A low-code development method and system based on general target recognition

Abstract

The present invention discloses a low-code development method and system based on general target recognition, which specifically relates to the field of low-code development technology, and is used to solve the problem that the existing model abstraction level cannot be flexibly adapted to the complexity and requirements of specific application scenarios to dynamically adjust the model abstraction level; it is through extracting parameters from the application scenarios input by developers, using the fuzzy hierarchical analysis method to evaluate the degree of abstraction level requirements of the model, and determining the initial abstraction level through the grey correlation analysis method; when the initial abstraction level is high, further evaluating the changing trend of the model prediction uncertainty through time series autoregression analysis, and evaluating the feature representation stability of the model through topological data analysis; after comprehensive evaluation, it is decided whether to maintain a high abstraction level or make adjustments, which can improve the adaptability and stability of the target recognition model, improve development efficiency, and optimize computing resource utilization.

Description

Technical Field

[0001] The present invention relates to the field of low-code development technology, and more specifically, the present invention relates to a low-code development method and system based on general target recognition. Background Art

[0002] With the rapid development of artificial intelligence and computer vision technology, target recognition has been widely used in various fields, such as autonomous driving, security monitoring, and medical image analysis. However, developing a high-performance target recognition model usually requires professional knowledge and a lot of code writing, which has a high technical threshold for developers who lack a deep learning background. To this end, low-code development platforms have gradually emerged, aiming to reduce the difficulty of development and enable developers to build applications with less code or even no coding. In actual application scenarios, the demand for target recognition is very extensive, usually including the recognition and application of targets such as graphics, text, and objects. For example, users may want to use target recognition technology to identify specific content in a picture, such as how many people, trees, or items are in the picture. This type of demand is particularly common in scenarios such as intelligent monitoring, resource statistics, and urban planning.

[0003] In low-code development platforms, existing technologies often highly abstract general target recognition models, but are unable to flexibly adapt to the complexity and needs of specific application scenarios. This lack of ability to dynamically adjust the model's abstraction level will lead to insufficient recognition accuracy in complex scenarios, or waste of computing resources in simple scenarios, reducing system efficiency and recognition effects.

[0004] In order to solve the above problems, a technical solution is now provided. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a low-code development method and system based on general target recognition to solve the problems raised in the above-mentioned background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A low-code development method based on general object recognition includes the following steps:

[0008] S1: extracting application scenario parameters from the specific application scenario input by the developer;

[0009] S2: Use fuzzy analytic hierarchy process to conduct multi-level fuzzy comprehensive evaluation on application scenario parameters to assess the abstract level requirement of general object recognition model;

[0010] S3: According to the abstraction level requirement of the general target recognition model, the grey relational analysis method is used to determine the initial abstraction level;

[0011] S4: When the initial abstraction level is high: perform time series autoregression analysis on the prediction entropy value of the general target recognition model in recent executions to evaluate the changing trend of the prediction uncertainty of the general target recognition model; use topological data analysis methods to evaluate the stability of the feature representation of the general target recognition model to determine the general target recognition model's ability to capture the intrinsic structure of the input data;

[0012] S5: Based on a comprehensive analysis of the changing trend of the uncertainty prediction of the general target recognition model and the general target recognition model's ability to capture the intrinsic structure of the input data, determine whether to maintain a high level of abstraction.

[0013] In a preferred embodiment, S1 specifically includes:

[0014] S101: The developer enters a description of a specific application scenario on the low-code development platform;

[0015] S102: automatically parsing the scene description, analyzing the scene using a preset complexity classification algorithm, and extracting scene complexity features;

[0016] S103: According to the task goal input by the developer, determine the requirement of the task on the target recognition accuracy, and extract the recognition accuracy requirement feature;

[0017] S104: Based on the system resource configuration or execution environment input by the developer, evaluate the currently available computing resources and extract computing resource limitation features;

[0018] S105: Summarize the extracted scene complexity features, recognition accuracy requirement features, and computing resource limitation features to form application scenario parameters.

[0019] In a preferred embodiment, S2 specifically includes:

[0020] S201: construct a fuzzy hierarchical structure according to the extracted application scenario parameters, divide it into a target layer, a criterion layer and a parameter layer, and establish a multi-level fuzzy decision model;

[0021] S202: In the hierarchical model, by performing pairwise comparisons between parameters of various application scenarios, a fuzzy comparison matrix is ​​constructed;

[0022] S203: Calculating the fuzzy weight of each application scenario parameter through fuzzy consistency judgment and weight allocation algorithm;

[0023] S204: Perform fuzzy comprehensive operation on the fuzzy contrast matrix, and aggregate the fuzzy evaluation values ​​of the application scenario parameters at each layer into a model abstract selection index by combining the weights:

[0024] According to the fuzzy weight and fuzzy evaluation value of each application scenario parameter, the model abstract selection index is calculated, and its expression is: ;in, Represents the model abstract selection index, is the calculated fuzzy weight of each application scenario parameter, is the corresponding fuzzy evaluation value, is the number of application scenario parameters, It is the number of the application scenario parameter.

[0025] In a preferred embodiment, S3 specifically includes:

[0026] Extract the predicted entropy value of each execution from the execution results of the general object recognition model: Set the model output probability distribution ,in is the class predicted by the model, is the input data; the model predicts the entropy value At the moment is defined as: ;in, Represents the model at time The predicted entropy value of Indicates that given the input data The model predicts the category The probability of represents the number of categories, Represents the category index;

[0027] Arrange the predicted entropy values ​​at each moment in chronological order to form a predicted entropy value sequence;

[0028] The autoregressive integrated moving average model is used to analyze the predicted entropy value sequence and further predict the entropy value change trend in the future;

[0029] Calculate the rate of change of the predicted entropy value: ;in, is the rate of change of entropy value;

[0030] when When , the prediction uncertainty increases; when When , the prediction uncertainty is judged to be reduced or unchanged.

[0031] In a preferred embodiment, the stability of the feature representation of the general object recognition model is evaluated by using a topological data analysis method to determine the general object recognition model's ability to capture the intrinsic structure of the input data, specifically including:

[0032] Extract high-dimensional feature representations from the middle layers of the general object recognition model to form a feature space dataset;

[0033] Based on the extracted high-dimensional feature representation, the topological data analysis method is used to construct persistent homology, generate the topological structure of the feature space, and capture the multi-scale morphological characteristics of the data;

[0034] The stability of the model's feature representation is quantified by computing topological invariants through persistent homology, and the model's ability to capture the intrinsic structure of the data is evaluated;

[0035] Compare the feature topologies at different time points or input data, and use the bottleneck distance to measure the changes in feature representation;

[0036] According to the degree of change of the topological invariant, the ability of the general target recognition model to capture the intrinsic structure of the input data is judged: if the model maintains a stable and consistent feature representation in different scenarios, the general target recognition model is judged to have a good ability to capture the intrinsic structure of the input data; if the model's feature representation fluctuates greatly under different inputs, the general target recognition model is judged to have a poor ability to capture the intrinsic structure of the input data.

[0037] In a preferred embodiment, based on a comprehensive analysis of the changing trend of the uncertainty prediction of the general target recognition model and the ability of the general target recognition model to capture the internal structure of the input data, it is determined whether to maintain a high level of abstraction, specifically:

[0038] When the prediction uncertainty decreases or remains unchanged, and the general object recognition model has a good ability to capture the intrinsic structure of the input data, it is determined to maintain a high level of abstraction; otherwise, it is determined not to maintain a high level of abstraction.

[0039] On the other hand, the present invention provides a low-code development system based on general target recognition, including a scenario parameter extraction module, an abstract demand evaluation module, an initial level determination module, a high abstraction evaluation module, and an abstract level decision module;

[0040] Scenario parameter extraction module: extracts application scenario parameters from specific application scenarios input by developers;

[0041] Abstract demand evaluation module: Use fuzzy analytic hierarchy process to conduct multi-level fuzzy comprehensive evaluation on application scenario parameters to evaluate the abstract level requirement of general target recognition model;

[0042] Initial level determination module: According to the abstract level requirement of the general target recognition model, the initial abstract level is determined by using the grey correlation analysis method;

[0043] High-abstraction evaluation module: When the initial abstraction level is high, a time series autoregressive analysis is performed on the prediction entropy value of the general target recognition model in recent execution to evaluate the changing trend of the prediction uncertainty of the general target recognition model; the stability of the feature representation of the general target recognition model is evaluated using the topological data analysis method to determine the general target recognition model's ability to capture the intrinsic structure of the input data;

[0044] Abstraction level decision module: Based on the comprehensive analysis of the changing trend of the uncertainty predicted by the general target recognition model and the ability of the general target recognition model to capture the intrinsic structure of the input data, it determines whether to maintain a high level of abstraction.

[0045] The technical effects and advantages of a low-code development method and system based on general target recognition in the present invention are as follows:

[0046] 1. By extracting key parameters from the application scenarios input by developers and using the fuzzy hierarchical analysis method and grey correlation analysis method to automatically determine the abstraction level of the model, the difficulty of model selection for developers is greatly reduced and development efficiency is improved. It is not only suitable for complex scenarios, but also can dynamically adjust the complexity of the model in simple scenarios to ensure that the model can achieve optimal performance in different application environments, effectively reducing resource waste.

[0047] 2. The prediction uncertainty of the model and the stability of feature representation are evaluated through time series autoregression analysis and topological data analysis. This in-depth analysis mechanism enables the model to be dynamically optimized according to its performance in actual execution. When the prediction uncertainty decreases, a high level of abstraction is maintained to ensure computational efficiency. When the prediction uncertainty increases or the feature capture capability is insufficient, the recognition accuracy is improved by lowering the level of abstraction, thereby achieving flexible adaptability and efficient operation of the model in an environment with limited computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A schematic diagram of a low-code development method based on general target recognition according to the present invention;

[0049] Figure 2 This is a structural diagram of a low-code development system based on general target recognition in the present invention. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Example

[0051] Figure 1 The present invention provides a low-code development method based on general target recognition, which includes the following steps:

[0052] S1: Extract application scenario parameters from the specific application scenario input by the developer.

[0053] S2: Use the fuzzy hierarchical analysis method to conduct a multi-level fuzzy comprehensive evaluation of the application scenario parameters to assess the abstract level requirement of the general target recognition model.

[0054] S3: According to the abstraction level requirement of the general target recognition model, the grey relational analysis method is used to determine the initial abstraction level.

[0055] S4: When the initial abstraction level is high: perform time series autoregression analysis on the prediction entropy value of the general target recognition model in recent executions to evaluate the changing trend of the prediction uncertainty of the general target recognition model; use topological data analysis methods to evaluate the stability of the feature representation of the general target recognition model to determine the general target recognition model's ability to capture the intrinsic structure of the input data.

[0056] S5: Based on a comprehensive analysis of the changing trend of the uncertainty prediction of the general target recognition model and the general target recognition model's ability to capture the intrinsic structure of the input data, determine whether to maintain a high level of abstraction.

[0057] S1 specifically includes:

[0058] S101: The developer enters a description of a specific application scenario on the low-code development platform:

[0059] Developers enter specific task requirements related to target identification through the user interface of the low-code platform. The input information includes but is not limited to the following aspects:

[0060] Target type: For example, the object to be identified is a person, vehicle, tree, text, etc.

[0061] Identify the scene: including the physical environment of the scene (such as indoors, outdoors, complex backgrounds), lighting conditions, weather conditions (such as sunny days, rainy days), etc.

[0062] Specific tasks: For example, developers may need to identify the number of people or trees in a picture, or classify and count certain items.

[0063] Application purpose: Specific application scenarios such as real-time monitoring, image analysis, object classification, and whether real-time processing is required.

[0064] S102: Automatically parse the scene description, analyze the scene through a preset complexity classification algorithm, and extract scene complexity features. Scene complexity features include but are not limited to the number of objects in the scene, background complexity, overlap and occlusion between objects, etc.

[0065] Among them, the platform will evaluate the expected number of targets in the scene. A larger number of targets means a higher complexity of the scene.

[0066] By analyzing image features such as background texture and color changes, the complexity of the background is quantified. Complex backgrounds usually have multiple different colors, materials, and dynamic elements.

[0067] Through the deep learning-based occlusion detection algorithm, the degree of occlusion between multiple targets in the scene is evaluated. When there is a lot of occlusion, the difficulty of recognition will increase.

[0068] Use an algorithm based on image segmentation to evaluate whether multiple objects overlap. Overlapping areas will affect the accurate detection of objects, thereby increasing the complexity of the scene.

[0069] S103: According to the task goal input by the developer, determine the requirement of the task on the target recognition accuracy, and extract the recognition accuracy requirement features. The recognition accuracy requirement features include classification accuracy, detection confidence, false alarm rate and missed alarm rate, etc.

[0070] Classification accuracy requirements: Based on the task's requirements for classification accuracy (such as the proportion of correctly classified targets in image classification), the platform analyzes whether a high-precision classification algorithm is needed.

[0071] Detection confidence analysis: By calculating the confidence distribution of each identified target, it is determined whether a high confidence (close to 1.0) is required to meet the developer's needs.

[0072] False Positive Rate Evaluation: The system evaluates the tolerance of false positives in the scene. Tasks with lower false positive rates require higher recognition accuracy.

[0073] False negative rate evaluation: By analyzing the false negatives (False Negative) in the target recognition process, the tolerance of the task to false negatives is evaluated. The lower the false negative rate, the higher the coverage capability required of the model.

[0074] S104: Based on the system resource configuration or execution environment input by the developer, currently available computing resources are evaluated and computing resource limitation features are extracted.

[0075] Hardware resource evaluation: Based on the computing resource information provided by the developer, such as available CPU / GPU computing power, memory size, etc., the system automatically evaluates the maximum processing capacity of the resources. For GPU computing power, it can be quantified by indicators such as the number of CUDA cores and memory bandwidth.

[0076] Real-time processing requirements: If the task requires real-time processing, the platform will calculate whether the system can complete image processing, recognition and output within the specified time window, including evaluations such as processing frame rate and latency.

[0077] Resource load assessment: Use a load prediction algorithm based on historical data to assess the current system load and ensure that running the model under existing resource conditions will not cause performance bottlenecks or resource overload.

[0078] S105: Summarize the extracted scene complexity features, recognition accuracy requirement features, and computing resource limitation features to form application scenario parameters.

[0079] S2 specifically includes:

[0080] S201: Construct a fuzzy hierarchical structure based on the extracted application scenario parameters, divide it into a target layer, a criterion layer, and a parameter layer, and establish a multi-level fuzzy decision model:

[0081] First, the extracted application scenario parameters are taken as input and divided into three levels to establish a clear fuzzy hierarchical decision model. The three levels are:

[0082] Target layer: The abstract level requirements of the evaluation model, which is the ultimate goal of this decision model. The target layer is used to define the ultimate goal to be achieved, that is, to determine the abstract level requirements of the general object recognition model through application scenario parameters.

[0083] Criteria layer: The criteria layer includes key scene feature parameters, such as scene complexity, recognition accuracy requirements, and computing resource constraints. These criteria reflect the core criteria that influence the choice of model abstraction level and are the intermediary between the target layer and the parameter layer.

[0084] Parameter layer: The parameter layer contains the actual scene parameter values. For example, scene complexity can be subdivided into background complexity, number of targets, and degree of occlusion; recognition accuracy requirements can be refined into detection confidence, false alarm rate, and missed alarm rate; computing resource limitations include the CPU / GPU computing power and memory size of the computing device.

[0085] Through this three-layer structure, the fuzzy hierarchical analysis method can analyze the impact of each parameter on the final decision (i.e., the model abstraction level) layer by layer. The weight and impact of the parameters in each layer are calculated and synthesized through subsequent steps.

[0086] S202: In the hierarchical model, by performing pairwise comparisons between parameters of various application scenarios, a fuzzy comparison matrix is ​​constructed:

[0087] For each parameter in the criterion layer and parameter layer, we need to compare them two by two. That is, we need to compare the scene complexity with the recognition accuracy requirements, the recognition accuracy requirements with the computing resource limitations, etc. When comparing, based on the developer's needs for the task, use qualitative language (such as "slightly important" or "very important") to express the relative importance of each parameter. Through the fuzzy quantification of language, quantitative values ​​are formed. For example, the following fuzzy values ​​can be defined: "equally important" = 1, "slightly important" = 3, "obviously important" = 5, "very important" = 7, "extremely important" = 9.

[0088] For example, assuming there are three criteria layer parameters A1, A2 and A3, the scene complexity parameter is A1, the recognition accuracy requirement is A2, and the computing resource limit is A3, the following fuzzy contrast matrix is ​​constructed: ;in, is the fuzzy contrast matrix, represents the relative importance between A1 and A2, Indicates the relative importance between A1 and A3, Indicates the relative importance between A2 and A3, and other elements are determined according to symmetry. The relative weights between the parameters of each application scenario are determined through expert or system automatic judgment, and the matrix is ​​filled with fuzzy values.

[0089] S203: Calculate the fuzzy weight of each application scenario parameter through fuzzy consistency judgment and weight allocation algorithm:

[0090] After constructing the fuzzy contrast matrix, check the consistency of the matrix, that is, determine whether the relative importance of the application scenario parameters is logical. The consistency index can be measured by calculating the consistency ratio (CR, Consistency Ratio). Specifically, for example, when the consistency ratio is less than 0.1, the fuzzy contrast matrix is ​​considered to have reasonable consistency, otherwise the contrast value needs to be readjusted.

[0091] The fuzzy comprehensive evaluation method is combined with the pairwise comparison matrix to calculate the fuzzy weight of each application scenario parameter. Assume that the fuzzy weights finally calculated are , and ;in, is the fuzzy weight of scene complexity, To identify the fuzzy weights of accuracy requirements, is the fuzzy weight for computing resource constraints.

[0092] S204: Perform fuzzy comprehensive operation on the fuzzy contrast matrix, and aggregate the fuzzy evaluation values ​​of the application scenario parameters at each layer into a model abstract selection index by combining the weights:

[0093] According to the fuzzy weight and fuzzy evaluation value of each application scenario parameter, the model abstract selection index is calculated, and its expression is: ;in, Represents the model abstract selection index, is the calculated fuzzy weight of each application scenario parameter, is the corresponding fuzzy evaluation value, is the number of application scenario parameters, It is the number of the application scenario parameter.

[0094] The influence of parameters at all levels is comprehensively considered through the model abstract selection index, and each fuzzy evaluation value is aggregated into a global evaluation result.

[0095] The fuzzy evaluation values ​​of the criterion layer and the parameter layer are integrated layer by layer, and the fuzzy evaluation values ​​of each layer are gradually accumulated to the final model abstract selection index, which can reflect the influence of parameters at different levels and ensure the integrity and logical rationality of the entire decision-making model.

[0096] The larger the model abstraction selection index is, the lower the abstraction level required for the general target recognition model is, indicating that the current scene has a high demand for the refinement of the recognition model and a lower level of abstraction needs to be selected so that the general target recognition model can process and capture the complex features in the scene more finely.

[0097] S3 specifically includes:

[0098] S301: construct a standard reference sequence according to the model abstract selection index, and convert the model abstract selection index into a corresponding comparison sequence:

[0099] The standard reference sequence is a reference value that reflects the performance of each abstraction level under ideal conditions. In order to ensure that different abstraction levels can be adapted to the application scenario, an ideal performance reference value sequence is constructed for each abstraction level. Set the standard reference sequence as ;in, to Corresponding to the ideal values ​​of each abstraction level.

[0100] The comparison sequence is based on the model abstract selection index of the actual scenario, which is converted into a corresponding sequence for comparison with the standard reference sequence. ;in, to The model abstraction selection indexes corresponding to the scenarios respectively.

[0101] S302: Using the grey correlation analysis method, the grey correlation values ​​of each abstraction level are calculated based on the standard reference sequence and the comparison sequence to reflect the degree of conformity between different abstraction levels and the requirements:

[0102] Grey correlation analysis is a method for evaluating the similarity between different objects. Here, we use this method to calculate the correlation value of each abstraction level to determine which abstraction level best meets the needs of the current application scenario. The specific process is as follows:

[0103] Calculate the absolute difference between the standard reference sequence and the comparison sequence. The smaller the difference, the closer the two sequences are, which means that the level of abstraction is more in line with the current needs. The calculation formula for the absolute difference between the standard reference sequence and the comparison sequence is: ;in, It is The absolute difference in the level of abstraction, is the first values, It is the first value.

[0104] In order to normalize the difference values, it is necessary to calculate the maximum and minimum possible differences in all sequences, setting: ; .

[0105] in, is the biggest difference between the various levels of abstraction; is the smallest difference between the levels of abstraction; Indicates the number of model abstraction levels, specifically the total number of different model abstraction levels that can be selected in different application scenarios; Indicates the index of the specific abstraction level currently being calculated, used to mark and process specific abstraction levels.

[0106] The grey relational value of each abstraction level is obtained through the grey relational calculation formula, and its expression is: ;in, It is Grey relational value of the abstract level; is the resolution coefficient, usually set to 0.5, which is used to adjust the impact of differences on association.

[0107] The larger the grey correlation value is, the more consistent the abstraction level is with the requirements of the current application scenario.

[0108] Through the above steps, the grey relational values ​​of all abstraction levels are obtained: ;in, Represents the set of grey relational values ​​of all abstraction levels, to are the grey relational values ​​of each abstraction level.

[0109] S303: According to the calculated grey correlation value, the abstraction level with the highest correlation is selected as the initial model abstraction level to ensure that the selected abstraction level best matches the current scenario requirements:

[0110] By comparing the grey relational values ​​of all abstraction levels, find the maximum value , the corresponding abstraction level is ,Right now .

[0111] is with The corresponding level of abstraction, It represents the abstraction level index with the highest grey relational value. Its corresponding abstraction level is the abstraction level selected as the initial model. is selected as the initial level of abstraction for the model; in this way, it is possible to determine the initial level of abstraction that best meets the needs of the current scenario.

[0112] It includes high abstraction level and low abstraction level; among them, high abstraction level usually means that the model structure is relatively simple and has fewer parameters, which is suitable for processing simple scenes or tasks that do not require high precision. It can process quickly, but the accuracy of complex scenes is limited. Low abstraction level means that the model structure is complex and has many parameters. It can process complex scenes with rich details and is suitable for high-precision recognition tasks, although it may require more computing resources and time.

[0113] The time series autoregression analysis of the prediction entropy values ​​in the recent execution of the general target recognition model is carried out to evaluate the changing trend of the prediction uncertainty of the general target recognition model, specifically:

[0114] Extract the prediction entropy value of each execution from the execution results of the general object recognition model. Each time the model is executed, it predicts the input data and outputs a probability distribution, which indicates the confidence of the model in the prediction result. The prediction entropy value is an indicator to measure the uncertainty of the probability distribution, which is used to indicate the uncertainty of the model in the prediction result at the current moment.

[0115] That is, setting the model output probability distribution ,in is the class predicted by the model, is the input data. The model predicts entropy value At the moment is defined as: ;in, Represents the model at time The predicted entropy value of Indicates that given the input data The model predicts the category The probability of represents the number of categories, Represents the category index.

[0116] The larger the prediction entropy value, the more uncertain the general object recognition model is.

[0117] Extract the predicted entropy value from the execution record of the model, and arrange the predicted entropy value at each moment in chronological order to form a predicted entropy value sequence: ;in, Represents the predicted entropy value sequence of the model at different time points, From the moment To time The predicted entropy value of Represents the total number of moments in a time series. This series reflects the dynamic changes in the model's prediction uncertainty of the input data over a period of time.

[0118] The autoregressive integrated moving average (ARIMA) model is used to analyze the predicted entropy value sequence and further predict the entropy value change trend in the future.

[0119] The Autoregressive Integrated Moving Average (ARIMA) model is a commonly used time series analysis model that can capture the characteristics of time series through a combination of autoregression and moving average. The mathematical expression of the ARIMA model is: ;in, is a constant term; is the autoregression coefficient, indicating the The impact of the entropy value before a moment on the current entropy value; is the autoregressive order, indicating how many past entropy values ​​are used to predict the current entropy value; is the moving average coefficient, which indicates the impact of past prediction errors on the current entropy value; is the moving average order, indicating how many past errors are used to correct the current forecast value; is the error term at the current moment; Represents the autoregressive order in the autoregressive model, which refers to the number of past data points used as reference when predicting the current value, that is, the number of past time points reviewed by the model; Represents the moving average order in the moving average model, which refers to the number of past forecast errors used as a reference when correcting the current forecast value.

[0120] By training the ARIMA model, the entropy value at future moments can be predicted based on the previous predicted entropy value sequence. The predicted future entropy value can help determine whether the prediction uncertainty of the model will continue to rise or fall.

[0121] After completing the autoregressive analysis and predicting the future entropy value series, it is necessary to evaluate the changing trend of the model predicted entropy value to determine whether there is a significant upward or downward trend.

[0122] By observing the changes in the predicted entropy value, we can identify whether the model's prediction uncertainty is showing an upward trend. If the model's predicted entropy value shows a significant increase, it indicates that the model's prediction uncertainty is increasing, and the model may no longer be stable when processing the current task. Conversely, if the predicted entropy value shows a downward trend, it indicates that the model's prediction of the input data is gradually stabilizing and the prediction uncertainty is decreasing.

[0123] By calculating the rate of change of the predicted entropy value, the trend of uncertainty change can be quantified. The formula is as follows: ;in, is the rate of change of entropy.

[0124] when When , it means that the predicted entropy value is increasing; when When , it means that the predicted entropy value is decreasing; when When , it means that the predicted entropy value at the next moment is equal to the entropy value at the current moment, and the prediction uncertainty of the model remains unchanged.

[0125] If the forecast results show that the model's prediction uncertainty continues to increase, it may be necessary to adjust the model's abstraction level (such as lowering the abstraction level and using a more sophisticated model). If the forecast results show that the uncertainty decreases or remains stable, the current abstraction level can be maintained.

[0126] when When , it is determined that the prediction uncertainty increases and the model abstraction level needs to be lowered.

[0127] when When , it is determined that the prediction uncertainty decreases or remains unchanged, and the current level of abstraction needs to be maintained.

[0128] Based on the rate of change of the predicted entropy value, the adaptability of the model to the current task can be dynamically judged, and it can be decided whether the abstraction level of the general object recognition model needs to be adjusted.

[0129] The topological data analysis method is used to evaluate the stability of the feature representation of the general object recognition model and determine the general object recognition model's ability to capture the intrinsic structure of the input data, including:

[0130] Extract high-dimensional feature representations from the middle layer of the general object recognition model to form a feature space dataset:

[0131] After processing the input data, the general object recognition model generates feature representations in the middle layer. This feature representation can be the output of different layers such as the convolutional layer and the fully connected layer of the model, reflecting the feature extraction results of the model on the input data. The extracted feature representation forms high-dimensional data, and each feature representation can be regarded as a high-dimensional vector.

[0132] By running the input data multiple times or through different batches of input data, these high-dimensional feature representations are collected and formed into a feature space dataset. The feature space dataset contains the feature capture results of the model on multiple input data.

[0133] Based on the extracted high-dimensional feature representation, the topological data analysis method is used to construct persistent homology, generate the topological structure of the feature space, and capture the multi-scale morphological characteristics of the data:

[0134] Persistent homology is a topological tool that can analyze the morphological characteristics of a dataset at different scales. By varying the scale of the connections between data points, topological properties such as connectivity, loops, and holes can be discovered in the feature space. This multi-scale analysis can reveal both the geometric and topological structure of the data.

[0135] At different scales, by calculating the distance between feature representations, a gradually connected topological structure is formed. At a smaller scale, the distance between data points is far, and only a small number of connections are formed; as the scale increases, the data points are gradually connected, generating more complex forms, such as topological features such as rings and holes.

[0136] The stability of the model's feature representation is quantified by persistent coherence calculation of topological invariants, which evaluates the model's ability to capture the intrinsic structure of the data:

[0137] Topological invariants are used to measure the persistence characteristics of data structures. For each topological feature (such as a connected component or a ring), its appearance and disappearance at different scales can be regarded as a "lifetime", which is used to measure the persistence of the topological feature. Topological invariants quantify the stability of feature representation by analyzing these lifetimes. Topological features with long lifetimes indicate that the structure remains stable at multiple scales and the model has a strong ability to capture data.

[0138] By calculating the topological invariant of each feature, the stability of the model feature representation can be quantified into one or more values. The larger the topological invariant, the more stable the model is in capturing the data structure and the more consistent the model features are; conversely, if the topological invariant is small, it means that the model feature representation is relatively unstable and lacks the ability to capture the structure of the input data.

[0139] Compare the feature topology at different time points or input data, and use the bottleneck distance to measure the change in feature representation:

[0140] Extract high-dimensional feature representations at different time points or input data and construct corresponding topological structures. Comparing these different topological structures can reveal whether the model performs consistently under different conditions.

[0141] The bottleneck distance is a distance metric used to compare the differences between two topological structures. By comparing the changes in topological features under different input data or time points, the bottleneck distance can quantify these differences. If the bottleneck distance is small, it means that the two topological structures are similar and the feature representation of the model remains stable; if the bottleneck distance is large, it means that the feature representation changes significantly and the model may not be able to capture the input data stably.

[0142] According to the degree of change of topological invariants, the ability of the general object recognition model to capture the intrinsic structure of the input data is judged:

[0143] The model's ability to capture the intrinsic structure of the input data is determined based on the calculated topological invariants and changes in the bottleneck distance.

[0144] If the topological invariant remains relatively stable under different input data or at different time points, it means that the feature representation of the model is consistent and can effectively capture the intrinsic structure of the input data; if the topological invariant fluctuates greatly, it means that the feature representation of the model is unstable in its ability to capture the input data.

[0145] The ability of general object recognition models to capture the structure of input data is evaluated through a comprehensive analysis of topological invariants and bottleneck distances.

[0146] If the model maintains a stable and consistent feature representation in different scenarios, it is judged that the general target recognition model has a good ability to capture the intrinsic structure of the input data, indicating that the model has good capture capabilities; if the model's feature representation fluctuates greatly under different inputs, it is judged that the general target recognition model has a poor ability to capture the intrinsic structure of the input data, indicating that the model's capture ability is poor and it is necessary to further adjust the model parameters or the level of abstraction to improve the data capture capability.

[0147] Based on the comprehensive analysis of the changing trend of the uncertainty of the general target recognition model prediction and the ability of the general target recognition model to capture the internal structure of the input data, it is judged whether to maintain a high level of abstraction, specifically:

[0148] If the prediction uncertainty decreases or remains the same, it means that the model has good prediction stability in the current task scenario and the model can confidently and reliably identify. This indicates that the model's performance is relatively stable and meets the conditions for maintaining a high level of abstraction. If the general object recognition model can better capture the inherent structure of the input data, it means that the model can correctly identify and reflect the characteristics of the data without more detailed adjustments. Therefore, the model's understanding and processing of the data at the current level of abstraction is effective.

[0149] When the prediction uncertainty decreases or remains unchanged, and the model has a good ability to capture the data structure, it means that the model can accurately handle the task at the current high level of abstraction, and it is reasonable to maintain a high level of abstraction. If the prediction uncertainty increases, or the model's ability to capture data is poor, it means that the model is not adaptable enough to the current task, and the level of abstraction needs to be lowered to improve the understanding and processing capabilities of complex data.

[0150] Therefore: when the prediction uncertainty decreases or remains unchanged, and the general target recognition model has a good ability to capture the intrinsic structure of the input data, it is determined to maintain a high level of abstraction; otherwise, it is determined not to maintain a high level of abstraction, and it is necessary to lower the level of abstraction to improve the understanding and processing capabilities of complex data.

[0151] Example 2

[0152] The difference between Example 2 of the present invention and Example 1 is that this example introduces a low-code development system based on general target recognition.

[0153] Figure 2 A structural schematic diagram of a low-code development system based on general target recognition is given in the present invention. A low-code development system based on general target recognition includes a scenario parameter extraction module, an abstract demand evaluation module, an initial level determination module, a high-abstraction evaluation module and an abstract level decision module.

[0154] Scenario parameter extraction module: extracts application scenario parameters from the specific application scenario input by the developer.

[0155] Abstract requirement evaluation module: Use the fuzzy hierarchical analysis method to conduct a multi-level fuzzy comprehensive evaluation of the application scenario parameters to evaluate the abstract level requirement of the general target recognition model.

[0156] Initial level determination module: According to the abstract level requirement of the general target recognition model, the initial abstraction level is determined by using the grey correlation analysis method.

[0157] High-abstraction evaluation module: When the initial abstraction level is high, a time series autoregression analysis is performed on the prediction entropy value in the recent execution of the general target recognition model to evaluate the changing trend of the prediction uncertainty of the general target recognition model; the topological data analysis method is used to evaluate the stability of the feature representation of the general target recognition model to determine the general target recognition model's ability to capture the intrinsic structure of the input data.

[0158] Abstraction level decision module: Based on the comprehensive analysis of the changing trend of the uncertainty predicted by the general target recognition model and the ability of the general target recognition model to capture the intrinsic structure of the input data, it determines whether to maintain a high level of abstraction.

[0159] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters and thresholds in the formula are set by technicians in this field according to actual conditions.

[0160] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented by software, the above embodiments may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or may be transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from one website, computer, server or data center to another website, computer, server or data center by wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that contains one or more available media sets. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium may be a solid-state hard disk.

[0161] Those of ordinary skill in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0162] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and modules described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0163] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the modules is only a logical function division. There may be other division methods in actual implementation, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0164] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical modules, and may be located in one place or distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0165] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0166] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program codes.

[0167] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0168] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A low-code development method based on general object recognition, characterized in that: The steps include: S1: extracting application scenario parameters from the specific application scenario input by the developer; S2: Use fuzzy analytic hierarchy process to conduct multi-level fuzzy comprehensive evaluation on application scenario parameters to assess the abstract level requirement of general object recognition model; S3: According to the abstraction level requirement of the general target recognition model, the grey relational analysis method is used to determine the initial abstraction level; S4: When the initial abstraction level is high: perform time series autoregression analysis on the prediction entropy value of the general target recognition model in recent executions to evaluate the changing trend of the prediction uncertainty of the general target recognition model; use topological data analysis methods to evaluate the stability of the feature representation of the general target recognition model to determine the general target recognition model's ability to capture the intrinsic structure of the input data; S5: Based on the comprehensive analysis of the changing trend of the uncertainty prediction of the general target recognition model and the ability of the general target recognition model to capture the internal structure of the input data, determine whether to maintain a high level of abstraction; S2 specifically includes: S201: construct a fuzzy hierarchical structure according to the extracted application scenario parameters, divide it into a target layer, a criterion layer and a parameter layer, and establish a multi-level fuzzy decision model; S202: In the hierarchical model, by performing pairwise comparisons between parameters of various application scenarios, a fuzzy comparison matrix is ​​constructed; S203: Calculating the fuzzy weight of each application scenario parameter through fuzzy consistency judgment and weight allocation algorithm; S204: Perform fuzzy comprehensive operation on the fuzzy contrast matrix, and aggregate the fuzzy evaluation values ​​of the application scenario parameters at each layer into a model abstract selection index by combining the weights: According to the fuzzy weight and fuzzy evaluation value of each application scenario parameter, the model abstract selection index is calculated, and its expression is: ;in, Represents the model abstract selection index, is the calculated fuzzy weight of each application scenario parameter, is the corresponding fuzzy evaluation value, is the number of application scenario parameters, is the number of the application scenario parameter; S3 specifically includes: S301: constructing a standard reference sequence according to the model abstract selection index, and converting the model abstract selection index into a corresponding comparison sequence; S302: using a grey relational analysis method to calculate the grey relational values ​​of each abstraction level based on the standard reference sequence and the comparison sequence; S303: According to the calculated grey correlation value, select the abstraction level with the highest correlation as the initial model abstraction level.

2. According to a low-code development method based on general target recognition according to claim 1, it is characterized in that: S1 specifically includes: S101: The developer enters a description of a specific application scenario on the low-code development platform; S102: automatically parsing the scene description, analyzing the scene using a preset complexity classification algorithm, and extracting scene complexity features; S103: According to the task goal input by the developer, determine the requirement of the task on the target recognition accuracy, and extract the recognition accuracy requirement feature; S104: Based on the system resource configuration or execution environment input by the developer, evaluate the currently available computing resources and extract computing resource limitation features; S105: Summarize the extracted scene complexity features, recognition accuracy requirement features, and computing resource limitation features to form application scenario parameters.

3. A low-code development method based on general target recognition according to claim 2, characterized in that: The time series autoregression analysis of the prediction entropy values ​​in the recent execution of the general target recognition model is carried out to evaluate the changing trend of the prediction uncertainty of the general target recognition model, specifically: Extract the predicted entropy value of each execution from the execution results of the general object recognition model: set the model output probability distribution ,in is the class predicted by the model, is the input data; the model predicts the entropy value At the moment is defined as: ;in, Represents the model at time The predicted entropy value of Indicates that given the input data The model predicts the category The probability of represents the number of categories, Represents the category index; Arrange the predicted entropy values ​​at each moment in chronological order to form a predicted entropy value sequence; The autoregressive integrated moving average model is used to analyze the predicted entropy value sequence and further predict the entropy value change trend in the future; Calculate the rate of change of the predicted entropy value: ;in, is the rate of change of entropy value; when When , the prediction uncertainty increases; when When , the prediction uncertainty is judged to be reduced or unchanged.

4. A low-code development method based on general target recognition according to claim 3, characterized in that: The topological data analysis method is used to evaluate the stability of the feature representation of the general object recognition model and determine the general object recognition model's ability to capture the intrinsic structure of the input data. Specifically, it includes: extracting high-dimensional feature representation from the middle layer of the general object recognition model to form a feature space data set; Based on the extracted high-dimensional feature representation, the topological data analysis method is used to construct persistent homology, generate the topological structure of the feature space, and capture the multi-scale morphological characteristics of the data; The stability of the model's feature representation is quantified by computing topological invariants through persistent homology, and the model's ability to capture the intrinsic structure of the data is evaluated; Compare the feature topologies at different time points or input data, and use the bottleneck distance to measure the changes in feature representation; According to the degree of change of the topological invariant, the ability of the general target recognition model to capture the intrinsic structure of the input data is judged: if the model maintains a stable and consistent feature representation in different scenarios, the general target recognition model is judged to have a good ability to capture the intrinsic structure of the input data; if the model's feature representation fluctuates greatly under different inputs, the general target recognition model is judged to have a poor ability to capture the intrinsic structure of the input data.

5. A low-code development method based on general target recognition according to claim 4, characterized in that: Based on the comprehensive analysis of the changing trend of the uncertainty of the general target recognition model prediction and the ability of the general target recognition model to capture the internal structure of the input data, it is judged whether to maintain a high level of abstraction, specifically: When the prediction uncertainty decreases or remains unchanged, and the general object recognition model has a good ability to capture the intrinsic structure of the input data, it is determined to maintain a high level of abstraction; otherwise, it is determined not to maintain a high level of abstraction.

6. A low-code development system based on general target recognition, used to implement a low-code development method based on general target recognition as described in any one of claims 1 to 5, characterized in that: It includes scenario parameter extraction module, abstract demand assessment module, initial level determination module, high abstraction assessment module and abstract level decision module; Scenario parameter extraction module: extracts application scenario parameters from specific application scenarios input by developers; Abstract demand evaluation module: Use fuzzy analytic hierarchy process to conduct multi-level fuzzy comprehensive evaluation on application scenario parameters to evaluate the abstract level requirement of general target recognition model; Initial level determination module: According to the abstract level requirement of the general target recognition model, the initial abstract level is determined by using the grey correlation analysis method; High-abstraction evaluation module: When the initial abstraction level is high, a time series autoregressive analysis is performed on the prediction entropy value of the general target recognition model in recent execution to evaluate the changing trend of the prediction uncertainty of the general target recognition model; the stability of the feature representation of the general target recognition model is evaluated using the topological data analysis method to determine the general target recognition model's ability to capture the intrinsic structure of the input data; Abstraction level decision module: Based on the comprehensive analysis of the changing trend of the uncertainty predicted by the general target recognition model and the ability of the general target recognition model to capture the intrinsic structure of the input data, it is judged whether to maintain a high level of abstraction.

Images