An adaptive deep learning model-based industrial vision dimensional measurement system

Abstract

This invention discloses an industrial visual dimensional measurement system based on an adaptive deep learning model, comprising the following steps: acquiring image sequence data of a target workpiece and establishing a time stamp to form temporally continuous visual input data; modeling the temporally continuous visual input data using a liquid time constant neural network to generate a visual feature representation reflecting the temporal changes of the target workpiece; generating a structured geometric expression of the target workpiece based on the visual feature representation, and further calculating a set of dimensional parameters; constraining the set of dimensional parameters by introducing an improved geometric consistency constraint model to ensure geometric consistency between the dimensional parameters; and finally outputting the dimensional measurement result of the target workpiece based on the geometrically consistent dimensional parameter set. This invention is applicable to visual dimensional measurement applications under continuous operation conditions in industrial production lines.

Description

Technical Field

[0001] This invention relates to the field of industrial vision inspection and intelligent measurement technology, and in particular to an industrial vision dimension measurement system based on an adaptive deep learning model. Background Technology

[0002] Industrial vision dimension measurement technology is widely used in industrial scenarios such as machinery manufacturing, electronic assembly and automated production lines. It achieves automatic measurement of key dimensions of workpieces by imaging the target workpiece and combining it with algorithm analysis. Existing industrial vision dimension measurement systems are usually built based on fixed imaging conditions and static image processing methods. They rely on edge extraction, template matching or dimension regression models based on convolutional neural networks to analyze single-frame images in order to obtain the dimension information of the workpiece.

[0003] In the continuous operation of actual industrial production lines, factors such as changes in lighting, equipment vibration, workpiece posture changes, and frequent changes in workpiece models make it difficult for single-frame or static models to stably represent the true geometric structure of the workpiece, resulting in large fluctuations in dimensional measurement results. Although some existing technologies have introduced deep learning models for dimensional prediction, these models mostly adopt discrete-time modeling methods, and the internal response characteristics of the neural network are fixed, making it difficult to adaptively adjust with changes in input. Under long-term operation or changing operating conditions, prediction drift is prone to occur.

[0004] Existing dimensional measurement methods largely rely on data-driven regression results, lacking systematic constraints on workpiece geometry. When the workpiece structure is complex or partially obscured, dimensional parameters are prone to non-geometrical relationships. Even when some technologies introduce geometric constraints, these constraints are typically based on static geometric relationships within a single frame, failing to fully utilize temporal information during continuous measurement and neglecting to hierarchically process the dependencies between different geometric constraints, thus making it difficult to effectively suppress cumulative errors.

[0005] Existing technologies generally suffer from the following shortcomings: First, they lack measurement models that can adaptively model time-continuous visual data, making it difficult to adapt to dynamic working conditions on industrial production lines; second, they rely excessively on single prediction results during dimensional measurement and lack a systematic geometric consistency constraint mechanism; and third, the geometric constraint models are mostly static and one-time constraints, failing to incorporate historical measurement results for recursive constraints, resulting in insufficient long-term operational stability.

[0006] Therefore, how to provide an industrial vision dimension measurement system with an adaptive deep learning model is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose an industrial vision dimension measurement system based on an adaptive deep learning model. This invention introduces a liquid time constant neural network to model the time-continuous visual data in industrial production lines, and combines an improved geometric consistency constraint model to constrain the workpiece dimension parameters, thereby realizing automatic measurement of the target workpiece dimension. It has the advantages of strong adaptability to complex working conditions, high stability of dimension measurement, and good measurement consistency under continuous operation conditions.

[0008] An industrial visual size measurement system based on an adaptive deep learning model according to an embodiment of the present invention includes the following steps: The visual timing acquisition module is used to acquire image sequence data of the target workpiece during the continuous operation of the industrial production line, and to establish a corresponding time marker for each frame of the image sequence data to generate time-continuous visual input data. The time-adaptive feature modeling module constructs a liquid time constant neural network, which models the continuous visual input data through a continuous time state evolution mechanism. This enables the time response characteristics of the neurons inside the liquid time constant neural network to adaptively adjust with the changes in the continuous visual input data, and outputs a time-dynamic visual feature representation. The geometric structure generation module generates a structured geometric representation of the target workpiece based on time-dynamic visual feature representation. The structured geometric representation is used to characterize the geometric boundary relationships and structural association relationships of the target workpiece. The dimension generation module generates a set of dimension parameters for the target workpiece based on structured geometric representation. The geometric consistency constraint module is used to introduce the set of dimensional parameters into the improved geometric consistency constraint model using a structured geometric expression as a geometric reference, perform geometric consistency constraint processing on the set of dimensional parameters, and generate a set of dimensional parameters subject to geometric consistency constraints. The dimension output module is used to output the dimension measurement results of the target workpiece based on a set of dimension parameters constrained by geometric consistency.

[0009] Optionally, the visual temporal acquisition module includes: The image sequence acquisition unit is used to continuously acquire images of the same target workpiece during the continuous operation of an industrial production line, and to obtain image sequence data arranged in the order of acquisition. The image sequence data consists of multiple frames of target workpiece images, and each frame of target workpiece image contains the imaging area of ​​the target workpiece and the corresponding image pixel data. An imaging parameter acquisition unit is used to synchronously acquire imaging parameter information corresponding to each frame of the target workpiece image during the image sequence acquisition process. The imaging parameter information includes at least exposure parameters, gain parameters, and focal length parameters, and establishes a correlation with the corresponding frame of the target workpiece image. The time stamp generation unit is used to generate a corresponding time stamp for each frame of the target workpiece image during the image sequence acquisition process. The time stamp is generated by the internal clock of the industrial camera or a unified system clock synchronized with the industrial camera, and is used to identify the acquisition time of the corresponding frame of the target workpiece image. The time binding unit is used to bind each frame of the target workpiece image, the corresponding imaging parameter information, and the corresponding time identifier to form an image data unit with time attributes; The temporally continuous data generation unit is used to organize image data units with time attributes according to the chronological order of time markers to generate temporally continuous visual input data.

[0010] Optionally, the time-adaptive feature modeling module includes: The network topology construction unit is used to configure the input layer, hidden layer, and output layer of the liquid time constant neural network based on the time-continuous visual input data, determine the number of neurons in the input layer, hidden layer, and output layer, and establish directed connections between neurons in the input layer and hidden layer, between neurons in the hidden layer and hidden layer, and between neurons in the hidden layer and output layer, thus forming the network topology of the liquid time constant neural network. The state variable initialization unit is used to configure continuous-time state variables for each neuron in the hidden layer based on the topology of the liquid time constant neural network, configure initial state values ​​and initial time constant parameters for each continuous-time state variable, and construct a network parameter set by combining the initial state values ​​and initial time constant parameters. The input encoding unit is used to vectorize and encode each frame of the target workpiece image and its corresponding imaging parameter information in the time-continuous visual input data to generate an input sequence vector corresponding to the acquisition time identifier. The continuous-time state evolution unit is used to receive the input sequence vector, the network parameter set, and the continuous-time state variables. Within the time interval corresponding to the adjacent time markers, it calculates the state update amount of the hidden layer neurons according to the network topology of the liquid time constant neural network, and performs continuous-time state updates on the continuous-time state variables to generate updated continuous-time state variables. The time constant adaptive update unit calculates the input change information based on the input sequence vector corresponding to adjacent time identifiers, updates the time constant parameters in the network parameter set according to the input change information, writes the updated time constant parameters into the network parameter set, and outputs them to the continuous time state evolution unit. The state readout unit is used to receive the updated continuous-time state variables, generate the output of the liquid time constant neural network based on the connection relationship between the hidden layer neurons and the output layer neurons, and output the output of the liquid time constant neural network as a time dynamic visual feature representation.

[0011] Optionally, the structured geometry representation generation module for the target workpiece includes: The feature response parsing unit analyzes the feature response region corresponding to the target workpiece based on the time-dynamic visual feature representation, and determines the candidate geometric region corresponding to the spatial position of the target workpiece. The geometric boundary generation unit is used to aggregate the temporal dynamic visual feature representations within the candidate geometric region to generate geometric boundary information that characterizes the contour of the target workpiece. A key geometric element extraction unit is used to extract key geometric elements of a target workpiece based on geometric boundary information. The key geometric elements include boundary point information and boundary connection relationships. The structured encapsulation unit is used to organize key geometric elements in a structured way, generating a structured geometric representation that includes geometric boundary information and their relationships.

[0012] Optionally, the size generation module includes: The geometric representation parsing unit is used to receive structured geometric representations, extract geometric boundary information, boundary point information, and boundary connection relationships from the structured geometric representations, and generate a set of geometric elements. The dimension feature construction unit is used to determine the dimension feature pairs for dimension calculation based on the geometric feature set and according to the boundary connection relationship. The dimension feature pairs include point pairs formed by boundary point information, line pairs determined by boundary connection relationship, and point-line pairs formed by combining boundary point information and boundary connection relationship, thereby generating a set of dimension feature pairs. The geometric calculation unit is used to perform distance calculations on each pair of dimension elements in the set of dimension element pairs, and generate a set of dimension calculation results. The dimension parameter encapsulation unit is used to collect and organize the dimension calculation results according to the dimension type, form the dimension parameter set of the target workpiece, and output the dimension parameter set of the target workpiece to the geometric consistency constraint module.

[0013] Optionally, the improved geometric consistency constraint model includes: The constraint input construction unit is used to receive the structured geometric expression and the set of size parameters. Based on the geometric boundary information, boundary point information and boundary connection relationship in the structured geometric expression, it establishes the correspondence between the set of size parameters and geometric elements, and generates the constraint input data corresponding to the current time identifier. The historical constraint result reading unit is used to read the set of dimensional parameters subject to geometric consistency constraints output by the previous time identifier, generate historical constraint result input, and align the historical constraint result input with the constraint input data with the same-name dimensional parameters to generate cross-time aligned data. The temporal recursive consistency constraint construction unit constructs temporal recursive consistency constraint terms with the same size parameters based on cross-time aligned data, and writes the temporal recursive consistency constraint terms into the constructed constraint term candidate set; The projection scale consistency constraint construction unit is used to select pairs of size features with the same imaging reference conditions from geometric features based on constraint input data, construct projection scale consistency constraint terms, and write the projection scale consistency constraint terms into the constraint term candidate set. The structural parallelism consistency constraint construction unit is used to select structural line segment pairs from the boundary connection relationship based on the constraint input data, construct structural parallelism consistency constraint terms based on the direction information of the structural line segment pairs, and write the structural parallelism consistency constraint terms into the constraint term candidate set. The length invariant consistency constraint construction unit is used to establish consistency relationships for multiple size parameters corresponding to the same physical size based on constraint input data, construct length invariant consistency constraint terms, and write length invariant consistency constraint terms into the constraint term candidate set; The constraint term set generation unit is used to collect and organize the candidate constraint term set to generate a constraint term set that includes time-series recursion consistency constraint terms, projection ratio consistency constraint terms, structural parallelism consistency constraint terms, and length invariant consistency constraint terms. The constraint hierarchical partitioning unit is used to generate a first-level constraint term set and a second-level constraint term set based on the constraint term set. The first-level constraint term set consists of projection ratio consistency constraint terms and length invariant consistency constraint terms, and the second-level constraint term set consists of structural parallelism consistency constraint terms and temporal recursion consistency constraint terms. The first-stage constraint solving unit is used to receive the set of size parameters and the set of first-level constraint terms, construct the first-stage constraint objective, perform the first-stage constraint update on the set of size parameters, generate an intermediate set of size parameters, and output the intermediate set of size parameters to the second-stage constraint solving unit. The second-stage constraint solving unit is used to receive the intermediate size parameter set and the second-level constraint term set, construct the second-stage constraint objective, perform the second-stage constraint update on the intermediate size parameter set, and generate the size parameter set subject to geometric consistency constraints corresponding to the current time identifier. The constraint result output unit is used to output the set of dimensional parameters subject to geometric consistency constraints corresponding to the current time marker, and write the set of dimensional parameters subject to geometric consistency constraints corresponding to the current time marker into the historical constraint result reading unit for input of historical constraint results for the next time marker.

[0014] Optionally, the size output module includes: The parameter set receiving unit is used to receive the set of dimension parameters subject to geometric consistency constraints, and extract the dimension parameter identifier, dimension type identifier and dimension value from the set of dimension parameters subject to geometric consistency constraints to generate a sequence of dimension parameters to be output. The dimension result item generation unit is used to group the dimension parameters based on the dimension type identifier in the sequence of dimension parameters to be output, encapsulate each group of dimension parameters into a corresponding dimension result item, and generate a set of dimension result items. The identifier binding unit is used to bind the acquisition time identifier and the target workpiece identifier to each dimension result item in the dimension result item set, respectively, to generate a set of dimension result items with identifiers; The result structuring unit is used to organize the set of identifiable dimension result items in a structured manner to generate a dimension measurement result data structure. The dimension measurement result data structure includes a target workpiece identification field, an acquisition time identification field, and a dimension result item set field. The result output unit is used to output the dimension measurement result data structure as the dimension measurement result of the target workpiece.

[0015] The beneficial effects of this invention are: (1) This invention introduces a liquid time constant neural network into the industrial vision dimension measurement system to perform continuous time modeling of the continuously collected visual data in the industrial production line, so that the time response characteristics inside the model can be adaptively adjusted with the changes in visual input, thereby effectively coping with complex working conditions such as light fluctuations, workpiece posture changes and unstable production line operation, and improving the adaptability of the dimension measurement process to the dynamic environment.

[0016] (2) This invention constructs an improved geometric consistency constraint model, introduces temporal recursive consistency constraint and hierarchical geometric constraint mechanism after the size parameters are generated, and performs systematic constraint processing on the size parameter set, so that the size parameters maintain consistency in spatial geometric relationship and time evolution, reduces the mutual contradiction and cumulative deviation between size prediction results, and improves the stability and reliability of size measurement results.

[0017] (3) This invention combines time-adaptive deep learning modeling with an improved geometric consistency constraint model to form a complete dimension measurement process of “time modeling - geometric expression - dimension generation - constraint optimization - result output”. Without the need for frequent manual intervention or repeated calibration, it can realize continuous dimension measurement of different workpieces and different operating conditions, thereby enhancing the engineering applicability of the industrial vision dimension measurement system in automated production lines. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Fig. 1 This is a flowchart of an industrial vision dimension measurement system based on an adaptive deep learning model proposed in this invention. Fig. 2 This is a schematic diagram of the constraint construction and hierarchical solution process of the improved geometric consistency constraint model in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figs. 1-2 An industrial vision dimension measurement system based on an adaptive deep learning model includes the following steps: The visual timing acquisition module is used to acquire image sequence data of the target workpiece during the continuous operation of the industrial production line, and to establish a corresponding time marker for each frame of the image sequence data to generate time-continuous visual input data. The time-adaptive feature modeling module constructs a liquid time constant neural network, which models the continuous visual input data through a continuous time state evolution mechanism. This enables the time response characteristics of the neurons inside the liquid time constant neural network to adaptively adjust with the changes in the continuous visual input data, and outputs a time-dynamic visual feature representation. The geometric structure generation module generates a structured geometric representation of the target workpiece based on time-dynamic visual feature representation. The structured geometric representation is used to characterize the geometric boundary relationships and structural association relationships of the target workpiece. The dimension generation module generates a set of dimension parameters for the target workpiece based on structured geometric representation. The geometric consistency constraint module is used to introduce the set of dimensional parameters into the improved geometric consistency constraint model using a structured geometric expression as a geometric reference, perform geometric consistency constraint processing on the set of dimensional parameters, and generate a set of dimensional parameters subject to geometric consistency constraints. The dimension output module is used to output the dimension measurement results of the target workpiece based on a set of dimension parameters constrained by geometric consistency.

[0021] In this embodiment, the visual temporal acquisition module includes: The image sequence acquisition unit is used to continuously acquire images of the same target workpiece during the continuous operation of an industrial production line, and to obtain image sequence data arranged in the order of acquisition. The image sequence data consists of multiple frames of target workpiece images, and each frame of target workpiece image contains the imaging area of ​​the target workpiece and the corresponding image pixel data. An imaging parameter acquisition unit is used to synchronously acquire imaging parameter information corresponding to each frame of the target workpiece image during the image sequence acquisition process. The imaging parameter information includes at least exposure parameters, gain parameters, and focal length parameters, and establishes a correlation with the corresponding frame of the target workpiece image. The time stamp generation unit is used to generate a corresponding time stamp for each frame of the target workpiece image during the image sequence acquisition process. The time stamp is generated by the internal clock of the industrial camera or a unified system clock synchronized with the industrial camera, and is used to identify the acquisition time of the corresponding frame of the target workpiece image. The time binding unit is used to bind each frame of the target workpiece image, the corresponding imaging parameter information, and the corresponding time identifier to form an image data unit with time attributes; The temporally continuous data generation unit is used to organize image data units with time attributes according to the chronological order of time markers to generate temporally continuous visual input data.

[0022] In this embodiment, the time-adaptive feature modeling module includes: The network topology construction unit is used to configure the input layer, hidden layer, and output layer of the liquid time constant neural network based on the time-continuous visual input data, determine the number of neurons in the input layer, hidden layer, and output layer, and establish directed connections between neurons in the input layer and hidden layer, between neurons in the hidden layer and hidden layer, and between neurons in the hidden layer and output layer, thus forming the network topology of the liquid time constant neural network. The state variable initialization unit is used to configure continuous-time state variables for each neuron in the hidden layer based on the topology of the liquid time constant neural network, configure initial state values ​​and initial time constant parameters for each continuous-time state variable, and construct a network parameter set by combining the initial state values ​​and initial time constant parameters. Continuous-time state variables refer to the internal state quantities that are set for each hidden layer neuron in a liquid-time constant neural network and change continuously with time. They are used to represent the comprehensive response state of the neuron to continuous visual input data and network connection input at any given moment. Their values ​​are updated in adjacent time intervals according to the continuous-time state evolution rules. The input encoding unit is used to vectorize and encode each frame of the target workpiece image and its corresponding imaging parameter information in the time-continuous visual input data to generate an input sequence vector corresponding to the acquisition time identifier. The continuous-time state evolution unit is used to receive the input sequence vector, the network parameter set, and the continuous-time state variables. Within the time interval corresponding to the adjacent time markers, it calculates the state update amount of the hidden layer neurons according to the network topology of the liquid time constant neural network, and performs continuous-time state updates on the continuous-time state variables to generate updated continuous-time state variables. Calculating the state update of hidden layer neurons refers to the weighted summation of each input signal within the time interval corresponding to adjacent time markers, based on the current continuous time state variable, the input sequence vector at the corresponding time, and the connection weights between hidden layer neurons, and combined with the continuous time evolution rule to obtain the change used to update the continuous time state variable. Performing continuous-time state update refers to updating the continuous-time state variable in integral form after obtaining the state update amount of the hidden layer neurons, based on the time interval between adjacent time markers, so that the continuous-time state variable transitions from the state of the previous time moment to the state of the current time moment within the current time interval. The time constant adaptive update unit calculates the input change information based on the input sequence vector corresponding to adjacent time identifiers, updates the time constant parameters in the network parameter set according to the input change information, writes the updated time constant parameters into the network parameter set, and outputs them to the continuous time state evolution unit for subsequent continuous time state variable updates. Calculating input change information refers to comparing the input sequence vectors corresponding to adjacent time markers, obtaining the amount of change in the input sequence vectors in the time dimension, and characterizing the changes of time-continuous visual input data in adjacent time intervals based on this amount of change; Updating the time constant parameters in the network parameter set means reassigning or adjusting the time constant parameters corresponding to each hidden layer neuron based on the calculated input change information, and writing the updated time constant parameters into the network parameter set. The state readout unit is used to receive the updated continuous-time state variables, generate the output of the liquid time constant neural network based on the connection relationship between the hidden layer neurons and the output layer neurons, and output the output of the liquid time constant neural network as a time dynamic visual feature representation.

[0023] In this embodiment, the structured geometric representation generation module for the target workpiece includes: The feature response parsing unit analyzes the feature response region corresponding to the target workpiece based on the time-dynamic visual feature representation, and determines the candidate geometric region corresponding to the spatial position of the target workpiece. Analyzing the feature response region corresponding to the target workpiece refers to performing spatial distribution analysis on the response values ​​of each feature channel in the time-dynamic visual feature representation, identifying regions with continuous response intensity and spatial connectivity, and determining such regions as the feature response regions corresponding to the spatial position of the target workpiece. The geometric boundary generation unit is used to aggregate the temporal dynamic visual feature representations within the candidate geometric region to generate geometric boundary information that characterizes the contour of the target workpiece. Aggregation processing refers to the weighted merging or statistical summarization of temporal dynamic visual feature representations in the spatial dimension within the candidate geometric region, integrating feature responses from multiple times or channels into a comprehensive feature representation for characterizing the contour of the target workpiece. A key geometric element extraction unit is used to extract key geometric elements of a target workpiece based on geometric boundary information. The key geometric elements include boundary point information and boundary connection relationships. The structured encapsulation unit is used to organize key geometric elements in a structured manner, generate a structured geometric expression containing geometric boundary information and its relationships, and output the structured geometric expression to the dimension generation module; The structured organization of key geometric elements refers to arranging and associating the extracted boundary point information and boundary connection relationships according to a preset data structure to form a structured data set that can clearly represent the geometric boundary composition and element relationships.

[0024] In this embodiment, the size generation module includes: The geometric representation parsing unit is used to receive structured geometric representations, extract geometric boundary information, boundary point information, and boundary connection relationships from the structured geometric representations, and generate a set of geometric elements. The dimension feature construction unit is used to determine the dimension feature pairs for dimension calculation based on the geometric feature set and according to the boundary connection relationship. The dimension feature pairs include point pairs formed by boundary point information, line pairs determined by boundary connection relationship, and point-line pairs formed by combining boundary point information and boundary connection relationship, thereby generating a set of dimension feature pairs. The geometric calculation unit is used to perform distance calculations on each pair of dimension elements in the set of dimension element pairs, and generate a set of dimension calculation results. Performing distance calculation refers to calculating the Euclidean distance or projected distance along a specified direction between each pair of dimensional features based on the spatial coordinates of the corresponding boundary points or the spatial coordinates of the endpoints of line segments, in order to obtain a distance value used to characterize the size of the target workpiece. The dimension parameter encapsulation unit is used to collect and organize the dimension calculation result set according to the dimension type, form the dimension parameter set of the target workpiece, and output the dimension parameter set of the target workpiece to the geometric consistency constraint module; Size type refers to the category based on the geometric meaning of the size calculation results, including length size, spacing size, diameter size, or angle size; collection and organization refers to classifying and storing size calculation results according to size type, and organizing size calculation results of the same type into corresponding size parameter subsets.

[0025] In this embodiment, the improved geometric consistency constraint model includes: The constraint input construction unit is used to receive the structured geometric expression and the set of size parameters. Based on the geometric boundary information, boundary point information and boundary connection relationship in the structured geometric expression, it establishes the correspondence between the set of size parameters and geometric elements, and generates the constraint input data corresponding to the current time identifier. Establishing the correspondence between a set of dimensional parameters and geometric features means specifying the geometric feature or pair of geometric features from which each dimensional parameter in the set of dimensional parameters originates, and binding the dimensional parameter with the corresponding boundary point, structural line segment or combination thereof, so as to clarify which geometric features the dimensional parameter is calculated from; The historical constraint result reading unit is used to read the set of dimensional parameters subject to geometric consistency constraints output by the previous time identifier, generate historical constraint result input, and align the historical constraint result input with the constraint input data with the same-name dimensional parameters to generate cross-time aligned data. Aligning dimension parameters with the same name means matching dimension parameters with the same identifier between the set of dimension parameters corresponding to the current time identifier and the set of dimension parameters corresponding to the previous time identifier, based on the identifier name or index number of the dimension parameter, to form a one-to-one correspondence across time. The temporal recursive consistency constraint construction unit constructs temporal recursive consistency constraint terms with the same size parameters based on cross-time aligned data, and writes the temporal recursive consistency constraint terms into the constructed constraint term candidate set; Constructing a temporal recursive consistency constraint term for the same-named size parameter refers to establishing a difference or change constraint between the size parameter of the current time marker and the corresponding size parameter of the previous time marker based on the aligned cross-time size parameter, and using the difference as a constraint term to limit the change range of the size parameter under adjacent time markers; The candidate set of constraint terms refers to the data set formed by uniformly collecting temporal recursive consistency constraint terms, projection ratio consistency constraint terms, structural parallelism consistency constraint terms, and length invariant consistency constraint terms in the geometric consistency constraint model. The projection scale consistency constraint construction unit is used to select pairs of size features with the same imaging reference conditions from geometric features based on constraint input data, construct projection scale consistency constraint terms, and write the projection scale consistency constraint terms into the constraint term candidate set. The projection scale consistency constraint refers to the constraint on the proportional relationship between the size parameters of a pair of size features located under the same imaging reference conditions, and the corresponding geometric feature projection distance ratio in the image, to limit the scale consistency between the relevant size parameters. The structural parallelism consistency constraint construction unit is used to select structural line segment pairs from the boundary connection relationship based on the constraint input data, construct structural parallelism consistency constraint terms based on the direction information of the structural line segment pairs, and write the structural parallelism consistency constraint terms into the constraint term candidate set. The directional information of a structural line segment refers to the direction vector or direction angle calculated from the coordinates of the two endpoints of each structural line segment. It is used to characterize the spatial orientation of the structural line segment in the image plane or measurement coordinate system, and to determine the directional relationship between two structural line segments. The structural parallelism consistency constraint term refers to the directional difference constraint constructed based on the directional information for a pair of structural line segments with parallel geometric relationship. The directional difference is used as a constraint condition to restrict the parallel relationship between the two structural line segments in the structured geometric expression. The length invariant consistency constraint construction unit is used to establish consistency relationships for multiple size parameters corresponding to the same physical size based on constraint input data, construct length invariant consistency constraint terms, and write length invariant consistency constraint terms into the constraint term candidate set; Establishing a consistency relationship for multiple dimensional parameters corresponding to the same physical dimension refers to establishing equal or limited difference constraints between these dimensional parameters when the same physical dimension is calculated from multiple dimensional parameters by combining different geometric elements, in order to uniformly represent the physical dimension; The length invariant consistency constraint term refers to the constraint condition that maintains the consistency of the length value for multiple dimension parameters obtained under different combinations of geometric elements for the same physical dimension, and is used to limit the consistency of the value of the physical dimension in different structural expressions. The constraint term set generation unit is used to collect and organize the candidate constraint term set to generate a constraint term set that includes time-series recursion consistency constraint terms, projection ratio consistency constraint terms, structural parallelism consistency constraint terms, and length invariant consistency constraint terms. The constraint hierarchical partitioning unit is used to generate a first-level constraint term set and a second-level constraint term set based on the constraint term set. The first-level constraint term set consists of projection ratio consistency constraint terms and length invariant consistency constraint terms, and the second-level constraint term set consists of structural parallelism consistency constraint terms and temporal recursion consistency constraint terms. The first-stage constraint solving unit is used to receive the set of size parameters and the set of first-level constraint terms, construct the first-stage constraint objective, perform the first-stage constraint update on the set of size parameters, generate an intermediate set of size parameters, and output the intermediate set of size parameters to the second-stage constraint solving unit. Performing the first-stage constraint update on the set of dimensional parameters means taking the set of dimensional parameters as the object to be updated, introducing only the constraint relationships corresponding to the first-level constraint term set, performing a constraint solution or iterative adjustment on the set of dimensional parameters once, obtaining an intermediate set of dimensional parameters that satisfies the first-level constraint conditions, and using it as the input for subsequent constraint updates; The second-stage constraint solving unit is used to receive the intermediate size parameter set and the second-level constraint term set, construct the second-stage constraint objective, perform the second-stage constraint update on the intermediate size parameter set, and generate the size parameter set subject to geometric consistency constraints corresponding to the current time identifier. Performing a second-stage constraint update on the intermediate dimension parameter set means taking the intermediate dimension parameter set obtained in the previous stage as input, introducing the constraint relationship corresponding to the second-level constraint term set, and performing constraint solving or iterative adjustment on the intermediate dimension parameter set again to generate a dimension parameter set that satisfies the second-level constraint conditions. The constraint result output unit is used to output the set of dimensional parameters subject to geometric consistency constraints corresponding to the current time marker, and write the set of dimensional parameters subject to geometric consistency constraints corresponding to the current time marker into the historical constraint result reading unit for input of historical constraint results for the next time marker.

[0026] In this embodiment, the size output module includes: The parameter set receiving unit is used to receive the set of dimension parameters subject to geometric consistency constraints, and extract the dimension parameter identifier, dimension type identifier and dimension value from the set of dimension parameters subject to geometric consistency constraints to generate a sequence of dimension parameters to be output. The dimension parameter identifier is a number or name field used to uniquely identify each dimension parameter; the dimension type identifier is used to characterize the dimension category to which the dimension parameter belongs; and the dimension value is used to record the specific measurement value of the corresponding dimension parameter. Together, the three constitute a complete description of a single dimension parameter. The dimension result item generation unit is used to group the dimension parameters based on the dimension type identifier in the sequence of dimension parameters to be output, encapsulate each group of dimension parameters into a corresponding dimension result item, and generate a set of dimension result items. Grouping dimensional parameters refers to classifying the dimensional parameters in the dimensional parameter set according to the dimensional type identifier corresponding to the dimensional parameter, and grouping dimensional parameters with the same dimensional type identifier into the same group, forming multiple subsets of dimensional parameters divided by dimensional type; The identifier binding unit is used to bind the acquisition time identifier and the target workpiece identifier to each dimension result item in the dimension result item set, respectively, to generate a set of dimension result items with identifiers; The target workpiece identifier is a unique identifier used to distinguish different workpieces being measured. It is used to indicate the specific workpiece object corresponding to the dimensional measurement result. It can be composed of production line workpiece number, batch number, or a unique identification code generated by the system. The result structuring unit is used to organize the set of identifiable dimension result items in a structured manner to generate a dimension measurement result data structure. The dimension measurement result data structure includes a target workpiece identification field, an acquisition time identification field, and a dimension result item set field. Structured organization refers to arranging and associating each dimension result item and its corresponding target workpiece identifier and acquisition time identifier in a field-based manner according to a preset data structure format, forming a dimension measurement result data structure with a fixed field order and hierarchical relationship; The result output unit is used to output the dimension measurement result data structure as the dimension measurement result of the target workpiece.

[0027] Example 1: To verify the feasibility of this invention in practice, its application scenario is online measurement of structural component dimensions in an industrial automated production line. The measurement object is a metal structural workpiece continuously conveyed in the production line. After processing, these workpieces require inspection of several key structural dimensions, including hole spacing, distance from the edge to the hole center, and spacing between symmetrical structures, to ensure structural matching accuracy during subsequent assembly. In actual production line operation, because the workpieces pass through the inspection station continuously, the imaging process is affected by changes in ambient lighting, mechanical vibration, and slight changes in workpiece posture. Traditional dimension measurement methods based on single-frame image processing or static deep learning models are prone to fluctuations in measurement results, inconsistencies between dimensional parameters, and gradual accumulation of errors over long-term operation, often requiring frequent manual intervention or recalibration to maintain measurement accuracy.

[0028] In this embodiment, an industrial camera is installed above the inspection station on the production line. It continuously captures images of the same target workpiece as it passes through the inspection area, forming an image sequence data consisting of multiple frames. Each frame is appended with a corresponding time stamp during acquisition, thus creating temporally continuous visual input data. This approach enables the system to acquire multi-moment visual information of the target workpiece within a measurement cycle, rather than relying solely on a single instantaneous image, providing a foundation for subsequent temporal modeling.

[0029] The acquired temporally continuous visual input data is fed into the time-adaptive feature modeling module. This module is built on a liquid time-constant neural network. By configuring continuous-time state variables for the neurons within the network and dynamically adjusting the time constant parameters according to input changes during network operation, the temporal response characteristics of the neurons can adaptively adjust with changes in the temporally continuous visual input data. Even if there are short-term noises or brightness changes in local images as the workpiece continuously passes through the detection area, the temporally dynamic visual feature representation output by the liquid time-constant neural network maintains continuity and stability, avoiding feature abrupt changes caused by single-frame anomalies.

[0030] After obtaining the temporal dynamic visual feature representation, the system models the geometric structure of the target workpiece in the geometric structure generation module. The temporal dynamic visual feature representation is first parsed to determine the feature response region corresponding to the target workpiece. Within this region, features are aggregated to generate geometric boundary information that stably reflects the workpiece's contour. Based on this geometric boundary information, the system extracts boundary point information and boundary connection relationships, and then organizes the extracted geometric elements in a structured manner to form a structured geometric expression of the target workpiece. This structured geometric expression not only describes the workpiece's external contour but also clarifies the spatial relationships between key structural elements, providing a reliable geometric basis for calculating dimensional parameters.

[0031] During the dimension generation stage, the system parses the structured geometric representation, extracts geometric boundary information, boundary point information, and boundary connection relationships, and constructs dimension feature pairs based on these relationships. The system then performs distance calculations between these dimension feature pairs, obtaining multiple initial dimension calculation results. These results are then aggregated and organized according to dimension type to form a set of dimension parameters for the target workpiece. Before constraint processing, this set of dimension parameters may still contain numerical deviations caused by local feature fluctuations or incomplete geometric relationships.

[0032] To address the issues of inconsistencies between dimensional parameters and error accumulation during continuous measurement, this embodiment introduces an improved geometric consistency constraint model to constrain the dimensional parameter set. At each time marker, the system first establishes a correspondence between the dimensional parameter set and geometric elements based on structured geometric representation. Simultaneously, it reads the dimensional parameter set already output under geometric consistency constraints at the previous time marker and aligns the corresponding dimensional parameters between the current and previous times. The system constructs a temporal recursive consistency constraint term for the corresponding dimensional parameters, ensuring the continuity of dimensional parameter changes across adjacent time markers.

[0033] The system constructs projection scale consistency constraints, structural parallelism consistency constraints, and length invariant consistency constraints based on the current structured geometric representation, and writes these constraints into a unified constraint candidate set. The system hierarchically divides the constraint candidate set. First, it uses the projection scale consistency constraints and length invariant consistency constraints to perform a first-stage constraint update on the dimensional parameter set, generating an intermediate dimensional parameter set. Then, it introduces the structural parallelism consistency constraints and temporal recursion consistency constraints to perform a second-stage constraint update on the intermediate dimensional parameter set, obtaining a dimensional parameter set constrained by geometric consistency. This constrained dimensional parameter set is output and stored for constructing recursive constraints under the next time marker, thus forming a closed-loop constraint mechanism in the continuous measurement process.

[0034] During the output stage of dimensional measurement results, the system organizes the set of dimensional parameters subject to geometric consistency constraints, groups the dimensional parameters according to dimensional type, and binds the target workpiece identifier and acquisition time identifier to each group of dimensional parameters, generating structured dimensional measurement results and outputting them to the production line quality management system.

[0035] To verify the practical application effect of this embodiment, comparative tests were conducted on multiple batches of workpieces under continuous operation conditions. Test results show that, under the same working conditions, when using traditional measurement methods based on single-frame deep learning models, the average measurement error of key dimensions such as hole spacing, distance from edge to hole center, and symmetrical structure spacing is typically above 0.1 mm, and the error fluctuates significantly during continuous operation. However, when using the system described in this embodiment, the average measurement error of the aforementioned key dimensions can be stably controlled within 0.1 mm, and the fluctuation range of the dimensional results during continuous measurement is significantly reduced. Further statistical analysis revealed that when workpiece models change or environmental conditions change, traditional methods often require multiple manual recalibrations to restore measurement stability, while the system described in this embodiment requires very little or no manual intervention to maintain the consistency of dimensional measurement results under similar conditions.

[0036] In summary, this invention effectively addresses the problems of insufficient measurement accuracy and poor stability in industrial visual dimensional measurement scenarios caused by environmental changes, workpiece variations, and static model characteristics. By introducing a liquid time constant neural network to achieve adaptive modeling of time-continuous visual data, and combining this with an improved geometric consistency constraint model to systematically constrain dimensional parameters, this invention enables stable and reliable measurement of target workpiece dimensions under continuous operation conditions in industrial production lines.

[0037] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An industrial vision dimension measurement system based on an adaptive deep learning model, characterized in that, include: The visual timing acquisition module is used to acquire image sequence data of the target workpiece during the continuous operation of the industrial production line, and to establish a corresponding time marker for each frame of the image sequence data to generate time-continuous visual input data. The time-adaptive feature modeling module constructs a liquid time constant neural network, which models the continuous visual input data through a continuous time state evolution mechanism. This enables the time response characteristics of the neurons inside the liquid time constant neural network to adaptively adjust with the changes in the continuous visual input data, and outputs a time-dynamic visual feature representation. The geometry generation module generates a structured geometric representation of the target workpiece based on time-dynamic visual feature representation. The dimension generation module generates a set of dimension parameters for the target workpiece based on structured geometric representation. The geometric consistency constraint module is used to introduce the set of dimensional parameters into the improved geometric consistency constraint model using a structured geometric expression as a geometric reference, perform geometric consistency constraint processing on the set of dimensional parameters, and generate a set of dimensional parameters subject to geometric consistency constraints. The dimension output module is used to output the dimension measurement results of the target workpiece based on a set of dimension parameters constrained by geometric consistency.

2. The industrial vision dimension measurement system based on an adaptive deep learning model according to claim 1, characterized in that, The visual temporal acquisition module includes: The image sequence acquisition unit is used to continuously acquire images of the same target workpiece during the continuous operation of an industrial production line, and to obtain image sequence data arranged in the order of acquisition. The image sequence data consists of multiple frames of target workpiece images, and each frame of target workpiece image contains the imaging area of ​​the target workpiece and the corresponding image pixel data. The imaging parameter acquisition unit is used to synchronously acquire imaging parameter information corresponding to each frame of the target workpiece image during the image sequence acquisition process; The time stamp generation unit is used to generate a corresponding time stamp for each frame of the target workpiece image during the image sequence acquisition process; The time binding unit is used to bind each frame of the target workpiece image, the corresponding imaging parameter information, and the corresponding time identifier to form an image data unit with time attributes; The temporally continuous data generation unit is used to organize image data units with time attributes according to the chronological order of time markers to generate temporally continuous visual input data.

3. The industrial vision dimension measurement system based on an adaptive deep learning model according to claim 1, characterized in that, The time-adaptive feature modeling module includes: The network topology construction unit is used to configure the input layer, hidden layer, and output layer of the liquid time constant neural network based on the time-continuous visual input data, determine the number of neurons in the input layer, hidden layer, and output layer, and establish directed connections between neurons in the input layer and hidden layer, between neurons in the hidden layer and hidden layer, and between neurons in the hidden layer and output layer, thus forming the network topology of the liquid time constant neural network. The state variable initialization unit is used to configure continuous-time state variables for each neuron in the hidden layer based on the topology of the liquid time constant neural network, configure initial state values ​​and initial time constant parameters for each continuous-time state variable, and construct a network parameter set by combining the initial state values ​​and initial time constant parameters. The input encoding unit is used to vectorize and encode each frame of the target workpiece image and its corresponding imaging parameter information in the time-continuous visual input data to generate an input sequence vector corresponding to the acquisition time identifier. The continuous-time state evolution unit is used to receive the input sequence vector, the network parameter set, and the continuous-time state variables. Within the time interval corresponding to the adjacent time markers, it calculates the state update amount of the hidden layer neurons according to the network topology of the liquid time constant neural network, and performs continuous-time state updates on the continuous-time state variables to generate updated continuous-time state variables. The time constant adaptive update unit calculates the input change information based on the input sequence vector corresponding to adjacent time identifiers, updates the time constant parameters in the network parameter set according to the input change information, writes the updated time constant parameters into the network parameter set, and outputs them to the continuous time state evolution unit. The state readout unit is used to receive the updated continuous-time state variables, generate the output of the liquid time constant neural network based on the connection relationship between the hidden layer neurons and the output layer neurons, and output the output of the liquid time constant neural network as a time dynamic visual feature representation.

4. The industrial vision dimension measurement system based on an adaptive deep learning model according to claim 1, characterized in that, The structured geometric representation generation module for the target workpiece includes: The feature response parsing unit analyzes the feature response region corresponding to the target workpiece based on the time-dynamic visual feature representation, and determines the candidate geometric region corresponding to the spatial position of the target workpiece. The geometric boundary generation unit is used to aggregate the temporal dynamic visual feature representations within the candidate geometric region to generate geometric boundary information that characterizes the contour of the target workpiece. A key geometric element extraction unit is used to extract key geometric elements of a target workpiece based on geometric boundary information. The key geometric elements include boundary point information and boundary connection relationships. The structured encapsulation unit is used to organize key geometric elements in a structured way, generating a structured geometric representation that includes geometric boundary information and their relationships.

5. The industrial vision dimension measurement system based on an adaptive deep learning model according to claim 1, characterized in that, The size generation module includes: The geometric representation parsing unit is used to receive structured geometric representations, extract geometric boundary information, boundary point information, and boundary connection relationships from the structured geometric representations, and generate a set of geometric elements. The dimension feature construction unit is used to determine the dimension feature pairs for dimension calculation based on the geometric feature set and according to the boundary connection relationship. The dimension feature pairs include point pairs formed by boundary point information, line pairs determined by boundary connection relationship, and point-line pairs formed by combining boundary point information and boundary connection relationship, thereby generating a set of dimension feature pairs. The geometric calculation unit is used to perform distance calculations on each pair of dimension elements in the set of dimension element pairs, and generate a set of dimension calculation results. The dimension parameter encapsulation unit is used to collect and organize the dimension calculation results according to the dimension type to form the dimension parameter set of the target workpiece.

6. The industrial vision dimension measurement system based on an adaptive deep learning model according to claim 1, characterized in that, The improved geometric consistency constraint model includes: The constraint input construction unit is used to receive the structured geometric expression and the set of size parameters. Based on the geometric boundary information, boundary point information and boundary connection relationship in the structured geometric expression, it establishes the correspondence between the set of size parameters and geometric elements, and generates the constraint input data corresponding to the current time identifier. The historical constraint result reading unit is used to read the set of dimensional parameters subject to geometric consistency constraints output by the previous time identifier, generate historical constraint result input, and align the historical constraint result input with the constraint input data with the same-name dimensional parameters to generate cross-time aligned data. The temporal recursive consistency constraint construction unit constructs temporal recursive consistency constraint terms with the same size parameters based on cross-time aligned data, and writes the temporal recursive consistency constraint terms into the constructed constraint term candidate set; The projection scale consistency constraint construction unit is used to select pairs of size features with the same imaging reference conditions from geometric features based on constraint input data, construct projection scale consistency constraint terms, and write the projection scale consistency constraint terms into the constraint term candidate set. The structural parallelism consistency constraint construction unit is used to select structural line segment pairs from the boundary connection relationship based on the constraint input data, construct structural parallelism consistency constraint terms based on the direction information of the structural line segment pairs, and write the structural parallelism consistency constraint terms into the constraint term candidate set. The length invariant consistency constraint construction unit is used to establish consistency relationships for multiple size parameters corresponding to the same physical size based on constraint input data, construct length invariant consistency constraint terms, and write length invariant consistency constraint terms into the constraint term candidate set; The constraint term set generation unit is used to collect and organize the candidate constraint term set to generate a constraint term set that includes time-series recursion consistency constraint terms, projection ratio consistency constraint terms, structural parallelism consistency constraint terms, and length invariant consistency constraint terms. The constraint hierarchical partitioning unit is used to generate a first-level constraint term set and a second-level constraint term set based on the constraint term set. The first-level constraint term set consists of projection ratio consistency constraint terms and length invariant consistency constraint terms, and the second-level constraint term set consists of structural parallelism consistency constraint terms and temporal recursion consistency constraint terms. The first-stage constraint solving unit is used to receive the set of size parameters and the set of first-level constraint terms, construct the first-stage constraint objective, perform the first-stage constraint update on the set of size parameters, generate an intermediate set of size parameters, and output the intermediate set of size parameters to the second-stage constraint solving unit. The second-stage constraint solving unit is used to receive the intermediate size parameter set and the second-level constraint term set, construct the second-stage constraint objective, perform the second-stage constraint update on the intermediate size parameter set, and generate the size parameter set subject to geometric consistency constraints corresponding to the current time identifier. The constraint result output unit is used to output the set of dimensional parameters subject to geometric consistency constraints corresponding to the current time marker, and write the set of dimensional parameters subject to geometric consistency constraints corresponding to the current time marker into the historical constraint result reading unit for input of historical constraint results for the next time marker.

7. The industrial vision dimension measurement system based on an adaptive deep learning model according to claim 1, characterized in that, The dimension output module includes: The parameter set receiving unit is used to receive the set of dimension parameters subject to geometric consistency constraints, and extract the dimension parameter identifier, dimension type identifier and dimension value from the set of dimension parameters subject to geometric consistency constraints to generate a sequence of dimension parameters to be output. The dimension result item generation unit is used to group the dimension parameters based on the dimension type identifier in the sequence of dimension parameters to be output, encapsulate each group of dimension parameters into a corresponding dimension result item, and generate a set of dimension result items. The identifier binding unit is used to bind the acquisition time identifier and the target workpiece identifier to each dimension result item in the dimension result item set, respectively, to generate a set of dimension result items with identifiers; The result structuring unit is used to organize the set of identifiable dimension result items in a structured manner to generate a dimension measurement result data structure. The dimension measurement result data structure includes a target workpiece identification field, an acquisition time identification field, and a dimension result item set field. The result output unit is used to output the dimension measurement result data structure as the dimension measurement result of the target workpiece.

Images