Spatial fringe target detection method under complex background based on semi-supervised learning
Through a semi-supervised learning method, using pseudo-label and consistency regularization loss to train the UNet network, the problem of spatial stripe object detection under strong stray light background is solved, reducing the annotation cost and improving the robustness and generalization of the model Ability to achieve accurate detection in strong stray light scenes.
Patent Information
- Application Number
- CN202411317286.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-20
- Publication Date
- 2025-01-03
AI Technical Summary
The prior art is difficult to effectively detect space fringe targets under the background of strong stray light, and the signal-to-noise ratio is low. Traditional methods such as Hough transformation and Rough transformation are not applicable, and the data annotation cost is high, resulting in insufficient training data sets.
Using a semi-supervised learning method, two UNet networks are initialized as teacher models and student models, and pre-trained the teacher model with a small amount of labeled data to generate pseudo-labels for unsupervised learning, and combined with consistency regularization loss to reduce pixel level The dependence of tags improves model generalization capabilities.
It significantly reduces labeling costs, improves model robustness and generalization capabilities, and can accurately detect spatial fringe targets in stray light-changing scenarios, improving the ability to utilize labelless data.
Smart Images
Figure CN119251519A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of deep space image target detection and segmentation, and in particular to a method for detecting spatial stripe targets under complex backgrounds based on semi-supervised learning. Background Art
[0002] Space target detection is currently divided into two main categories: one is stripe-shaped space target detection, and the other is point-shaped space target detection. The morphological differences of space targets depend on the working mode and exposure time of the camera. Among them, in star tracking mode and when the long exposure is long, the space target will appear striped. At this time, space stripe target detection refers to the process of detecting stripe-tailed space targets from point-shaped stars and noise background. In the case of strong noise and stray light interference, due to the extremely low signal-to-noise ratio of the stripe space target area, the Hough transform and Rough transform designed for linear objects will no longer be applicable.
[0003] According to Guy Nir et al., "Optimal and Efficient Streak Detection in Astronomical Images" (see "The Astronomical Journal", 2018, 156(11): 229-244), rough transform is combined to detect spatial stripe targets, but it is not suitable for situations with strong stray light background interference.
[0004] According to "Space Debris Automation Detection and Extraction Based on a Wide-field Surveillance System" by Ping Jiang et al. (see "The Astrophysical Journal Supplement Series", 2022, 259(03): 4-16), the paper first uses an improved median filter for denoising, and then uses Hough transform for spatial stripe target detection, but it is also not suitable for situations under strong stray light background interference. Summary of the invention
[0005] In order to detect the above-mentioned stripe-shaped space targets, the present invention proposes a method for detecting space stripe targets in complex backgrounds based on semi-supervised learning. This method significantly reduces the dependence on pixel-level labeled images and improves the overall generalization ability of the model by automatically discovering new stripe segmentation knowledge from unlabeled images during the training process.
[0006] To achieve the above object, the technical solution adopted by the present invention is as follows: A method for detecting spatial stripe targets under complex backgrounds based on semi-supervised learning includes the following steps: Step 1: Initialize two UNet networks as the teacher model and the student model respectively, use a small amount of labeled spatial image data to pre-train the teacher model, and obtain a teacher model with preliminary stripe target recognition ability; Step 2: Input a large amount of unlabeled spatial image data into the teacher model and the student model at the same time, and use the pseudo labels generated by the teacher model to supervise the student model, and calculate the unsupervised loss and consistency regularization loss; Step 3: Input a small amount of labeled spatial image data only into the student model to obtain the supervised loss; weighted combination of the unsupervised loss, consistency regularization loss and supervised loss is used to train the final loss of the student model; Step 4: Only use the trained student model as the model in the inference stage for spatial stripe target detection.
[0007] The beneficial effects of the present invention compared with the prior art are: (1) Reduce annotation costs: Only a small amount of labeled spatial image data is needed to improve the ability to utilize unlabeled data and solve the problem of insufficient training data sets due to the scarcity of spatial image data; (2) Improved robustness: Consistency regularization in the loss when training the student model ensures the robustness of the model in scenarios with variable stray light.
[0008] (3) Improving generalization ability: By combining labeled data and unlabeled data, the proposed semi-supervised learning method can better capture the distribution characteristics of stripe targets in spatial image data, thereby improving the generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS
[0009] Figure 1 It is a flow chart of a method for detecting spatial stripe targets under complex backgrounds based on semi-supervised learning of the present invention; Figure 2 It is the detection result diagram; among them, (a) is the input spatial image containing stripe target, (b) is the stripe target benchmark real data, and (c) is the spatial stripe target detection result. DETAILED DESCRIPTION
[0010] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0011] The present invention is further described in detail below in conjunction with specific embodiments.
[0012] The training data of the present invention includes 1000 spatial images containing different stray light interferences, including 20% labeled spatial image data and 80% unlabeled spatial image data.
[0013] like Figure 1 As shown, a method for detecting spatial stripe targets under complex backgrounds based on semi-supervised learning of the present invention specifically includes the following steps: Step 1: Initialize two UNet networks and use them as the teacher model and student model respectively. Use a small amount of labeled spatial image data to pre-train the teacher model to obtain a teacher model with preliminary stripe target recognition capability. Step 2: Input a large amount of unlabeled spatial image data into the teacher model and the student model at the same time, and use the pseudo labels generated by the teacher model to supervise the student model, and calculate the unsupervised loss and consistency regularization loss; Step 3: Input a small amount of labeled spatial image data into the student model to obtain the supervised loss. The unsupervised loss, consistency regularization loss, and supervised loss are weighted together to train the final loss of the student model.
[0014] Step 4: Only use the trained student model as the model in the inference stage for spatial stripe target detection.
[0015] Specifically, the step 1 includes: Step 11: Initialize the teacher model T and student model S, both of which use UNet as the backbone, as follows: ; in, and Represent the initialization network parameters of teachers and students respectively, N( ) represents the normal distribution function, and They represent the variance of the teacher network initialization parameter distribution and the variance of the student network initialization parameter distribution respectively.
[0016] Step 12: Use a small amount of labeled spatial image data to train the teacher model to obtain a pre-trained teacher model. , its weight parameter The update process is shown as follows: ; Among them, Train( ) represents the network training process, is an input image training set containing a small amount of labeled spatial image data, is the learning rate, B is the batch size, and E is the total number of training iterations.
[0017] Specifically, the step 2 includes: Step 21: Pre-trained teacher model For a large amount of unlabeled spatial image data, pseudo labels are generated as follows: ; in, is the generated pseudo-label set, It is an input image training set containing a large amount of unlabeled spatial image data.
[0018] Step 22: The student model S uses pseudo labels for unsupervised learning and calculates the unsupervised loss as follows: ; Among them, S( ) is the predicted output of the student model for unlabeled spatial image data.
[0019] Step 23: Calculate the consistency regularization loss of the teacher model and the student model output as follows: ; in, It means to find the second norm.
[0020] Specifically, the step 3 includes: Step 31: The student model uses a small amount of labeled spatial image data for supervised learning and calculates the supervised loss. as follows: ; Among them, S( ) is the predicted output of the student model for labeled spatial image data, is a collection of tags. is the input image training set containing label-space image data.
[0021] Step 32, there will be supervised loss , unsupervised loss and consistency regularization loss Combined to form the final weighted loss for training the student model S The formula is as follows: ; in, They are the weight coefficients of supervised loss, consistency regularization loss, and unsupervised loss, respectively.
[0022] Specifically, the step 4 includes: Step 41, the final weighted loss will be used Trained student model As the final inference model, the stripe target detection result is obtained as follows: ; Among them, Binary() is a binarization function. is the binarization threshold. is the input image training set.
[0023] Figure 2 The detection results of the method proposed in this invention on four spatial image sequences are shown in FIG. Figure 2 (a) is the input spatial image containing stripe targets. Figure 2 (b) is the real data of the stripe target benchmark, Figure 2 (c) is the spatial stripe target detection result. Figure 2 (b) and Figure 2 It can be seen from (c) that the present invention can better preserve the integrity and continuity of the target while detecting the spatial stripe target.
[0024] The parts not described in detail in the present invention belong to the known technologies in the technical field. The above description is only a specific example of the present invention and is not intended to limit the present invention. Any modification, equivalent replacement and improvement made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A method for detecting spatial stripe targets in complex backgrounds based on semi-supervised learning, characterized in that: The steps include: Step 1: Initialize two UNet networks as the teacher model and the student model respectively, and use a small amount of labeled spatial image data to pre-train the teacher model to obtain a teacher model with preliminary stripe target recognition capabilities; Step 2: Input a large amount of unlabeled spatial image data into the teacher model and the student model at the same time, and use the pseudo labels generated by the teacher model to supervise the student model, and calculate the unsupervised loss and consistency regularization loss; Step 3: Input a small amount of labeled spatial image data only into the student model to obtain the supervised loss; weighted combination of the unsupervised loss, consistency regularization loss and supervised loss is used to train the final loss of the student model; Step 4: Only use the trained student model as the model in the inference stage for spatial stripe target detection.
2. According to claim 1, a method for detecting spatial stripe targets under complex backgrounds based on semi-supervised learning is characterized in that: The step 1 comprises: Step 11, let the teacher model be T and the student model be S. Initialize the teacher model T and the student model S with UNet as the backbone as follows: ； in, and Respectively represent the network parameters initialized by the teacher model and the network parameters initialized by the student model, N( ) represents the normal distribution function, and They represent the variance of the teacher model network initialization parameter distribution and the variance of the student model network initialization parameter distribution respectively; Step 12: Use a small amount of labeled spatial image data to train the teacher model to obtain a pre-trained teacher model. , its weight parameter The update process is shown as follows: ； Among them, Train( ) represents the network training process, is an input image training set containing a small amount of labeled spatial image data, is the learning rate, B is the batch size, and E is the total number of training iterations.
3. According to claim 2, a method for detecting spatial stripe targets under complex backgrounds based on semi-supervised learning is characterized in that: The step 2 comprises: Step 21: Pre-trained teacher model For a large amount of unlabeled spatial image data, the pseudo label set is generated as follows: ； in, is the generated pseudo-label set, It is an input image training set containing a large amount of unlabeled spatial image data; Step 22: The student model uses pseudo labels for unsupervised learning and calculates the unsupervised loss as follows: ； Among them, S( ) is the prediction output of the student model for unlabeled spatial image data; Step 23: Calculate the pre-trained teacher model and the consistency regularization loss of the output of the student model S : ； in, It means to find the second norm.
4. The method for detecting spatial stripe targets under complex backgrounds based on semi-supervised learning according to claim 3, characterized in that: The step 3 comprises: Step 31: The student model S uses a small amount of labeled spatial image data for supervised learning and calculates the supervised loss. as follows: ； Among them, S( ) is the predicted output of the student model for labeled spatial image data, is the input image training set containing label space image data, is a collection of tags; Step 32, there will be supervised loss , unsupervised loss and consistency regularization loss Combined to form the final weighted loss for training the student model S : ； in, They are the weight coefficients of supervised loss, consistency regularization loss, and unsupervised loss, respectively.
5. The method for detecting spatial stripe targets under complex backgrounds based on semi-supervised learning according to claim 4, characterized in that: The step 4 comprises: Step 41, the final weighted loss will be used Trained student model As the final inference model, the stripe target detection result is obtained as follows: ； Among them, Binary() is a binarization function. is the binarization threshold; represents the input image training set.
6. The method for detecting spatial stripe targets in complex backgrounds based on semi-supervised learning according to claim 1, characterized in that: The ratio of a small amount of labeled spatial image data to a large amount of unlabeled spatial image data is 1:4.
Citation Information
Cited By
Method and device for training polypropylene film defect detection model and polypropylene film defect detection method
CN120783149A