Decoupling-based weak supervision airport image semantic segmentation method
By building a weakly supervised semantic segmentation model in the airport environment, and using the pseudo-tagging module and multi-task learning module for decoupling operations, the problems of airport object coupling and poor data quality are solved, and efficient semantic segmentation performance is achieved, which is suitable for monitoring in the airport environment.
Patent Information
- Application Number
- CN202411968358.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-30
- Publication Date
- 2025-04-25
AI Technical Summary
The phenomenon of object coupling in airport environments is serious, resulting in limited performance of weakly supervised semantic segmentation algorithms, and it is difficult to obtain high-quality airport datasets and pixel-level labels.
The weakly supervised airport image semantic segmentation method is adopted based on decoupling. By constructing a weakly supervised semantic segmentation model, the airport pseudo-tagging module and multi-task learning module are used for decoupling operations, new training images and pseudo-tagging are generated, auxiliary learning tasks and main tasks are trained, and finally fully supervised to obtain the final segmentation result.
It effectively decouples the relationship between airport objects, improves the performance of semantic segmentation, and significantly improves the monitoring efficiency of aircraft and ground equipment, and is suitable for this special environment of airports.
Smart Images
Figure CN119888228A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of semantic segmentation of computer vision, and in particular relates to a decoupling-based weakly supervised airport image semantic segmentation method. Background Art
[0002] In airports, the use of semantic segmentation technology to monitor airport targets is crucial to the efficient operation of airports. However, to achieve high-precision semantic segmentation, a large amount of data with precise labels is usually required. The acquisition process of such data is complex and time-consuming, and requires a lot of human resources for manual annotation. Therefore, how to use limited data resources to improve the performance of semantic segmentation has become an urgent problem to be solved. In this context, making full use of imprecise samples that are easy to obtain in airports, such as image-level labels, to improve the performance of semantic segmentation has become a direction of great research significance. Therefore, it is very necessary to study weakly supervised airport semantic segmentation based on image-level labels.
[0003] The task of semantic segmentation is to segment an image into different semantic regions, such as people, cars, buildings, roads, etc. Airport semantic segmentation can accurately identify and locate various types of targets in the airport, such as aircraft, vehicles, etc., thereby improving the safety and operational efficiency of the airport. Existing weakly supervised semantic segmentation algorithms such as IRNet, SEAM, W-OoD, etc. can also be applied to airports. However, due to the coupling phenomenon of objects in airports, the segmentation effect is not good. Coupling refers to the fact that some objects often appear together in the same picture. This phenomenon is very serious in airports. This is caused by the unique structure of the airport itself. For example, airplanes often appear on the airport scene, and the terminal is often next to the boarding bridge. In airports, there is often coupling between different categories, and they often appear together in the same picture. This coupling relationship can mislead semantic segmentation, especially weakly supervised semantic segmentation algorithms, because the classifier is only trained on image labels, so it will be affected by the false correlation between coupled objects, which fundamentally limits the performance of weakly supervised semantic segmentation methods. For semantic segmentation, this coupled contextual information may be misleading, especially for weakly supervised semantic segmentation based on class activation maps. Since weakly supervised semantic segmentation needs to obtain information from the classification model to obtain the seed area, that is, a rough segmentation result, this coupling relationship will mislead the model to expand one category into another category, such as segmenting the airport scene as part of the airplane. Therefore, this coupling relationship needs to be decoupled.
[0004] The main problems of existing semantic segmentation algorithms in airport scenes are as follows:
[0005] 1) Coupling phenomenon of airport objects: In airports, different categories are often coupled and often appear together in the same picture. This coupling relationship will mislead semantic segmentation, especially weakly supervised semantic segmentation algorithms, because the classifier is only trained on image labels and will be affected by the false correlation between coupled objects, which fundamentally limits the performance of weakly supervised semantic segmentation methods.
[0006] 2) Poor quality of airport data: When studying and applying semantic segmentation techniques in airport environments, obtaining high-quality airport data is a significant challenge. There are few datasets dedicated to collecting airport images, and pixel-level labels are even rarer. Weakly supervised semantic segmentation methods only require image-level labels, which is more economical.
[0007] When studying and applying semantic segmentation techniques in airport environments, obtaining high-quality airport data is a significant challenge. There are few datasets dedicated to collecting airport images, and pixel-level labels are even rarer. Weakly supervised semantic segmentation methods only require image-level labels, which is more economical. Summary of the invention
[0008] In order to solve the above technical problems, the present invention provides a decoupling-based weakly supervised airport image semantic segmentation method, which can effectively cope with the special environment of the airport and significantly improve the monitoring efficiency of aircraft and ground equipment.
[0009] The technical solution adopted by the present invention is: a method for semantic segmentation of airport images based on weak supervision of decoupling, and the specific steps are as follows:
[0010] S1. Build a weakly supervised semantic segmentation model;
[0011] The model includes: an airport pseudo-labeling module, an airport auxiliary learning and main task segmentation module, and a semantic segmentation module.
[0012] The airport auxiliary learning and main task segmentation module is a multi-task learning module, including: an auxiliary learning module and a main task module. The semantic segmentation module is a weakly supervised semantic segmentation model.
[0013] S2. Use weakly supervised semantic segmentation method to collect instance images of various airport categories and input them into the airport pseudo-label module. Then input the training images into the airport pseudo-label module for decoupling operation to generate new training images and airport pseudo-labels.
[0014] S3, based on step S2, new images and airport pseudo labels are obtained, and input into the airport auxiliary learning and main task segmentation module, and the airport-based auxiliary learning tasks and supervised training image classification main tasks are trained to generate new airport pseudo labels with better effects;
[0015] S4. Based on the new uncoupled airport pseudo-labels obtained in step S3, a weakly supervised semantic segmentation model is trained, that is, the segmentation model is trained on the pseudo-labels obtained in step S3 and fully supervised to obtain the final segmentation result.
[0016] Among them, the weakly supervised semantic segmentation model includes: DeepLab, FCN, U-Net; the model backbone includes: ResNet38, VGG, ResNet.
[0017] Furthermore, the step S2 is specifically as follows:
[0018] First, a weakly supervised semantic segmentation method is used to collect instance graphs of segmented objects, and a classification network is trained using image-level labels. The classification network selects ResNet as the backbone network and is followed by a fully connected layer. This classification network is used to generate a localization map, which is a rough segmentation map, that is, the seed area.
[0019] The instance graph of the segmented object is the instance graph of each category of the airport.
[0020] After obtaining the positioning map, the relationship between pixels is used to predict the boundary of the object. That is, the same side of the object boundary has similarity, and the opposite side of the object boundary has no similarity. The object boundary and the seed area can be combined to remove the part outside the boundary, and the expansion can obtain a more accurate segmentation result, that is, a pseudo label.
[0021] Based on the pseudo-labels obtained above, object instances with good segmentation effects are selected to decouple the relationship between objects, that is, the training images are input into the airport pseudo-label module and the decoupling operation is performed to generate new airport pseudo-labels.
[0022] The object instance with good segmentation effect, that is, the current image has only one category. The decoupling operation is as follows:
[0023] Assume that the training data set has a total of N images. In each training round, N / 2 images are extracted from the training data set, and the same number of object instances are extracted from the generated pseudo labels. The object instances are obtained from the pseudo labels.
[0024] Then, the extracted objects are randomly pasted into the input image to generate N / 2 batches of new images, and then the generated N / 2 new images are combined with the N / 2 images extracted from the dataset. Then, a batch of training images of size N is obtained each time training. When N new training images are obtained, the pseudo-labels generated above are updated to generate new airport pseudo-labels.
[0025] Among them, the extracted objects are randomly pasted into the input image, that is, the above object instances are merged into the input image, that is, by randomly adjusting the size of the object, adding random rotation operations to change the direction of the object, adding Gaussian smoothing to help the added object boundaries blend together more naturally, and pasting the moving objects in the airport to other categories of other pictures. And the object size does not exceed 50% of the image and is not less than 5%. The moving objects in the airport include: airplanes, cars, and people.
[0026] Furthermore, the step S3 is specifically as follows:
[0027] First, a multi-task learning module is trained based on the airport pseudo-labels obtained in step S2 and N new training images, as follows:
[0028] Among them, the auxiliary learning module and the main task module in the multi-task learning module share an encoder to extract features, and the main task module generates a rough segmentation map based on the classification network, namely the seed area, and the supervision of the main task module comes from the image-level label of the weakly supervised semantic segmentation in step S2; the auxiliary learning module is based on the significant parts of the objects that are easily coupled with the decoder segmentation part, including: the head, tail, and car of the aircraft, and the supervision of the auxiliary learning module comes from the new airport pseudo-label generated in step S2.
[0029] Set the function f(x) to represent the encoder, whose parameter is the function θ that accepts the input image x. In the final feature layer, add a decoder to the auxiliary learning module for segmentation, add a fully connected layer to the main task module for classification, and use the SoftMax function to output the corresponding prediction results for each task, as follows:
[0030] For the true labels of training, the main task labels are y pri Represents the generated auxiliary task label y aux For the main task and auxiliary task modules, a fixed focusing parameter γ = 2 is used for the focusing loss L(y ^ ,y) training, its definition expression is as follows:
[0031] L(y ^ ,y)=-y(1-y ^ ) γ log(y ^ )
[0032] Among them, y ^ represents the predicted label and y represents the true label.
[0033] Then update the parameters θ of the multi-task network, and the focus loss function of the multi-task target is defined as follows:
[0034]
[0035] in, Represents the prediction result of the main task module, Represents the prediction results of the auxiliary task module and Represent the loss functions of the main task and auxiliary task respectively, arg θ Represents the minimization loss function. -
[0036] Based on the training of the above-mentioned multi-task learning module, that is, using uncoupled pictures and auxiliary learning to help decoupling, the trained classification network generates new seed areas, and the positioning map is refined or the particularity of the airport environment is used to obtain a new uncoupled airport pseudo-label, that is, a deep neural network model is used to predict the semantic similarity between a pair of adjacent coordinates in the image, and the predicted affinity is used as the transition probability in the random walk, so that the random walk propagates the activation score of the positioning map to the nearby area of the same semantic entity, significantly improving the quality of the positioning map, and finally obtaining a new uncoupled airport pseudo-label.
[0037] Beneficial effects of the present invention: The method of the present invention constructs a weakly supervised semantic segmentation model, collects instance images of various categories of airports using a weakly supervised semantic segmentation method, decouples the training images, generates new training images and airport pseudo labels, and then trains the auxiliary learning tasks based on the airport and supervises the main task of training image classification to generate new airport pseudo labels with better effects. Finally, the weakly supervised semantic segmentation model is trained and fully supervised to obtain the final segmentation result. The existing methods have not considered the problem of airport object coupling. Even some decoupling methods are only suitable for some commonly used image databases, but cannot be applied to the real and unique environment of the airport. Through the generation and auxiliary learning of airport pseudo labels, the method of the present invention can generate images without coupled objects, and at the same time allow the model to better understand the details of different objects, which can effectively avoid it from segmenting the two objects together, and finally achieve the purpose of decoupling. If the existing fully supervised algorithm wants to be applied to the airport, it needs manually annotated pixel-level labels, which is very time-consuming and labor-intensive, so it is more economical to apply the weakly supervised algorithm. Other weakly supervised algorithms are not effective in the special environment of the airport, so the method of the present invention is more effective. The existing methods do not consider how to efficiently generate new airport pictures. The method of the present invention uses a pseudo-label module to combine instances of different objects, which is different from the previous data enhancement methods such as direct image synthesis and cropping, which do not change the context information and structure of the object. The method of the present invention can destroy unnecessary connections between objects and promote model training and prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1The present invention is a flowchart of a method for semantic segmentation of airport images based on weak supervision based on decoupling.
[0039] Figure 2 Schematic diagram of the object coupling problem in an embodiment of the present invention.
[0040] Figure 3 Schematic diagram of the principle of generating airport pseudo-labels in an embodiment of the present invention.
[0041] Figure 4 Schematic diagram of the airport assisted learning principle in an embodiment of the present invention.
[0042] Figure 5 Schematic diagram of the segmentation result of the method of the present invention in an embodiment of the present invention. DETAILED DESCRIPTION
[0043] The method of the present invention is further described below in conjunction with the accompanying drawings and embodiments.
[0044] like Figure 1 As shown, a flow chart of a method for semantic segmentation of airport images based on weak supervision of decoupling of the present invention is shown, and the specific steps are as follows:
[0045] S1. Build a weakly supervised semantic segmentation model;
[0046] The model includes: an airport pseudo-labeling module, an airport auxiliary learning and main task segmentation module, and a semantic segmentation module.
[0047] The airport auxiliary learning and main task segmentation module is a multi-task learning module, including: an auxiliary learning module and a main task module. The semantic segmentation module is a weakly supervised semantic segmentation model.
[0048] S2. Use weakly supervised semantic segmentation method to collect instance images of various airport categories and input them into the airport pseudo-label module. Then input the training images into the airport pseudo-label module for decoupling operation to generate new training images and airport pseudo-labels.
[0049] S3, based on step S2, new images and airport pseudo labels are obtained, and input into the airport auxiliary learning and main task segmentation module, and the airport-based auxiliary learning tasks and supervised training image classification main tasks are trained to generate new airport pseudo labels with better effects;
[0050] S4. Based on the new uncoupled airport pseudo-labels obtained in step S3, a weakly supervised semantic segmentation model is trained, that is, the segmentation model is trained on the pseudo-labels obtained in step S3 and fully supervised to obtain the final segmentation result.
[0051] Among them, the weakly supervised semantic segmentation model includes: DeepLab, FCN, U-Net; the model backbone includes: ResNet38, VGG, ResNet.
[0052] In this embodiment, step S2 is specifically as follows:
[0053] In this embodiment, since there is no pixel-level label training semantic segmentation for the airport, pseudo labels need to be generated. In addition, in order to solve the problem of airport coupling, this embodiment generates new airport pseudo labels that help decoupling. This also helps to increase the diversity of airport data. The inherent contextual information in the airport data set reduces the quality of image segmentation. In this embodiment, the pictures in the airport data set are decoupled, and the well-segmented object instances are randomly found from the results generated by other models, and pasted to other positions of the current picture, so that a large area of contextual information in the original image can be cut off with a high probability, forcing the network to dig out the real location of the target and improve the quality of the pseudo labels generated by the classifier.
[0054] First, a weakly supervised semantic segmentation method is used to collect instance graphs of segmented objects (IRNet is selected in this embodiment), and a classification network is trained using image-level labels. The classification network selects ResNet as the backbone network and is followed by a fully connected layer. This classification network is used to generate a localization map, which is a rough segmentation map, that is, the seed area.
[0055] The instance graph of the segmented object is the instance graph of each category of the airport.
[0056] After obtaining the localization map, the relationship between pixels is used to predict the object boundary, that is, the object boundary has similarity on the same side, and the object boundary has no similarity on the opposite side. The object boundary and the seed area can be combined to remove the part outside the boundary, and the expansion can obtain a more accurate segmentation result, that is, the pseudo label (class activation map);
[0057] Based on the pseudo-labels obtained above, object instances with good segmentation effects are selected to decouple the relationship between objects, that is, the training images are input into the airport pseudo-label module and the decoupling operation is performed to generate new airport pseudo-labels.
[0058] Among them, the object instance with good segmentation effect is that the current image has only one category. (When the image contains only one category, the image information is usually simpler and lacks complex semantic environment. At this time, the segmentation result of the model should be more accurate.) Figure 2As shown in the figure, the object coupling problem diagram, coupling means that some objects often appear together in the same picture, this phenomenon is very serious in the airport, which is caused by the unique structure of the airport itself, such as airplanes often appear on the airport scene, boarding bridges are often next to the terminal building, etc. The decoupling operation is as follows:
[0059] The principle of airport pseudo-label generation is as follows Figure 3 As shown in FIG. 1 , the training data set is assumed to have a total of N images. In each training round, N / 2 images are extracted from the training data set, and the same number of object instances are extracted from the generated pseudo labels, and the object instances are obtained from the pseudo labels.
[0060] Then, the extracted objects are randomly pasted into the input image to generate N / 2 batches of new images, and then the generated N / 2 new images are combined with the N / 2 images extracted from the data set, so that each training obtains a batch of training images of size N. When the new N training images are obtained, the pseudo labels generated above are updated to generate new airport pseudo labels, which is also very useful for subsequent model training. (Since some objects obtained from some images are pasted onto other images, the generated images have many new category combinations, such as cars in the sky, airplanes in buildings, etc., and there will be no coupling between objects. When the model is trained using such images without coupled objects, the surrounding contextual information will not mislead the model too much).
[0061] Among them, the extracted objects are randomly pasted into the input image, that is, the above object instances are merged into the input image, that is, by randomly adjusting the size of the object, adding random rotation operations to change the direction of the object, adding Gaussian smoothing to help the added object boundaries blend together more naturally, and pasting the moving objects in the airport to other categories of other pictures. The object size does not exceed 50% of the image (to avoid blocking the original information of the image) and is not less than 5% (to avoid being too small to be difficult to segment). The moving objects in the airport include: airplanes, cars, and people.
[0062] In this embodiment, step S3 is specifically as follows:
[0063] In order to solve the problem of object coupling not only at the dataset level but also at the algorithm level, this embodiment proposes airport-based auxiliary learning to help the model better understand the characteristics of each category of airports. Auxiliary learning is a method to improve the performance of the main task by training other auxiliary tasks alongside the main task. This embodiment adds an auxiliary branch to segment the salient parts of the coupled objects after feature extraction, so that the model can better understand the aircraft, thereby assisting the main task, and this auxiliary task is equivalent to refining the contours of the coupled objects, avoiding the intrusion of the segmentation mask of a certain object into other objects. Compared with other weakly supervised semantic segmentation methods, decoupled images and algorithms can train semantic segmentation models with good performance.
[0064] First, a multi-task learning module is trained based on the airport pseudo-labels obtained in step S2 and N new training images, as follows:
[0065] Among them, the auxiliary learning module and the main task module in the multi-task learning module share an encoder to extract features, and the main task module generates a rough segmentation map based on the classification network, namely the seed area, and the supervision of the main task module comes from the image-level label of the weakly supervised semantic segmentation in step S2; the auxiliary learning module is based on the significant parts of the objects that are easily coupled with the decoder segmentation part, including: the head, tail, and car of the aircraft, and the supervision of the auxiliary learning module comes from the new airport pseudo-label generated in step S2.
[0066] Set the function f(x) to represent the encoder, whose parameter is the function θ that accepts the input image x. In the final feature layer, add a decoder to the auxiliary learning module for segmentation, add a fully connected layer to the main task module for classification, and use the SoftMax function to output the corresponding prediction results for each task, as follows:
[0067] For the true labels of training, the main task labels are y pri Represents the generated auxiliary task label y aux For the main task and auxiliary task modules, a fixed focusing parameter γ = 2 is used for the focusing loss L(y ^ ,y) training, its definition expression is as follows:
[0068] L(y ^ ,y)=-y(1-y ^ ) γ log(y ^ )
[0069] Among them, y ^ represents the predicted label and y represents the true label.
[0070] Then update the parameters θ of the multi-task network, and the focus loss function of the multi-task target is defined as follows:
[0071]
[0072] in, Represents the prediction result of the main task module, Represents the prediction results of the auxiliary task module and Represent the loss functions of the main task and auxiliary task respectively, arg θ It represents minimizing the loss function so that the predicted result is close to the real result. From the loss function, we can see that in the whole module, when the auxiliary branch is trained together with the main task, the prediction ability learned by the auxiliary task can also be learned by the main task, which helps to avoid the segmentation error of the main task and improve the performance of the main task. It is very useful for decoupling between classes. The principle of airport auxiliary learning is as follows Figure 4 shown.
[0073] Based on the training of the above-mentioned multi-task learning module, that is, using uncoupled pictures and auxiliary learning to help decoupling, a trained classification network is obtained to generate a new seed area (localization map), and the localization map is refined or the particularity of the airport environment is used to obtain a brand new uncoupled airport pseudo-label, that is, a deep neural network model (affinity network) is used to predict the semantic similarity between a pair of adjacent coordinates in the image, and the predicted affinity is used as the transition probability in the random walk, so that the random walk propagates the activation score of the localization map to the nearby area of the same semantic entity, significantly improving the quality of the localization map, and finally obtaining a brand new uncoupled airport pseudo-label.
[0074] After decoupling, the differences between different objects in the airport are very large, and they all have their own unique structures, colors, etc. (such as airplanes, cars, high pole lights, etc.). Therefore, the pixels belonging to the same object have a high semantic similarity, while the similarity between pixels belonging to different objects is very low.
[0075] In this embodiment, step S4 is specifically as follows:
[0076] Based on the new uncoupled airport pseudo-labels obtained in step S3, a weakly supervised semantic segmentation model is trained, that is, a segmentation model DeepLab with ResNet38 as the backbone is trained on the pseudo-labels obtained in step S3, and fully supervised to obtain the final segmentation results.
[0077] The weakly supervised semantic segmentation model uses the original image corresponding to the pseudo-label obtained in step S3 as the input of the model, and regards the pseudo-label as the real label for model training (since weakly supervised semantic segmentation does not have training segmentation labels, only pseudo-labels can be used). Finally, the trained segmentation model can be used to predict various scenes and pictures. The prediction results of the experiment in this embodiment are as follows: Figure 5As shown in the figure, it can be seen that it can not only segment larger objects in the sky and on the ground, but also process small objects such as airplanes and cars, showing good experimental performance.
[0078] In summary, the airport pseudo-label generation proposed by the method of the present invention can generate new data images and increase the diversity of data. After solving the coupling problem on the data set, the method of the present invention uses airport-based auxiliary task learning to solve the coupling problem. It uses the generated pseudo-labels to train airport-based auxiliary tasks and supervise the training of image classification main tasks, and finally generates better pseudo-labels to train the real semantic segmentation model. The method of the present invention enhances the robustness of the system in the airport environment. In the airport, it can effectively cope with the special environment of the airport, significantly improve the monitoring efficiency of aircraft and ground equipment, and has broad application prospects in the fields of autonomous driving, drone supervision and intelligent transportation, providing important support for the intelligent and automated operation of the airport.
[0079] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.
Claims
1. A decoupling-based weakly supervised airport image semantic segmentation method, the specific steps are as follows: S1. Build a weakly supervised semantic segmentation model; The model includes: Airport pseudo-labeling module, airport auxiliary learning and main task segmentation module, and semantic segmentation module; The airport auxiliary learning and main task segmentation module is a multi-task learning module, including: an auxiliary learning module and a main task module; the semantic segmentation module is a weakly supervised semantic segmentation model; S2. Use weakly supervised semantic segmentation method to collect instance images of various airport categories and input them into the airport pseudo-label module. Then input the training images into the airport pseudo-label module for decoupling operation to generate new training images and airport pseudo-labels. S3, based on step S2, new images and airport pseudo labels are obtained, and input into the airport auxiliary learning and main task segmentation module, and the airport-based auxiliary learning tasks and supervised training image classification main tasks are trained to generate new airport pseudo labels with better effects; S4, based on the new uncoupled airport pseudo-labels obtained in step S3, a weakly supervised semantic segmentation model is trained, that is, a segmentation model is trained on the pseudo-labels obtained in step S3, and fully supervised to obtain the final segmentation result; Among them, the weakly supervised semantic segmentation model includes: DeepLab, FCN, U-Net; the model backbone includes: ResNet38, VGG, ResNet.
2. According to the decoupling-based weakly supervised airport image semantic segmentation method of claim 1, it is characterized in that: The step S2 is specifically as follows: First, a weakly supervised semantic segmentation method is used to collect instance graphs of segmented objects, and a classification network is trained using image-level labels. The classification network uses ResNet as the backbone network and is followed by a fully connected layer. This classification network is used to generate a localization map, which is a rough segmentation map, that is, the seed region. The instance graph of the segmented object is the instance graph of each category of the airport; After obtaining the positioning map, the relationship between pixels is used to predict the boundary of the object. That is, the same side of the object boundary has similarity, and the opposite side of the object boundary has no similarity. The object boundary and the seed area can be combined to remove the part outside the boundary, and the expansion can obtain a more accurate segmentation result, that is, a pseudo label. Based on the pseudo labels obtained above, object instances with good segmentation effects are selected to decouple the relationship between objects. That is, the training images are input into the airport pseudo label module and decoupled to generate new airport pseudo labels. The object instance with good segmentation effect, that is, the current image has only one category; the decoupling operation is specifically as follows: Assume that the training data set has a total of N images. In each training round, N / 2 images are extracted from the training data set, and the same number of object instances are extracted from the generated pseudo labels. The object instances are obtained from the pseudo labels. Then, the extracted objects are randomly pasted into the input image to generate N / 2 batches of new images, and then the generated N / 2 new images are combined with the N / 2 images extracted from the dataset. Then, each training session obtains a batch of training images of size N. When N new training images are obtained, the pseudo labels generated above are updated to generate new airport pseudo labels. Among them, the extracted objects are randomly pasted into the input image, that is, the above object instances are merged into the input image, that is, by randomly adjusting the size of the objects, adding random rotation operations to change the direction of the objects, adding Gaussian smoothing processing to help the added object boundaries blend together more naturally, and pasting the moving objects in the airport to other categories of other pictures; and the object size does not exceed 50% of the image and is not less than 5%; the moving objects in the airport include: airplanes, cars, and people.
3. The method for semantic segmentation of airport images based on weak supervision based on decoupling according to claim 1, characterized in that: The step S3 is specifically as follows: First, a multi-task learning module is trained based on the airport pseudo-labels obtained in step S2 and N new training images, as follows: Among them, the auxiliary learning module and the main task module in the multi-task learning module share an encoder to extract features, and the main task module generates a rough segmentation map based on the classification network, that is, the seed area, and the supervision of the main task module comes from the image-level label of the weakly supervised semantic segmentation in step S2; the auxiliary learning module is based on the significant parts of the objects that are easily coupled to the decoder segmentation part, including: the head, tail, and car of the aircraft, and the supervision of the auxiliary learning module comes from the new airport pseudo-label generated in step S2; Set the function f(x) to represent the encoder, whose parameter is the function θ that accepts the input image x. In the final feature layer, add a decoder to the auxiliary learning module for segmentation, add a fully connected layer to the main task module for classification, and use the SoftMax function to output the corresponding prediction results for each task, as follows: For the true labels of training, the main task labels are y pri Represents the generated auxiliary task label y aux Represents; then for the main task and auxiliary task modules, a fixed focusing parameter γ = 2 focusing loss L(y^, y) is used for training, and its definition expression is as follows: L(y^,y)=-y(1-y^) γ log(y^) Among them, y^ represents the predicted label, and y represents the true label; Then update the parameters θ of the multi-task network, and the focus loss function of the multi-task target is defined as follows: in, Represents the prediction result of the main task module, Represents the prediction results of the auxiliary task module and Represent the loss functions of the main task and auxiliary task respectively, arg θ Represents the minimization loss function; Based on the training of the above-mentioned multi-task learning module, that is, using uncoupled pictures and auxiliary learning to help decoupling, the trained classification network generates new seed areas, and the positioning map is refined or the particularity of the airport environment is used to obtain a new uncoupled airport pseudo-label, that is, a deep neural network model is used to predict the semantic similarity between a pair of adjacent coordinates in the image, and the predicted affinity is used as the transition probability in the random walk, so that the random walk propagates the activation score of the positioning map to the nearby area of the same semantic entity, significantly improving the quality of the positioning map, and finally obtaining a new uncoupled airport pseudo-label.