Remote sensing small target detection label distribution method with label self-correction function
By building a remote sensing small target detection model and dynamically adjusting the potential target threshold and sample label allocation, the problem of insufficient model performance caused by missing data set annotations is solved, and the small-scale target detection effect is improved.
Patent Information
- Application Number
- CN202510656010.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-21
- Publication Date
- 2025-09-09
AI Technical Summary
The problem of insufficient learning of small-scale target features due to omissions in remote sensing dataset annotations, making it impossible to achieve optimal model training results.
By building a remote sensing small target detection model, dynamically adjusting the potential target threshold and sample label allocation, adjusting the loss function weight, and optimizing the model parameters to correct label missing, the model performance is improved.
It effectively improves the performance of small-scale target detection models, reduces the negative impact of incorrect labeling on detector performance, and achieves a more reasonable distribution of learning targets.
Smart Images
Figure CN120612466A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of target detection, and in particular relates to a remote sensing small target detection label assignment method with a label self-correction function. Background Art
[0002] The development of remote sensing technology has provided a unique perspective and data resources for Earth observation. In the massive amount of remote sensing data, small-scale targets are numerous and widely distributed. Compared with standard-sized targets, small-scale targets contain fewer feature pixels, their appearance is blurred and they are easily submerged by the complex background environment. This leads to the inevitable phenomenon of insufficient annotation when annotating small-scale target datasets, such as Figure 1 As shown, green boxes represent true annotations, and red boxes represent missing annotations. In actual research, it has been found that the quality of small-scale object detection datasets is strongly correlated with the final performance of object detection models. Therefore, improving the annotation quality of small-scale object detection datasets or making targeted adjustments to address these issues during model training can effectively improve the performance of small-scale object detection models.
[0003] Label assignment refers to the process by which the model divides each sample on the feature map into positive and negative samples and assigns a learning objective during training. Label assignment influences the model's training performance by determining the learning objective of each sample. When a dataset contains missing labels, traditional label assignment strategies define samples in the missing label region as negative samples, or background regions. This inappropriate label assignment causes the optimizer to suppress the target classification scores of samples in the missing label region, forcing the detection model to classify the target features in the current region as background features, thereby limiting the detection model's optimal performance. By adjusting the label assignment strategy, the detection performance of small-scale objects can be improved without increasing the model's capacity or inference burden. Summary of the Invention
[0004] (1) Technical issues to be resolved
[0005] The technical problem to be solved by the present invention is to provide a remote sensing small target detection label assignment method with label self-correction function, so as to solve the problem that the learning targets of some samples are unreasonable due to the omission of data set annotation, the small-scale target feature learning is insufficient, and the optimal model training effect cannot be achieved.
[0006] (2) Technical solution
[0007] In order to solve the above technical problems, the present invention provides a remote sensing small target detection label assignment method with label self-correction function, comprising the following steps:
[0008] Step 1: Build a remote sensing small-scale target detection model, which includes three parts: backbone network, neck network, and detection head network;
[0009] Step 2: Select a suitable small-scale object detection dataset for model parameter tuning based on task requirements;
[0010] Step 3: Determine the potential target threshold based on the data of the small-scale target detection dataset and the pre-training effect of the target detection model;
[0011] Step 4: Pre-assign sample labels and determine the rationality of the learning objectives of negative samples. Then, calculate the loss function of the target detection model by adjusting the weights of positive and negative samples.
[0012] Step 5: Select a suitable optimizer to adjust the parameters of the target detection model according to the loss function;
[0013] Step 6: Use the validation set in the small-scale object detection dataset to test the detection performance of the small-scale object detection model for small-scale objects;
[0014] Step 7: Select the target detection model with the best test results and deploy it.
[0015] Preferably, step three is specifically:
[0016] At the beginning of the target detection model training, the potential target threshold is set to 1. As the training proceeds iteratively, the potential target threshold is gradually reduced to the preset value α(t). The adjustment method is determined by the following formula:
[0017]
[0018] in, epochs is the total number of training iterations, and epoch is the current number of training iterations.
[0019] Preferably, the rationality of the learning objectives of pre-allocating sample labels and determining negative samples in step 4 is specifically as follows:
[0020] 1) Sample label pre-assignment
[0021] According to the label assignment strategy, the samples on the feature map of the input target detection model are divided into positive samples and negative samples, and a learning target is assigned to each sample;
[0022] 2) Determine the rationality of the learning objectives of negative samples
[0023] Based on the label pre-assignment results and the target category scores output by the target detection model, it is determined whether the current negative sample area may contain missed labeled targets; when the target category score output by the negative sample exceeds the potential target threshold, it is considered that the area where the negative sample is located may have potential missed labeled targets and is an unreasonable negative sample. At the same time, masks are added to these unreasonable negative samples so that they do not participate in the optimization of model parameters related to feature extraction.
[0024] Preferably, in step 4, the loss function of the target detection model is calculated by adjusting the weights of positive and negative samples as follows:
[0025] By changing the weight of each sample in the loss function, the impact of each sample on the detection performance of the target detection model is adjusted; the loss function calculation formula is as follows:
[0026]
[0027] Among them, w n is the weight of negative samples that potentially miss labeled targets, L p is the positive sample loss, L n is the negative sample loss, N is the number of positive samples, are the positive and negative sample sets respectively; for negative samples whose target category scores are lower than the potential target threshold, w n is set to 1; for negative samples whose target category scores are higher than the potential target threshold, w n is set to 0.
[0028] Preferably, in the target detection model, the backbone network is used to extract global and local features of the input image; the neck network is located between the backbone network and the detection head network, and is used to further process and integrate the features extracted by the backbone network, fuse the multi-scale features extracted by the backbone network to obtain a fused feature map, and pass it to the detection head network to perform the final detection or classification task; the detection head network is integrated with an edge perception enhancement module, which is used to map the feature map information fused by the neck network to the final output space after enhanced perception, and generate the network detection results, wherein the edge perception enhancement module is used in the P3 feature map branch of small-scale target detection. The edge perception enhancement module enhances the feature extraction ability and boundary perception ability of the remote sensing image small-scale target detection model for small-scale targets by establishing a global correlation between row elements and column elements in the remote sensing image feature map.
[0029] The present invention also provides a remote sensing small target detection method implemented based on the method.
[0030] (3) Beneficial effects
[0031] This paper provides a convenient and effective label assignment method for small-scale target detection in optical remote sensing images. By dynamically identifying potential areas of missing target labels on feature maps during detector training, this method assigns more reasonable learning targets to samples, reducing the negative impact of incorrect labeling on detector performance. This method addresses the problem of inadequate learning targets for some samples due to missing dataset labels, insufficient small-scale target feature learning, and the inability to achieve optimal model training. The method can be used for label assignment and parameter tuning of small-scale target detection models in remote sensing images. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] Figure 1 A schematic diagram of an incomplete annotation of a remote sensing small-scale target detection dataset;
[0033] Figure 2 This is a flow chart of the remote sensing small target detection label assignment method of the present invention. DETAILED DESCRIPTION
[0034] In order to make the purpose, content and advantages of the present invention more clear, the specific implementation methods of the present invention are further described in detail below with reference to the accompanying drawings and examples.
[0035] The present invention provides a convenient and effective label assignment method for small-scale target detection in optical remote sensing images, which can be used for label assignment and parameter tuning of small-scale target detection models in remote sensing images. This method solves the problem of unreasonable learning targets for some samples due to missing data set annotations, insufficient small-scale target feature learning (in small target data sets, there is a phenomenon of incomplete target annotation, and the features of unannotated targets cannot be learned by the target detection model), and the inability to achieve optimal model training results. Specifically, this method dynamically searches for potential target annotation missing areas on the feature map during detector training, giving samples more reasonable learning targets and reducing the negative impact of incorrect annotations on detector performance improvement.
[0036] This invention is mainly used for parameter optimization of small-scale target detection models in remote sensing images. The main technical contents are as follows:
[0037] 1) Define potential target thresholds
[0038] The potential target threshold parameter is determined based on the data of the small-scale target detection dataset (targets with pixels smaller than 32×32) and the model pre-training results. At the beginning of model training, the potential target threshold is set to 1. As training progresses iteratively, the potential target threshold is gradually reduced to the preset value α(t). The adjustment method is determined by the following formula:
[0039]
[0040] The design principle of the above formula is as follows: In the initial stage, the target confidence score judgment accuracy is low, so raising the target threshold retains only high-confidence target areas and reduces the number of incorrectly judged targets; as training proceeds iteratively, the network's judgment of the target area becomes more and more accurate, and the confidence score is more feasible at this time, so lowering the threshold increases the number of samples.
[0041] in, epochs is the total number of training iterations, and epoch is the current number of training iterations.
[0042] 2) Sample label pre-assignment
[0043] According to the label assignment strategy (the specific strategy is determined by the model and dataset), the samples (feature pixels) on the feature map of the input target detection model are divided into positive samples and negative samples, and a learning target is assigned to each sample (the learning target indicates in which direction this sample is optimized).
[0044] 3) Determine the rationality of the learning objectives of negative samples
[0045] Based on the label pre-assignment results and the target category scores output by the target detection model, we determine whether the current negative sample region may contain missing labeled targets. When the target category score output by the negative sample exceeds the potential target threshold, this method considers the region containing the negative sample to be potentially missing (missing) labeled targets (these negative samples are unreasonable negative samples). We also add masks to these unreasonable negative samples to exclude them from the feature extraction-related model parameter optimization (i.e., steps 4 and 5 below).
[0046] 4) Calculate the loss function of the target detection model
[0047] This method adjusts the impact of each sample on the detection performance of the target detection model by changing the weight of each sample in the loss function. The loss function calculation formula is as follows:
[0048]
[0049] Among them, w n is the weight of the negative sample of the potential missed labeled target, and the rest of the parameters follow the original loss function definition: L p is the positive sample loss, L n is the negative sample loss, N is the number of positive samples, are the sets of positive and negative samples respectively. For negative samples whose target category scores are lower than the potential target threshold, w n is set to 1; for negative samples whose target category scores are higher than the potential target threshold, w nis set to 0. Among them, negative samples that potentially miss labeled targets are unreasonable negative samples and do not participate in model optimization, so the model does not learn target features as background features. Because they lack accurate label information, these samples are not considered positive samples, and the learning target cannot be determined. Features of these unlabeled targets are not used in model parameter optimization to prevent the model from learning target features as background features.
[0050] 5) Optimize and adjust model parameters
[0051] The optimizer adjusts the model parameters according to the loss function.
[0052] Steps 3) to 5) above are performed iteratively. The number of iterations is affected by many factors such as the model and dataset. The target category score is also dynamically adjusted based on the training effect.
[0053] 6) Test and verify the performance of the target detection model
[0054] Use the dataset validation set to test the detection performance of the current target detection model for small-scale targets.
[0055] The operation process of the remote sensing small target detection label assignment method with label self-correction function implemented based on the above main technical content is as follows: Figure 2 As shown, each step is described as follows:
[0056] Step 1: Construct a remote sensing small-scale target detection model (small-scale targets refer to targets with pixels smaller than 32×32). The target detection model includes three parts: a backbone network, a neck network, and a detection head network. The backbone network is used to extract global and local features of the input image. The neck network is located between the backbone network and the detection head network, and is used to further process and integrate the features extracted by the backbone network, fuse the multi-scale features extracted by the backbone network to obtain a fused feature map, and pass it to the detection head network to perform the final detection or classification task. The edge-aware enhancement module is integrated in the detection head network to map the feature map information fused by the neck network to the final output space after enhanced perception, and generate the network detection result. The edge-aware enhancement module is used in the P3 feature map branch of small-scale target detection. The edge-aware enhancement module enhances the feature extraction and boundary perception capabilities of the remote sensing image small-scale target detection model for small-scale targets by establishing a global correlation between row and column elements in the remote sensing image feature map.
[0057] Step 2: Select a suitable small-scale object detection dataset: According to the task requirements, select a suitable small-scale object detection dataset for model parameter tuning;
[0058] Step 3: Define the potential target threshold: Determine the potential target threshold parameters (corresponding to the content of step 1) based on the data of the small-scale target detection dataset and the model pre-training effect;
[0059] Step 4: Divide the positive and negative samples (corresponding to the content of steps 2) and 3) and calculate the loss function of the target detection model (corresponding to the content of step 4)): adjust the weights of the positive and negative samples according to the method of the present invention, and calculate the loss function of the current target detection model;
[0060] Step 5: Optimize detection model parameters: Select a suitable optimizer to adjust the parameters of the target detection model according to the loss function;
[0061] Step 6: Test and verify the performance of the target detection model: Use the validation set in the small-scale target detection dataset to test the detection performance of the small-scale target detection model for small-scale targets;
[0062] Step 7: Deploy and use the target detection model: Select the target detection model with the best test results and deploy it.
[0063] Example
[0064] This example uses the AI-TOD dataset as a small-scale target detection dataset for remote sensing images for experimental verification. The AI-TOD dataset contains 700,621 target instances in 8 categories, and α(t) is set to 0.6. AP is the most commonly used target detection performance evaluation metric. The closer the evaluation result is to 1, the better the detection performance of the model. vt 、AP t 、AP s 、AP m It is the AP index distinguished by target size, where AP vt Indicates size less than or equal to 8 2 The target AP, AP t Indicates size between (8 2 , 16 2 ] target AP, AP s Indicates a size between (16 2 , 32 2 ] target AP, AP m Indicates that the size is greater than 32 2 The experimental results are shown in Table 1 below, and the optimal indicators are marked in bold.
[0065] Table 1 Experimental results record
[0066]
[0067] As can be seen, first, this method improves the detection performance of the baseline model YOLOX for small-scale objects, with APvt increasing by 3.8% and APt by 6.0%, a significant improvement. Second, compared to some small-object detection methods that previously achieved the best performance on the AI-TOD dataset, this method also performs better, with an AP50 of 57.7%, and other metrics also showing significant improvements. Compared with current mainstream object detection models, this method has a more significant advantage in small-object detection. In summary, the effectiveness and advancement of this method have been fully verified.
[0068] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.
Claims
1. A remote sensing small target detection label assignment method with label self-correction function, characterized in that: The following steps are involved: Step 1: Build a remote sensing small-scale target detection model, which includes three parts: backbone network, neck network, and detection head network; Step 2: Select a suitable small-scale object detection dataset for model parameter tuning based on task requirements; Step 3: Determine the potential target threshold based on the data of the small-scale target detection dataset and the pre-training effect of the target detection model; Step 4: Pre-assign sample labels and determine the rationality of the learning objectives of negative samples. Then, calculate the loss function of the target detection model by adjusting the weights of positive and negative samples. Step 5: Select a suitable optimizer to adjust the parameters of the target detection model according to the loss function; Step 6: Use the validation set in the small-scale object detection dataset to test the detection performance of the small-scale object detection model for small-scale objects; Step 7: Select the target detection model with the best test results and deploy it.
2. The method according to claim 1, wherein Step three is as follows: At the beginning of the target detection model training, the potential target threshold is set to 1. As the training proceeds iteratively, the potential target threshold is gradually reduced to the preset value α(t). The adjustment method is determined by the following formula: in, epochs is the total number of training iterations, and epoch is the current number of training iterations.
3. The method according to claim 1, wherein In step 4, the rationality of the learning objectives of pre-assigning sample labels and determining negative samples is as follows: 1) Sample label pre-assignment According to the label assignment strategy, the samples on the feature map of the input target detection model are divided into positive samples and negative samples, and a learning target is assigned to each sample; 2) Determine the rationality of the learning objectives of negative samples Based on the label pre-assignment results and the target category scores output by the target detection model, it is determined whether the current negative sample area may contain missed labeled targets; when the target category score output by the negative sample exceeds the potential target threshold, it is considered that the area where the negative sample is located may have potential missed labeled targets and is an unreasonable negative sample. At the same time, masks are added to these unreasonable negative samples so that they do not participate in the optimization of model parameters related to feature extraction.
4. The method according to claim 1, wherein In step 4, the loss function of the target detection model is calculated by adjusting the weights of positive and negative samples: By changing the weight of each sample in the loss function, the impact of each sample on the detection performance of the target detection model is adjusted; the loss function calculation formula is as follows: Among them, w n is the weight of negative samples that potentially miss labeled targets, L p is the positive sample loss, L n is the negative sample loss, N is the number of positive samples, are the positive and negative sample sets respectively; for negative samples whose target category scores are lower than the potential target threshold, w n is set to 1; for negative samples whose target category scores are higher than the potential target threshold, w n is set to 0.
5. The method according to claim 1, wherein Small-scale objects refer to objects with pixels smaller than 32×32.
6. The method according to claim 3, wherein The label assignment strategy used in step 4 is determined based on the target assignment model and the small-scale target detection dataset.
7. The method according to claim 2, wherein Epochs are determined based on the target detection model and small-scale target detection dataset.
8. The method according to claim 1, wherein In the target detection model, the backbone network is used to extract global and local features of the input image; the neck network is located between the backbone network and the detection head network, and is used to further process and integrate the features extracted by the backbone network, fuse the multi-scale features extracted by the backbone network to obtain a fused feature map, and pass it to the detection head network to perform the final detection or classification task; the detection head network is integrated with an edge-perceived enhancement module, which is used to map the feature map information fused by the neck network to the final output space after enhanced perception, and generate the network's detection results, wherein the edge-perceived enhancement module is used in the P3 feature map branch of small-scale target detection. The edge-perceived enhancement module enhances the feature extraction capability and boundary perception capability of the remote sensing image small-scale target detection model for small-scale targets by establishing a global correlation between row elements and column elements in the remote sensing image feature map.
9. The method according to claim 1, wherein The small-scale target detection dataset is the AI-TOD dataset.
10. A remote sensing small target detection method based on the method according to any one of claims 1 to 7.