An optimization method based on target detection data containing no labeled frame

Abstract

The application discloses an optimization method based on target detection data containing no labeled frame, and comprises the following steps: obtaining a target detection data set, preprocessing the target detection data set to obtain a to-be-detected target data set; inputting the to-be-detected target data set into a target detection model for training to generate a test model, wherein the test model is used for detecting the to-be-detected target data set and outputting a test result; determining whether the target detection model needs to be optimized according to the test result; when the target detection model needs to be optimized, determining an optimization method of the target detection model according to the test result. The test result of the to-be-detected target data set is obtained through the test model, and whether the target detection model needs to be optimized and the optimization method thereof are determined according to the test result, so that the test result of the to-be-detected target data set can be evaluated by using a unified evaluation index, the target detection model can be optimized according to the detection result, and the accuracy of target detection is improved.

Description

Technical Field

[0001] This specification relates to the fields of computer vision technology and information processing computing, specifically to an optimization method based on target detection data containing unlabeled bounding boxes. Background Technology

[0002] In the field of computer vision technology, object detection has become one of the main tasks and is widely used in smart cities, smart agriculture, smart logistics and other fields. Object detection directly affects the test results, so the evaluation and optimization of object detection tasks are also very important.

[0003] In existing technologies, the MS-COCO evaluation method is a commonly used object detection evaluation method. However, in practical applications, the MS-COCO evaluation method has certain limitations. For example, when the dataset contains negative samples (i.e., no labeled bounding boxes), the object detection algorithm can still detect predicted bounding boxes. This may be because the labeled bounding boxes were missed but the model can still recognize them, or the model is generating false detections. Alternatively, if the dataset does not contain negative samples, but the model does not detect any predicted bounding boxes, this may be because the object detection algorithm used is effective. Summary of the Invention

[0004] This invention overcomes the problem of inaccurate detection in existing detection datasets containing negative sample data, and provides an optimization method based on target detection data containing unlabeled bounding boxes. It enables the evaluation of different target detection datasets using a unified evaluation index, so as to analyze target detection models with application prospects and promote the application of algorithm engineering.

[0005] To solve the above technical problems, the technical solution of the present invention is as follows:

[0006] This invention provides an optimization method for target detection data containing unlabeled bounding boxes, comprising:

[0007] Obtain the target detection dataset, preprocess the target detection dataset to obtain the target dataset to be detected;

[0008] The dataset of targets to be detected is input into the target detection model for training, and a test model is generated. The test model is used to detect the dataset of targets to be detected and output the test results.

[0009] Determine whether the target detection model needs optimization based on the test results;

[0010] When the object detection model needs optimization, the optimization method is determined based on the test results.

[0011] By training on the object detection dataset and optimizing the object detection model, different optimization methods can be adopted according to the difference between the test results and the preset threshold, and the detection performance of object detection can be optimized in a targeted manner. This improves the detection accuracy of object detection datasets containing unlabeled bounding boxes and is suitable for various application scenarios.

[0012] As a further improvement of the present invention, determining whether the target detection model needs optimization based on the test results includes:

[0013] Obtain a preset threshold and determine the relationship between the test results and the preset threshold;

[0014] The target detection model is optimized based on the relationship between the test results and the preset threshold.

[0015] By identifying multiple optimization methods for the target detection model and adopting different optimization methods based on the relationship between the test results and the preset threshold, redundant computational operations can be reduced, computational costs can be lowered to some extent, the performance of the target detection model can be improved, and the detection accuracy of the target detection dataset can be increased.

[0016] As a further improvement of the present invention, the target detection model is optimized based on the relationship between the test results and a preset threshold, including:

[0017] When the test result is less than the preset threshold, the optimization method for the target detection model is determined as the first optimization method;

[0018] When the test result is greater than the preset threshold, the optimization method for the target detection model is determined to be the second optimization method;

[0019] When the test result equals the preset threshold, the target detection model does not need to be optimized.

[0020] As a further improvement of the present invention, the first optimization method includes: increasing positive sample data.

[0021] As a further improvement of the present invention, the second optimization method includes: training the target detection model using a deep neural network and adding implicit regularization parameters.

[0022] By adopting different optimization methods based on different test results, computational costs can be reduced to some extent, and the detection accuracy of the target detection dataset can be improved.

[0023] As a further improvement of the present invention, the target detection dataset is preprocessed, including:

[0024] The target detection dataset is sequentially processed with data augmentation, data cleaning, data deduplication, and image augmentation.

[0025] The target dataset to be detected is obtained by manually annotating the target detection dataset;

[0026] Input the dataset of the target objects to be detected into the target detection model.

[0027] By preprocessing the object detection dataset, the features in the object data can be aggregated into the same dimension, enabling the object detection model to effectively identify and extract information, eliminating the situation where different features in the original data cannot be compared, reducing the interference of invalid data features, and improving the reliability of the object detection model.

[0028] As a further improvement of the present invention, after obtaining the target dataset to be detected, the method further includes:

[0029] The dataset of targets to be detected is divided into three sets according to the preset partitioning rules: training set, validation set, and test set. The training set is used to train the target detection model, the validation set is used to evaluate the performance of the target detection model, and the test set is used to evaluate the generalization ability of the target detection model.

[0030] As a further improvement of the present invention, the target detection dataset includes: first target detection data, second target detection data and third target detection data, wherein the first target detection data is positive sample data, the second target detection data is negative sample data without prediction boxes, and the third target detection data is negative sample data with prediction boxes.

[0031] The object detection dataset can be classified according to whether it contains bounding boxes. The proportions of the three types of object test data in the training set, validation set, and test set can be set separately to conduct targeted training and testing of the object detection model.

[0032] As a further improvement of the present invention, the method for calculating the test results includes:

[0033] The test model calculates AP new and AP coco The value of AP coco To test the model, the MS-COCO evaluation method was used to calculate the output score, AP new Calculate the output score for the test model using the following formula:

[0034]

[0035] Where GT is the bounding box, len(pred FalseBoxes ) represents the number of predicted bounding boxes detected in the negative sample data, len(pred) AllBoxes ) represents the number of all predicted boxes, len(Imgs) WithGt ) represents the number of positive samples, len(Imgs) WithoutGt) represents the number of negative samples, len(pred) noboxImgsWithoutGt ) represents the number of negative samples in the negative sample data that do not have prediction boxes or label boxes, len(pred) boxImgsWithoutGt The number of negative samples in the negative sample data that have prediction boxes or label boxes.

[0036] It should be noted that the above formula is a derivation of the evaluation calculation formula proposed in the embodiments of the present invention, and the above variables are all obtained by statistical analysis of the contents of the annotation box file and the prediction box file.

[0037] As a further improvement of the present invention, the target detection algorithm in the target detection model includes, but is not limited to, any one of the following algorithms: YOLOv5 detection algorithm, Faster RCNN algorithm, and Mask RCNN algorithm.

[0038] Compared with the prior art, the beneficial effects that at least one technical solution adopted in the embodiments of this specification can achieve include at least:

[0039] By training a test model on the dataset of targets to be detected, the test results of the dataset are obtained. Based on the test results, it is determined whether the target detection model needs optimization and the optimization method. The test results of the dataset of targets to be detected can be evaluated using a unified evaluation index. The target detection model can be optimized in a targeted manner based on the detection results. Target detection models with application prospects can be analyzed to improve the accuracy of target detection and promote the application of algorithm engineering. Attached Figure Description

[0040] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings.

[0041] Figure 1 The flowchart illustrates a method for optimizing target detection data containing unlabeled bounding boxes, as provided by this invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only one preferred embodiment of this invention and are only used to explain this invention. They do not limit the scope of protection of this invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0043] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the operations (or steps) as sequential processes, many of the operations (or steps) can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but it may also have additional steps not included in the figures; the process may correspond to a method, function, procedure, subroutine, subroutine, etc.

[0044] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0045] Example: Figure 1 As shown, this embodiment of the invention provides a method for optimizing target detection data containing unlabeled bounding boxes, including:

[0046] Obtain the target detection dataset, preprocess the target detection dataset to obtain the target dataset to be detected;

[0047] The dataset of targets to be detected is input into the target detection model for training, and a test model is generated. The test model is used to detect the dataset of targets to be detected and output the test results.

[0048] Determine whether the target detection model needs optimization based on the test results;

[0049] When the object detection model needs optimization, the optimization method is determined based on the test results.

[0050] In practice, the dataset of targets to be detected is divided into positive sample data and negative sample data based on whether the image data contains bounding boxes. Positive sample data contains bounding boxes, while negative sample data does not contain any bounding boxes.

[0051] In practice, the image data information to be detected is collected from the data stream and denoted as the target detection dataset. The target detection dataset includes image data and annotation files. The annotation files include location information and target category information of the targets of interest in the image data. After preprocessing the target detection dataset, the target dataset to be detected is obtained.

[0052] In practice, the target detection model is a commonly used target detection model based on the MS-COCO algorithm. The dataset of targets to be detected is divided into multiple groups of data, which are then input into the target detection model for training. The training results of the multiple groups of data are combined to generate a test model, and the test results are output.

[0053] In practice, the test model integrates detection data from multiple datasets of target objects and outputs test results, which reflect the overall detection performance of the target detection model. Specifically, the target detection performance is the overall performance considering factors such as TP (True Positive) detection of positive samples correctly, FP (False Positive) detection of positive samples incorrectly, FN (False Negative) detection of negative samples incorrectly, precision, and recall.

[0054] During implementation, a preset threshold can be set. Based on the relationship between the preset threshold and the test results, it can be determined whether the target detection model needs to be optimized. When the test results are equal to or close to the preset threshold, it indicates that the target detection model has excellent detection performance and the accuracy of the data output by the detection model is high. When the test results are greater than or less than the preset threshold, it indicates that the detection performance of the target detection model does not meet the actual application requirements and the detection model needs to be improved and optimized.

[0055] By training on the object detection dataset and optimizing the object detection model, different optimization methods can be adopted according to the difference between the test results and the preset threshold, and the detection performance of object detection can be optimized in a targeted manner. This improves the detection accuracy of object detection datasets containing unlabeled bounding boxes and is suitable for various application scenarios.

[0056] In some implementations, determining whether the target detection model needs optimization based on test results includes:

[0057] Obtain a preset threshold and determine the relationship between the test results and the preset threshold;

[0058] The target detection model is optimized based on the relationship between the test results and the preset threshold.

[0059] In practice, the preset threshold is the value at which the target detection model achieves optimal performance. It can be set according to the actual situation and factors such as the dataset of the target to be detected, and there are no restrictions here.

[0060] In practice, a comparator can be used to compare the test results with a preset threshold. The test results and the preset threshold are input into the comparator and the comparison results are output at the comparator to determine whether the data values ​​are equal or to determine the size relationship between the two.

[0061] During implementation, the relationship between test results and preset thresholds is used to determine whether the performance of the target detection model meets the needs of practical applications. When the performance of the target detection model is determined to be good, no optimization is required; when the target detection model is determined to need optimization, optimization is required.

[0062] In practice, in order to improve the detection performance of the target detection model, multiple optimization methods can be determined for the target detection model. Based on the relationship between the test results and the preset threshold, different optimization methods can be adopted to reduce redundant calculation operations, reduce computational costs to a certain extent, improve the performance of the target detection model, and improve the detection accuracy of the target detection dataset.

[0063] In some implementations, the target detection model is optimized based on the relationship between the test results and a preset threshold, including: when the test results are less than the preset threshold, the optimization method for the target detection model is determined as the first optimization method;

[0064] When the test result is greater than the preset threshold, the optimization method for the target detection model is determined to be the second optimization method;

[0065] When the test result equals the preset threshold, the target detection model does not need to be optimized.

[0066] In some implementations, the first optimization method includes: increasing positive sample data.

[0067] During implementation, when the test result is less than the preset threshold, it is determined that there are too many negative sample data in the target dataset to be detected, and the optimization method for the target detection model is determined as the first optimization method. The first optimization method includes: increasing positive sample data.

[0068] In one specific embodiment, annotations can be added to the data in the target dataset to be detected through manual annotation or algorithms, thereby increasing the number of positive samples in the target dataset.

[0069] In one specific embodiment, depending on the actual application scenario, the number of positive samples in the target dataset can be increased during the preprocessing process by performing image enhancement on the target data.

[0070] It should be noted that adding positive sample data can improve the detection accuracy of the target detection model. The specific number of positive samples to be added can be increased according to the needs of the actual application scenario, and there is no specific limit to the number.

[0071] In practice, after adding positive sample data, the target dataset to be detected is input into the target detection model again for calculation. When the test result is equal to the preset threshold, the target detection model is determined to meet the optimal performance and the training ends.

[0072] In some implementations, the second optimization method includes training the target detection model using a deep neural network and incorporating implicit regularization parameters.

[0073] In practice, when the test result is greater than the preset threshold, it is determined that the generalization ability and robustness of the target detection algorithm in the target detection model are poor, and the optimization method for the target detection model is determined as the second optimization method. The second optimization method includes: training the target detection model using a deep neural network and adding implicit regularization parameters.

[0074] In practice, a neural network composed of convolutional and pooling layers is used. After the target data is convolved, it enters the pooling layer to compress the target data, aggregates and statistically analyzes the target data with the surrounding data, and takes the average value of the adjacent regions. This reduces the number of parameters in the target data, improves the adaptability of the target data, prevents overfitting, and prevents the target detection model from performing well on the training set but having a large error on the test set.

[0075] In practice, implicit regularization parameters are added to the cost function during training to make all weights obtained from training the object detection model tend to be smooth, so that the parameters of all layers of the model are controlled within a certain range, thereby improving the model's generalization ability and robustness.

[0076] In practice, after optimizing the target detection model, the target dataset to be detected is input into the target detection model again for calculation. When the test result is equal to the preset threshold, the target detection model is determined to meet the optimal performance and the training ends.

[0077] By adopting different optimization methods based on different test results, computational costs can be reduced to some extent, the performance of the object detection model can be improved, and the detection accuracy of the object detection dataset can be increased.

[0078] In some implementations, the object detection dataset is preprocessed, including:

[0079] The target detection dataset is sequentially processed with data augmentation, data cleaning, data deduplication, and image augmentation.

[0080] The target dataset to be detected is obtained by manually annotating the target detection dataset;

[0081] Input the dataset of the target objects to be detected into the target detection model.

[0082] In practice, the image data in the object detection dataset is preprocessed. The preprocessing process includes: removing distortion from the image data, randomly flipping it vertically and horizontally, randomly cropping it, and merging four cropped segments into one image. Through the above series of operations, data augmentation, data cleaning, data deduplication, and image augmentation can be achieved on the image data of the object detection dataset. This improves the display effect of the image data, preserves or enhances the feature information in the image data, thereby improving the accuracy of object detection and effectively reducing the error rate.

[0083] During implementation, targets in the target detection data are labeled to outline the target information that needs to be detected in the current application scenario, forming a dataset of targets to be detected. Specifically, targets can be labeled manually or using data labeling software.

[0084] In one specific implementation, if mobile phones need to be identified, then all mobile phones in the target dataset are considered to be detected, while others are not. Each target in the target dataset is labeled, and targets with mobile phone features are highlighted. Specifically, mobile phone features include information such as mobile phone shape, size, and brightness.

[0085] In practice, the dataset of targets to be detected is converted into an input format that conforms to the calculation of the target detection model, and then the dataset of targets to be detected is input into the target detection model for data processing.

[0086] By preprocessing the object detection dataset, the features in the object data can be aggregated into the same dimension, enabling the object detection model to effectively identify and extract information, eliminating the situation where different features in the original data cannot be compared, reducing the interference of invalid data features, and improving the reliability of the object detection model.

[0087] In some implementations, after obtaining the dataset of the target to be detected, the method further includes:

[0088] The dataset of targets to be detected is divided into three sets according to the preset partitioning rules: training set, validation set, and test set. The training set is used to train the target detection model, the validation set is used to evaluate the performance of the target detection model, and the test set is used to evaluate the generalization ability of the target detection model.

[0089] In practice, data from the training set is input into the target training model for training; data from the validation set and the test machine are input into the test model for processing, and the performance of the target detection model is evaluated based on the test results.

[0090] In implementation, the preset partitioning rules include, but are not limited to: dividing positive and negative sample data according to a preset ratio. It can be understood that the same target data can only be assigned to one of the training, validation, or test sets.

[0091] In some implementations, the object detection dataset includes: first object detection data, second object detection data, and third object detection data, wherein the first object detection data is positive sample data, the second object detection data is negative sample data without prediction boxes, and the third object detection data is negative sample data with prediction boxes.

[0092] During implementation, based on the information in the annotation files, the target detection dataset is divided into three categories: first target detection data, second target detection data, and third target detection data. The first target detection data consists of positive sample data, the second target detection data consists of negative sample data without prediction boxes, and the third target detection data consists of negative sample data with prediction boxes.

[0093] In practice, positive sample data includes data with labeled boxes, while negative sample data does not include any labeled boxes. The predicted box is a dataset containing target data information calculated and output by the target detection model. Specifically, the predicted box includes: target location information, target category information, and information on the degree of overlap with the labeled boxes.

[0094] The object detection dataset can be classified according to whether it contains bounding boxes. The proportions of the three types of object test data in the training set, validation set, and test set can be set separately to train and test the object detection model.

[0095] In some implementations, the methods for calculating test results include:

[0096] The test model calculates AP new and AP coco The value of AP coco To test the model, the MS-COCO evaluation method was used to calculate the output score, AP new Calculate the output score for the test model using the following formula:

[0097]

[0098] Where GT is the bounding box, len(pred FalseBoxes ) represents the number of predicted bounding boxes detected in the negative sample data, len(pred) AllBoxes ) represents the number of all predicted boxes, len(Imgs) WithGt ) represents the number of positive samples, len(Imgs) WithoutGt ) represents the number of negative samples, len(pred) noboxImgsWithoutGt ) represents the number of negative samples in the negative sample data that do not have prediction boxes or label boxes, len(pred) boxImgsWithoutGt The number of negative samples in the negative sample data that have prediction boxes or label boxes.

[0099] It should be noted that the above formula is a derivation of the evaluation calculation formula proposed in the embodiments of the present invention, and the above variables are all obtained by statistical analysis of the contents of the annotation box file and the prediction box file.

[0100] len(pred noboxImgsWithoutGt) represents the number of negative samples in the negative sample data that have predicted bounding boxes or labeled bounding boxes, i.e., the number of negative samples in the negative sample data where no bounding boxes were detected. len(pred) FalseBoxes ) represents the number of predicted bounding boxes detected in the negative sample image; len(pred) in the above formula noboxImgsWithoutGt Taking the partial derivative, we get len(pred) noboxImgsWithoutGt When )>0, Among them, AP new It's a decreasing function, meaning the more negative samples in the negative data that have prediction boxes or labeled boxes, the lower the AP. new The smaller. Similarly, in len(pred FalseBoxes When )>0, Among them, AP new It's a decreasing function, meaning the more predicted bounding boxes are detected in a negative sample image, the higher the AP (Average Per Second). new The smaller it is.

[0101] In some implementations, the object detection algorithm in the object detection model includes, but is not limited to, any of the following algorithms: YOLOv5 detection algorithm, Faster RCNN algorithm, and Mask RCNN algorithm.

[0102] Example: In one specific implementation, the target detection data contains 1459 test data points, of which positive sample data (len(Imgs)) WithGt There are 1409 negative samples (len(Imgs)). WithoutGt There are 50 images. The AP (Average Per Count) is calculated using the object detection model based on the test data. coco The value is 0.297, representing the total number of predicted boxes (len(pred)). AllBoxes The value is 436763, of which the number of predicted bounding boxes detected in the negative sample data (len(pred)) is 436763. FalseBoxes The number of images in the negative sample data where no bounding boxes were detected was 14946 (len(pred)). noboxImgsWithoutGt The result is 0. Substituting the above data into the formula above, we get:

[0103]

[0104] Calculate AP new It equals 0.277. AP_coco and AP new The scores show a significant gap, indicating that the model has a high false detection rate for negative sample data and needs optimization.

[0105] Set the preset threshold to 1 and calculate AP. coco With AP newThe ratio of the two values, if greater than 1, indicates poor generalization ability and robustness of the object detection algorithm on the test set. To address this, the second optimization method is chosen: training the object detection model using a deep neural network with implicit regularization parameters. This keeps the parameters of all layers within a certain range, improving the model's generalization ability.

[0106] Example: In one specific implementation, the target detection data includes 859 test data images, of which positive sample data (len(Imgs)) WithGt There are 845 negative samples (len(Imgs)). WithoutGt There are 14 images. The AP obtained by inputting the test data into the object detection model... coco The value is 0.303, representing the total number of predicted boxes (len(pred)). AllBoxes The value is 18003, of which the predicted bounding boxes (len(pred)) were detected in the negative sample data. FalseBoxes The number of images where the number of pred (len(pred)) is 0, and the number of images where no boxes were detected in the negative samples (len(pred)) is 0. noboxImgsWithoutGt The value is 14. The above data is then used in the above formula for calculation, i.e.:

[0107]

[0108] Calculate AP coco It is 0.303.

[0109] Set the preset threshold to 1 and calculate AP. coco With AP new The ratio of the two values ​​is 1, indicating that the model has a low false detection rate for negative sample data, and the presence of negative sample data has little impact on the algorithm model. This suggests that the object detection model has excellent performance and does not require optimization.

[0110] This invention provides an optimization method for target detection data containing unlabeled bounding boxes. By training the target detection dataset and optimizing the target detection model, different optimization methods can be adopted according to the difference between the test results and the preset threshold, reducing redundant calculation operations, reducing computational costs to a certain extent, improving the performance of the target detection model, and improving the detection accuracy of the target detection dataset.

[0111] The specific embodiments described above are preferred embodiments of the present invention based on a target detection data optimization method containing unlabeled bounding boxes. They are not intended to limit the specific scope of the present invention. The scope of the present invention includes, but is not limited to, these specific embodiments. All equivalent changes made in accordance with the shape and structure of the present invention are within the protection scope of the present invention.

Claims

1. A method for optimizing target detection data containing unlabeled bounding boxes, characterized in that, include: The image data information to be detected is collected from the data stream and denoted as the target detection dataset. The target detection dataset is preprocessed to obtain the target dataset to be detected. The dataset of targets to be detected is input into the target detection model for training, and a test model is generated. The test model is used to detect the dataset of targets to be detected and output the test results. Based on the test results, determine whether the target detection model needs optimization; When the target detection model needs to be optimized, the optimization method for the target detection model is determined based on the test results; The calculation method for the test results includes: The test model calculates AP new and AP coco The value, according to AP coco With AP new The ratio determines the test result, where AP coco To test the model, the MS-COCO evaluation method was used to calculate the output score, AP new Calculate the output score for the test model using the following formula: ； Where GT is the bounding box, len(pred FalseBoxes ) represents the number of predicted bounding boxes detected in the negative sample data, len(pred) AllBoxes ) represents the number of all predicted boxes, len(Imgs) WithGt ) represents the number of positive samples, len(Imgs) WithoutGt ) represents the number of negative sample data, len(pred noboxImgsWithoutGt ) represents the number of negative samples in the negative sample data that do not have prediction boxes or label boxes, len(pred ImgsWithoutGt ) represents the number of negative sample data points in the negative sample data that have prediction boxes but no label boxes.

2. The method for optimizing target detection data containing unlabeled bounding boxes according to claim 1, characterized in that, Determining whether the target detection model needs optimization based on the test results includes: Obtain a preset threshold and determine the relationship between the test result and the preset threshold; The target detection model is optimized based on the relationship between the test results and the preset threshold.

3. The method for optimizing target detection data containing unlabeled bounding boxes according to claim 2, characterized in that, The optimization of the target detection model based on the relationship between the test results and the preset threshold includes: When the test result is less than the preset threshold, the optimization method for the target detection model is determined to be the first optimization method; When the test result is greater than the preset threshold, the optimization method for the target detection model is determined to be the second optimization method; When the test result equals the preset threshold, the target detection model does not need to be optimized.

4. The method for optimizing target detection data containing unlabeled bounding boxes according to claim 3, characterized in that, The first optimization method includes: adding positive sample data.

5. The method for optimizing target detection data containing unlabeled bounding boxes according to claim 3, characterized in that, The second optimization method includes: training the target detection model using a deep neural network and adding implicit regularization parameters.

6. The method for optimizing target detection data containing unlabeled bounding boxes according to claim 1, characterized in that, Preprocessing of the target detection dataset includes: The target detection dataset is subjected to data augmentation, data cleaning, data deduplication, and image augmentation processes in sequence. The target detection dataset is obtained by manually annotating the target detection dataset; The dataset of the target to be detected is input into the target detection model.

7. The method for optimizing target detection data containing unlabeled bounding boxes according to claim 1, characterized in that, After obtaining the dataset of the target to be detected, the method further includes: The dataset of targets to be detected is divided into a training set, a validation set, and a test set according to a preset partitioning rule. The training set is used to train the target detection model, the validation set is used to evaluate the performance of the target detection model, and the test set is used to evaluate the generalization ability of the target detection model.

8. The method for optimizing target detection data containing unlabeled bounding boxes according to claim 1, characterized in that, The target detection dataset includes: first target detection data, second target detection data, and third target detection data. The first target detection data is positive sample data, the second target detection data is negative sample data without prediction boxes, and the third target detection data is negative sample data with prediction boxes.

9. The method for optimizing target detection data containing unlabeled bounding boxes according to claim 1, characterized in that, The target detection algorithm in the target detection model includes any one of the following algorithms: YOLOv5 detection algorithm, Faster R-CNN algorithm, and Mask R-CNN algorithm.

Images