shortest connected path length between adjacent node pairs in the graph node that does not pass through the pixels inside the second bounding box is determined. The reciprocal of the shortest connected path length is determined as the edge feature weight of the virtual bridge edge. The graph matching unit is used to calculate the graph edit distance between the spatial semantic topology graph and the standard prototype topology graph, and outputs the image style matching similarity.

[0022] Preferably, the semantic anchor extraction unit is used to: input two-dimensional real-scene image data into a deep convolutional neural network to extract global texture features and local semantic features; use a candidate region proposal module to locate bounding boxes containing decorative components in the local semantic features; label the bounding boxes through a classification branch to establish a first bounding box set and a second bounding box set respectively, and determine the vertex coordinate sequence of each bounding box in the two-dimensional pixel coordinate system.

[0023] Preferably, the local pixel mask verification unit includes the following process: Step S31: Identify the intersecting pixel region where the first bounding box and the second bounding box geometrically overlap in the two-dimensional pixel coordinate system; Step S32: Extract the first pixel mask corresponding to the first bounding box set and the second pixel mask corresponding to the second bounding box set within the intersecting pixel region; Step S33: Calculate the intersection-union ratio (IUU) of the first pixel mask and the second pixel mask; Step S34: When the IUU is lower than a preset ratio threshold, determine that the intersecting pixel region is in a non-physical occlusion state and retain the first bounding box.

[0024] Preferably, the spatial topology repair unit is used to: obtain the center pixel coordinates and category probability values ​​of each first bounding box, and map the category probability values ​​to the node weights of graph nodes; calculate the Euclidean distance between any two graph nodes and generate a connection vector; associate the connection vector with its angular offset in a preset reference coordinate system to generate edge features with rotation invariance.

[0025] Preferably, the graph matching unit is used to map the spatial semantic topological graph to a high-dimensional feature space to generate a feature vector to be identified; retrieve standard topological features from a preset library to obtain the corresponding standard feature vector; and calculate the geodesic distance between the feature vector to be identified and the standard feature vector.

[0026] Preferably, the image matching unit is also used to introduce a deviation correction factor and to perform numerical compensation on the geodesic distance according to the initial screening ratio; when there are cluttered interference signals in the two-dimensional real scene image data, the weight ratio of the overlapping area in the matching process is reduced by the deviation correction factor.

[0027] Preferably, when the semantic anchor extraction unit outputs the second bounding box set, it is also used to perform pixel-level annotation of light sources and shadows in the indoor environment, and generate a style interference distribution map based on the annotation results.

[0028] Preferably, the system further includes: a memory for storing a preset style database and a standard prototype topology map; and a processor connected to the memory for implementing the functions of an image acquisition unit, a semantic anchor extraction unit, a cross-semantic space interference screening unit, a local pixel mask verification unit, a spatial topology repair unit, and a graph matching unit by calling program instructions stored in the memory.

[0029] Example 1: In an industrial application scenario where indoor real-world images are processed to determine the similarity of decorative style matching, the system faces the problem of visual occlusion and topological cutting caused by high-frequency semantic entities without style attributes on low-frequency style anchor points. The real-world images contain a large number of style-neutral entities. These entities cut off the visual continuity between decorative style anchor points such as wall panels of specific materials and fixed decorative lines on the two-dimensional pixel projection plane. This causes shape projection deviations in the bounding boxes output by conventional object detection models. Specifically, the bounding boxes of irregular neutral entities and curved decorative anchor points geometrically overlap, but the underlying pixels do not physically overlap. The coordinate offset caused by two-dimensional perspective distortion is converted into incorrect graph network topology edge weights, leading to image distortion. The manifold undergoes structural distortion in multidimensional space; after the image acquisition unit acquires two-dimensional real-scene image data, the semantic anchor extraction unit extracts a first set of bounding boxes with decorative style semantic labels and a second set of bounding boxes with style-neutral semantic labels from the two-dimensional real-scene image data; the cross-semantic space interference screening unit calculates the initial screening ratio for each first bounding box in the first bounding box set in the two-dimensional pixel coordinate system based on the overlap area of ​​each first bounding box and each second bounding box and the total area of ​​the first bounding boxes, and determines the corresponding initial attenuation weight to filter background nodes with no interference or slight interference, and concentrates computational resources on local matrix regions with visual occlusion risk; when the initial screening ratio is greater than the preset interference threshold, At that time, the local pixel mask verification unit extracts the local image matrix corresponding to the overlapping pixel region. Within the local image matrix, the local pixel mask verification unit uses a preset classification threshold to determine the probability value of each channel in the local image matrix. The processor generates corresponding single-channel binary mask matrices for the decorative style semantic label and the style neutral semantic label, and calculates the dot product of the two single-channel binary mask matrices at the same coordinate position. The system accumulates the total number of pixels whose dot product is not equal to 0, and determines the total number as the intersecting pixel area. The overlapping area is replaced by the intersecting pixel area, and the initial attenuation weight is updated to obtain the real node weight. The coordinate offset error caused by the projection contour of irregular objects is removed through the local pixel intersection counting process, and the spatial position state of the underlying feature resolution is maintained under the condition of using two-dimensional image hardware. When the real node weight is equal to 0, the spatial topology repair unit removes the corresponding graph node and constructs a spatial semantic topology graph with the center coordinates of the first bounding box as the graph node.

[0030] The system constructs a pixel adjacency graph in the 2D real-scene image data and marks pixels inside the second bounding box as impassable nodes. The processor uses an eight-neighbor search algorithm to find the shortest pixel sequence connecting the edges of adjacent graph nodes within the non-occluded pixel region. The length of the shortest connected path is determined by counting the number of pixel units contained in the shortest pixel sequence. Based on the principle of pixel continuity in discrete grid space, the distance between physically adjacent pixels has a non-zero baseline. The physical point-to-point distance parameter of the image acquisition unit is extracted as input, and the corresponding minimum pixel truncation value is calculated. Specifically, the calculation process of the minimum pixel truncation value establishes a mapping path from the physical surface scale of the sensor to the overall pixel scale of the digital image. The processor determines the optical magnification based on the focal length parameter of the image acquisition unit and the preset typical indoor working distance ratio, and then converts the physical point-to-point distance of the sensor (e.g., the physical width of a single pixel unit) to the equivalent pixel step size in the two-dimensional real-scene image data. The minimum pixel... The truncation value is the minimum pixel redundancy required to maintain the physical coherence of interior decorative components. By introducing this physical parameter, the system can constrain the convergence boundary of the eight-neighborhood search algorithm within a discrete grid space using the sampling resolution limitations of the underlying physical layer. This cross-scale clamping mechanism ensures that even when complex occlusion causes visual feature fragmentation, the search algorithm can still use the non-overlapping physical entities as a logical bottom line, forcibly excluding false ultra-short connected paths caused by imaging noise. This guarantees the geometric authenticity of the shortest connected path length Lsp. During the determination process, when the shortest connected path length is statistically obtained... When the value is less than the minimum pixel truncation value, the processor will use the shortest connected path length. The numerical value is assigned the minimum pixel truncation value. After completing the numerical lower limit clamping operation, the shortest connected path length is... The reciprocal of the value is used to determine the edge feature weight of the virtual bridge edge. The edge feature weights The calculation formula is: Before outputting the constructed graph structure data, based on the principle of the independence of the argument and angle of two-dimensional vectors in Euclidean space, the processor extracts the Euclidean distance of each graph node to generate a connection vector. It calculates the angle between the connection vector and the preset two-dimensional pixel horizontal axis and extracts it as an angular offset. Substituting the Euclidean distance and angular offset into the polar coordinate to complex plane standard transformation formula, it extracts the real and imaginary scalars from the calculation results to construct a two-dimensional structure matrix. This matrix is ​​then confirmed as having rotation-invariant edge features and appended to the corresponding virtual bridging edge data segment. In the specific calculation of constructing the two-dimensional structure matrix, the angular offset is not an absolute angle in the global coordinate system, but rather a relative rotation phase determined by using the connection vector between the current graph node and its adjacent nodes as a local reference axis. This is achieved by mapping the magnitude of the connection vector (i.e., the Euclidean distance) to... The radial magnitude of the complex plane is represented by the relative rotation phase, which is mapped to the argument. The real and imaginary scalars after the complex plane transformation essentially represent the geometric rigidity of the local topology. Since this mapping process is based on the relative position vectors between nodes, when the entire real-world image undergoes a global rotation, the relative displacement vectors between node pairs rotate synchronously with the local reference axis, and their mapping projection in the complex plane remains numerically constant. This gives the two-dimensional structure matrix rotation invariance against global affine transformations in algebraic logic, ensuring the stability of the subsequent graph matching process to changes in the shooting angle. The constructed spatial semantic topology graph is input to the graph matching unit, which calculates the graph editing distance between the spatial semantic topology graph and the standard prototype topology graph and outputs the image style matching similarity.

[0031] Example 2: This experiment verifies the recognition accuracy and topological stability of an automatic recognition system for matching decorative styles under conditions of high-density clutter obstruction. The experimental platform uses a graphics processing core with a floating-point computing power of ≥10 TFLOPS and a memory capacity of 32GB. The experimental data source consists of 1000 sets of real-world indoor images including French retro, American classic, and modern minimalist styles, with the image resolution uniformly adjusted to 1024×1024 pixels. To simulate imaging quality fluctuations in an industrial environment, Gaussian white noise with a signal-to-noise ratio of 25dB is actively superimposed on the original image data, and a 15% brightness grayscale deviation caused by non-uniform indoor lighting is simulated. The core parameters are preset with an interference threshold. The value of is affected by the bounding box localization accuracy output by the multi-scale feature extraction network and the pixel classification confidence of the semantic segmentation operator; its trade-off lies in the balance between the computational load of pixel-level verification and the accuracy of occlusion determination. If the value is too low, a large amount of slight geometric overlap will trigger a high load of pixel mask calculation; if the value is too high, deep occlusions that substantially damage the topology cannot be identified. By modeling the false positive rate under different occlusion ratios, the error rate can be determined. The following judgment model is met: when the positioning accuracy error is within the range of 3% to 5% and the classification confidence is greater than 0.85, it is selected to control the response time of the control system and suppress the coordinate offset deviation caused by boundary overlap. The initial weight was set to 0.15. During the experiment, a baseline was set for a typical working condition where potted plants obscured the wall lines. The original input data showed that due to the projection interference of the irregular leaves of the potted plants, the ratio of the rectangular geometric overlap area between the first and second bounding boxes in the two-dimensional coordinate system was 0.42. In the control group, the initial attenuation weight of the decorative style anchor point was determined directly using the rectangular overlap ratio, which was 0.58, resulting in the graph node being assigned a lower weight in the subsequent graph matching stage. In the experimental group, the system called the local pixel mask verification unit to determine the pixel-level classification of the overlapping area. The measured intermediate data showed that the rectangular bounding box overlap ratio reached 0.42, but the actual intersection pixel area ratio of the underlying semantic labels was 0.064. Based on this measurement, the system corrected the initial attenuation weight to 0.936 to obtain the true node weight. To verify the nonlinear compensation effect of the spatial topology repair unit, a gradient experimental group with an occlusion ratio from 10% to 90% was constructed.

[0032] When the occlusion ratio is between 10% and 50%, the image style matching similarity of the experimental group remains stable above 0.92, while the similarity of the control group decreases linearly with the increase of the occlusion ratio, dropping to 0.68 when the occlusion ratio is 50%. When the occlusion ratio enters the high-intensity interference region of 60% to 85%, the system exhibits a clear performance inflection point. The weight of the real node in the core style anchor point decays to 0 due to complete occlusion. The spatial topology repair unit determines the shortest connected path length between adjacent node pairs that does not pass through the pixels inside the second bounding box. And calculate the edge feature weights of the virtual bridge edges according to the aforementioned formula. Under extreme conditions with an occlusion rate of 80%, the experimental group maintained a graph editing distance offset of 12.5% ​​through the topology repair mechanism, while the control group experienced matching failure due to topology graph breakage. When the occlusion rate exceeded 85%, the number of remaining effective anchor points was insufficient to support global structure reconstruction, resulting in a decrease in system performance. The experimental results confirmed that the system solves the topology distortion problem caused by false occlusion at the two-dimensional image stream level through a heterogeneous semantic interference verification mechanism and a pixel trajectory obstacle avoidance algorithm. In complex real-world environments containing 25dB noise and 15% brightness disturbance, the experimental group achieved a recognition accuracy of greater than or equal to 89.6% in the range where the occlusion rate was less than 80%, which is 35.4 percentage points higher than the existing technology's 54.2%.

[0033] Example 3: When the system faces the task of comparing a dynamically reconstructed spatial semantic topology graph with a preset standard specification library, the system calls a prototype library containing structured features and cost calculation logic to eliminate spatial divergence during feature matching. The standard prototype topology graph is constructed using offline specification data and includes a prototype node feature set and a prototype edge adjacency matrix. Each node in the prototype node feature set is accompanied by a semantic label for basic decorative components, and the off-diagonal elements of the prototype edge adjacency matrix record the reciprocal of the standard spatial spacing of the basic decorative components under the standard two-dimensional physical projection plane. The graph matching unit receives the spatial semantic topology graph output by the spatial topology repair unit and the extracted standard prototype topology graph, traverses the source nodes of the spatial semantic topology graph and the target nodes of the standard prototype topology graph, and establishes a sequence of combined mapping relationships. The graph matching unit calculates the node editing cost and the edge editing cost respectively to determine the graph editing distance. For node editing costs, the graph matching unit extracts the source semantic label of the source node and the target semantic label of the target node. When the source semantic label and the target semantic label are inconsistent, the graph matching unit presets a penalty constant based on the node replacement cost. When the two semantic labels are consistent, the graph matching unit assigns the node replacement cost to 0. According to the distribution law of node potential energy conservation in non-Euclidean manifold space, the single node replacement penalty threshold covers the average maintenance cost of the surrounding associated structure during topology editing. The system's built-in pre-calibration process extracts a set of interference-free real-world samples from the standard prototype topology graph. The statistical expectation value of the edge editing cost of the sample set under the correct matching state is calculated. Based on this, the boundary coefficient is truncated according to the preset normal distribution confidence interval rule. According to the data output of the pre-calibration process, the preset penalty constant is determined according to 1.5 times the mean of the non-zero elements in the prototype edge adjacency matrix, so that semantic mismatch has a judgment weight in the total cost function.

[0034] In the cost calculation logic, the mean of edge feature weights is introduced as a benchmark for the penalty constant. A dimensionless weight benchmark based on the current spatial structure feature scale is established. Since the edge feature weights Wedge reflect the strength of topological associations between decorative components, their mean represents the average potential energy for maintaining geometric coherence of the global topological network. The penalty constant is set to 1.5 times this mean, which essentially sets a non-linear hierarchical threshold in the cost function of graph editing distance. This ensures that the node replacement cost caused by node semantic label errors (i.e., incorrect identification of decorative component categories) must exceed the local deformation cost Ce caused by geometric offset of any single topological edge. This mapping rule normalizes and associates semantic mismatches and geometric deformations of different natures through the global structural mean, enabling the algorithm to prioritize the semantic correctness of topological nodes when searching for the minimum overall cost. This prevents the system from sacrificing the semantic recognition accuracy of key decorative anchor points in order to fit local geometric deformations. For edge editing costs, the graph matching unit extracts the edge feature weights in the spatial semantic topological graph. Target weights at corresponding positions in the prototype edge adjacency matrix Calculate the absolute difference between the two and determine it as the cost of local polygon deformation. Its calculation formula is The graph matching unit uses a bipartite graph matching algorithm to accumulate the node replacement cost and local edge deformation cost under each combined mapping relationship sequence, and finds the target mapping combination that minimizes the overall cost. The minimum overall cost is determined as the graph editing distance. The graph matching unit converts the reciprocal of the graph editing distance into the image style matching similarity. The system transforms the heterogeneous spatial graph comparison into a process of obtaining the difference of determined matrix elements and discrete optimization, and outputs the state determination result while maintaining the uniformity of visual feature measurement standards.

[0035] Example 4: When the system faces the task of determining a standard reference benchmark for a specific decorative style system, the system determines the data distribution of the standard prototype topology map through an offline calibration program; in a controlled environment, it acquires 100 sets of sample images containing standard French carvings and specific partition wall panels. The image acquisition unit collects image data, and the semantic anchor point extraction unit determines the geometric centroid coordinates of each basic decorative component in the two-dimensional pixel coordinate system based on the vertex coordinate sequence and calculates the standard Euclidean distance between any two geometric centroids. To address the scaling differences in pixel distance at different imaging distances, the system will Convert to normalized spatial spacing And calculate The statistical mean and variance of the test sample set are used. The system fills the reciprocal of the normalized spatial spacing mean into the prototype edge adjacency matrix as the baseline value of the edge feature weight. At the same time, three times the variance is determined as the penalty constant of the node replacement cost in the graph matching unit, so that the values ​​of each element of the prototype edge adjacency matrix correspond to the physical interval probability distribution of a specific decorative style.

[0036] When the system faces on-site environmental conditions that include unknown imaging noise and dynamic lighting changes, a preset interference threshold is determined through a pre-calibration procedure. The parameter values; the calibration procedure acquires 10 reference images containing known grayscale distributions through the image acquisition unit, and the processor calculates the variance of the pixel brightness values ​​of the reference images. To quantize the noise power density of the imaging channel; the system based on variance And the current hardware's quantization bit depth determines the interference noise correction coefficient. Correction factor The determination logic aims to offset the impact of signal-to-noise ratio fluctuations in the imaging link on semantic segmentation confidence. Its calculation formula is defined as follows: ,in, The image quantization bit depth for the current hardware (e.g., 8-bit or 10-bit) is specified in the formula. This formula normalizes the noise intensity by ratioing the standard deviation of the noise to the full dynamic range supported by the current hardware. In the actual calibration procedure, the system uses 10 acquired reference images to calculate the fluctuation range of the average brightness. Enlargement leads to During the upgrade, the system will proportionally increase the preset interference threshold in the local pixel mask verification unit. This compensates for bounding box positioning drift caused by high photosensitive noise at the algorithm level, ensuring that the high-load pixel-level verification process is not mistakenly triggered due to background thermal noise in low signal-to-noise ratio environments. The local pixel mask verification unit extracts the reference occlusion test sequence from the memory and traverses the correction coefficients. The pixel classification accuracy variation trend within a step range of 0.1 to 0.3 is used to determine the target parameter that maximizes the intersection-union ratio (IU) of the semantic segmentation operator, and this target parameter is then set based on a preset interference threshold. The system establishes a monotonic correspondence between hardware-perceived noise and algorithm-judged thresholds through physical parameter mapping, so that the actual node weights output by the system maintain numerical characteristics consistent with the physical occlusion state.

[0037] Example 5: In integrated scenarios requiring adaptation to different imaging module sampling rates and quantization accuracies, the system invokes an offline semantic distribution calibration procedure to establish classification criteria for a multi-scale feature extraction network; it extracts 50,000 labeled images covering irregular everyday objects and standard decorative components to construct a training set, and establishes a dual-branch classification branch after the residual layer of a deep convolutional neural network. The first branch determines a first bounding box set with specific weights based on the response intensity of the decorative style feature map, and the second branch determines a second bounding box set for non-style-sensitive pixel clusters; the system calculates the mutual exclusivity of each semantic label by calculating the normalized classification entropy value of each branch, and maps the negative exponent of this entropy value to the basic interference metric value of the cross-semantic space interference screening unit. The system generates a semantic weight distribution table with cross-scene transfer capabilities without relying on manual intervention.

[0038] When the system is deployed in an environment with a camera exhibiting wide-angle distortion, the system invokes an automatic geometric mapping calibration procedure to eliminate nonlinear projection errors between the two-dimensional pixel coordinate system and the real spatial topology; the image acquisition unit acquires a preset checkerboard feature image, and the processor determines the lens distortion coefficient by detecting the pixel deviation matrix. The local pixel mask verification unit checks the coefficients. Perform an affine transformation on the coordinate space of the overlapping region to restore the true physical occupancy state; calculate the shortest connected path length under different distortion levels in the spatial semantic topology graph. The residual fluctuations are assessed, and a corrected gain operator for the current hardware environment is obtained by fitting the least squares method. The system establishes a compensation path from the underlying optical sensor parameters to the high-level graph matching weights, and establishes the engineering reproduction accuracy of the multi-mode imaging device while keeping the computational framework unchanged.

[0039] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.

Claims

1. An automatic recognition system for matching decorative styles, characterized in that, The system includes: Image acquisition unit, used to acquire two-dimensional real-scene image data; The semantic anchor extraction unit is used to extract a first set of bounding boxes with decorative style semantic labels and a second set of bounding boxes with style-neutral semantic labels from two-dimensional real-scene image data. The cross-semantic space interference screening unit is used to calculate the screening ratio and determine the corresponding initial attenuation weight in a two-dimensional pixel coordinate system based on the overlap area of ​​each first bounding box and each second bounding box, as well as the total area of ​​the first bounding boxes. The local pixel mask verification unit is used to determine the intersecting pixel area of ​​decorative style semantic tags and style neutral semantic tags in the overlapping pixel area when the initial screening ratio exceeds the preset interference threshold. The overlapping area is replaced with the intersecting pixel area, and the initial attenuation weight is updated to obtain the real node weight. The spatial topology repair unit is used to construct a spatial semantic topology graph with the center coordinates of the first bounding box as the graph node. When the weight of the real node is 0, the corresponding graph node is removed. In the two-dimensional real scene image data, the shortest connected path length between adjacent node pairs in the graph node that does not pass through the pixels inside the second bounding box is determined. The reciprocal of the shortest connected path length is determined as the edge feature weight of the virtual bridge edge. The graph matching unit is used to calculate the graph edit distance between the spatial semantic topology graph and the standard prototype topology graph, and outputs the image style matching similarity.

2. The automatic recognition system for matching decorative styles according to claim 1, characterized in that, The semantic anchor extraction unit is used to input two-dimensional real-scene image data into a deep convolutional neural network to extract global texture features and local semantic features; the candidate region proposal module is used to locate the bounding box containing the decorative component in the local semantic features; the bounding box is labeled by the classification branch, and the first bounding box set and the second bounding box set are established respectively, and the vertex coordinate sequence of each bounding box in the two-dimensional pixel coordinate system is determined.

3. The automatic recognition system for matching decorative styles according to claim 1, characterized in that, The local pixel mask verification unit includes the following process: Step S31: Identify the intersecting pixel region where the first bounding box and the second bounding box geometrically overlap in the two-dimensional pixel coordinate system; Step S32: Extract the first pixel mask corresponding to the first bounding box set and the second pixel mask corresponding to the second bounding box set within the intersecting pixel region; Step S33: Calculate the intersection-union ratio (IUGR) of the first pixel mask and the second pixel mask; Step S34: When the IUGR is lower than a preset ratio threshold, determine that the intersecting pixel region is in a non-physical occlusion state and retain the first bounding box.

4. The automatic recognition system for matching decorative styles according to claim 1, characterized in that, The spatial topology repair unit is used to: obtain the center pixel coordinates and category probability values ​​of each first bounding box, map the category probability values ​​to the node weights of graph nodes; calculate the Euclidean distance between any two graph nodes and generate a connection vector; associate the connection vector with its angular offset in a preset reference coordinate system to generate edge features with rotation invariance.

5. The automatic recognition system for matching decorative styles according to claim 1, characterized in that, The graph matching unit is used to map the spatial semantic topology graph to a high-dimensional feature space to generate a feature vector to be identified. Retrieve standard topological features from the preset library and obtain the corresponding standard feature vectors; calculate the geodesic distance between the feature vector to be identified and the standard feature vector.

6. The automatic recognition system for matching decorative styles according to claim 5, characterized in that, The image matching unit is also used to introduce a deviation correction factor and to numerically compensate for the geodesic distance based on the initial screening ratio. When there are cluttered interference signals in the two-dimensional real-scene image data, the deviation correction factor reduces the weight ratio of the overlapping area in the matching process.

7. The automatic recognition system for matching decorative styles according to claim 1, characterized in that, When outputting the second bounding box set, the semantic anchor extraction unit is also used to perform pixel-level annotation of light sources and shadows in the indoor environment, and generate a style interference distribution map based on the annotation results.

8. The automatic recognition system for matching decorative styles according to claim 1, characterized in that, The system also includes: a memory for storing a preset style database and a standard prototype topology map; and a processor connected to the memory for implementing the functions of an image acquisition unit, a semantic anchor extraction unit, a cross-semantic space interference screening unit, a local pixel mask verification unit, a spatial topology repair unit, and a graph matching unit by calling program instructions stored in the memory.

Images