Intelligent generation method based on layered recognition of painting intention

Abstract

The present disclosure provides an intelligent generation method based on layered recognition of painting intention, which comprises: obtaining a painting description text input by a user, and pre-processing the painting description text to extract layered semantic features. A geometric constraint expression model is constructed according to the layered semantic features, and geometric state parameters are generated through the geometric constraint expression model. The state transition stage is determined according to the geometric state parameters, and a spatial state evolution model of Markov chain is constructed based on the state transition stage. The optimal geometric state is determined through the spatial state evolution model, and a layered generation model driven by spatial state constraint is constructed according to the optimal geometric state, so as to generate an initial target image through the layered generation model. The initial target image is subjected to consistency test and optimization, and the optimized target image is output after the test is passed. The present application improves the consistency of the generated image in visual expression and engineering implementation by introducing geometric constraints into the state transition process.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an intelligent generation method based on hierarchical recognition of painting intentions. Background Technology

[0002] In high-precision industrial design sketching scenarios, intelligent drawing is prone to technical problems such as geometric inconsistencies. This is because in industrial product design, mechanical structure sketching, or architectural sketching generation scenarios, intelligent drawing not only requires semantic accuracy but also strict consistency in geometric proportions, dimensional relationships, and spatial topological relationships.

[0003] However, in the process of layered recognition of painting intentions, the independent modeling of the structural layer and the attribute layer can easily lead to a break in geometric constraints. For example, the perspective relationships determined by the structural layer may not match the dimensional proportions described by the attribute layer, resulting in a final generated result that is visually plausible but not practically feasible. The above problems belong to the category of missing cross-layer constraints and involve spatial continuity, which are difficult to compensate for using the simple rules of existing intelligent painting technologies. Summary of the Invention

[0004] Therefore, it is necessary to propose an intelligent generation method based on hierarchical recognition of painting intent to address the aforementioned technical problems.

[0005] The present invention adopts the following technical solution.

[0006] The first aspect of this invention discloses an intelligent generation method based on hierarchical recognition of drawing intent, the method comprising:

[0007] Obtain the painting description text input by the user, and preprocess the painting description text to extract hierarchical semantic features;

[0008] A geometric constraint expression model is constructed based on the hierarchical semantic features, and geometric state parameters are generated through the geometric constraint expression model.

[0009] The state transition stage is determined based on the geometric state parameters, and a spatial state evolution model of the Markov chain is constructed based on the state transition stage.

[0010] The optimal geometric state is determined by the spatial state evolution model, and a hierarchical generation model driven by spatial state constraints is constructed based on the optimal geometric state to generate the initial target image through the hierarchical generation model.

[0011] The initial target image is subjected to consistency check and optimization, and the optimized target image is output after the check passes;

[0012] The hierarchical semantic features are composed of structural semantic units and attribute semantic units.

[0013] Furthermore, the step of obtaining the user-input painting description text and preprocessing the painting description text to extract hierarchical semantic features includes:

[0014] The painting description text is standardized to unify it into a preset view vocabulary and measurement system, resulting in a standardized text sequence.

[0015] The adhesion strength of adjacent semantic segments is determined based on the text symbols and contextual relationships in the standardized text sequence, and the semantic segments are segmented and merged based on the adhesion strength to generate multiple semantic units.

[0016] Furthermore, the step of obtaining the user-input painting description text and preprocessing the painting description text to extract hierarchical semantic features also includes:

[0017] The frequency of use of each semantic unit in the project is calculated, and the semantic unit is encoded into a feature vector through a semantic coding model;

[0018] Based on the frequency of use of the project and the feature vector corresponding to each semantic unit, the semantic units are classified to obtain the structural semantic units and attribute semantic units, and the structural semantic units and attribute semantic units are normalized to obtain the hierarchical semantic features.

[0019] Furthermore, the step of constructing a geometric constraint expression model based on the hierarchical semantic features and generating geometric state parameters through the geometric constraint expression model includes:

[0020] Multiple types of engineering calculation information are extracted from the structural semantic units, and the engineering calculation information is mapped to corresponding parameter items to generate spatial relationship parameters for each structural semantic unit based on the parameter items.

[0021] The quantized size is extracted from the attribute class semantic unit, and the quantized size is subjected to unit unification and completion processing to obtain the size parameter;

[0022] Geometric constraint terms are constructed for the spatial relationship parameters and size parameters that satisfy the preset pairing relationship, and a standard scale is set for the geometric constraint terms to calculate the geometric deviation between the geometric constraint terms and the corresponding standard scale.

[0023] The geometric deviations are logarithmically compressed and then summarized using preset constraint weights to obtain the overall consistency score corresponding to the geometric deviations.

[0024] The geometric state parameters are composed of the geometric constraint terms, the constraint weights corresponding to the geometric constraint terms, and the overall consistency score.

[0025] Furthermore, the step of determining the state transition stage based on the geometric state parameters and constructing a spatial state evolution model of the Markov chain based on the state transition stage includes:

[0026] The geometric state parameters are split into multiple components to be compared, and each component to be compared is bound to a preset painting generation stage. The painting generation stage and the bound components to be compared are written into a Markov chain to obtain a painting generation model.

[0027] The painting generation model is transformed into multiple geometric correction actions, and candidate painting images are generated based on the current painting image when the geometric correction actions are executed. At the same time, the state transition stage from the current painting image to the candidate painting image is determined.

[0028] Furthermore, the step of determining the state transition stage based on the geometric state parameters and constructing a spatial state evolution model of the Markov chain based on the state transition stage also includes:

[0029] Calculate the geometric correction deviation of the state transition stage, determine the transition probability based on the geometric correction deviation, and determine a prohibition flag when the geometric correction deviation exceeds a set threshold;

[0030] The transition probability distribution is obtained by normalizing the transition probability using a normalization factor.

[0031] Based on the transition probability distribution, the geometric state of each candidate painting image is iteratively generated starting from the geometric state parameters, and the candidate painting images corresponding to the prohibited transition markers are removed in each iteration to construct a spatial state evolution model.

[0032] Furthermore, the step of determining the optimal geometric state through the spatial state evolution model and constructing a spatial state constraint-driven hierarchical generation model based on the optimal geometric state to generate an initial target image through the hierarchical generation model includes:

[0033] The optimal geometric state with the smallest geometric correction deviation is selected from the spatial state evolution model, and the optimal geometric state is decomposed into geometric constraint instructions to be executed, so as to convert the geometric constraint instructions into structural constraint vectors.

[0034] Based on the structural constraint vector and the matched attribute class semantic units, the structural constraint instructions and attribute constraint instructions required by the hierarchical generative model in the current state transition stage are determined by the Markov chain, and the initial target image is generated in response to the structural constraint instructions and attribute constraint instructions.

[0035] Furthermore, the step of performing consistency checks and optimizations on the initial target image, and outputting the optimized target image after the checks pass, includes:

[0036] The initial target image is compared with preset standard indicators, and the various constraints of the initial target image are reviewed to obtain the consistency verification result of the initial target image.

[0037] The consistency verification result and the optimal geometric state are aligned and compared to obtain the state difference degree. When the state difference degree is not lower than a set threshold, a difference driving quantity is generated to optimize the initial target image through the difference driving quantity.

[0038] A second aspect of this invention discloses an intelligent generation device based on hierarchical recognition of painting intent, used to implement the intelligent generation method based on hierarchical recognition of painting intent as described in any one of the first aspects, the device comprising:

[0039] The semantic analysis module is used to obtain the painting description text input by the user and preprocess the painting description text to extract hierarchical semantic features;

[0040] The state parameter generation module is used to construct a geometric constraint expression model based on the hierarchical semantic features, and generate geometric state parameters through the geometric constraint expression model.

[0041] The state evolution construction module is used to determine the state transition stage based on the geometric state parameters, and to construct the spatial state evolution model of the Markov chain based on the state transition stage.

[0042] The target image generation module is used to determine the optimal geometric state through the spatial state evolution model, and to construct a hierarchical generation model driven by spatial state constraints based on the optimal geometric state, so as to generate an initial target image through the hierarchical generation model.

[0043] The target image optimization module is used to perform consistency checks and optimizations on the initial target image, and output the optimized target image after the checks are passed.

[0044] The hierarchical semantic features are composed of structural semantic units and attribute semantic units.

[0045] A third aspect of the present invention discloses a terminal, including a processor and a storage medium;

[0046] The storage medium is used to store instructions;

[0047] The processor is configured to operate according to the instructions to perform the steps of the method described in the first aspect.

[0048] A fourth aspect of the present invention discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in the first aspect.

[0049] The present invention has the following advantages:

[0050] (1) This invention obtains the painting description text input by the user and preprocesses the painting description text to extract hierarchical semantic features. Then, a geometric constraint expression model is constructed based on the extracted hierarchical semantic features, and geometric state parameters are generated through the geometric constraint expression model. Then, the state transition stage is determined based on the generated geometric state parameters, and a spatial state evolution model of a Markov chain is constructed based on the state transition stage. This fundamentally divides the structural layer and attribute layer in the painting intention, thereby avoiding the problems of spatial proportion distortion and inconsistent topological relationships caused by the break of geometric constraints between the structural layer and the attribute layer.

[0051] (2) Based on the division of the structural layer and attribute layer in the painting intention and the determination of the state transition stage, this invention determines the optimal geometric state through a spatial state evolution model, and constructs a hierarchical generation model driven by spatial state constraints based on the optimal geometric state to generate the initial target image. Finally, the initial target image is subjected to consistency verification and optimization, and the optimized target image is output after the verification is passed. Geometric constraints are introduced into the state transition process, realizing the continuity and consistency of semantics at each level in the generation process. Combined with the verification and optimization of the generated target image, the consistency of the generated image in visual expression and engineering implementation is further improved. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0053] Figure 1 This is a flowchart illustrating the intelligent generation method based on hierarchical recognition of drawing intent provided by the present invention.

[0054] Figure 2 This is a schematic diagram of the intelligent generation device based on layered recognition of drawing intent provided by the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] like Figure 1 As shown, in one embodiment, an intelligent generation method based on hierarchical recognition of drawing intent includes the following steps:

[0057] Step S110: Obtain the painting description text input by the user, and preprocess the painting description text to extract hierarchical semantic features.

[0058] Among them, the hierarchical semantic features are composed of structural semantic units and attribute semantic units.

[0059] In some embodiments, the intelligent generation method based on hierarchical recognition of drawing intent provided by the present invention includes the following steps in step S110:

[0060] Step S111: Standardize the painting description text to unify it into a preset viewpoint vocabulary and measurement system, thus obtaining a standardized text sequence.

[0061] Step S112: Determine the adhesion strength of adjacent semantic segments based on text symbols and contextual relationships in the standardized text sequence, and perform semantic segmentation and merging based on adhesion strength to generate multiple semantic units.

[0062] In some embodiments, the intelligent generation method based on hierarchical recognition of drawing intent provided by the present invention further includes the following steps in step S110:

[0063] Step S113: Calculate the engineering usage frequency corresponding to each semantic unit, and encode the semantic unit into a feature vector through a semantic coding model. The feature vector is a numerical representation of the semantic unit obtained through the semantic coding model. It is used for semantic unit classification to obtain structural semantic units and attribute semantic units, and participates in hierarchical semantic feature normalization.

[0064] Step S114: Based on the engineering usage frequency and feature vector corresponding to each semantic unit, the semantic units are classified to obtain structural semantic units and attribute semantic units. The structural semantic units and attribute semantic units are then normalized to obtain hierarchical semantic features.

[0065] In a specific embodiment, the intelligent generation method based on hierarchical recognition of drawing intent provided by the present invention includes steps 1 to 5:

[0066] Step 1: Obtain the painting description and construct a hierarchical semantic feature set.

[0067] Includes the following sub-steps:

[0068] Sub-step 1.1: Obtaining the drawing description and standardizing the text.

[0069] Specifically, firstly, the system obtains the user-input drawing description text, which consists of one or more sentences containing engineering terms related to dimensions, perspective, and materials. For example, it could be "two-point perspective chassis sketch, front view, 120 mm wide, 60 mm high, rounded corners" or "top view gearbox structure sketch, equidistant hole spacing, coaxial assembly, aluminum alloy brushed finish on the outer shell," etc. Next, the input text undergoes symbol unification, unit unification, and synonym merging; "millimeters, centimeters, meters" are unified to the same measurement system, and "front view or front view" and "top view or top view" are unified to the same perspective vocabulary. Then, candidate semantic segments are generated based on punctuation and contextual dependencies. Furthermore, to avoid fragmenting dimension phrases, the adhesion strength of adjacent semantic segments is calculated; segments with low adhesion strength are split, and segments with high adhesion strength are merged, resulting in a sequence of candidate semantic segments.

[0070] In this embodiment, the expression for adhesion strength is: ; In the formula, For the first The adhesion strength of adjacent semantic segments, dimensionless, with a value range of 0.35-0.65; For the first The mutual information strength of adjacent semantic segments, dimensionless, is obtained by statistical analysis using a sliding window, with the sliding window ranging from 5 to 15 words. For the first The shortest path length of each adjacent semantic segment in the dependency tree, in the number of edges, with a value range of 1-10; This is the dependency distance penalty coefficient, which is dimensionless and ranges from 0.2 to 1.2.

[0071] Sub-step 1.2: Semantic unit segmentation and semantic unit set construction.

[0072] Specifically, based on the adhesion strength, after merging or segmenting the segments, semantic units are formed. Then, the frequency of use in engineering is calculated for each semantic unit. This frequency not only considers the number of occurrences, but also the phrase length and ambiguity, in order to avoid short but generalized terms from having too much weight.

[0073] The expression for the frequency of use in engineering projects is as follows: ; In the formula, For the first The engineering usage frequency of each semantic unit, dimensionless; The number of times the semantic unit appears in the input text is an integer, ranging from 1 to 20. This is a smoothing coefficient, with a value range of 0.2-2.0; This is the frequency amplification index, with a value range of 0.8-1.6; The word count length of this semantic unit is an integer ranging from 1 to 12. This is the length penalty coefficient, with a value ranging from 0.3 to 1.2; The ambiguity level is dimensionless and is obtained by combining the number of synonym hits and the number of part-of-speech conflicts. Its value ranges from 0 to 5. This is the ambiguity penalty coefficient, with a value ranging from 0.2 to 0.9.

[0074] Sub-step 1.3: Semantic coding model configuration and feature vector set generation.

[0075] Specifically, a semantic encoding model is used to encode each semantic unit into a vector. In this example, a structure combining word segmentation embedding layer, multi-layer self-attention encoder, and unit-level pooling is adopted: the embedding dimension is 256-1024; the number of encoder layers is 4-24; and the maximum input length is 32-256 words. During the encoding process, the encoding results of multiple words within the same semantic unit are weighted and pooled to obtain the vector features of that semantic unit.

[0076] Sub-step 1.4: Divide the structure class and attribute class and calculate the hierarchical weights to form an initial hierarchical semantic feature set.

[0077] Specifically, firstly, semantic units are divided into two categories based on their engineering meaning: structural units, used to express perspective, relative position, and topological relationships; and attribute units, used to express dimensions, materials, processes, chamfers, apertures, and tolerances. The division process employs a joint decision-making method combining engineering dictionary rules and a classifier with a small number of labeled samples: the engineering dictionary provides a safety net for stability; the classifier covers new terms. Subsequently, normalized weights are calculated for both structural and attribute units. These normalized weights are not only related to the aforementioned frequency of engineering usage but also need to suppress overly common but low-information terms. This example uses a normalization method with exponential and logarithmic representations. Logarithmic representations suppress extreme frequencies, while exponential representations achieve comparable normalized weight outputs. Simultaneously, specialized and scarce terms are brought back to their proper attention to prevent key geometric semantics from being overwhelmed by generalized semantics.

[0078] Step S120: Construct a geometric constraint expression model based on hierarchical semantic features, and generate geometric state parameters through the geometric constraint expression model.

[0079] The geometric constraint expression model pairs spatial relationship parameters generated by structural semantic units with dimensional parameters generated by attribute semantic units to form dimensionless proportional constraints. Each constraint is assigned a standard proportion, constraint deviation, and overall consistency score. Its construction mechanism relies on the pairing relationship between spatial relationship parameters and dimensional parameters, standard proportion reference, logarithmic compression, and constraint weight aggregation. Key parameter settings include: setting the saturation slope and quality threshold of the spatial relationship parameter confidence level to a moderate range to suppress weak semantics; prioritizing industry standards and standard parts libraries for the standard proportion; and ensuring the overall consistency score threshold and subsequent state transition threshold are of the same order of magnitude. Attribute semantic units are used to express dimensions, materials, processes, chamfers, apertures, and tolerances. Example: Width 120 mm; Radius 3.5 mm; Aluminum alloy brushed finish; Chamfer.

[0080] In some embodiments, the intelligent generation method based on hierarchical recognition of drawing intent provided by the present invention includes the following steps in step S120:

[0081] Step S121 involves extracting multiple types of engineering calculation information from structural semantic units and mapping this information to corresponding parameter items. Spatial relationship parameters for each structural semantic unit are then generated based on these parameter items. The engineering calculation information consists of structural information that can be directly mapped to parameter items, such as one-point perspective, two-point perspective, parallel or perpendicular or coaxial alignment, relative orientation, and scale reference. This information is obtained by mapping structural semantic units using a structural semantic dictionary and then filtering based on structural semantic quality scores and confidence levels.

[0082] Step S122: Extract the quantized size from the semantic unit of the attribute class, and perform unit unification and completion processing on the quantized size to obtain the size parameter.

[0083] Step S123: Construct geometric constraint terms for spatial relationship parameters and size parameters that satisfy the preset pairing relationship, and set a standard scale for the geometric constraint terms to calculate the geometric deviation between the geometric constraint terms and the corresponding standard scale.

[0084] Step S124: Logarithmically compress the geometric deviations and summarize them using preset constraint weights to obtain the overall consistency score corresponding to the geometric deviations.

[0085] The geometric state parameters are composed of geometric constraint terms, their corresponding constraint weights, and an overall consistency score. Understandably, the geometric constraint terms define the dimensionless proportional relationship ontology; the constraint weights determine the contribution of each constraint to the overall consistency score; and the overall consistency score provides a quantitative result of the overall geometric consistency and serves as the basis for state transition thresholds and prohibition criteria. These three elements together constitute the geometric state parameters.

[0086] In a specific embodiment, the intelligent generation method based on hierarchical recognition of painting intent provided by the present invention, step 2, constructing a geometric constraint expression model and extracting spatial parameters, includes the following sub-steps:

[0087] Sub-step 2.1: Extraction and parameterization encoding of structural semantics into spatial relation parameters.

[0088] Specifically, from structural semantic units, three types of calculable engineering information are extracted first: viewpoint and perspective type, relative positional relationships of objects, and scale reference relationships. During the extraction process, a structural semantic dictionary table needs to be established, mapping common industrial sketch descriptions to calculable parameter items, such as one-point perspective, two-point perspective, three-point perspective; front view, top view, side view; symmetrical, parallel, perpendicular, coaxial; relative orientation (front, back, left, right, up, down); and scale references with the same width, height, or distance as a reference. Then, a spatial relationship parameter entry is formed for each structural semantic unit, and the strength of textual support (i.e., confidence level) of this entry is calculated, which is used to suppress unreliable items when combined with dimensional parameters later.

[0089] The confidence expression for the spatial relationship parameter is as follows: ; In the formula, For the first The confidence level of the spatial relationship parameter is dimensionless and ranges from 0 to 1. This is the normalized weight of the corresponding structural semantic unit within the set of structural classes. It is dimensionless and ranges from 0 to 1. This is the weighting amplification index, with a value range of 0.8-1.8; The scarcity coefficient is dimensionless and is obtained from the inverse occurrence rate in the industry thesaurus. Its value ranges from 0 to 1000. The scarcer the product, the more professional it is and the more it should be emphasized. The saturation slope coefficient is dimensionless and ranges from 0.8 to 3.0. The structural semantic quality score is dimensionless and is obtained by weighting three factors: "whether it contains explicit perspective words", "whether it contains computable geometric relation words", and "whether it contains benchmark words". The value range is 0-5. The quality threshold is dimensionless and ranges from 1.5 to 3.5.

[0090] Sub-step 2.2 involves extracting attribute semantics into size parameters, unifying units, and completing dimensions.

[0091] Specifically, quantifiable dimensions are extracted from attribute semantics. In this example, dimensions include numerical values ​​and units (e.g., width 120 mm, radius 3.5 mm) or proportional relationships (e.g., height is twice the width, hole spacing is one-third of the outer diameter). First, units are standardized; in this example, millimeters or meters are used to avoid inconsistencies in subsequent proportional calculations. Then, missing items are completed: if only the width-to-height ratio is given without absolute dimensions, a scale reference is used as the entry point for completion, for example, using the reference side length in structural parameters or the dimensions of known standard parts as anchors. Finally, for each dimension item, a usability score (range 0-1, dimensionless) is set so that clearly defined and implementable dimensions are prioritized when combining constraints later.

[0092] Sub-step 2.3: Construct a geometric constraint expression model and generate standard scale reference terms.

[0093] Specifically, for each pair of spatial relationship parameters and dimensional parameters that satisfy the pairing relationship, geometric constraint terms are constructed. During the constraint construction process, spatial relationship parameters must provide a reference scale or perspective scale coefficient, and dimensional parameters must provide absolute dimensions or landable proportional dimensions. Only after pairing can they constitute a dimensionless proportional constraint. Furthermore, to ensure dimensional consistency, constraints should be uniformly in a dimensionless form, with standard proportions referenced from industry standards (e.g., common chamfer proportions), standard parts libraries (e.g., the relationship between threaded holes and outer diameters), or proportions explicitly specified by the user.

[0094] In this embodiment, the constructed geometric constraint function expression is: ; In the formula, For the first The geometric constraint function values ​​are dimensionless. The dimensions after pairing are expressed in millimeters or meters. This is the spatial reference dimension after pairing, and the unit is consistent with the unit of the paired dimension. To prevent tiny constants with a denominator of zero; Score the availability of the dimensional parameters corresponding to the current constraint. The confidence level of the spatial relation parameter corresponding to the current constraint term is dimensionless. This is the corresponding saturation slope coefficient, dimensionless, with a value range of 1.0-4.0; The confidence threshold is dimensionless and ranges from 0.35 to 0.65.

[0095] In addition, a standard scale reference needs to be generated for each constraint item, and the standard reference is saved in the form of a dimensionless scale. If the source is a standard parts library or specification library, the standard scale can be directly found; if the source is user-specified or derived from a scale statement, it should be used as a reference, and the standard reference should be uniformly recorded as the target scale value of the constraint item.

[0096] Sub-step 2.4: Calculate the geometric constraint deviation and form the geometric state vector.

[0097] Specifically, the deviation between each geometric constraint and the standard scale reference is calculated. This deviation must remain dimensionless to facilitate cross-constraint comparisons. Furthermore, to prevent an excessively large deviation from leading to overall instability, the deviations are logarithmically compressed and aggregated using constraint weights to obtain an overall consistency score.

[0098] The expression for the constraint deviation is: ; In the formula, For the first The constraint deviation of the geometric constraint term is dimensionless, and the smaller the value, the higher the consistency. For the first The geometric constraint function values ​​of the geometric constraint terms; For standard proportion reference; It is a reliability enhancement index, dimensionless, with a value range of 0.6-1.8; Score the availability of the dimensional parameters corresponding to the current constraint. This represents the confidence level of the spatial relation parameter corresponding to the current constraint term, and is dimensionless.

[0099] In this embodiment, the overall consistency score (dimensionless, ranging from 0 to 0.2) is aggregated in a weighted manner so that it can be used as a constraint threshold index during subsequent state transitions. The deviations and reliability of all constraint items are packaged with the overall score into a geometric state vector.

[0100] Step S130: Determine the state transition stage based on the geometric state parameters, and construct the spatial state evolution model of the Markov chain based on the state transition stage.

[0101] In some embodiments, the intelligent generation method based on hierarchical recognition of drawing intent provided by the present invention includes the following steps in step S130:

[0102] Step S131: The geometric state parameters are split into multiple components to be compared, and each component to be compared is bound to a preset painting generation stage. The painting generation stage and the bound components to be compared are written into a Markov chain to obtain the painting generation model.

[0103] The components to be compared are dimensionless components of the geometric state parameters that can be used for state comparison. They are obtained by decomposing them according to the overall geometric consistency score, the average deviation of the subset of structure-related constraints, the average deviation of the subset of size-related constraints, and the maximum deviation of the key constraints. The painting generation model is a generative-driven model that binds the painting generation stage with the geometric state parameters and writes them into a Markov chain: using the geometric state parameters as state representations and geometric correction actions as controllable inputs, it generates candidate painting images and simultaneously generates candidate geometric states. Its settings mainly include: the action amplitude adopts a dimensionless proportional adjustment amount and is gated by stage; perspective intensity correction and topology-related actions are given priority in the structural stage; local detail correction is added in the detail stage; and consistency evaluation must be performed on the geometric state parameters after each generation.

[0104] Step S132: The painting generation model is transformed into multiple geometric correction actions, and candidate painting images are generated based on the current painting image when the geometric correction actions are executed. At the same time, the state transition stage from the current painting image to the candidate painting image is determined.

[0105] In some embodiments, the intelligent generation method based on hierarchical recognition of drawing intent provided by the present invention further includes the following steps in step S130:

[0106] Step S133: Calculate the geometric correction deviation during the state transition phase, determine the transition probability based on the geometric correction deviation, and simultaneously determine the prohibition of transition flag when the geometric correction deviation exceeds a set threshold.

[0107] The geometric correction bias is the total geometric deviation of the candidate next state relative to the standard proportion, obtained by summing the geometric deviations of each candidate geometric state through constraint weights.

[0108] Step S134: Normalize the transition probability using a normalization factor to obtain the transition probability distribution.

[0109] Step S135: Based on the transition probability distribution, the geometric state of each candidate painting image is iteratively generated starting from the geometric state parameters, and the candidate painting images corresponding to the prohibited transition markers are removed in each iteration to construct a spatial state evolution model.

[0110] Among them, the spatial state evolution model is a spatial state sequence generation mechanism based on Markov chains. It uses geometric state parameters as states, generates candidate states using geometric correction actions, iterates according to the transition probability distribution to obtain a spatial state sequence that satisfies the continuity constraint, and then outputs the optimal geometric state.

[0111] In a specific embodiment, the intelligent generation method based on hierarchical recognition of painting intent provided by the present invention, step 3, constructing a spatial state evolution model based on a Markov chain, includes the following sub-steps:

[0112] Sub-step 3.1: Unpacking the initial geometric state vector, defining the state space, and mapping it to the drawing generation stage.

[0113] Specifically, the initial geometric state vector is decomposed into several comparable dimensionless components, and each component is bound to a specific painting generation stage, thus substituting the painting model into a Markov chain. In this example, painting generation is divided into several stages: the sketch structure stage (constraining only perspective and topology), the form stage (constraining main size proportions), the detail stage (constraining local apertures, chamfers, and tolerance performance), and the rendering stage (not adding new geometric constraints, only maintaining the aforementioned constraints). Therefore, the state is defined as a geometric deviation profile under the current stage, and a set of states is constructed. For any geometric state, a combination of overall score, hierarchical score, and key constraint deviation is used, where all components are dimensionless, facilitating unified threshold control. Furthermore, the implementation form of the geometric state includes the overall geometric consistency score, the average deviation of the structurally related constraint subset, the average deviation of the size-related constraint subset, and the maximum deviation of the key constraints. To ensure the controllability of subsequent transition probability calculations, stage weights and stage target deviations need to be defined first, and then written into the state machine. Among them, the stage weights need to satisfy the condition that the structural stage places more emphasis on perspective and topology, and the form stage places more emphasis on main size proportions.

[0114] Sub-step 3.2: Generation of state transition candidates.

[0115] Specifically, Markov chains require that the next state depends only on the current state. Therefore, in this example, the controllable input of the painting generation model needs to be abstracted into a finite number of geometric correction actions, so that executing a certain action can generate a candidate next deviation image from the current geometric deviation image. In this process, the entire image is not directly used as the state, but rather the geometric constraint deviations that can be resolved in the image are used as the state; at the same time, the control variables of the generation model are abstracted into geometric correction actions. The sources of these geometric correction actions include perspective intensity correction (changing the degree of perspective convergence), scale reference correction (adjusting the reference length and the ratio of the main scale), and local detail correction (only for aperture, chamfer, and spacing constraints). Each type of action is assigned an action amplitude, forming an action set. The action amplitude is a dimensionless proportional adjustment to avoid dimensional inconsistencies.

[0116] Subsequently, based on the influence of actions on constraint deviations, candidate next states are constructed. This example does not require precise physical modeling, but it must remain monotonic and controllable. For example, perspective correction actions mainly affect perspective and relative position constraint deviations, scale reference correction mainly affects principal scale ratio deviations, and local detail correction mainly affects aperture and chamfering deviations. Finally, these influences are recorded as an action influence coefficient table to update each constraint deviation component when generating candidate next states, and to output a set of candidate next states.

[0117] Sub-step 3.3 defines the transition probability, normalization factor, and transition prohibition condition.

[0118] Specifically, for each candidate next state, its total geometric deviation needs to be calculated, and the transition probability is defined accordingly. The total geometric deviation should remain dimensionless and reflect the principle that key constraints are more important; it cannot be simply averaged. Therefore, this example uses a reliability-weighted power summation to enhance sensitivity to large deviations, while introducing an action magnitude penalty to avoid achieving short-term deviation reductions through excessively large actions but disrupting continuity. Finally, the transition probability is defined based on the total deviation and normalized to a probability distribution using a normalization factor.

[0119] In this embodiment, transfer is prohibited when the total deviation or its equivalent total deviation exceeds the threshold. The threshold value should correspond to the magnitude of the overall consistency score formed in step 2, and the value range is 0.05-0.20. In implementation, the threshold can be bound to the stage: the threshold is stricter in the structural stage and slightly more lenient in the detailed stage.

[0120] Sub-step 3.4: State sequence generation, stability determination, and output of the optimal stable state.

[0121] Specifically, starting from the initial state, a state sequence is iteratively generated. In each iteration, prohibited candidate transitions are first eliminated, then sampling is performed based on the transition probability, or the action with the highest probability is selected to generate the next state. After generating the next state, the stage weights and action magnitudes are updated before proceeding to the next round. Furthermore, to ensure continuity, a stability criterion is introduced. That is, when the overall geometric consistency score no longer decreases significantly over multiple consecutive steps, and the maximum deviation of the key constraints is below a threshold, a stable state is determined. This stability criterion includes both convergence speed and upper bound constraints; for example, the process stops if the improvement is insufficient for a set number of consecutive steps and the maximum deviation is controlled.

[0122] Step S140 involves determining the optimal geometric state through a spatial state evolution model and constructing a hierarchical generation model driven by spatial state constraints based on this optimal geometric state. This hierarchical generation model then generates the initial target image. The spatial state constraints are perspective constraints, topological constraints, and scale reference constraints obtained by decoding the optimal geometric state. These constraints are set through a structural constraint list, structural control vectors, and stage gating variables, and are executed according to stage threshold constraints. In essence, the hierarchical generation model is a model that performs staged generation under the drive of spatial state constraints. It uses structural control vectors to constrain spatial topological relationships, attribute control vectors to control detail attributes, and performs geometric deviation evaluation and constraint correction at each generation stage to output the initial target image.

[0123] In some embodiments, the intelligent generation method based on hierarchical recognition of drawing intent provided by the present invention includes the following steps in step S140:

[0124] Step S141: Select the optimal geometric state with the smallest geometric correction deviation from the spatial state evolution model, and decompose the optimal geometric state into geometric constraint instructions to be executed, so as to transform the geometric constraint instructions into structural constraint vectors.

[0125] Step S142: Based on the structural constraint vector and the matched attribute class semantic units, the structural constraint instructions and attribute constraint instructions required by the hierarchical generative model in the current state transition stage are determined through a Markov chain, and the initial target image is generated in response to the structural constraint instructions and attribute constraint instructions.

[0126] In a specific embodiment, the intelligent generation method based on hierarchical recognition of painting intent provided by the present invention, step 4, driving the hierarchical generation model based on spatial state constraints, includes the following sub-steps:

[0127] Sub-step 4.1 decodes the optimal geometry into an executable list of structural constraints and aligns it with the Markov chain stage index.

[0128] Specifically, the optimal geometric state of a Markov chain is essentially the state with the smallest geometric deviation and stable transition at a certain generation stage. Therefore, to substitute it into the painting model, this state with the smallest geometric deviation and stable transition must be decomposed into constraint instructions that the painting model can execute. In this example, the optimal geometric state is decomposed into three types of structural constraints: perspective constraints (e.g., one- or two-point perspective and convergence strength), topological constraints (e.g., parallel, perpendicular, coaxial, or symmetric relationships), and scale reference constraints (e.g., reference edge length, main width-to-height ratio). Simultaneously, the stage index obtained in step 3 is used as the current stage switch of the painting model to ensure that the painting model is controlled gradually in stages, rather than forcing all constraints at once. The stage index is obtained by mapping from the state number, and this mapping table is read and aligned. Furthermore, to ensure that subsequent control quantities are dimensionless, scale-related items in the structural constraint list are uniformly converted into proportional constraints, i.e., using the reference scale as the denominator to form a dimensionless ratio.

[0129] Sub-step 4.2 generates structural control vectors from the structural constraint list and explicitly preserves the engineering meaning of the vector elements.

[0130] Specifically, the structural control vector must be consumable by the generative model, and each element within it must correspond to a clear engineering meaning to avoid black-box vectors that cannot be tuned. In this example, the structural control vector is concatenated into four segments: perspective segment, topology segment, scale segment, and stage segment. The perspective segment contains perspective type encoding and convergence strength; the topology segment contains the strength of parallel, perpendicular, coaxial, and symmetric relationships; the scale segment contains the main aspect ratio and critical length ratio; and the stage segment is the gating variable for the current generation stage. All elements in each segment are dimensionless, ranging from 0 to 1 or -1 to 1 (used to represent directionality). The generation of the structural control vector uses a combination of constraint strength weighted summation and normalization compression to ensure numerical stability.

[0131] Sub-step 4.3: Integrate attribute features to form an attribute control vector, and use stage gating to implement the substitution method of first structure and then details.

[0132] Specifically, the key to incorporating the painting model into a Markov chain lies in stage gating. That is, the Markov chain represents the most stable geometric state at a given stage. Therefore, the generated model must prioritize satisfying structural constraints at that stage before gradually introducing attribute details; otherwise, details may be introduced prematurely, leading to structural failure. In this process, the attribute control vector is first split into two parts: dimensional details (apertures, chamfers, and spacing strongly correlated with geometry) and appearance details (material textures and colors weakly correlated with geometry). Then, stage gating controls the introduction ratio; that is, the structural stage primarily introduces structural control and dimensional details, while the rendering stage introduces appearance details. The stage gating value is directly taken from the stage segment elements of the structural control vector or calculated from the stage index.

[0133] Sub-step 4.4: Jointly generate, evaluate and correct deviations within the stage, and output the initial generated image.

[0134] Specifically, the structural control vector and attribute control vector are input into the generative model to perform hierarchical guided generation. The generative model used is either diffusion-type or adversarial-type and must support conditional control input. Furthermore, to ensure the Markov chain is applied, execution must be strictly phased; that is, each phase generates an intermediate image and calculates the geometric deviation for that image. The definition of the deviation must maintain the same dimensionality (dimensionless proportional deviation) as in steps 2 or 3. To avoid simply calculating absolute differences, a deviation form combining reliability weighting, logarithmic compression, and action penalties is used, and then compared with a threshold: exceeding the threshold triggers constraint correction, i.e., increasing the structural control strength, decreasing the attribute gating, or reducing the generation step size, so that the next round is closer to the optimal geometric state.

[0135] In this embodiment, correction is triggered when the geometric deviation exceeds a set threshold, which ranges from 0.05 to 0.15. The threshold is adjusted in stages, with a stricter threshold in the structural stage and a slightly wider threshold in the detail stage. The corresponding correction amount is a combination of enhanced structural control, weakened attribute gating, and reduced step size, ensuring that the correction magnitude is controllable. During correction, the structural control vector is increased proportionally with the structural segment elements corresponding to the correction intensity, and the gating amount of the attribute control vector is decreased proportionally with the correction intensity, taking effect in the next generation round. Finally, after each stage iteration, the initial target image is output.

[0136] Step S150: Perform consistency check and optimization on the initial target image, and output the optimized target image after the check passes.

[0137] In some embodiments, the intelligent generation method based on hierarchical recognition of drawing intent provided by the present invention includes the following steps in step S150:

[0138] Step S151: Compare the initial target image with the preset standard indicators and review the various constraints of the initial target image to obtain the consistency verification result of the initial target image.

[0139] Step S152: Align and compare the consistency verification result with the optimal geometric state to obtain the state difference degree. When the state difference degree is not lower than a set threshold, generate the difference driving quantity to optimize the initial target image through the difference driving quantity.

[0140] In a specific embodiment, the intelligent generation method based on hierarchical recognition of drawing intent provided by the present invention, step 5, performing geometric consistency feedback optimization and outputting the final result, includes the following sub-steps:

[0141] Sub-step 5.1: Extract reference scale and size from the initial generated image to form a recalcible parameter set.

[0142] Specifically, the measurable geometric quantities in the initial target image are restored to reference scales and dimensions consistent with those in step 2 above, thus ensuring dimensional consistency in subsequent deviation calculations. In this process, the reference scale parameter set is used to carry the perspective datum, the length of the principal datum side, and the datum plane scale; the dimension parameter set is used to carry the external dimensions (width and height), aperture, chamfer, and aperture spacing. To ensure dimensional consistency, all measurable lengths are unified to the same unit system (e.g., pixel length), and then converted to a dimensionless scale using the reference scale. If absolute size anchoring exists (e.g., the user-defined width is 120 mm and is labeled in the image), a pixel-to-millimeters scaling factor is simultaneously output, but subsequent constraints still primarily use dimensionless scales to avoid unit drift.

[0143] Sub-step 5.2: Calculate the dimensionless proportional constraint value and deviation, and construct the feedback state.

[0144] Specifically, for each constraint, the current dimensionless scale value is obtained by dividing the size by the reference scale value. Then, the deviation is calculated by subtracting this from the standard scale reference. This must be completely consistent with the expression in step 2 above; that is, both the scale value and the deviation are dimensionless. Furthermore, considering the uncertainty in the extraction process, confidence levels are used to weight the deviations, and logarithmic compression is applied to extreme deviations to prevent a single constraint from deviating from the feedback drive direction due to local analysis failure. Finally, all constraint deviations and the overall summary value are packaged to form the feedback state, and the corresponding overall deviation is obtained by weighted averaging.

[0145] Sub-step 5.3: Quantify the difference between the feedback state and the optimal state and determine convergence.

[0146] Specifically, the feedback state is aligned and compared with the optimal geometric state. The alignment principle is to compare using the same constraint term at the same location, without changing the sign or meaning. The difference degree should reflect both the overall deviation gap and the maximum gap of key constraints (e.g., perspective and master scale). Simultaneously, the difference is adjusted by confidence level to prevent unreliable extraction from causing false convergence or false divergence. The difference degree is obtained by combining a weighted generalized norm with a maximum term penalty. When the difference degree is less than a set threshold, convergence is determined; otherwise, transition probability updates and regeneration are performed. This convergence threshold is consistent with the magnitude of the geometric deviation, taking a small range based on the target interval of the overall consistency score, and using a more stringent value during the structural stage.

[0147] Sub-step 5.4 updates the Markov transition probabilities and triggers regeneration until the final image is output.

[0148] Specifically, if the result of sub-step 5.3 fails to converge, the current feedback state is reintegrated into the Markov chain's transition framework. That is, the feedback state is considered the current state, and the optimal state is used as the target attractor. By updating the temperature coefficient and action amplitude penalty, the transition probability is made more biased towards corrective actions that reduce the difference. Here, it's emphasized how the painting model is incorporated into the Markov chain; that is, the actions of the Markov chain still correspond to the geometric correction actions defined in step 3 (i.e., perspective correction, scale reference correction, and local detail correction). Furthermore, after each action selection, the state is not directly changed, but rather the action is converted into adjustments to the structural control vector and attribute control vector. The generative model is then called to regenerate the target image, and the generated image is then parsed into a new feedback state. Simultaneously, to achieve difference-driven transition updates, the temperature coefficient needs adaptive adjustment: the greater the difference, the more inclined towards exploration; the smaller the difference, the more inclined towards greedy convergence. At the same time, the action penalty is increased to avoid final-stage jitter. Finally, the new transition probability is calculated using the updated temperature coefficient, the action is selected and mapped to the control vector for adjustment, the feedback state is regenerated, and the process continues until convergence, generating and outputting the target image that meets the engineering requirements.

[0149] The intelligent generation device based on hierarchical recognition of painting intentions provided by the present invention will be described below. The intelligent generation device based on hierarchical recognition of painting intentions described below can be referred to in correspondence with the intelligent generation method based on hierarchical recognition of painting intentions described above.

[0150] like Figure 2 As shown, in one embodiment, an intelligent generation device based on hierarchical recognition of drawing intent includes a semantic analysis module, a state parameter generation module, a state evolution construction module, a target image generation module, and a target image optimization module.

[0151] The semantic analysis module is used to obtain the painting description text input by the user and preprocess the painting description text to extract hierarchical semantic features.

[0152] The state parameter generation module is used to construct a geometric constraint expression model based on hierarchical semantic features, and to generate geometric state parameters through the geometric constraint expression model.

[0153] The state evolution building module is used to determine the state transition stages based on the geometric state parameters, and to build a spatial state evolution model of the Markov chain based on the state transition stages.

[0154] The target image generation module is used to determine the optimal geometric state through a spatial state evolution model, and to construct a hierarchical generation model driven by spatial state constraints based on the optimal geometric state, so as to generate the initial target image through the hierarchical generation model.

[0155] The target image optimization module is used to perform consistency checks and optimizations on the initial target image, and outputs the optimized target image after the check passes.

[0156] Among them, the hierarchical semantic features are composed of structural semantic units and attribute semantic units.

[0157] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0158] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. An intelligent generation method based on hierarchical recognition of drawing intent, characterized in that, The method includes: Obtain the painting description text input by the user, and preprocess the painting description text to extract hierarchical semantic features; A geometric constraint expression model is constructed based on the hierarchical semantic features, and geometric state parameters are generated through the geometric constraint expression model. The state transition stage is determined based on the geometric state parameters, and a spatial state evolution model of the Markov chain is constructed based on the state transition stage. The optimal geometric state is determined by the spatial state evolution model, and a hierarchical generation model driven by spatial state constraints is constructed based on the optimal geometric state to generate the initial target image through the hierarchical generation model. The initial target image is subjected to consistency check and optimization, and the optimized target image is output after the check passes; The hierarchical semantic features are composed of structural semantic units and attribute semantic units.

2. The intelligent generation method based on hierarchical recognition of drawing intent according to claim 1, characterized in that, The process of obtaining the user-input painting description text and preprocessing the painting description text to extract hierarchical semantic features includes: The painting description text is standardized to unify it into a preset view vocabulary and measurement system, resulting in a standardized text sequence. The adhesion strength of adjacent semantic segments is determined based on the text symbols and contextual relationships in the standardized text sequence, and the semantic segments are segmented and merged based on the adhesion strength to generate multiple semantic units.

3. The intelligent generation method based on hierarchical recognition of drawing intent according to claim 2, characterized in that, The step of obtaining the user-input painting description text and preprocessing the painting description text to extract hierarchical semantic features also includes: The frequency of use of each semantic unit in the project is calculated, and the semantic unit is encoded into a feature vector through a semantic coding model; Based on the frequency of use of the project and the feature vector corresponding to each semantic unit, the semantic units are classified to obtain the structural semantic units and attribute semantic units, and the structural semantic units and attribute semantic units are normalized to obtain the hierarchical semantic features.

4. The intelligent generation method based on hierarchical recognition of drawing intent according to claim 1, characterized in that, The step of constructing a geometric constraint expression model based on the hierarchical semantic features and generating geometric state parameters through the geometric constraint expression model includes: Multiple types of engineering calculation information are extracted from the structural semantic units, and the engineering calculation information is mapped to corresponding parameter items to generate spatial relationship parameters for each structural semantic unit based on the parameter items. The quantized size is extracted from the attribute class semantic unit, and the quantized size is subjected to unit unification and completion processing to obtain the size parameter; Geometric constraint terms are constructed for the spatial relationship parameters and size parameters that satisfy the preset pairing relationship, and a standard scale is set for the geometric constraint terms to calculate the geometric deviation between the geometric constraint terms and the corresponding standard scale. The geometric deviations are logarithmically compressed and then summarized using preset constraint weights to obtain the overall consistency score corresponding to the geometric deviations. The geometric state parameters are composed of the geometric constraint terms, the constraint weights corresponding to the geometric constraint terms, and the overall consistency score.

5. The intelligent generation method based on hierarchical recognition of drawing intent according to claim 1, characterized in that, The step of determining the state transition stage based on the geometric state parameters and constructing a spatial state evolution model of the Markov chain based on the state transition stage includes: The geometric state parameters are split into multiple components to be compared, and each component to be compared is bound to a preset painting generation stage. The painting generation stage and the bound components to be compared are written into a Markov chain to obtain a painting generation model. The painting generation model is transformed into multiple geometric correction actions, and candidate painting images are generated based on the current painting image when the geometric correction actions are executed. At the same time, the state transition stage from the current painting image to the candidate painting image is determined.

6. The intelligent generation method based on hierarchical recognition of drawing intent according to claim 5, characterized in that, The step of determining the state transition stage based on the geometric state parameters and constructing a spatial state evolution model of the Markov chain based on the state transition stage further includes: Calculate the geometric correction deviation of the state transition stage, determine the transition probability based on the geometric correction deviation, and determine a prohibition flag when the geometric correction deviation exceeds a set threshold; The transition probability distribution is obtained by normalizing the transition probability using a normalization factor. Based on the transition probability distribution, the geometric state of each candidate painting image is iteratively generated starting from the geometric state parameters, and the candidate painting images corresponding to the prohibited transition markers are removed in each iteration to construct a spatial state evolution model.

7. The intelligent generation method based on hierarchical recognition of drawing intent according to claim 6, characterized in that, The step of determining the optimal geometric state through the spatial state evolution model and constructing a spatial state constraint-driven hierarchical generation model based on the optimal geometric state to generate an initial target image through the hierarchical generation model includes: The optimal geometric state with the smallest geometric correction deviation is selected from the spatial state evolution model, and the optimal geometric state is decomposed into geometric constraint instructions to be executed, so as to convert the geometric constraint instructions into structural constraint vectors. Based on the structural constraint vector and the matched attribute class semantic units, the structural constraint instructions and attribute constraint instructions required by the hierarchical generative model in the current state transition stage are determined by the Markov chain, and the initial target image is generated in response to the structural constraint instructions and attribute constraint instructions.

8. The intelligent generation method based on hierarchical recognition of drawing intent according to claim 1, characterized in that, The process of performing consistency checks and optimizations on the initial target image, and outputting the optimized target image after the checks pass, includes: The initial target image is compared with preset standard indicators, and the various constraints of the initial target image are reviewed to obtain the consistency verification result of the initial target image. The consistency verification result and the optimal geometric state are aligned and compared to obtain the state difference degree. When the state difference degree is not lower than a set threshold, a difference driving quantity is generated to optimize the initial target image through the difference driving quantity.

9. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-8.

Images