Target detection label distribution optimization system based on feedback mechanism

Abstract

The invention discloses a target detection label distribution optimization system based on a feedback mechanism, and particularly relates to the technical field of target detection, and the system comprises a data input module which is used for receiving image data and initial information of a target frame and an anchor frame; the matching score calculation module calculates the matching score of the anchor frame and the target frame based on the input information and the prediction result of the classification head and the regression head; a dynamic sample screening module screens positive sample anchor frames through matching scores to generate a primary positive sample candidate set and a preferred positive sample set; the negative sample generation module generates a negative sample set after eliminating the positive sample anchor frame; the dynamic feedback optimization module is used for optimizing the matching score calculation module and the dynamic sample screening module based on a network prediction result in a training process; the method dynamically adjusts the division strategy of the positive and negative samples according to the network prediction result, deeply digs the potential of the hard negative samples, enables the model to better distinguish the target from the complex background, and improves the robustness of the classification head to the difficult-to-classify samples.

Description

Technical Field

[0001] The present invention relates to the technical field of target detection, and more specifically, to a target detection label allocation optimization system based on a feedback mechanism. Background Art

[0002] In the field of object detection, label assignment is a crucial step in the model training process. Traditional label assignment methods mainly include anchor-based and anchor-free strategies: Label assignment strategy based on anchor boxes: positive and negative samples are divided by setting a fixed intersection-over-union (IoU) threshold, for example, IoU greater than 0.5 is a positive sample, and less than 0.4 is a negative sample. Although this method is simple and intuitive, the accuracy of label assignment is limited because the fixed threshold cannot dynamically adapt to the shape, size and scene complexity of different targets.

[0003] Label assignment strategy without anchor boxes: Select positive samples based on the prior knowledge of the center point of the target box. Although the regional characteristics of the target box can be fully utilized, it is impossible to dynamically adjust the uncertainty of the target box such as deformation and occlusion.

[0004] The main defects of existing methods are: Limitations of fixed thresholds: The fixed IoU threshold ignores the diversity of targets and dynamic changes during training, resulting in uneven distribution of positive and negative samples. Lack of feedback optimization mechanism: Traditional label assignment methods cannot dynamically adjust the division strategy of positive and negative samples according to network prediction results. Insufficient utilization of negative samples: Ordinary negative samples are difficult to provide valuable supervision information, and the potential of hard negative samples has not been fully explored. Therefore, a dynamic, adaptive target detection label assignment system with feedback optimization mechanism is needed to improve the training efficiency and detection performance of the model. Summary of the invention

[0005] To achieve the above object, the present invention provides the following technical solutions: The target detection label allocation optimization system based on feedback mechanism includes data input module, matching score calculation module, dynamic sample screening module, negative sample generation module and dynamic feedback optimization module; The data input module is used to receive image data and initial information of the target frame and the anchor frame; The matching score calculation module is used to calculate the matching score between the anchor box and the target box based on the data information received by the data input module and the prediction results of the classification head and the regression head; The dynamic sample screening module is used to screen the positive sample anchor frames according to the matching scores, and obtain the initial positive sample candidate set and the preferred positive sample set in turn; The negative sample generation module is used to generate a negative sample set after removing the positive sample anchor frames from all anchor frames; The dynamic feedback optimization module is used to optimize the matching score calculation module and the dynamic sample screening module based on the network prediction results during the training process.

[0006] The matching score calculation module includes: The intersection-over-union calculation unit is used to calculate the degree of overlap between the anchor box and the target box; A shape matching calculation unit, used to calculate the shape matching degree based on the aspect ratio of the anchor box and the target box; A contour matching calculation unit is used to extract the coverage degree of the anchor frame covering the target frame contour information through pseudo labels; The comprehensive score generating unit is used to generate a matching score by combining the above results according to a weighted formula.

[0007] In a preferred embodiment, the dynamic sample screening module uses the Top-k method to select the optimal positive sample set from the initial positive sample candidate set. positive samples, and then summarize them to get the optimal positive sample set. The dynamic initialization is optimized by the dynamic feedback module. Optimized.

[0008] In a preferred embodiment, the dynamic feedback optimization module includes a network prediction feedback unit and a sliding average adjustment unit; The network prediction feedback unit is used to optimize the currently used network prediction model based on the network prediction results in the training process and the partitioned sets of positive and negative samples; The sliding average adjustment unit is used to smoothly adjust the Top-k method through historical matching data. Numeric value.

[0009] In a preferred embodiment, the matching score calculation module uses the following logic: The degree of overlap is obtained by the following formula: ; represents the anchor box, represents the target box, Represents the overlapping area between the anchor box and the target box, Represents the total coverage area of ​​the anchor box and the target box, Indicates the degree of overlap; The degree of shape matching is obtained by the following formula: ; , Represents the width and height of the anchor box, , Indicates the width and height of the target box, is the preset non-zero shape matching coefficient, Indicates the degree of shape matching; The coverage is obtained by the following formula: ; Represents the total number of pixels of the anchor box, Represents the probability value of the pseudo label at the pixel in the anchor box; The above results are combined to generate a matching score using the following weighted formula: ; , , are all preset non-zero weight coefficients, Score for the match.

[0010] In a preferred embodiment, the dynamic sample screening module is used to screen the positive sample anchor frames according to the matching scores, and obtaining the initial positive sample candidate set refers to: Get the matching score corresponding to each anchor box, summarize the anchor boxes corresponding to all non-zero matching scores, and get the initial positive sample candidate set.

[0011] In the Top-k method The dynamic initialization is optimized by the dynamic feedback module. The optimized logic is: Calculate the feature influence value of the target box: ; Represents the characteristic influence value of the target frame, represents the area of ​​the target box, is the preset area ratio coefficient; Calculate the impact value of network prediction results: ; Indicates the impact value of network prediction results, represents the classification prediction score of the i-th anchor box, n represents the total number of anchor boxes, is a preset non-zero prediction adjustment coefficient; Calculate the anchor box distribution influence value: ; represents the influence value of anchor box distribution, Represents the variance of the matching score set between the anchor box and the target box, Represents the mean of the matching score set between the anchor box and the target box, is a preset non-zero distribution adjustment coefficient; Dynamic Initialization The formula is: ; In the Top-k method The formula for determining is: ; Represents the dynamic initialization obtained from the previous training value, It is the preset smoothing coefficient, and its value range is [0,1].

[0012] In a preferred embodiment, the network prediction feedback unit is used to optimize the currently used network prediction model based on the network prediction results in the training process and the partitioned sets of positive and negative samples, which means: Get the sample set in the current training process: Positive sample set: the positive sample set obtained during the current training process. The anchor boxes with the highest matching scores; Negative sample set: all other anchor boxes that are not selected as positive samples; From the current negative sample set, select hard negative samples, i.e., the remaining samples in the initial positive sample candidate set that are not selected into the preferred positive sample set; The current hard negative samples are recombined with the historical training set used in the last training according to the following ratios: ; is the recombined training combination set, which is used to optimize the currently used network prediction model. is the historical training set used in the last training. is the set of hard negative samples obtained from the current training, , These are all preset recombination ratio coefficients.

[0013] Technical effects and advantages of the present invention: The present invention combines multi-dimensional matching indicators such as intersection over union, shape matching, and contour matching to dynamically generate a set of positive and negative samples that are more suitable for training, avoiding the limitations of traditional fixed threshold allocation, supporting the diverse shapes, sizes, and position characteristics of different target boxes, and realizing the refinement of label allocation. The dynamic adjustment mechanism of the number of positive samples ensures a reasonable ratio of positive and negative samples according to the target distribution and the current prediction state of the model, avoiding the problem of unbalanced sample allocation. Prioritizing challenging hard negative samples enables the model to better distinguish between targets and complex backgrounds, and improves the robustness of the classification head to difficult-to-classify samples.

[0014] By extracting the local contour features of the anchor frame and the target frame through the contour matching unit, the model can capture the key areas of the target more accurately and improve the adaptability to complex scenes such as deformation and occlusion. The present invention introduces a network prediction feedback unit, and dynamically optimizes the sample selection logic by analyzing the positive and negative sample division results in real time during the training process. By combining historical training data with current data, the stability and continuity of training are maintained, and the training efficiency is effectively improved. By optimizing the positive sample set to eliminate redundant anchor frames, the interference of invalid samples is reduced; the negative sample set is optimized to enhance the contribution of training samples to the classification task. Through the dynamic feedback mechanism, the samples generated by the current training process are combined with historical data to generate a diversified training set, which effectively avoids model overfitting. A sliding average adjustment unit is introduced in the sample quantity and allocation strategy, so that the model can still maintain training stability in dynamically changing sample allocation. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to facilitate understanding by those skilled in the art, the present invention is further described below in conjunction with the accompanying drawings; Figure 1 It is a schematic diagram of the target detection label allocation optimization system based on the feedback mechanism in the present invention.

[0016] Figure 2 It is a schematic diagram of the matching score calculation module in the present invention.

[0017] Figure 3 It is a schematic diagram of the dynamic feedback optimization module in the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] Reference Figure 1-3 The following examples are obtained: Example 1

[0020] The target detection label allocation optimization system based on feedback mechanism includes data input module, matching score calculation module, dynamic sample screening module, negative sample generation module and dynamic feedback optimization module; The data input module is used to receive image data and the initial information of the target frame and anchor frame; provide basic training data for the system, including image data and the initial information of the target frame and anchor frame. It provides the initial matching information of the anchor frame and the target frame, laying the foundation for the calculation of subsequent modules, supporting diversified input, and adapting to different scenes and target detection requirements.

[0021] The matching score calculation module is used to calculate the matching score between the anchor frame and the target frame based on the data information received by the data input module and the prediction results of the classification head and regression head; the matching score between the anchor frame and the target frame is calculated through the prediction results of the classification head and regression head. The matching degree between the anchor frame and the target frame is dynamically quantified by comprehensively considering the overlap degree, shape similarity and prediction accuracy of the anchor frame and the target frame, providing a basic matching score and providing a basis for the dynamic screening of positive and negative samples.

[0022] The dynamic sample screening module is used to screen the positive sample anchor frames according to the matching scores, and obtain the initial positive sample candidate set and the preferred positive sample set in turn; screen the positive sample anchor frames according to the matching scores, and gradually optimize the positive sample set. The initial positive sample candidate set quickly screens out anchor frames with a high degree of match with the target frame to avoid interference from low-quality samples; the preferred positive sample set further eliminates redundant or inefficient samples to ensure the quality of the final positive sample set; dynamically adjust the number and distribution of positive samples to adapt to the characteristics of different targets.

[0023] The negative sample generation module is used to generate a set of negative samples after removing the positive sample anchor frames from all anchor frames. After removing the positive sample anchor frames, a set of negative samples is generated for training the classification head to ensure that the classification head can effectively distinguish between the target and the background. It supports dynamic sampling of negative samples and selects "hard negative samples" to improve the model's adaptability to complex backgrounds.

[0024] The dynamic feedback optimization module is used to optimize the matching score calculation module and the dynamic sample screening module based on the network prediction results during the training process. According to the network prediction results during the training process, the matching score calculation and the positive and negative sample division strategy are optimized to realize the feedback learning mechanism of the model, and the algorithm is dynamically adjusted according to the historical prediction results and training effects; the allocation efficiency of positive and negative samples is improved to avoid imbalanced sample distribution, and the sliding average mechanism smoothes the sample adjustment process to enhance the stability of training.

[0025] The matching score calculation module includes: The intersection over union (IoU) calculation unit is used to calculate the degree of overlap between the anchor box and the target box. IoU is the most basic matching indicator in target detection, which can directly measure whether the anchor box coincides with the target box. If IoU is too small, the anchor box and the target box are almost irrelevant and can be directly used as negative samples to improve calculation efficiency. The higher the IoU between the anchor box and the target box, the higher the matching score, which becomes an important basis for positive sample screening.

[0026] The shape matching calculation unit is used to calculate the degree of shape matching based on the aspect ratio of the anchor box and the target box; based on the aspect ratio of the anchor box and the target box, the shape matching degree of the two is calculated to quantify the adaptability of the anchor box to the geometric characteristics of the target box. Enhance shape sensitivity: The degree of aspect ratio matching can reflect the shape characteristics of the target box, such as whether it is elongated, square, etc. Solve the problem of size inconsistency: Even if the IoU between the anchor box and the target box is high, but the shape is seriously mismatched (such as a large difference in aspect ratio), it will also affect the accuracy of the prediction. The shape matching score can effectively screen out anchor boxes that are more in line with the characteristics of the target box. Supplementary IoU limitations: IoU only focuses on the degree of spatial overlap and cannot fully describe the geometric relationship between the anchor box and the target box. Shape matching provides a more comprehensive measurement indicator.

[0027] The contour matching calculation unit is used to extract the coverage degree of the anchor box covering the target box contour information through pseudo labels; using pseudo labels (such as the target contour probability map output by the model), extract the matching degree of the anchor box covering the target box contour, and quantify the matching relationship between the anchor box and the real shape of the target. Improve detection accuracy: The more accurate the contour information of the anchor box covering the target box, the closer the prediction result is to the true value. Enhance target adaptability: The target may be occluded, deformed or have a complex background. Contour matching captures this information through pseudo labels to improve the model's adaptability to complex scenes. Optimize positive sample allocation: Even if the IoU and shape matching scores are low, if the contour matching degree is high, it means that the anchor box may still be a high-quality positive sample, avoiding missing potential high-quality samples.

[0028] The comprehensive score generation unit is used to generate a matching score based on the above results according to a weighted formula. The comprehensive score integrates spatial, geometric and contour characteristics into a unified quantitative index to fully describe the matching relationship between the anchor box and the target box. In different scenarios, the influence weight of each score may be different. The comprehensive score can be adapted to specific scenarios by adjusting the weight parameters (such as training stage or target type). Guide positive and negative sample screening: The comprehensive score provides a clear basis for subsequent positive sample screening, giving priority to anchor boxes with high scores to ensure the high quality of the positive sample set.

[0029] When generating the optimal positive sample set, the dynamic sample screening module uses the Top-k method to select the optimal positive sample set from the initial positive sample candidate set. positive samples, and then summarize them to get the optimal positive sample set. The dynamic initialization is optimized by the dynamic feedback module. The dynamic sample screening module selects a number of positive sample candidates from the initial positive sample candidate set through the Top-k method. positive samples; then these positive samples are aggregated to generate an optimized positive sample set, and finally The value is initialized dynamically by the initial value in the dynamic feedback optimization module Make dynamic adjustments.

[0030] Through the Top-k screening method, we can ensure that the best samples are selected from the initial positive sample candidate set. positive samples. The quality of positive samples plays a decisive role in the training of classification and regression heads. It avoids irrelevant or redundant anchor frames from entering the positive sample set, improving training efficiency. Dynamically adjust the number of samples The dynamic adjustment mechanism enables the number of positive samples to adapt to different images and target characteristics. It avoids the limitation of fixed number of samples and performs better in multi-target and complex scenarios. It enhances the robustness and flexibility of the system. The dynamic sample screening module generates an optimized positive sample set by aggregating and can provide more representative positive samples for training that are more in line with the current state of the model.

[0031] The dynamic feedback optimization module includes a network prediction feedback unit and a sliding average adjustment unit; the network prediction feedback unit is used to optimize the currently used network prediction model based on the network prediction results during the training process and the partition set of positive and negative samples; the sliding average adjustment unit is used to smoothly adjust the Top-k method through historical matching data. The network prediction feedback unit feeds back the current positive and negative sample division to the network prediction model, adjusts the model parameters, and improves the classification and regression performance.

[0032] The sample screening logic is optimized using network prediction results to ensure that the system always uses the best samples for training. The sliding average adjustment unit is dynamically adjusted based on historical data. , avoiding unstable sample allocation caused by fluctuations in single training data. This smoothing mechanism improves the stability and robustness of the training process. Dynamic adjustment to meet the needs of different training stages: In the early stage of training, and The initial value of may be low to ensure that the model learns simple targets first; as training progresses, the number of positive samples gradually increases through feedback optimization and sliding average adjustment, and the model gradually adapts to complex scenarios and target characteristics.

[0033] The matching score calculation module uses the following logic: The overlap between the anchor box and the target box is calculated, that is, the intersection-over-union ratio, to quantify the spatial overlap between the anchor box and the target box. The overlap degree is obtained by the following formula: ; represents the anchor box, represents the target box, Represents the overlapping area between the anchor box and the target box, Represents the total coverage area of ​​the anchor box and the target box, Indicates the degree of overlap; The shape matching degree is based on the aspect ratio of the anchor box and the target box. The shape matching degree between the two is calculated to quantify the adaptability of the anchor box to the geometric characteristics of the target box. It is obtained by the following formula: ; , Represents the width and height of the anchor box, , Indicates the width and height of the target box, is the preset non-zero shape matching coefficient, Indicates the degree of shape matching; Using pseudo labels (such as the target contour probability map output by the model), we extract the matching degree of the anchor box covering the target box contour and quantify the matching relationship between the anchor box and the real shape of the target. The coverage degree is obtained by the following formula: ; Represents the total number of pixels of the anchor box, Represents the probability value of the pseudo label at the pixel in the anchor box; The above results are combined to generate a matching score using the following weighted formula: ; , , are all preset non-zero weight coefficients, The larger the matching score, the higher the matching quality. Accurate matching score calculation can improve the quality of positive samples and reduce the number of misclassified negative samples, ultimately improving the training effect of the classification head and regression head, thereby improving the target detection performance of the model.

[0034] The dynamic sample screening module is used to screen the positive sample anchor boxes according to the matching scores. The initial positive sample candidate set is: Get the matching score corresponding to each anchor box, summarize the anchor boxes corresponding to all non-zero matching scores, and get the initial positive sample candidate set.

[0035] In the Top-k method The dynamic initialization is optimized by the dynamic feedback module. The optimized logic is: The size, shape, and complexity of the target box will affect Generally, large targets require more anchor boxes to cover, and small targets require fewer anchor boxes. Calculate the feature influence value of the target box: ; Represents the characteristic influence value of the target frame, represents the area of ​​the target box, is the preset area ratio coefficient; Based on the output of the prediction head, dynamically adjust , targets with higher prediction scores need more anchor box allocations, and targets with lower prediction scores need fewer allocations. Calculate the impact value of the network prediction result: ; Indicates the impact value of network prediction results, represents the classification prediction score of the i-th anchor box, n represents the total number of anchor boxes, is a preset non-zero prediction adjustment coefficient; The distribution of matching scores between the anchor frame and the target frame (i.e., the variance and mean of the matching scores) reflects the adaptation of the anchor frame to the target frame. The more concentrated the distribution, the more likely it is that a small number of anchor frames can cover the target frame well, which can reduce , calculate the anchor box distribution influence value: ; represents the influence value of anchor box distribution, Represents the variance of the matching score set between the anchor box and the target box, Represents the mean of the matching score set between the anchor box and the target box, is a preset non-zero distribution adjustment coefficient; Dynamic Initialization The formula is: ; During the training process, as the network is optimized, the number and distribution of matching anchor boxes may change dynamically. This can be achieved by introducing a sliding average mechanism. Perform dynamic smooth adjustment and finally get , to ensure the stability during training, the Top-k method The formula for determining is: ; Represents the dynamic initialization obtained from the previous training value, It is the preset smoothing coefficient, and its value range is [0,1].

[0036] The data input module provides basic input data (initial information of anchor boxes and target boxes), which is the basis for the prediction of the classification head and regression head. The prediction results of the classification head and regression head are dynamically generated during the network training phase, reflecting the current network's understanding and prediction of the input data. The classification head and regression head are network structure components of the target detection model. Their functions are as follows: The classification head outputs the category prediction probability of each anchor box, reflecting the model's prediction of the target category. The regression head outputs the regression offset value of each anchor box, which is used to predict the specific position and size difference between the anchor box and the target box.

[0037] More specifically, the classification head mainly predicts the probability distribution of each anchor box belonging to a specific target category. For example, for an image containing N anchor boxes, the classification head will output a matrix of size N×C, where C is the number of categories. Result: Each anchor box will have a classification prediction probability value; these probability values ​​will be used to determine whether the anchor box matches the target box (i.e. whether it is a positive sample anchor box).

[0038] Prediction results of the regression head: The regression head predicts the position information between the anchor box and the target box (such as the center point offset, width and height changes). The output size of the regression head is N×4, where each anchor box corresponds to four regression values, namely: center point x offset; center point y offset; width scaling ratio; height scaling ratio. Results: The regression results describe the specific differences between the anchor box and the target box in space, and are used to adjust the position and size of the anchor box to make it closer to the target box.

[0039] The prediction results of the classification head and regression head correspond to the prediction information of each specific anchor box, and the relationship between the anchor box and the target box is established through the intersection-over-union ratio and other matching algorithms in the prior art known to those skilled in the art: the result of the classification head: used to determine whether each anchor box is a positive sample, that is, whether it matches a certain target box; the result of the regression head: used to calculate the position information that needs to be adjusted for the anchor box, so as to fit the target box more accurately. Finally, the joint results of classification and regression are used to optimize the target detection model: the classification prediction results of the anchor box determine which anchor boxes are positive samples (associated with the target box). The regression prediction results of the anchor box are used to fine-tune the position of the anchor box to make it closer to the corresponding target box.

[0040] The network prediction feedback unit is used to optimize the currently used network prediction model based on the network prediction results in the training process and the partition set of positive and negative samples. Get the sample set in the current training process: Positive sample set: the positive sample set obtained during the current training process. The anchor boxes with the highest matching scores; Negative sample set: all other anchor boxes that are not selected as positive samples; From the current negative sample set, select hard negative samples, i.e., the remaining samples in the initial positive sample candidate set that are not selected into the preferred positive sample set; The current hard negative samples are recombined with the historical training set used in the last training according to the following ratios: ; is the recombined training combination set, which is used to optimize the currently used network prediction model. is the historical training set used in the last training. is the set of hard negative samples obtained from the current training, , These are all preset recombination ratio coefficients.

[0041] The remaining anchor boxes are selected as negative samples, and the hard negative sample strategy is used to prioritize more challenging negative samples to enhance the classification ability of the model. The historical data is combined with the current samples to maintain the diversity and representativeness of the training set and improve the robustness and generalization ability of the model. , to achieve dynamic optimization of samples: introduce challenging hard negative samples in the current training to enhance the learning ability of complex scenes or backgrounds. Retain some historical training data to avoid overfitting the model to the current batch data, enhance the prediction ability of unseen data, and improve the adaptability and detection performance of the model. And combined with hard negative samples, ensure that sample allocation meets the current network training requirements. Through the sliding window sample update method, the combination of historical data and current data reduces the instability of training. This mechanism strengthens the classification head's ability to distinguish hard negative samples, and at the same time improves the accuracy of the regression head through dynamic screening of positive samples.

[0042] The significance of the feedback mechanism based on positive and negative sample division: Dynamically adjust sample distribution: The positive and negative samples generated during the current training process reflect the prediction ability of the model at a specific stage. By redistributing the positive and negative sample sets through the feedback mechanism, the sample distribution can be gradually optimized. For example, hard negative samples (negative samples with a high probability of misclassification) can significantly improve the classification head's ability to distinguish.

[0043] Avoid sample selection bias: In a single training, the current data may be noisy or have distribution bias. By combining the current sample with historical data, the impact of single training data on model optimization can be reduced.

[0044] The significance of the reorganization mechanism combined with historical data: Maintain data diversity: Simply using current samples may cause the model to focus too much on local information and be difficult to generalize. The introduction of historical data can make up for the shortcomings of current samples and enhance the diversity of training data.

[0045] Smooth transition: reorganization scale factor 𝜗, Controlling the combination ratio of historical data and current data allows the model to smoothly transition between historical knowledge and new knowledge, avoiding training instability.

[0046] Dynamic trade-off: When the model gradually stabilizes, appropriately increasing the proportion of historical data can prevent the model from forgetting the key features learned earlier. In the early stages of training, appropriately reducing the weight of historical data can help the model quickly adapt to the current scenario.

[0047] The significance of the priority selection of hard negative samples: Strengthen negative sample training: Hard negative samples are those anchor boxes that have a high match with the target box but do not meet the positive sample conditions. These samples have greater optimization value for the classification head. By giving priority to hard negative samples, the model's adaptability to complex backgrounds or high-interference scenes can be improved.

[0048] Reduce the redundancy of negative samples: Negative samples are usually large in number and of varying quality. The priority selection of hard negative samples avoids the introduction of invalid negative samples and improves training efficiency.

[0049] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0050] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0051] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0052] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0053] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. The target detection label allocation optimization system based on feedback mechanism is characterized by: It includes data input module, matching score calculation module, dynamic sample screening module, negative sample generation module and dynamic feedback optimization module; The data input module is used to receive image data and initial information of the target frame and the anchor frame; The matching score calculation module is used to calculate the matching score between the anchor box and the target box based on the data information received by the data input module and the prediction results of the classification head and the regression head; The dynamic sample screening module is used to screen the positive sample anchor frames according to the matching scores, and obtain the initial positive sample candidate set and the preferred positive sample set in turn; The negative sample generation module is used to generate a negative sample set after removing the positive sample anchor frames from all anchor frames; The dynamic feedback optimization module is used to optimize the matching score calculation module and the dynamic sample screening module based on the network prediction results during the training process.

2. The target detection label allocation optimization system based on feedback mechanism according to claim 1 is characterized in that: The matching score calculation module includes: The intersection-over-union calculation unit is used to calculate the degree of overlap between the anchor box and the target box; A shape matching calculation unit, used to calculate the shape matching degree based on the aspect ratio of the anchor box and the target box; A contour matching calculation unit is used to extract the coverage degree of the anchor frame covering the target frame contour information through pseudo labels; The comprehensive score generating unit is used to generate a matching score by combining the above results according to a weighted formula.

3. The target detection label allocation optimization system based on feedback mechanism according to claim 2 is characterized in that: When generating the optimal positive sample set, the dynamic sample screening module uses the Top-k method to select the optimal positive sample set from the initial positive sample candidate set. positive samples, and then summarize them to get the optimal positive sample set. The dynamic initialization is optimized by the dynamic feedback module. Optimized.

4. The target detection label allocation optimization system based on feedback mechanism according to claim 3 is characterized in that: The dynamic feedback optimization module includes a network prediction feedback unit and a sliding average adjustment unit; The network prediction feedback unit is used to optimize the currently used network prediction model based on the network prediction results in the training process and the partitioned sets of positive and negative samples; The sliding average adjustment unit is used to smoothly adjust the Top-k method through historical matching data. Numeric value.

5. The target detection label allocation optimization system based on feedback mechanism according to claim 4 is characterized in that: The matching score calculation module uses the following logic: The degree of overlap is obtained by the following formula: ; represents the anchor box, represents the target box, Represents the overlapping area between the anchor box and the target box, Represents the total coverage area of ​​the anchor box and the target box, Indicates the degree of overlap; The degree of shape matching is obtained by the following formula: ; , represents the width and height of the anchor box, , Indicates the width and height of the target box, is the preset non-zero shape matching coefficient, Indicates the degree of shape matching; The coverage is obtained by the following formula: ; Represents the total number of pixels of the anchor box, Represents the probability value of the pseudo label at the pixel in the anchor box; The above results are combined to generate a matching score using the following weighted formula: ; , , are all preset non-zero weight coefficients, Score for the match.

6. The target detection label allocation optimization system based on feedback mechanism according to claim 5 is characterized in that: The dynamic sample screening module is used to screen the positive sample anchor boxes according to the matching scores. The initial positive sample candidate set is: Get the matching score corresponding to each anchor box, summarize the anchor boxes corresponding to all non-zero matching scores, and get the initial positive sample candidate set.

7. The target detection label allocation optimization system based on feedback mechanism according to claim 6 is characterized in that: In the Top-k method The dynamic initialization is optimized by the dynamic feedback module. The optimized logic is: Calculate the feature influence value of the target box: ; Represents the characteristic influence value of the target frame, represents the area of ​​the target box, is the preset area ratio coefficient; Calculate the impact value of network prediction results: ; Indicates the impact value of network prediction results, represents the classification prediction score of the i-th anchor box, n represents the total number of anchor boxes, is a preset non-zero prediction adjustment coefficient; Calculate the anchor box distribution influence value: ; represents the influence value of anchor box distribution, Represents the variance of the matching score set between the anchor box and the target box, Represents the mean of the matching score set between the anchor box and the target box, is a preset non-zero distribution adjustment coefficient; Dynamic Initialization The formula is: ; In the Top-k method The formula for determining is: ; Represents the dynamic initialization obtained from the previous training value, It is the preset smoothing coefficient, and its value range is [0,1].

8. The target detection label allocation optimization system based on feedback mechanism according to claim 7 is characterized in that: The network prediction feedback unit is used to optimize the currently used network prediction model based on the network prediction results in the training process and the partition set of positive and negative samples. Get the sample set in the current training process: Positive sample set: the positive sample set obtained during the current training process. The anchor boxes with the highest matching scores; Negative sample set: all other anchor boxes that are not selected as positive samples; From the current negative sample set, select hard negative samples, i.e., the remaining samples in the initial positive sample candidate set that are not selected into the preferred positive sample set; The current hard negative samples are recombined with the historical training set used in the last training according to the following ratios: ; is the recombined training combination set, which is used to optimize the currently used network prediction model. is the historical training set used in the last training. is the set of hard negative samples obtained from the current training, , These are all preset recombination ratio coefficients.

Images