Micro-model self-optimization visual recognition system and method based on negative sample feedback
By using a micro-model self-optimizing visual recognition system based on negative sample feedback, a closed-loop feedback mechanism is formed by leveraging lightweight neural networks and transfer learning. This automatically labels false alarm samples and performs incremental training, solving the problems of high false alarm rate and lagging model update in visual recognition systems, and achieving efficient deployment and improved robustness of edge devices.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- VIETADATA TECH (SHENZHEN) CO LTD
- Filing Date
- 2026-03-17
- Publication Date
- 2026-06-05
AI Technical Summary
Existing visual recognition systems suffer from problems such as high false alarm rates, lagging model updates, lack of closed-loop feedback mechanisms, and large model sizes that make deployment difficult, especially on resource-constrained edge devices where they cannot run in real time.
A micro-model self-optimizing visual recognition system based on negative sample feedback is adopted, including an image push module, a secondary verification module, a model training module, and an automatic deployment module. By combining lightweight neural networks and transfer learning, a closed-loop feedback mechanism is formed, which automatically labels false alarm samples and performs incremental training to achieve hot-swappable deployment.
Significantly reduce the false alarm rate of visual recognition systems, shorten model iteration cycles, improve system robustness and scene adaptability, and ensure efficient operation of edge devices.
Smart Images
Figure CN122156858A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of visual technology, specifically to a micro-model self-optimizing visual recognition system and method based on negative sample feedback. Background Technology
[0002] Currently, deep learning models are widely used in scenarios such as intelligent security, retail analytics, and human-computer interaction for object recognition, abnormal behavior detection, and emotion assessment. However, existing visual recognition systems generally suffer from the following problems:
[0003] 1. High false alarm rate: General models are prone to generating a large number of false alarms in specific scenarios (such as changes in lighting, occlusion, and small sample classes);
[0004] 2. Lagging model updates: Traditional methods require manual collection of a large number of samples, retraining, and manual deployment, which is time-consuming and costly;
[0005] 3. Lack of closed-loop feedback mechanism: Existing systems usually use the algorithm output directly as the final result, without establishing a closed loop of "false alarm → labeling → learning → optimization";
[0006] 4. Large model size and difficult deployment: Many high-precision models are difficult to run in real time on edge devices (such as IPC cameras and embedded terminals).
[0007] In summary, existing visual recognition systems suffer from high false alarm rates, delayed model updates, lack of closed-loop feedback mechanisms, and large model sizes that make deployment difficult, and therefore require improvement. Summary of the Invention
[0008] The purpose of this invention is to provide a micro-model self-optimizing visual recognition system and method based on negative sample feedback, so as to solve the problems mentioned in the background art.
[0009] To achieve the above objectives, the present invention provides the following technical solution:
[0010] A micro-model self-optimizing visual recognition system based on negative sample feedback includes:
[0011] The image push module is used to push the corresponding image frame of the alarm event to the manual review end when the visual recognition system (object recognition / behavior analysis / emotion recognition) triggers an alarm event.
[0012] The secondary verification module is used to annotate the corresponding image frame of the alarm event (such as "non-target object", "normal behavior", "non-angry expression", etc.) and mark its micro-model category (object / behavior / emotion) if the result of manual review of the alarm event is determined to be a false alarm.
[0013] The model training module is used to automatically trigger the incremental training task of the micro-model when the accumulation of negative samples of any one of the multiple micro-models reaches a set threshold.
[0014] The automatic deployment module is used to automatically replace the corresponding micro-model in the online visual recognition system after the micro-model training is completed, and to record the version log.
[0015] The output of the image push module is connected to the input of the secondary verification module, the output of the secondary verification module is connected to the input of the model training module, and the output of the model training module is connected to the input of the automatic deployment module (the task scheduler coordinates the work of each module to ensure that the data flow and training tasks are executed in an orderly manner).
[0016] As a further aspect of the present invention, the micro-model is a lightweight neural network (such as MobileNetV2, Tiny-YOLO variant), and different micro-models are applied to different sub-tasks.
[0017] As a further aspect of the present invention, the training of the micro-model adopts a strategy that combines transfer learning and negative sample fine-tuning: first, pre-trained weights are loaded as initial parameters, and then incremental training is triggered when the accumulated false negative samples reach a set threshold. Through feature distillation or residual learning, the micro-model focuses only on learning the sample distribution that the main model (the deep learning model of the conventional visual recognition system) failed to handle correctly, thereby suppressing false alarms while retaining the original positive sample recognition performance.
[0018] As a further aspect of the present invention: the micro-model is embedded in the edge device and hot-swappable, without affecting the operation of the visual recognition system.
[0019] As a further aspect of the present invention: in the image push module, RequestData (a component used in programming to request and transmit data) is selected to be responsible for pushing the alarm images generated by the main model to the review platform.
[0020] As a further solution of the present invention: In the secondary verification module, RequestCheck (an interface module for manual review and annotation) is selected as the manual review interface, which supports the annotation of false alarm types and true labels.
[0021] As a further aspect of the present invention: in the model training module, ModelTrain (a model training engine) is selected as the micro-model training engine to receive labeled data and perform incremental training.
[0022] The micro-model self-optimization visual recognition method based on negative sample feedback includes the following steps:
[0023] Step S1: When the main model (object recognition / behavior analysis / emotion recognition) in the visual recognition system detects a preset event and generates an alarm, the image push module immediately captures the image frame corresponding to the alarm event and attaches contextual metadata (such as timestamp, scene type, confidence level, etc.), and pushes the data to the manual review platform through the internal message queue for manual verification.
[0024] Step S2, (The reviewer opens the pushed image on the review platform and performs a secondary verification by the module) If the reviewer confirms that the alarm is a false alarm (e.g., mistaking car lights for fireworks, bending over to tie shoelaces for falling down, or frowning in thought for anger), and marks the specific type of false alarm (e.g., "non-target object", "normal behavior", "non-angry expression") and the micro-model category to which the false alarm belongs (object recognition micro-model / behavior analysis micro-model / emotion recognition micro-model); the image frame corresponding to the alarm event and its contextual metadata are stored in the negative sample library of the corresponding micro-model category, and the sample count of the micro-model category is updated at the same time;
[0025] Step S3: The task scheduler continuously monitors the number of samples in each micro-model category in the negative sample library. When the accumulation of negative samples in a certain category (e.g., object recognition micro-model) reaches the preset training trigger threshold, the task scheduler automatically wakes up the model training module and transmits the negative sample data of that category and the associated positive sample data to the model training module. At the same time, computing resources are allocated to start the incremental training task.
[0026] Step S4: After the model training module receives the trigger command, it loads the current online weights of the corresponding micro-model (lightweight backbone network such as MobileNetV2), and adopts the transfer learning + negative sample fine-tuning strategy to perform incremental training using newly accumulated negative samples and historical positive samples. During the training process, through feature distillation or residual learning, the micro-model focuses only on the sample distribution where the output of the main model is inconsistent with the true label, thereby effectively suppressing false positives of the same type while maintaining the original recognition performance.
[0027] Step S5: After training is completed, the model training module pushes the newly generated micro-model file to the model version repository and notifies the automatic deployment module. The automatic deployment module immediately distributes the new version of the micro-model to the corresponding edge inference node or cloud service, seamlessly replacing the old version of the micro-model in a hot-swappable manner. At the same time, it records the version log of this update (including training sample size, timestamp, performance metrics, etc.) to ensure that the system continues to run without being affected.
[0028] Step S6: The updated micro-model works in collaboration with the main model in the production environment. When a similar scenario occurs again, while the main model outputs an alarm, the micro-model performs a second judgment on the input: if the micro-model determines that it is a false alarm (output confidence is lower than the set value or a specific suppression signal), it automatically intercepts the alarm or reduces the alarm priority; if the micro-model confirms that it is a real event, it outputs normally.
[0029] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention forms a closed-loop feedback mechanism through four modules: image push, secondary verification, model training, and automatic deployment, which significantly reduces the false alarm rate of the visual recognition system; it utilizes manual review to accurately label false alarm samples and automatically triggers incremental training of micro-models, shortening the model iteration cycle from weekly to hourly; it adopts a lightweight micro-model design with less than 2MB of parameters, effectively suppressing similar false alarms while retaining the original recognition performance, enabling hot-swappable deployment of edge devices, and greatly improving the system's robustness and scene adaptability. Attached Figure Description
[0030] Figure 1 This is a schematic diagram of a micro-model self-optimizing visual recognition system based on negative sample feedback. Detailed Implementation
[0031] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0032] Please see Figure 1 A micro-model self-optimizing visual recognition system based on negative sample feedback includes:
[0033] The image push module is used to push the corresponding image frame of the alarm event to the manual review end when the visual recognition system (object recognition / behavior analysis / emotion recognition) triggers an alarm event.
[0034] The secondary verification module is used to annotate the corresponding image frame of the alarm event (such as "non-target object", "normal behavior", "non-angry expression", etc.) and mark its micro-model category (object / behavior / emotion) if the result of manual review of the alarm event is determined to be a false alarm.
[0035] The model training module is used to automatically trigger the incremental training task of the micro-model when the accumulation of negative samples of any one of the multiple micro-models reaches a set threshold.
[0036] The automatic deployment module is used to automatically replace the corresponding micro-model in the online visual recognition system after the micro-model training is completed, and to record the version log.
[0037] The output of the image push module is connected to the input of the secondary verification module, the output of the secondary verification module is connected to the input of the model training module, and the output of the model training module is connected to the input of the automatic deployment module.
[0038] The task scheduler coordinates the work of each module to ensure that the data flow and training tasks are executed in an orderly manner.
[0039] In this embodiment: Please refer to Figure 1 Micro-models are lightweight neural networks (such as MobileNetV2 and Tiny-YOLO variants), and different micro-models are applied to different sub-tasks.
[0040] Visual recognition systems are often deployed on resource-constrained edge devices (such as IPC cameras and embedded terminals), where traditional large models are difficult to run in real time. Using lightweight neural networks can significantly reduce computational resource consumption (parameter size <2MB), and at the same time, dedicated micro-models can be trained for different sub-tasks such as object recognition, behavior analysis, and emotion recognition, achieving fine-grained optimization of "one model for one task" and avoiding the performance degradation caused by a single model handling multiple tasks.
[0041] In this embodiment: Please refer to Figure 1 The training of the micro-model adopts a strategy that combines transfer learning and negative sample fine-tuning: first, pre-trained weights are loaded as initial parameters, and then incremental training is triggered when the accumulated false negative samples reach a set threshold. Through feature distillation or residual learning, the micro-model focuses only on learning the sample distribution that the main model (the deep learning model of the conventional visual recognition system) failed to handle correctly, thereby suppressing false alarms while retaining the original positive sample recognition performance.
[0042] Retraining directly on the full dataset is costly, time-consuming, and may lead to catastrophic forgetting (loss of the original ability to recognize positive samples). Transfer learning utilizes pre-trained weights to provide good initialization, negative sample fine-tuning focuses on correcting the errors of the main model, and feature distillation or residual learning ensures that the micro-model only learns the "residual" (i.e. the part where the main model made a mistake), thereby accurately suppressing false alarms while minimizing computational overhead and maintaining the original recognition accuracy.
[0043] In this embodiment: Please refer to Figure 1 The micro-model is embedded in the edge device and can be hot-swapped for deployment without affecting the operation of the visual recognition system.
[0044] Real-world production environments demand extremely high system stability. Traditional model updates require downtime for maintenance, impacting business continuity. Hot-swappable deployment allows new models to seamlessly replace old ones at runtime, without restarting services or interrupting video stream analysis, ensuring high availability in 24 / 7 real-time monitoring scenarios.
[0045] In this embodiment: Please refer to Figure 1 In the image push module, RequestData (a component used in programming to request and transmit data) is selected to push the alarm images generated by the main model to the review platform.
[0046] In this embodiment: Please refer to Figure 1 In the secondary verification module, RequestCheck (an interface module for manual review and annotation) is selected as the manual review interface, which supports the annotation of false alarm types and real labels.
[0047] In this embodiment: Please refer to Figure 1 In the model training module, ModelTrain (a model training engine) is selected as the micro-model training engine to receive labeled data and perform incremental training.
[0048] Please see Figure 1 A micro-model self-optimizing visual recognition method based on negative sample feedback includes the following steps:
[0049] Step S1: When the main model (object recognition / behavior analysis / emotion recognition) in the visual recognition system detects a preset event and generates an alarm, the image push module immediately captures the image frame corresponding to the alarm event and attaches contextual metadata (such as timestamp, scene type, confidence level, etc.), and pushes the data to the manual review platform through the internal message queue for manual verification.
[0050] Step S2, (The reviewer opens the pushed image on the review platform and performs a secondary verification by the module) If the reviewer confirms that the alarm is a false alarm (e.g., mistaking car lights for fireworks, bending over to tie shoelaces for falling down, or frowning in thought for anger), and marks the specific type of false alarm (e.g., "non-target object", "normal behavior", "non-angry expression") and the micro-model category to which the false alarm belongs (object recognition micro-model / behavior analysis micro-model / emotion recognition micro-model); the image frame corresponding to the alarm event and its contextual metadata are stored in the negative sample library of the corresponding micro-model category, and the sample count of the micro-model category is updated at the same time;
[0051] Step S3: The task scheduler continuously monitors the number of samples in each micro-model category in the negative sample library. When the accumulation of negative samples in a certain category (e.g., object recognition micro-model) reaches the preset training trigger threshold, the task scheduler automatically wakes up the model training module and transmits the negative sample data of that category and the associated positive sample data to the model training module. At the same time, computing resources are allocated to start the incremental training task.
[0052] Step S4: After the model training module receives the trigger command, it loads the current online weights of the corresponding micro-model (lightweight backbone network such as MobileNetV2), and adopts the transfer learning + negative sample fine-tuning strategy to perform incremental training using newly accumulated negative samples and historical positive samples. During the training process, through feature distillation or residual learning, the micro-model focuses only on the sample distribution where the output of the main model is inconsistent with the true label, thereby effectively suppressing false positives of the same type while maintaining the original recognition performance.
[0053] Step S5: After training is completed, the model training module pushes the newly generated micro-model file to the model version repository and notifies the automatic deployment module. The automatic deployment module immediately distributes the new version of the micro-model to the corresponding edge inference node or cloud service, seamlessly replacing the old version of the micro-model in a hot-swappable manner. At the same time, it records the version log of this update (including training sample size, timestamp, performance metrics, etc.) to ensure that the system continues to run without being affected.
[0054] Step S6: The updated micro-model works in collaboration with the main model in the production environment. When a similar scenario occurs again, while the main model outputs an alarm, the micro-model performs a second judgment on the input: if the micro-model determines that it is a false alarm (output confidence is lower than the set value or a specific suppression signal), it automatically intercepts the alarm or reduces the alarm priority; if the micro-model confirms that it is a real event, it outputs normally.
[0055] For example, let's take an object recognition micro-model as an illustration:
[0056] Main model alarm: The main object recognition algorithm deployed on the edge device detects the presence of "smoke and fire recognition alarm" in the image and triggers a safety alarm event;
[0057] Image push: The system pushes the reported image and context information (time, location, confidence level) corresponding to the alarm frame to the manual review platform via RequestData;
[0058] Manual verification: After reviewing the image, the auditor confirmed that the target was a "vehicle headlight" and that it was a false alarm. The auditor marked it as a "false alarm" in RequestCheck and categorized it under "object recognition".
[0059] Negative sample storage: The image is stored in the "Object Recognition Negative Sample Library" and associated with metadata (such as scene type, lighting conditions, occlusion, etc.).
[0060] Trigger training: When the cumulative number of newly added negative samples in the "Object Recognition Negative Sample Library" reaches a preset threshold, the task scheduler starts ModelTrain;
[0061] Micro-model retraining: The object recognition micro-model (using a lightweight MobileNetV2 backbone network) is loaded with the current online weights and incrementally fine-tuned by combining historical positive samples and newly added negative samples, with a focus on optimizing the discrimination boundary for easily confused categories;
[0062] Automatic deployment: After training is complete, the new version of the micro-model seamlessly replaces the original model through the automatic deployment module;
[0063] Performance verification: In subsequent scenarios with the same or similar conditions, the system no longer misidentifies such targets as "fireworks", and the false alarm rate has decreased significantly.
[0064] Using a behavioral analysis micro-model as an example:
[0065] Main model alarm: The main behavior analysis algorithm determines that a person has fallen and triggers a "person falls to the ground" alarm.
[0066] Image push: The system extracts keyframes from video clips containing this behavior and pushes them to the review end via RequestData;
[0067] Manual verification: The auditor reviewed the video and confirmed that the action was "bending over to tie shoelaces", which was not an abnormal event. It was marked as "false alarm" in RequestCheck and classified as "behavioral analysis".
[0068] Negative sample storage: The sample is stored in the "Behavioral Analysis Negative Sample Library" and auxiliary features such as action duration and posture key points are recorded at the same time;
[0069] Triggered training: When the number of negative samples meets the training conditions, the system automatically starts the behavioral micro-model training task;
[0070] Micro-model retraining: The behavior analysis micro-model (based on the lightweight TemporalConvolutionNetwork) enhances the robustness to "normal fall-like actions" by using new negative samples while retaining the performance of the original positive samples (real falls).
[0071] Automatic deployment: The updated model is deployed to edge inference nodes;
[0072] Performance verification: Subsequently, actions such as bending over and squatting no longer triggered false alarms, the system's recall rate for real fall events remained unchanged, and the accuracy improved.
[0073] Using an emotion recognition micro-model as an example:
[0074] Main model alarm: The main emotion recognition model outputs the emotion "anger" with a confidence level higher than the threshold, triggering a high-risk emotion alarm;
[0075] Image push: The system sends the face image to the manual review platform via RequestData;
[0076] Manual verification: The reviewer, based on the context (e.g., the user is focused on working), determines that it is actually "frowning in thought," not a negative emotion, marks it as a "false alarm" in RequestCheck, and designates it as an "emotion recognition" category;
[0077] Negative sample storage: The image is stored in the "emotion recognition negative sample library" and labeled with auxiliary tags (such as lighting, head posture, etc.).
[0078] Triggering training: When the number of negative emotion samples accumulates to a set number, the task scheduler activates ModelTrain;
[0079] Micro-model retraining: The emotion recognition micro-model (using the EfficientNet-Lite structure) is fine-tuned through transfer learning to enhance its ability to distinguish easily confused expressions such as "frowning ≠ anger";
[0080] Automated deployment: New models are automatically deployed, replacing older versions;
[0081] Performance verification: In subsequent operation, the system significantly reduced the misjudgment rate of neutral expressions such as focus and confusion, while maintaining a high recognition accuracy of genuine emotions such as anger and sadness.
[0082] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and not restrictive.
[0083] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.
Claims
1. A micro-model self-optimizing visual recognition system based on negative sample feedback, characterized in that, The micro-model self-optimizing visual recognition system based on negative sample feedback includes: The image push module is used to push the corresponding image frame of the alarm event to the manual review end when the visual recognition system triggers an alarm event. The secondary verification module is used to annotate the corresponding image frame of the alarm event and mark its micro-model category if the result of manual review of the alarm event is determined to be a false alarm. The model training module is used to automatically trigger the incremental training task of the micro-model when the accumulation of negative samples of any one of the multiple micro-models reaches a set threshold. The automatic deployment module is used to automatically replace the corresponding micro-model in the online visual recognition system after the micro-model training is completed, and to record the version log. The output of the image push module is connected to the input of the secondary verification module, the output of the secondary verification module is connected to the input of the model training module, and the output of the model training module is connected to the input of the automatic deployment module.
2. The micro-model self-optimizing visual recognition system based on negative sample feedback according to claim 1, characterized in that, Micro-models are lightweight neural networks, and different micro-models are applied to different sub-tasks.
3. The micro-model self-optimizing visual recognition system based on negative sample feedback according to claim 2, characterized in that, The training of the micro-model adopts a strategy that combines transfer learning and negative sample fine-tuning: first, pre-trained weights are loaded as initial parameters, and then incremental training is triggered when the accumulated false negative samples reach a set threshold. Through feature distillation or residual learning, the micro-model focuses only on learning the sample distribution that the main model failed to handle correctly, thereby suppressing false alarms while retaining the original positive sample recognition performance.
4. The micro-model self-optimizing visual recognition system based on negative sample feedback according to claim 3, characterized in that, Micro-models are embedded in edge devices and can be hot-swapped for deployment without affecting the operation of the visual recognition system.
5. The micro-model self-optimizing visual recognition system based on negative sample feedback according to claim 1, characterized in that, In the image push module, RequestData is selected to push the alarm images generated by the main model to the review platform.
6. The micro-model self-optimizing visual recognition system based on negative sample feedback according to claim 1, characterized in that, In the secondary verification module, RequestCheck is selected as the manual review interface, which supports labeling false alarm types and true labels.
7. The micro-model self-optimizing visual recognition system based on negative sample feedback according to claim 1, characterized in that, In the model training module, ModelTrain is selected as the micro-model training engine to receive labeled data and perform incremental training.
8. A micro-model self-optimizing visual recognition method based on negative sample feedback, applied to the micro-model self-optimizing visual recognition system based on negative sample feedback as described in any one of claims 1 to 7, characterized in that, The micro-model self-optimizing visual recognition method based on negative sample feedback includes the following steps: Step S1: When the main model in the visual recognition system detects a preset event and generates an alarm, the image push module immediately captures the image frame corresponding to the alarm event, attaches context metadata, and pushes the data to the manual review platform through the internal message queue for manual verification. Step S2: If the reviewer confirms that the alarm is a false alarm and marks the specific false alarm type and the micro-model category to which the false alarm belongs; the image frame corresponding to the alarm event and its context metadata are stored in the negative sample library of the corresponding micro-model category, and the sample count of the micro-model category is updated at the same time. Step S3: The task scheduler continuously monitors the number of samples in each micro-model category in the negative sample library. When the accumulation of negative samples in a certain category reaches the preset training trigger threshold, the task scheduler automatically wakes up the model training module and transmits the negative sample data of that category and the associated positive sample data to the model training module. At the same time, computing resources are allocated to start the incremental training task. Step S4: After the model training module receives the trigger command, it loads the current online weights of the corresponding micro-model and adopts the transfer learning + negative sample fine-tuning strategy to perform incremental training using newly accumulated negative samples and historical positive samples. During the training process, through feature distillation or residual learning, the micro-model focuses only on the sample distribution where the output of the main model is inconsistent with the true label, thereby effectively suppressing false alarms of the same type while maintaining the original recognition performance. Step S5: After training is completed, the model training module pushes the newly generated micro-model file to the model version repository and notifies the automatic deployment module. The automatic deployment module immediately distributes the new version of the micro-model to the corresponding edge inference node or cloud service to seamlessly replace the old version of the micro-model in a hot-swappable manner. At the same time, it records the version log of this update to ensure that the system continues to run without being affected. Step S6: The updated micro-model works in collaboration with the main model in the production environment. When a similar scenario occurs again, the main model outputs an alarm while the micro-model performs a second judgment on the input: if the micro-model determines that it is a false alarm, it automatically blocks the alarm or reduces the alarm priority; if the micro-model confirms that it is a real event, it outputs normally.