An object detection online learning dynamic sample selection method, system, computer device and storage medium

Abstract

The application discloses a target detection online learning dynamic sample selection method and system, computer equipment and storage medium, and the method comprises the steps of: obtaining classification uncertainty and positioning uncertainty of a to-be-screened sample through Monte Carlo Dropout sampling; constructing a multi-dimensional feature vector comprising classification uncertainty, positioning uncertainty, knowledge gap matching degree and the like; constructing a dynamic weight learning network based on an attention mechanism, and calculating the weight of each feature in combination with a model validation set performance index; performing multi-dimensional value scoring on the sample according to the feature weight, and screening out an optimal sample subset in combination with a computing power limit, so as to complete online updating of the model. The application comprehensively evaluates the sample value through multi-dimensional feature fusion, adapts to different scene requirements by using a dynamic weight, and balances resource consumption and updating effect by combining with computing power sensing to select a sample, thereby effectively improving the adaptability and detection precision of the model in a dynamic scene, and reducing the dependence on artificial labeling.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of computer vision and artificial intelligence, and particularly relates to a target detection online learning dynamic sample selection method and system, a computer device and a storage medium. BACKGROUND

[0002] Target detection is one of the core tasks in the field of computer vision, and is widely used in scenarios such as autonomous driving, intelligent monitoring, and industrial quality inspection. With the increasing dynamics and complexity of application scenarios, it is required that the target detection model has the ability of online learning, i.e., it can continuously optimize its performance using newly collected data without interrupting service, so as to adapt to dynamic factors such as environmental changes, light changes, and new target categories.

[0003] Traditional online learning methods often rely on a single and fixed evaluation index when selecting samples for model updating, such as only relying on the classification confidence score of the model. Samples with low confidence are considered to contain unknown information to the model and are therefore preferentially selected. However, this strategy has significant defects:

[0004] 1. Single dimension of sample value evaluation: Only considering classification confidence ignores other key information. For example, a sample may have high classification confidence, but its bounding box regression accuracy is poor, and such samples are also of great value to model optimization. In addition, whether the scene represented by the sample is rare and the quality of the sample annotation should be important dimensions for evaluating its value.

[0005] 2. Fixed feature weight: In some methods that fuse multiple features for sample evaluation, the weight of each feature is usually pre-set or fixed by experience. However, in different life cycle stages of the model or when facing different application scenarios, the focus of performance improvement may be different. For example, in the early stage of the model, it may be more necessary to improve the overall recognition rate; while in the mature stage of the model, it may focus more on solving the detection problem in specific difficult scenarios. Fixed feature weights cannot be adjusted adaptively, limiting the flexibility and effectiveness of the sample selection strategy.

[0006] 3. Ignoring the limitation of computing resources: Online learning is usually performed on edge devices with limited resources or servers that require high timeliness. Most existing methods do not fully consider the actual constraints of computing resources, which may cause update timeout due to the selection of too many samples, or insufficient model performance improvement due to the selection of too few samples, failing to achieve the best balance between resource utilization and model effectiveness.

[0007] 4. Insufficient consideration of labeling quality: The samples of online learning may come from automatic labeling (weak labeling) or manual labeling, with varying quality. Low-quality or incorrect labeling can seriously mislead model updating and reduce the reliability of the model. Existing technologies generally lack quantitative evaluation mechanisms for sample labeling quality.

[0008] Therefore, how to design an online learning sample strategy that can comprehensively and dynamically evaluate sample value and efficiently select samples under resource constraints is a technical problem to be solved in the current target detection field. SUMMARY

[0009] The purpose of the present application is to provide a target detection online learning dynamic sample selection method, system, computer device and storage medium, which aims to achieve more efficient, more robust and more adaptive online updating of target detection models by comprehensively evaluating sample value, dynamically adjusting evaluation weights and perceiving algorithmic constraints.

[0010] To achieve the above purpose, the present application adopts the following technical solutions:

[0011] A target detection online learning dynamic sample selection method, comprising the following steps:

[0012] S1: inputting a to-be-screened sample into a target detection model, obtaining classification uncertainty (U c ) and positioning uncertainty (U l ) of the to-be-screened sample through Monte Carlo Dropout sampling;

[0013] S2: constructing a multi-dimensional feature vector of the to-be-screened sample, the multi-dimensional feature vector comprising at least classification uncertainty (U c ), positioning uncertainty (U l ), knowledge gap matching degree (Im), labeling confidence (Sa) and scene scarcity (Sc);

[0014] S3: constructing a dynamic weight learning network based on an attention mechanism, combining the performance indicators of the validation set after online updating of the target detection model, and calculating the weights of each feature in the multi-dimensional feature vector;

[0015] S4: according to the feature weights calculated in step S3, performing multi-dimensional value scoring on the to-be-screened sample, screening the Top-K value scored samples to form an optimal sample subset, and the K is the number of selected samples determined based on the algorithmic constraints of online learning;

[0016] S5: inputting the optimal sample subset into the target detection model to complete online updating.

[0017] Further, in step S1:

[0018] Uncertainty of classification (U c ) is calculated by the entropy of the probability distribution of the class, formula is , where p k is the probability of the k-th class, and K is the total number of classes of the target detection model.

[0019] Uncertainty of localization (U l ) is calculated by the mean of the sampling variance of the bounding box coordinates, including the center coordinates (x, y) and the width and height (w, h), formula is , , where var(x) and var(y) are the sampling variances of the corresponding coordinates.

[0020] Further, in step S2:

[0021] The knowledge gap matching degree (Im) is calculated by the mutual information of information theory, formula is Im=(X;M), where X is the feature of the sample to be screened, specifically the high-dimensional global feature vector output by the backbone network (ResNet-50 or YOLOv8-backbone) of the target detection model; M is the feature distribution of the historical training samples of the target detection model, fitted by Gaussian Mixture Model (GMM), and the number of components of the GMM is determined by Bayesian Information Criterion (BIC).

[0022] The determination method of the labeling confidence (Sa) is: if the sample to be screened is manually labeled, Sa=1; if it is weakly labeled, Sa=1-labeling error rate, and the labeling error rate is calculated by the historical labeling verification set, specifically: select N weakly labeled samples as the verification set (N≥100), determine the number of incorrectly labeled samples n by manual review, and the labeling error rate is n / N.

[0023] The scene scarcity (Sc) is calculated by the reciprocal of the historical proportion of the scene to which the sample to be screened belongs, formula is Sc=1 / count(scene(X)), and count(scene(X)) is the number of samples of the scene in the historical training set of the target detection model.

[0024] Further, the structure of the dynamic weight learning network in step S3 includes a feature mapping layer, an attention calculation layer, and a weight output layer, and the specific process is as follows:

[0025] 1. Feature mapping layer: after standardizing each feature value (U c , U l , Im, Sa, Sc) of the multi-dimensional feature vector to the interval [0, 1], it is mapped to a 128-dimensional feature vector through a fully connected layer, and the activation function uses ReLU;

[0026] 2. Attention Calculation Layer: Inputs performance metrics of the target detection model's validation set (including mAP improvement rate ΔmAP, loss reduction rate ΔLoss, and generalization error change ΔE). gen The attention score is obtained by performing a dot product operation with the 128-dimensional feature vector output by the feature mapping layer. , where F perf For the performance metric feature vector (dimension 128), F feat Output of the feature mapping layer;

[0027] 3. Weighted Output Layer: This layer calculates the attention score A. f Normalization is performed to obtain the weights of each feature. , satisfying ∑w=1.

[0028] Furthermore, the training method for the dynamic weight learning network is as follows:

[0029] The optimization objective is to maximize the validation set mAP of the object detection model after the next round of updates. The Adam optimizer is used for end-to-end training with a learning rate of 1e-4 and 30 iterations. The training samples in each round are paired data of "historical multi-dimensional feature vectors + corresponding model performance indicators", with a sample size of no less than 200 sets.

[0030] Furthermore, in step S4:

[0031] The multi-dimensional value score is calculated as follows: for each sample i to be screened, a value score is calculated. ,in Let be the standardized feature values ​​of sample i;

[0032] The logic for determining the sample size K is as follows: T max M is the maximum allowed duration for a single online update. GPU T represents the GPU memory capacity. single M represents the training time for a single sample. single This refers to the memory usage of a single sample. This indicates rounding down to the nearest integer.

[0033] Furthermore, the online update strategy for the object detection model in step S5:

[0034] An incremental learning strategy is adopted: only the fully connected layers and the detection head layer of the model are updated, the backbone network layers are frozen, the learning rate is set to 1 / 10 of the initial training learning rate, and the number of iterations is 5-10 rounds.

[0035] An online learning dynamic sample selection system for object detection includes the following modules:

[0036] Uncertainty Sampling Module: Used to perform step S1 above and output the classification uncertainty (U) of the sample to be screened. c ) and positioning uncertainty (U l );

[0037] Multi-dimensional feature construction module: used to perform step S2 above and generate multi-dimensional feature vectors;

[0038] Attention weight learning module: used to perform step S3 above, outputting the real-time weights of each feature based on the attention mechanism and model validation set performance metrics;

[0039] Value scoring and sampling module: used to perform step S4 above, calculate multi-dimensional value scores for samples and select the optimal sample subset;

[0040] Model update and feedback module: This module is used to execute step S5 above, realize online model updates, and feed back the updated validation set performance metrics to the attention weight learning module.

[0041] This application also proposes a computer device, including a processor, a memory, and a computer program stored in the memory, wherein the processor executes the computer program stored in the memory to implement the method described in any of the above.

[0042] This application also proposes a computer-readable storage medium storing computer program instructions that, when executed by a processor, implement the steps of the method described in any one of the above-described embodiments.

[0043] This computer-readable storage medium, by executing the computer program instructions, stores the collected data on the user terminal itself, preventing data leakage. The collected data is added to the training process, and new data is integrated into a training batch and then added to the global model for training. This increases the amount of data and prevents overfitting, making the dynamically learned model more generalizable.

[0044] Beneficial effects

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] (1) Comprehensive sample value assessment: By integrating the features of five dimensions, namely classification uncertainty, positioning uncertainty, knowledge gap, annotation quality and scenario scarcity, this invention can more accurately and comprehensively assess the real value of samples for model updates than single index methods, avoiding the opportunity cost caused by one-sided selection.

[0047] (2) Dynamically adaptive weight strategy: A dynamic weight learning network based on attention mechanism is introduced, which enables the sample selection strategy to automatically adjust the weights of each feature according to the real-time performance feedback of the model (such as changes in indicators such as mAP and Loss). This allows the method to intelligently adapt to the needs of different scenarios and model lifecycles, and achieve "on-demand sampling".

[0048] (3) Resource-aware and efficient sampling: This invention uses computing power limitation as the key parameter to determine the number of samples, and dynamically calculates the optimal number of samples K through a formulaic approach. This ensures that model updates can make full use of available resources to achieve maximum performance improvement, while avoiding update failures or timeouts due to resource overload, thus achieving the best balance between model benefits and resource costs.

[0049] (4) Improved robustness and generalization ability: By introducing label confidence (Sa), the negative impact of low-quality or mislabeled samples on the model is effectively reduced. At the same time, by using the scene scarcity (Sc) feature, scene samples that the model has not seen or rarely see are selected first, which helps the model learn richer environmental features and significantly improves its generalization ability and robustness in complex and ever-changing real environments.

[0050] (5) Efficient incremental update strategy: Combining the incremental learning method of freezing the backbone network, the computational overhead and time cost in the online update process are greatly reduced, making high-frequency model iteration possible and enhancing the real-time adaptability of the model. Attached Figure Description

[0051] Figure 1 This is a flowchart of the method provided by the present invention;

[0052] Figure 2 This is a system module diagram provided by the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention will be further described in detail below with reference to specific embodiments.

[0054] This embodiment takes target detection (such as pedestrian and vehicle recognition) in an intelligent monitoring scenario as an example to illustrate the specific implementation process of the present invention. The target detection model used is Faster R-CNN based on the ResNet-50 backbone network. The hardware environment is a single NVIDIA Tesla V100 GPU (32GB of video memory), and the software environment is the PyTorch 1.13 framework.

[0055] This embodiment provides an online learning dynamic sample selection method for object detection. The method execution flow is as follows: Figure 1 The specific execution flow of this invention is as follows:

[0056] 1. Data preparation and initial model training

[0057] ① Dataset Construction: Historical labeled data from intelligent monitoring scenarios were selected as the initial training set, containing 30,000 images covering 10 scenarios such as "daytime streets", "nighttime parking lots", and "rainy day intersections". The labeled categories include four types: pedestrians, cars, buses, and trucks. The labeled format is the COCO standard format (including detection box coordinates and category labels).

[0058] ② Initial Model Training: The Faster R-CNN model was trained using the dataset described above, with a ResNet-50 backbone network, an initial learning rate of 1e-3, and an SGD optimizer (momentum 0.9) for 120 iterations. After training, the model achieved a mean average precision (mAP) of 78.5% on the validation set (5000 images), serving as a benchmark model for online learning.

[0059] 2. Input of samples to be screened and calculation of uncertainty (corresponding to step S1)

[0060] ① Source of samples to be screened: 20,000 unlabeled images generated in real time by the intelligent monitoring system, covering the newly added scenarios of "foggy highways" and "congested road sections", as candidate samples for online learning.

[0061] ② Monte Carlo Dropout Sampling: Enable Dropout (dropout rate = 0.5) in the backbone network and detection head layer of the target detection model, perform 10 independent forward propagations for each sample to be screened, and obtain 10 sets of class probability distributions and detection box coordinates.

[0062] ③ Classification uncertainty (U c Calculate: For the prediction of the "pedestrian" category in a certain sample, the probability distribution obtained from 10 samplings is [p1, p2, ..., p...]. 10 ], where the class probability of the k-th sample is p k =[0.82, 0.05, 0.10, 0.03] (corresponding to pedestrians, cars, buses, and trucks), then the classification uncertainty of this sample is:

[0063]

[0064] ④ Positioning uncertainty (U l Calculate: For 10 samples of the "pedestrian" detection box, the variance of the center coordinate x. =0.002, the variance of y =0.003, variance of width w =0.001, variance of height h =0.002, then the positioning uncertainty is:

[0065]

[0066] 3. Construction of multi-dimensional feature vectors (corresponding to step S2)

[0067] ① Knowledge gap matching degree (Im):

[0068] Extracting high-dimensional features X from the samples to be screened: Outputting a 512-dimensional global feature vector through the avgpool layer of ResNet-50.

[0069] Fitting the historical feature distribution M: A Gaussian mixture model (GMM) is trained using the features of 30,000 samples from the initial training set, and the number of GMM components is determined to be 8 using the Bayesian information criterion (BIC).

[0070] Calculate mutual information Im=I(X;M): Estimate the joint probability distribution of X and M using the Monte Carlo method, and obtain Im=0.32 for a certain "foggy highway" sample (the lower the value, the greater the difference between the sample and the historical distribution, and the greater the knowledge gap).

[0071] ② Label the confidence level (Sa):

[0072] Of the samples to be screened, 10,000 were manually labeled (Sa=1), and 10,000 were weakly labeled (automatically labeled using a traditional object detection model).

[0073] Weak labeling error rate calculation: 100 weakly labeled samples are randomly selected as the validation set. After manual review, 8 labels are found to be incorrect (such as mislabeling a truck as a bus). The labeling error rate is 8 / 100 = 0.08. Therefore, Sa = 1 - 0.08 = 0.92 for the weakly labeled samples.

[0074] ③Scenario scarcity (Sc):

[0075] The number of samples for the "foggy highway" scene in the historical training set is 0, while the number of samples for the "daytime street" scene is 5000.

[0076] The Sc=1 / 0=1 for the "Foggy Highway" sample (the default value for scarce scenarios is 1), and the Sc=1 / 5000=0.0002 for the "Daytime Street" sample.

[0077] 4. Dynamic weight learning network training and weight calculation (corresponding to step S3)

[0078] ①Network structure and training data:

[0079] The input to the dynamic weight learning network is a 5-dimensional feature vector (U c U l(Im, Sa, Sc), the output is a 5-dimensional weight vector.

[0080] The training data consists of paired data of "historical features + performance metrics": 1000 sets of sample features from the past 5 model updates are collected, along with the corresponding updated mAP improvement rate ∆mAP, loss reduction rate ∆Loss, and generalization error change ΔE. gen .

[0081] Training configuration: With the goal of "maximizing mAP in the next round", the Adam optimizer (learning rate 1e-4) is used, with 30 rounds of iteration and a batch size of 32 per round.

[0082] Example of weight calculation:

[0083] Feature mapping layer: Maps the standardized features of a sample [0.68, 0.002, 0.32, 0.92, 1] to a 128-dimensional vector F through a fully connected layer. feat .

[0084] Attention computation layer: Input the current model's validation set performance metrics ∆mAP=0.05, ∆Loss=-0.12, ΔE gen =-0.03, mapped to a 128-dimensional vector F feat Calculate the attention score by calculating the dot product.

[0085] ;

[0086] Weighted output layer: After normalization, the weights are obtained as w = [0.25, 0.15, 0.20, 0.10, 0.30] (∑w = 1).

[0087] 5. Sample selection and model update (corresponding to steps S4 and S5)

[0088] ① Multi-dimensional value rating (S):

[0089] For a weakly labeled sample of "foggy highway", its standardized feature is U c =0.68, U l =0.002, Im=0.32, Sa=0.92, Sc=1, then the value score is:

[0090] Si=0.25×0.68+0.15×0.002+0.20×0.32+0.10×0.92+0.30×1=0.17+0.0003+0.064+0.092+0.3=0.6263

[0091] ② Determine the sample size K:

[0092] Given T max=60 minutes (3600 seconds), M GPU =32GB, T single =0.8 seconds / sample, M single =1024MB (1GB) / sample, then:

[0093] ;

[0094] Since there are only 20,000 samples to be screened, the actual number of samples screened is Top-20,000.

[0095] ③ Online model updates:

[0096] An incremental learning strategy was adopted: the ResNet-50 backbone network was frozen, and only the fully connected layers and the detector head were updated. The learning rate was set to 1 / 10 (1e-4) of the initial learning rate, and the process was repeated for 8 rounds.

[0097] Performance improvements after the update: validation set mAP increased to 82.3% (an improvement of 3.8%), and detection accuracy in the "foggy highway" scenario increased from 65% to 88%.

[0098] 6. System Module Interaction Flow

[0099] Reference Figure 2 This invention can be implemented as an online learning dynamic sample selection system for target detection, comprising:

[0100] Uncertainty sampling module: Receives samples to be screened and outputs U through Monte Carlo Dropout. c and U l The data is then transmitted to the multi-dimensional feature construction module.

[0101] Multi-dimensional feature construction module: Combines historical data to calculate Im, Sa, and Sc, generates a complete feature vector, and sends it to the attention weight learning module.

[0102] Attention weight learning module: Calculates feature weights based on the model's current performance metrics (such as ∆mAP) and feeds them back to the value scoring and sampling module.

[0103] Value scoring and sampling module: Calculates sample value scores, selects Top-K samples, and submits them to the model update and feedback module.

[0104] Model update and feedback module: After completing the incremental update of the model, the new performance metrics are fed back to the attention weight learning module to achieve dynamic iteration.

[0105] This application also proposes a computer device, including a processor, a memory, and a computer program stored in the memory, wherein the processor executes the computer program stored in the memory to implement the above-described target detection online learning dynamic sample selection method.

[0106] This application also proposes a computer-readable storage medium storing computer program instructions that, when executed by a processor, implement the steps of the above-described target detection online learning dynamic sample selection method.

[0107] This computer-readable storage medium, by executing the computer program instructions, stores the collected data on the user terminal itself, preventing data leakage. The collected data is added to the training process, and new data is integrated into a training batch and then added to the global model for training. This increases the amount of data and prevents overfitting, making the dynamically learned model more generalizable.

[0108] This embodiment demonstrates that the present invention, through multi-dimensional feature evaluation and dynamic weight adjustment, can select the most valuable samples under computational constraints, significantly improving the adaptability and detection accuracy of the target detection model in dynamic scenes.

[0109] In summary, this invention provides a systematic, intelligent, and resource-aware online learning sample selection scheme for target detection, effectively solving many pain points in the prior art and providing strong technical support for building a high-performance target detection system that can continuously adapt to dynamic environments.

[0110] 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 method for online learning dynamic sample selection for object detection, characterized in that, Includes the following steps: S1: The sample to be screened is input into the target detection model, and the classification uncertainty U of the sample to be screened is obtained through Monte Carlo Dropout sampling. c With positioning uncertainty U l ; The classification uncertainty U c Calculated using the category probability distribution entropy, the formula is as follows: , where p k Let K be the probability of the k-th class, and K be the total number of classes in the object detection model. The positioning uncertainty U l The detection box coordinates, including the center coordinates (x, y) and width and height (w, h), are calculated using the mean of the sampling variance of the detection box coordinates. The calculation formula is as follows: , These are the sampling variances for the corresponding coordinates; S2: Construct a multi-dimensional feature vector for the samples to be screened. The multi-dimensional feature vector includes at least the classification uncertainty U. c Positioning uncertainty U l Knowledge gap matching degree Im, annotation confidence Sa, and scenario scarcity Sc; The knowledge gap matching degree Im is calculated using mutual information theory, with the formula Im=(X;M), where X represents the features of the sample to be screened, specifically the backbone network of the object detection model, which is either ResNet-50 or YOLOv8-backbone, and outputs a high-dimensional global feature vector; M represents the feature distribution of the historical training samples of the object detection model, fitted by a Gaussian mixture model (GMM), and the number of components of the GMM is determined by the Bayesian information criterion (BIC). The labeling confidence level Sa is determined as follows: if the sample to be screened is manually labeled, Sa = 1; if it is weakly labeled, Sa = 1 - labeling error rate. The labeling error rate is calculated through the historical labeling validation set. Specifically, N weakly labeled samples are selected as the validation set, N ≥ 100, and the number of incorrectly labeled samples n is determined by manual review. The labeling error is n / N. The scene scarcity Sc is calculated by the reciprocal of the historical proportion of the scene to which the sample to be screened belongs, and the formula is Sc=1 / count(scene(X)), where count(scene(X)) is the number of samples of that scene in the historical training set of the object detection model; S3: Construct a dynamic weight learning network based on the attention mechanism, and calculate the weight of each feature in the multi-dimensional feature vector by combining the performance index of the validation set after the online update of the object detection model; S4: Based on the feature weights calculated in step S3, perform multi-dimensional value scoring on the samples to be screened, and select the Top-K samples with the highest value scores to form the optimal sample subset, where K is the number of samples determined based on the online learning computing power limit. S5: Input the optimal sample subset into the target detection model to complete the online update.

2. The method according to claim 1, characterized in that, The structure of the dynamic weight learning network described in step S3 includes a feature mapping layer, an attention calculation layer, and a weight output layer. The execution flow is as follows: Feature mapping layer: This layer maps the eigenvalues ​​U of a multi-dimensional feature vector. c U l After being standardized to the [0, 1] interval, Im, Sa, and Sc are mapped to a 128-dimensional feature vector through a fully connected layer, and the activation function is ReLU. Attention computation layer: Inputs the validation set performance metrics of the target detection model, including mAP improvement rate ΔmAP, loss reduction rate ΔLoss, and generalization error change ΔE. gen The attention score is obtained by performing a dot product operation with the 128-dimensional feature vector output by the feature mapping layer. F perf The performance metric feature vector has a dimension of 128 and F. feat Output of the feature mapping layer; Weighted output layer: This layer calculates the attention score A. f Normalization is performed to obtain the weights of each feature. , satisfying ∑w=1.

3. The method according to claim 2, characterized in that, The training method of the dynamic weight learning network is as follows: with the optimization objective of maximizing the validation set mAP after the next round of object detection model update, the Adam optimizer is used for end-to-end training, the learning rate is set to 1e-4, and the number of iterations is 30 rounds; the training samples in each round are paired data of "historical multi-dimensional feature vectors + corresponding model performance indicators", and the sample size is no less than 200 sets.

4. The method according to claim 1, characterized in that, The multi-dimensional value score calculation method in step S4 is as follows: For each sample i to be screened, calculate the value score. ; in Let be the standardized feature value of sample i.

5. The method according to claim 1, characterized in that, The logic for determining the sample quantity K in step S4 is as follows: ； T max M is the maximum allowed duration for a single online update. GPU T represents the GPU memory capacity. single M represents the training time for a single sample. single For single sample memory usage, This indicates rounding down to the nearest integer.

6. The method according to claim 1, characterized in that, The online update of the target detection model in step S5 adopts an incremental learning strategy: only the fully connected layers and the detection head layer of the model are updated, the backbone network layers are frozen, the learning rate is set to 1 / 10 of the initial training learning rate, and the number of iterations is 5-10 rounds.

7. A target detection online learning dynamic sample selection system, characterized in that, include: Uncertainty sampling module: used to execute step S1 in claim 1, and output the classification uncertainty U of the sample to be screened. c With positioning uncertainty U l ; Multi-dimensional feature construction module: used to execute step S2 in claim 1 to generate a multi-dimensional feature vector; Attention weight learning module: used to execute step S3 in claim 1, and output the real-time weights of each feature based on the attention mechanism and the performance index of the model validation set; Value scoring and sampling module: used to perform step S4 in claim 1, calculate multi-dimensional value scores of samples and select the optimal sample subset; Model update and feedback module: used to execute step S5 in claim 1, realize online update of the model, and feed back the updated validation set performance index to the attention weight learning module.

8. A computer device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory, wherein the processor executes the computer program stored in the memory to implement the method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions that, when executed by a processor, implement the steps of the method described in any one of claims 1 to 6.

Images