A method for extracting a color image contour by deep learning fusion of color histograms

Abstract

The application discloses a method for realizing autonomous training and updating model based on deep learning, belongs to the field of artificial intelligence and intelligent data processing technology, and aims at the problems that existing model updating relies on manual intervention, the updating process lacks closed-loop control and is difficult to adapt to dynamic data environment, proposes that data generated in the running process is collected and uniformly processed to form a standardized data set, quality determination and screening are performed on the standardized data set to form a training sample set, parameter updating is performed on the current model based on the training sample set to obtain a candidate model, performance evaluation is performed on the candidate model to form a model updating decision result, model version management and deployment are performed based on the model updating decision result to obtain the current running model, and updating trigger determination is performed based on new data in the model running process and is fed back to the standardized data set to form a closed-loop update, and the method is suitable for intelligent application scenarios in which models need to continuously learn and automatically update in a dynamic environment.

Description

Technical Field

[0001] It involves the autonomous iterative optimization of deep learning models and is suitable for the automatic detection of agricultural machinery appearance labels. Background Technology

[0002] Currently, deep learning-based model training and updating technologies are widely used in various fields such as image recognition, speech processing, industrial inspection, and intelligent decision-making. Existing technologies typically employ a combination of offline training and online inference. This involves pre-building a training dataset, training the model in a centralized computing environment, and then deploying the trained model to terminals or business systems for inference applications. In practical applications, when data distribution changes or business scenarios shift, it is usually necessary to recollect data and perform centralized retraining to update and optimize model performance.

[0003] Furthermore, to improve the adaptability and real-time performance of models, some studies have begun to introduce online learning or incremental learning mechanisms, enabling models to update parameters based on new data during operation. For example, by continuously receiving new samples and performing mini-batch training, dynamic adjustment of model parameters can be achieved; or transfer learning methods can be used to perform local fine-tuning on existing models to reduce training costs. In addition, distributed training methods such as federated learning are increasingly being applied in privacy-sensitive scenarios, achieving model updates through multi-node collaborative training and avoiding the centralized transmission of original datasets.

[0004] However, the aforementioned existing technologies still have certain limitations. On the one hand, offline centralized training methods rely on manual intervention, have long update cycles, and struggle to respond promptly to data changes. On the other hand, online or incremental learning methods often lack stable training triggering mechanisms and quality control measures in actual deployment, easily introducing noisy data or causing model performance fluctuations. Furthermore, most solutions fail to form a closed-loop adaptive update system from data collection, sample selection, model training to model deployment, resulting in a fragmented and low-automation model update process. In addition, in multi-scenario or complex environments, model update strategies lack targeted scheduling and dynamic optimization capabilities, making it difficult to balance training efficiency and model performance.

[0005] In summary, existing technologies suffer from several drawbacks, including a high degree of reliance on manual intervention for model updates, a lack of closed-loop adaptive mechanisms in the update process, difficulty in effectively controlling the quality of training data, and insufficient adaptability of models in dynamic environments. Summary of the Invention

[0006] To address the shortcomings of existing technologies, such as high reliance on manual intervention in model updates, lack of closed-loop adaptive mechanisms in the update process, difficulty in effectively controlling training data quality, and insufficient model adaptability in dynamic environments, the technical solution provided by this invention is as follows: A method for extracting color image contours by fusing color histograms using deep learning includes: The steps involved in collecting operational data from the target system and performing unified processing to form a standardized dataset. The step of performing quality assessment and screening on the standardized dataset to form a training sample set; The step of updating the parameters of the current model based on the training sample set to obtain a candidate model; The steps include evaluating the performance of the candidate models and comparing them with the current model to form a model update decision result; Based on the model update decision results, perform version management and deployment of the model to obtain the currently running model; The steps involve performing an update trigger decision based on the data generated during the operation of the current running model, and feeding the new data back to the standardized data set to form a loop in order to achieve continuous adaptive updating of the model.

[0007] Furthermore, in a preferred embodiment, when processing the standardized dataset in a unified manner, the multi-source data generated during the operation is cleaned, format aligned, and tag-completed.

[0008] Furthermore, in a preferred embodiment, when performing quality assessment and screening on the standardized dataset, abnormal data, duplicate data, and low-confidence data are identified and removed, and the data are scored and screened based on the stability of the data source and the consistency of the labels.

[0009] Furthermore, in a preferred embodiment, when updating the parameters of the current model based on the training sample set, the model parameters are iteratively adjusted by inputting the training sample set in batches, and the existing model capabilities are kept stable during the parameter update process.

[0010] Furthermore, in a preferred embodiment, when evaluating the performance of the candidate model, the degree of matching between the candidate model output and the real result is statistically analyzed based on an independent validation dataset, and compared with the performance of the current model.

[0011] Furthermore, in a preferred embodiment, when performing version management and deployment of the model based on the model update decision results, a unique identifier is assigned to the candidate model that passes the evaluation and the currently running model is replaced, while historical models are retained.

[0012] Based on the same inventive concept, the present invention also provides an apparatus for extracting color image contours by fusing color histograms using deep learning, comprising: A module that collects operational data from the target system and performs unified processing to form a standardized data set; A module for performing quality assessment and screening on the standardized dataset to form a training sample set; The current model is updated with parameters based on the training sample set to obtain a module of candidate models; A module that evaluates the performance of the candidate models and compares them with the current model to form a model update decision result; Based on the model update decision results, version management and deployment are performed on the model to obtain the modules of the currently running model; The module that performs update triggering judgment based on the data generated during the operation of the current running model and feeds the new data back to the standardized data set to form a loop in order to achieve continuous adaptive updating of the model.

[0013] Based on the same inventive concept, the present invention also provides a computer storage medium for storing a computer program, wherein when the computer program is read by a computer, the computer executes the method described thereon.

[0014] Based on the same inventive concept, the present invention also provides a computer, including a processor and a storage medium, wherein when the processor reads a computer program stored in the storage medium, the computer executes the method described thereon.

[0015] Based on the same inventive concept, the present invention also provides a computer program product, which, when executed, implements the method described.

[0016] Compared with the prior art, the advantages of the technical solution provided by the present invention are as follows: By constructing a data-driven model self-training trigger mechanism, the input data is continuously monitored during operation, and the system determines whether to enter the training process based on preset conditions. This eliminates the need for manual intervention in model updates, thereby improving the automation and timeliness of model updates. This effect stems from the real-time data stream acquisition and training trigger condition setting process in the solution. Compared to the existing method that relies on periodic offline retraining, this significantly shortens the time it takes for the model to respond to data changes.

[0017] By screening and quality assessment of the collected data, and removing abnormal samples, low-confidence samples, or noisy data before entering the training process, the data participating in model updates has higher reliability, thereby avoiding performance fluctuations or even degradation of the model during the update process. This effect comes from the data preprocessing and sample screening steps in the solution. Compared with the existing online learning method of directly using real-time data for training, it effectively improves training stability and result credibility.

[0018] By constructing an incremental training mechanism, the model is locally updated on new data based on the existing model instead of being fully retrained. This allows the model to gradually adapt to new scenarios while maintaining its existing capabilities, thereby reducing computational resource consumption and improving training efficiency. This effect comes from the continuous iterative training process based on historical models in the solution. Compared with the existing full retraining method, it can significantly reduce training time and improve resource utilization.

[0019] By introducing a model performance evaluation and feedback mechanism, the model's performance is verified after the update, and a decision is made on whether to replace or roll back the model based on the evaluation results. This enables the model update process to have closed-loop control capabilities, thereby avoiding system performance degradation due to erroneous updates. This effect stems from the steps in the solution to evaluate the updated model and execute policy decisions. Compared with existing automatic update methods that lack verification mechanisms, this significantly improves the security and stability of system operation.

[0020] By establishing a model version management and dynamic deployment mechanism, different versions of the model can be switched in an orderly manner within the system, and the optimal model can be selected for application based on the actual running effect, thereby improving the model's adaptability in complex environments. This effect comes from the process of version control and deployment management of the model in the solution. Compared with the existing single model overlay update method, it can avoid the risks caused by the instability of the new model.

[0021] By integrating data collection, sample screening, model training, performance evaluation, and model deployment into a unified whole, a complete closed-loop self-updating system is formed. This enables the model to continuously self-optimize during operation, thereby improving the overall system's intelligence level and long-term performance. This effect stems from the collaborative processing mechanism between the various functional modules in the solution, which significantly improves the system's automation level and collaborative efficiency compared to the existing decentralized implementation of each link.

[0022] By adaptively scheduling the model update strategy in multiple scenarios or dynamic environments, and adjusting the training frequency and update strategy according to different data distributions or task requirements, the model can maintain good performance for different application scenarios. This effect comes from the control logic of dynamically adjusting the update strategy in the solution. Compared with the existing fixed strategy update method, it can improve the model's generalization ability and adaptability in complex environments.

[0023] It is suitable for intelligent application scenarios that require models to continuously learn, optimize, and automatically update and deploy in dynamic data environments. Attached Figure Description

[0024] Figure 1 This is the logic diagram of the method. Detailed Implementation

[0025] To make the advantages and benefits of the technical solution provided by the present invention clearer, the technical solution provided by the present invention will now be described in further detail with reference to the accompanying drawings, specifically: Implementation Method 1: This implementation method provides a method for extracting color image contours by fusing color histograms using deep learning, including: The steps involved in collecting operational data from the target system and performing unified processing to form a standardized dataset. The step of performing quality assessment and screening on the standardized dataset to form a training sample set; The step of updating the parameters of the current model based on the training sample set to obtain a candidate model; The steps include evaluating the performance of the candidate models and comparing them with the current model to form a model update decision result; Based on the model update decision results, perform version management and deployment of the model to obtain the currently running model; The steps involve performing an update trigger decision based on the data generated during the operation of the current running model, and feeding the new data back to the standardized data set to form a loop in order to achieve continuous adaptive updating of the model.

[0026] When processing the standardized dataset in a unified manner, the multi-source data generated during the process is cleaned, format aligned, and labeled.

[0027] When performing quality assessment and screening on the standardized dataset, abnormal data, duplicate data, and low-confidence data are identified and removed, and the data are scored and screened based on the stability of the data source and the consistency of the labels.

[0028] When updating the parameters of the current model based on the training sample set, the model parameters are iteratively adjusted by inputting the training sample set in batches, and the existing model capabilities are kept stable during the parameter update process.

[0029] When evaluating the performance of the candidate models, the degree of matching between the output results of the candidate models and the actual results is statistically analyzed based on the independent validation dataset, and compared with the performance of the current model.

[0030] When performing version management and deployment of the model based on the model update decision results, a unique identifier is assigned to the candidate model that passes the evaluation and the currently running model is replaced, while historical models are retained.

[0031] A device for extracting color image contours by fusing color histograms using deep learning is also provided, comprising: A module that collects operational data from the target system and performs unified processing to form a standardized data set; A module for performing quality assessment and screening on the standardized dataset to form a training sample set; The current model is updated with parameters based on the training sample set to obtain a module of candidate models; A module that evaluates the performance of the candidate models and compares them with the current model to form a model update decision result; Based on the model update decision results, version management and deployment are performed on the model to obtain the modules of the currently running model; The module that performs update triggering judgment based on the data generated during the operation of the current running model and feeds the new data back to the standardized data set to form a loop in order to achieve continuous adaptive updating of the model.

[0032] A computer storage medium is also provided for storing a computer program, which, when read by the computer, executes the method.

[0033] A computer is also provided, including a processor and a storage medium, wherein the computer executes the method when the processor reads a computer program stored in the storage medium.

[0034] A computer program product is also provided, which, when executed, implements the method described.

[0035] Implementation Method Two: This implementation method is a further detailed description of the technical solution provided in Implementation Method One, specifically: This embodiment provides a method for autonomously training and updating a model based on deep learning. By constructing a sequential connection between data collection, sample selection, model training, model evaluation, and model deployment, the model can continuously perform adaptive updates based on new data during operation.

[0036] First, the raw data generated during the operation of the target system is continuously collected, and the collected data is processed uniformly to form a data set with a consistent structure, which serves as input data for subsequent processing. During system operation, input data is received from business interfaces, log recording units, or sensing devices. The received data is organized according to a preset data structure, missing fields are filled in, abnormal fields are corrected, data from different sources are uniformly encoded, and data is classified and stored according to time or source, thus obtaining a data set with a consistent format. This data set is then output to the sample screening stage.

[0037] Subsequently, sample screening and quality assessment are performed on the dataset to obtain a sample set for model training. The system performs validity checks on each data point in the dataset, removing data with errors, conflicts, or incompleteness, deduplicating duplicate data, and calculating the credibility of each data point. By comprehensively judging the stability of the data source, label consistency, and historical performance, a quality score is assigned to each data point. Data that meets the preset score requirements are selected as training samples, thus forming a high-quality sample set, which is then output to the model training stage.

[0038] Next, the existing model's parameters are updated based on the sample set to obtain a candidate model. The system divides the sample set into multiple data subsets according to batches, and inputs each data subset into the current model sequentially. By calculating the difference between the model's output and the corresponding label, the internal parameters of the model are gradually adjusted so that the model can adapt to the feature changes reflected by the new data. During training, the stable part of the original model parameters is retained, and the parts with large changes are adjusted to obtain the updated candidate model, which is then output to the model evaluation stage.

[0039] Then, the candidate models are subjected to performance evaluation to obtain evaluation results and form model update decisions. The system applies the candidate models to the validation dataset, and by statistically analyzing the matching degree between the output results of the candidate models and the actual results of the validation dataset, the accuracy, stability, and consistency indicators of the models are obtained. These indicators are then compared with the corresponding indicators of the currently running model. When a candidate model meets preset conditions on multiple evaluation dimensions, it is determined to be a replaceable model; otherwise, it is determined to be a non-replaceable model, thus obtaining the model update decision result, which is output to the model management and deployment stage.

[0040] Subsequently, based on the model update decision results, version management and deployment operations are performed on the model to obtain the currently running model. The system assigns a unique identifier to each model that passes the evaluation and stores it in the model repository. When a candidate model is determined to be replaceable, it is loaded into the runtime environment to replace the original model, and the original model is transferred to the historical version sequence for rollback. When a candidate model is determined to be irreplaceable, the original model continues to run, thereby realizing dynamic model switching and version management. The currently running model is output to the business execution stage.

[0041] Finally, during model operation, continuous data acquisition and update triggering are performed, creating a closed-loop model update process. While the model performs its tasks, the system continues to collect new input data, detects the distribution differences between new and historical data, and statistically analyzes the model's output performance in actual tasks. When a change in data distribution exceeds a preset range or the model's performance shows a downward trend, a new round of model training is triggered, re-inputting the newly collected data into the data processing and sample selection stages. This cyclical execution of model training, evaluation, and deployment enables continuous adaptive updates of the model in dynamic environments.

[0042] Implementation Method 3, in conjunction with Appendix Figure 1 This embodiment describes the technical solution provided above in further detail through specific examples. Specifically: The purpose of this implementation is to overcome the shortcomings of existing technologies, support operators in completing label drawing through standardized label drawing software, and realize a closed-loop process of label verification, sample screening, model training and parameter updating by combining dynamic intervention of algorithm programs. This improves label quality and model training and update efficiency, optimizes model performance, provides a convenient method for updating the model, and enables flexible and low-cost automatic identification of expanded samples.

[0043] 1. Image import and preprocessing: The operator starts the label drawing software through the terminal device, imports the image to be labeled, and the software preprocesses the imported image. The preprocessing includes size normalization, noise reduction and grayscale correction to obtain a standardized image to be labeled. 2. Manual label drawing: The operator uses the drawing tools of the label drawing software to draw target labels on the standardized image to be labeled. The target labels include target area outline labels, category labels and attribute labels. The software automatically performs format standardization processing on the drawn labels to generate standardized label files. The standardized label files are uniquely associated with the corresponding images to be labeled. 3. Algorithm-based label verification: The algorithm program calls a preset label verification model to verify the standardized label files. The verification includes label integrity verification, label accuracy verification, and label format standardization verification, filtering out qualified and unqualified labels. For unqualified labels, the algorithm program generates verification prompt information and feeds it back to the label drawing software. The operator corrects the label according to the prompt information and re-verifies until qualified labels are obtained. 4. Training Sample Construction: The algorithm integrates images with valid labels and their corresponding standardized label files to construct the initial training sample set; 5. Algorithm-interventional model training: The algorithm program calls the target model to be updated, inputs the expanded training sample set into the target model, and uses a preset training algorithm to train the model. During the training process, the algorithm program monitors the model training indicators in real time, including loss value, accuracy, and recall. 6. Model Validation and Feedback: The algorithm calls the validation sample set to validate the updated target model. If the validation result meets the preset validation criteria, a model update completion signal is output. If the validation result does not meet the preset validation criteria, the algorithm returns to step 3, whereby the algorithm re-validates and filters the labels, or the operator adds labels and then re-executes the subsequent training and update steps.

[0044] Compared with the prior art, this embodiment has the following beneficial effects: 1. Achieving efficient collaboration between manual label drawing and algorithmic intervention, balancing label accuracy and efficiency: Standardized label drawing software ensures that the labels drawn by operators are uniform and standardized, facilitating algorithm recognition; at the same time, the algorithm program intervenes in label verification, quickly filtering out unqualified labels and prompting corrections, which not only ensures label accuracy but also avoids the tediousness of purely manual verification, improves label acquisition efficiency, and solves the problem of difficulty in balancing label quality and efficiency in existing technologies.

[0045] 2. The model training and updates are highly targeted, improving model performance and training efficiency: The algorithm can dynamically adjust training parameters and training strategies based on manually drawn label features, and expand the training sample set through data augmentation to make model training more targeted. During the training process, training indicators are monitored in real time, and the model is updated based on parameter differences to avoid ineffective training. This significantly improves model training efficiency and model performance, and solves the shortcomings of existing models that have weak targeting in updates and low training efficiency.

[0046] 3. Establish a closed-loop training and update process to improve practicality and operability: A closed-loop process of "image import - manual label drawing - algorithm verification - sample construction - model training - parameter update - model validation" has been constructed. Each link is seamlessly connected. Operators only need to complete the label drawing through label drawing software. The subsequent training and update process is automatically intervened by the algorithm program. The operation is simple, reduces manual costs, and ensures the stability and reliability of model updates. It is suitable for training and update scenarios of various image recognition, semantic segmentation and other models.

[0047] 4. Excellent scalability and compatibility: The label drawing module supports drawing various label types, the algorithm intervention module can adapt to different training algorithms and target models, and the storage module supports massive data storage. It can flexibly adjust relevant parameters and functions according to the needs of actual application scenarios, adapt to the model training and update needs of different fields, and has strong scalability and compatibility.

[0048] The implementation path in specific implementation work is as follows: Input Images: Collect images of new machine models from the production line.

[0049] Algorithm intervention: Execute algorithms for automated annotation.

[0050] Preprocessing: Based on this, the operator classifies and draws boxes for the new labels, and the program checks the names to prevent duplicate input.

[0051] Program summary: The program categorizes and outputs the labels, and the operator checks for any incorrect labels, correcting them if necessary.

[0052] Algorithm training: Configure the path of the deployed training program and start training with one click.

[0053] Results viewing: The training results are displayed graphically, indicating to the operator whether the training was successful.

[0054] The above description of several specific embodiments further details the technical solution provided by the present invention in order to highlight the advantages and benefits of the technical solution provided by the present invention. However, the above-described specific embodiments are not intended to limit the present invention. Any reasonable modifications and improvements to the present invention, combinations of embodiments, and equivalent substitutions based on the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for extracting color image contours by fusing color histograms using deep learning, characterized in that, include: The steps involved in collecting operational data from the target system and performing unified processing to form a standardized dataset. The step of performing quality assessment and screening on the standardized dataset to form a training sample set; The step of updating the parameters of the current model based on the training sample set to obtain a candidate model; The steps include evaluating the performance of the candidate models and comparing them with the current model to form a model update decision result; Based on the model update decision results, perform version management and deployment of the model to obtain the currently running model; The steps involve performing an update trigger decision based on the data generated during the operation of the current running model, and feeding the new data back to the standardized data set to form a loop in order to achieve continuous adaptive updating of the model.

2. The method for extracting color image contours by fusing color histograms using deep learning according to claim 1, characterized in that, When processing the standardized dataset in a unified manner, the multi-source data generated during the process is cleaned, format aligned, and labeled.

3. The method for extracting color image contours by fusing color histograms using deep learning according to claim 1, characterized in that, When performing quality assessment and screening on the standardized dataset, abnormal data, duplicate data, and low-confidence data are identified and removed, and the data are scored and screened based on the stability of the data source and the consistency of the labels.

4. The method for extracting color image contours by fusing color histograms using deep learning according to claim 1, characterized in that, When updating the parameters of the current model based on the training sample set, the model parameters are iteratively adjusted by inputting the training sample set in batches, and the existing model capabilities are kept stable during the parameter update process.

5. The method for extracting color image contours by fusing color histograms using deep learning according to claim 1, characterized in that, When evaluating the performance of the candidate models, the degree of matching between the output results of the candidate models and the actual results is statistically analyzed based on the independent validation dataset, and compared with the performance of the current model.

6. The method for extracting color image contours by fusing color histograms using deep learning according to claim 1, characterized in that, When performing version management and deployment of the model based on the model update decision results, a unique identifier is assigned to the candidate model that passes the evaluation and the currently running model is replaced, while historical models are retained.

7. An apparatus for extracting color image contours by fusing color histograms using deep learning, characterized in that, include: A module that collects operational data from the target system and performs unified processing to form a standardized data set; A module for performing quality assessment and screening on the standardized dataset to form a training sample set; The current model is updated with parameters based on the training sample set to obtain a module of candidate models; A module that evaluates the performance of the candidate models and compares them with the current model to form a model update decision result; Based on the model update decision results, version management and deployment are performed on the model to obtain the modules of the currently running model; The module that performs update triggering judgment based on the data generated during the operation of the current running model and feeds the new data back to the standardized data set to form a loop in order to achieve continuous adaptive updating of the model.

8. A computer storage medium for storing computer programs, characterized in that, When the computer program is read by the computer, the computer executes the method of claim 1.

9. A computer, comprising a processor and a storage medium, characterized in that, When the processor reads the computer program stored in the storage medium, the computer executes the method of claim 1.

10. A computer program product, as a computer program, is characterized by: When the computer program is executed, it implements the method of claim 1.

Images