Small sample remote sensing image semantic segmentation system based on multilevel feature aggregation and loss weighting

Abstract

The invention discloses a small sample remote sensing image semantic segmentation system based on multi-level feature aggregation and loss weighting. The small sample remote sensing image semantic segmentation system based on multi-level feature aggregation and loss weighting comprises a multi-level feature extraction module, a multi-level feature aggregation module and an accuracy loss weighting module. According to the small sample remote sensing image semantic segmentation system based on multi-level feature aggregation and loss weighting provided by the embodiment of the invention, multi-level and scale features are generated for the system through the multi-level feature extraction module, and more segmentation information is provided to adapt to target objects of different sizes; features of different levels are aggregated through a multi-level feature aggregation module, multi-level and scale segmentation results are output, rich information is provided for final accurate segmentation, a loss weight is calculated by using a segmentation accuracy-intersection-union ratio as a standard through an accuracy loss weighting module, so that the training condition under the condition of multiple segmentation tasks is balanced, and the training efficiency is improved. And the segmentation performance of the model is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a small sample remote sensing image semantic segmentation system based on multi-level feature aggregation and loss weighting. Background Art

[0002] In the field of remote sensing image semantic segmentation, efforts have long been made to perform pixel-level semantic segmentation of remote sensing images to support applications in geographic information systems, environmental monitoring, and military reconnaissance. The development of this technology provides strong support for our precise understanding of surface information, allowing us to understand changes on the earth's surface from a distant perspective. However, traditional remote sensing image semantic segmentation is often limited by the cost of data acquisition and the complexity of annotation, resulting in a limited number of samples available for training, which limits its accuracy and universality in practical applications.

[0003] In view of the limitations of traditional remote sensing image semantic segmentation, small sample remote sensing image semantic segmentation has emerged. It aims to achieve more accurate pixel-level segmentation using limited training data. This method attempts to better understand and segment remote sensing images even when data is scarce through a new learning paradigm. Small sample remote sensing image semantic segmentation provides stronger generalization capabilities for categories or samples that have never been seen before, and is expected to become a new way to overcome traditional limitations.

[0004] However, the current semantic segmentation of small sample remote sensing images still faces some challenges. Due to the limited data, the model may perform poorly on unknown categories or samples, resulting in a decrease in the accuracy of the segmentation results. The scarcity of data limits the model's ability to learn new sample features, which may affect its application in complex scenarios. Therefore, although the semantic segmentation technology of small sample remote sensing images has obvious advantages, it still needs more in-depth research and improvement to overcome its limitations in practical applications.

[0005] Based on this, there is an urgent need for a small-sample remote sensing image semantic segmentation system based on multi-level feature aggregation and loss weighting. Summary of the invention

[0006] In view of the above problems, the present invention is proposed to provide a small sample remote sensing image semantic segmentation system based on multi-level feature aggregation and loss weighting to overcome the above problems or at least partially solve the above problems.

[0007] Other features and advantages of the present invention will become apparent from the following detailed description, or may be learned in part by practice of the present invention.

[0008] According to a first aspect of an embodiment of the present invention, a small sample remote sensing image semantic segmentation system based on multi-level feature aggregation and loss weighting is provided, and the small sample remote sensing image semantic segmentation system based on multi-level feature aggregation and loss weighting comprises: a multi-level feature extraction module, used to extract feature information of a support data set and a query data set, the support data set is a set of support image data with existing labels, the query data set is a set of query image data, and the feature information includes support image features and query image features;

[0009] A multi-level feature aggregation module is used to aggregate the feature information extracted by the multi-level feature extraction module from top to bottom to obtain a feature aggregation result;

[0010] The accuracy loss weighting module is used to calculate the corresponding segmentation accuracy of the feature aggregation result output by the multi-level feature aggregation module, and calculate the loss weight corresponding to the feature aggregation result according to the segmentation accuracy to determine the overall training loss of the model in the system.

[0011] In some embodiments of the present invention, the multi-level feature extraction module is composed of 5 feature extraction layers connected in sequence, and each of the feature extraction layers is composed of basic neurons including a stacked convolution layer, a pooling layer, and a normalization layer.

[0012] In some embodiments of the present invention, the supporting image data is sequentially passed through five sequentially connected feature extraction layers to extract supporting image features of the supporting data set, and except for the feature extraction layer of the first layer, the input of each feature extraction layer is the output of the feature extraction layer of the previous layer.

[0013] In some embodiments of the present invention, the query image data is sequentially passed through five sequentially connected feature extraction layers to extract query image features of the query data set, and except for the feature extraction layer of the first layer, the input of each feature extraction layer is the output of the feature extraction layer of the previous layer.

[0014] In some embodiments of the present invention, the multi-level feature aggregation module is specifically used to:

[0015] Extracting a supporting prototype for semantic segmentation of input features according to supporting picture features, the supporting prototype comprising an upsampler, a convolver, and a segmenter;

[0016] The resolution of the query image features of the fourth layer corresponding to the query image data is enlarged by an upsampler;

[0017] The query image features of the third layer corresponding to the query image data and the query image features of the fourth layer after the resolution is enlarged are concatenated, and the feature dimensions are aligned through a convolution device to obtain a first aggregation result, and a segmentation result of the first aggregation result is outputted using a segmenter;

[0018] After concatenating the query image features of the second layer and the first aggregation result, the feature dimensions are aligned through a convolution device to obtain a second aggregation result, and a segmentation result of the second aggregation result is output using a segmenter;

[0019] After concatenating the query image features of the first layer and the second aggregation result, the feature dimensions are aligned through a convolution device to obtain a third aggregation result, and a segmentation result of the third aggregation result is output using a segmenter;

[0020] The segmentation result of the first aggregation result, the segmentation result of the second aggregation result, and the segmentation result of the third aggregation result are taken as the feature aggregation result.

[0021] In some embodiments of the present invention, the accuracy loss weighting module is specifically used to:

[0022] The feature aggregation result is enlarged in resolution by an upsampler so that each segmentation result corresponding to the feature aggregation result is consistent with the label size of the query image data;

[0023] Calculate the segmentation accuracy of each segmentation result and express it as intersection-over-union ratio;

[0024] The loss weight corresponding to the feature aggregation result is calculated according to the segmentation accuracy to determine the overall training loss of the model in the system.

[0025] The technical solution provided in the embodiments of the present invention has at least the following technical effects or advantages:

[0026] The small-sample remote sensing image semantic segmentation system based on multi-level feature aggregation and loss weighting described in the embodiment of the present invention generates multi-level and scale features for the system through a multi-level feature extraction module, provides more segmentation information to adapt to target objects of different sizes, aggregates features of different levels through a multi-level feature aggregation module, outputs multi-level and scale segmentation results, and provides rich information for the final accurate segmentation. The segmentation accuracy-intersection-over-union ratio is used as the standard to calculate the loss weight through the accuracy loss weighting module to balance the training conditions in the case of multiple segmentation tasks and improve the segmentation performance of the model.

[0027] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0029] Figure 1 A flowchart of a small sample remote sensing image semantic segmentation system based on multi-level feature aggregation and loss weighting provided by an embodiment of the present invention;

[0030] Figure 2 Reference example diagram for the few-shot segmentation task. DETAILED DESCRIPTION

[0031] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings.

[0032] Various structural schematic diagrams according to embodiments of the present disclosure are shown in the accompanying drawings. These figures are not drawn to scale, and some details are magnified and some details may be omitted for the purpose of clear expression. The shapes of various regions and layers shown in the figures and the relative sizes and positional relationships therebetween are only exemplary, and may deviate in practice due to manufacturing tolerances or technical limitations, and those skilled in the art may further design regions / layers with different shapes, sizes, and relative positions according to actual needs.

[0033] In the context of the present disclosure, when a layer / element is referred to as being "on" another layer / element, the layer / element may be directly on the other layer / element, or there may be an intervening layer / element therebetween. Additionally, if a layer / element is "on" another layer / element in one orientation, the layer / element may be "below" the other layer / element when the orientation is reversed. In the context of the present disclosure, similar or identical components may be denoted by the same or similar reference numerals.

[0034] In order to better understand the above-mentioned technical scheme, the above-mentioned technical scheme will be described in detail below in combination with specific implementation methods. It should be understood that the embodiments of the present disclosure and the specific features in the embodiments are detailed descriptions of the technical scheme of the present invention, rather than limitations on the technical scheme of the present invention. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0035] Figure 1 is a flowchart of a small sample remote sensing image semantic segmentation system based on multi-level feature aggregation and loss weighting provided by an embodiment of the present invention. Figure 1 As shown, the small sample remote sensing image semantic segmentation system based on multi-level feature aggregation and loss weighting includes a multi-level feature extraction module 100, a multi-level feature aggregation module 200 and an accuracy loss weighting module 300.

[0036] The multi-level feature extraction module 100 is used to extract feature information of a support data set and a query data set. The support data set is a set of support image data with labels, and the query data set is a set of query image data. The feature information includes support image features and query image features.

[0037] Since remote sensing images contain rich ground object information, remote sensing images may involve different scenes, landforms or ground object types, and the ground object boundary and texture information in remote sensing images are usually complex, remote sensing images present different features at different scales. Based on this, the embodiment of the present invention implements multi-level feature extraction of remote sensing images through the multi-level feature extraction module 100. The embodiment of the present invention first defines two remote sensing image data sets with multiple segmentation categories, including a training set D train and the test set D test , the segmentation categories contained in the two do not overlap, that is, Train the model on the training set D train After training, it will be directly transferred to the test set D test The system training of the present invention adopts a meta-learning method and is carried out in a batch training mode. Each batch is a small sample segmentation task, including a support data set S and a query data set Q.

[0038] Specifically, according to the N-way-K-shot setting of the small sample training model, each support dataset S includes K samples with annotations M, that is, The query dataset Q should include 1 query image data X q and the label M which is only used to calculate the loss during training q , Q = {X q , M q}. The training set after batch sampling can also be expressed as The test set is Where N train and N test are the number of sampling batches during training and testing, respectively. During training, the labels of query images are used to calculate the loss, while the labels of query images in the testing phase are not available.

[0039] Optionally, in an embodiment of the present invention, the multi-level feature extraction module 100 is composed of 5 feature extraction layers connected in sequence, and each of the feature extraction layers is composed of basic neurons including a stacked convolution layer, a pooling layer, and a normalization layer.

[0040] The feature extraction layer is defined as Layer i , the feature information is defined as f i , then:

[0041] f i =Layer i (f i-1 ), i={0, 1, 2, 3, 4};

[0042] Among them, f -1 is the input image, which can be supporting image data or query image data. Here, for simplicity of description, the number of channels, height, and width of all feature maps (feature information of each layer) are defined as: c i ,h i , w i .

[0043] When the input image passes through each feature extraction layer in turn, each layer will output the extracted features for the image, and the resolution will be reduced to half of the original.

[0044] For example, the supporting image data is sequentially passed through five sequentially connected feature extraction layers to extract supporting image features of the supporting data set, and except for the first feature extraction layer, the input of each feature extraction layer is the output of the previous feature extraction layer.

[0045] For example, the query image data is sequentially passed through five sequentially connected feature extraction layers to extract query image features of the query data set, and except for the feature extraction layer of the first layer, the input of each feature extraction layer is the output of the feature extraction layer of the previous layer.

[0046] The multi-level feature aggregation module 200 is used to aggregate the feature information extracted by the multi-level feature extraction module 100 from top to bottom to obtain a feature aggregation result.

[0047] In an embodiment of the present invention, the multi-level feature aggregation module 200 is specifically used to: extract a support prototype for semantic segmentation of input features based on support image features, the support prototype including an upsampler, a convolver and a segmenter; expand the resolution of the query image features of the fourth layer corresponding to the query image data through an upsampler; splice the query image features of the third layer corresponding to the query image data and the query image features of the fourth layer after the resolution is expanded, align the feature dimensions through a convolver to obtain a first aggregation result, and use a segmenter to output the segmentation result of the first aggregation result; splice the query image features of the second layer and the first aggregation result, align the feature dimensions through a convolver to obtain a second aggregation result, and use a segmenter to output the segmentation result of the second aggregation result; splice the query image features of the first layer and the second aggregation result, align the feature dimensions through a convolver to obtain a third aggregation result, and use a segmenter to output the segmentation result of the third aggregation result; use the segmentation result of the first aggregation result, the segmentation result of the second aggregation result and the segmentation result of the third aggregation result as the feature aggregation result.

[0048] The multi-level feature aggregation module 200 aggregates the multi-level features obtained by the feature extraction module from top to bottom, so that the segmentation results of the model can combine features at different levels, thereby improving the segmentation performance. The improvement in segmentation performance lies in that the aggregation module is designed based on the characteristics of the image to be segmented. First, the sizes of the objects to be segmented in remote sensing images are different, and the resolution features required for the optimal segmentation are also different; in addition, segmentation requires high-level semantically clear features, but remote sensing images also require low-level, semantically simple features due to their simple semantics.

[0049] Specifically, the process of multi-level feature aggregation is as follows. In the embodiment of the present invention, the supporting prototype is first extracted using supporting image features and labels:

[0050]

[0051] Among them, P represents the supported prototype, (x, y) represents the spatial coordinates, and the function δ[.] is an indicator function, which takes 1 when the parameter is true, otherwise it takes 0. The upsampler, convolution, and segmenter are defined as The upsampler is used to double the input feature resolution, the convolutional layer is used to align the feature dimensions and expand the receptive field, and the segmenter is used to perform semantic segmentation on the input features.

[0052] Then the fourth layer of features of the query image data is enlarged in resolution through an upsampler:

[0053]

[0054] Then, the query image features of the third layer corresponding to the query image data and the query image features of the fourth layer after the resolution is enlarged are concatenated and the feature dimensions are aligned through the convolution device to obtain the first aggregation result. The segmentation result of the first aggregation result is output using the segmenter:

[0055]

[0056] in, It means concatenating the input features in the feature dimension. Indicates that the resolution size of the input feature is expanded to

[0057] Finally, the segmentation result of the third and fourth layer features is output by the segmenter:

[0058]

[0059] Similarly, the upsampler, convolutional unit, and segmenter are used to concatenate the query image features of the second layer and the first aggregation result, and then the feature dimensions are aligned through the convolutional unit to obtain the second aggregation result. The segmentation result of the second aggregation result is output using the segmenter:

[0060]

[0061] Because Layer 0 The features of are relatively low-level and have a small receptive field, so the features of this layer do not participate in the aggregation process. Therefore, the upsampler, convolutional unit, and segmenter are used to concatenate the query image features of the first layer and the second aggregation result, and then the feature dimensions are aligned through the convolutional unit to obtain the third aggregation result. The segmentation result of the third aggregation result is output by the segmenter;

[0062] The multi-level feature aggregation module 200 outputs three different levels of out 34 ,out 234 ,out 1234 The final output of the model is:

[0063]

[0064] The segmentation result of the first aggregation result, the segmentation result of the second aggregation result, and the segmentation result of the third aggregation result are taken as the feature aggregation result.

[0065] The accuracy loss weighting module 300 is used to calculate the corresponding segmentation accuracy of the feature aggregation result output by the multi-level feature aggregation module 200, and calculate the loss weight corresponding to the feature aggregation result according to the segmentation accuracy to determine the overall training loss of the model in the system.

[0066] In a system with multiple segmentation tasks, there is an imbalance in the accuracy and loss values ​​of each segmentation task, which can easily cause the optimization process of the model to fall into a local optimum and affect the segmentation effect. Therefore, in order to balance the loss values ​​of multiple tasks and improve the segmentation performance of the model, an accuracy-based accuracy loss weighting module 300 is provided in the embodiment of the present invention. The accuracy loss weighting module 300 calculates the accuracy of the intersection over union (IoU) of the output results of multiple segmentation tasks and sorts them. If the accuracy of a segmentation task is higher, it means that the level and scale of the output features of the segmentation task are more suitable for remote sensing images, so the segmentation task is given a higher weight, otherwise, a low weight is given.

[0067] In an embodiment of the present invention, the accuracy loss weighting module 300 is specifically used to: expand the resolution of the feature aggregation result through an upsampler so that the corresponding segmentation results in the feature aggregation result are consistent with the label size of the query image data; calculate the segmentation accuracy corresponding to each segmentation result respectively, and express it as an intersection-over-union ratio; calculate the loss weight corresponding to the feature aggregation result according to the segmentation accuracy, and determine the overall training loss of the model in the system.

[0068] Specifically, the embodiment of the present invention defines the feature aggregation result as OUT, then:

[0069] OUT=[out 34 ,out 234 ,out 1234 ];

[0070] Then all the outputs are expanded to the input size through upsampling operation, that is, expanded to R 2×H×W , which is consistent with the label size of the query image. Next, calculate the segmentation results of all outputs, and the segmentation accuracy is expressed as the intersection over union ratio. The accuracy set IoU is:

[0071] IoU = {iou i}; i = {1, 2, 3};

[0072] in:

[0073]

[0074] In the formula Indicates the intersection-over-union ratio of the output calculated based on the label.

[0075] Then calculate the respective loss weights according to the accuracy value:

[0076]

[0077] Finally, the cross entropy loss function is calculated to supervise the optimization process of the model. The multi-level loss value is:

[0078]

[0079] Among them, N is the sum of the pixels of the image involved in calculating the loss, and the loss value of the final output out is:

[0080]

[0081] The overall training loss of the model is:

[0082]

[0083] Among them, α and β are loss weights.

[0084] The embodiment of the present invention uses a trained and deployed system based on a test data set D test , use the multi-level feature extraction module 100 to extract the multi-level features of the input image; use the multi-level feature aggregation module 200 to aggregate the multi-level features to obtain a multi-level output; use the accuracy loss weighting module 300 to calculate the training loss; use the final output result to perform accuracy testing and visualization testing; thereby verifying the aforementioned model based on the test process.

[0085] On the basis of the above embodiment, the embodiment of the present invention provides an implementation example based on an actual data set. This embodiment uses a public data set DLRSD, with a total number of categories of 20. In addition, pytorch is used as the artificial intelligence model framework.

[0086] Specifically, the multi-level feature extraction module 100 is improved from the deep convolutional neural network VGG16, and its composition is shown in Table 1. In this example, the VGG16 network is divided into 5 layers, and during training, the pre-trained parameters of the VGG16 network on the ImageNet dataset are loaded to speed up the training progress.

[0087] Table 1 Feature extraction module structure in this implementation example

[0088]

[0089]

[0090] In this embodiment, The structure is as shown in the following table, where Interpolate (scale_fact or = 2) means that the input feature resolution is magnified to twice the original value, and the input and output sizes of this component are not fixed.

[0091] Table 2 Upsampler structure in this implementation example

[0092]

[0093] in, The structure is shown in Table 3 below:

[0094] Table 3 Convolution structure in this embodiment

[0095]

[0096] in, The structure is shown in Table 4 below:

[0097] Table 4 The structure of the splitter in this embodiment

[0098]

[0099] As for the training details and hyperparameters, in the training and testing phases of the embodiment of the present invention, the input image resolution is 256×256 pixels (i.e., H and W are both 256), and the size of the feature map output by the multi-level feature extraction module 100 is 1 / 32 of the input image size. During training, the training images are randomly cropped and flipped. The SGD optimizer is used to train the entire system, and the training lasts for 1000 iterations, with a learning rate of 0.001, a batch size of 8, a momentum of 0.99, and a weight decay rate of 0.0005. The α of the loss weight is 0.5 and the β is 0.5.

[0100] After the training is completed, the system is tested on a randomly selected small sample segmentation task. The test results are as follows: Figure 2 As shown in the figure, the segmentation results have been marked with different colors in the picture. The experimental results show that the system trained in this example can achieve accurate semantic segmentation of small sample remote sensing images.

[0101] The small sample remote sensing image semantic segmentation system based on multi-level feature aggregation and loss weighting described in the embodiment of the present invention generates multi-level and scale features for the system through the multi-level feature extraction module 100, provides more segmentation information to adapt to target objects of different sizes, aggregates features of different levels through the multi-level feature aggregation module 200, outputs multi-level and scale segmentation results, and provides rich information for the final accurate segmentation. The accuracy loss weighting module 300 uses the segmentation accuracy-intersection-over-union ratio as the standard to calculate the loss weight, so as to balance the training conditions in the case of multiple segmentation tasks and improve the segmentation performance of the model.

[0102] In the description provided herein, a large number of specific details are described. However, it is understood that embodiments of the present invention can be practiced without these specific details. In some instances, well-known methods, structures and techniques are not shown in detail so as not to obscure the understanding of this description.

[0103] Similarly, it should be understood that in order to streamline the present disclosure and aid in understanding one or more of the various inventive aspects, in the above description of exemplary embodiments of the present invention, various features of the present invention are sometimes grouped together into a single embodiment, figure, or description thereof. However, this disclosed method should not be interpreted as reflecting the intention that the claimed invention requires more features than those expressly recited in each claim. Rather, as reflected in the claims, inventive aspects lie in less than all of the features of the individual embodiments previously disclosed. Therefore, the claims that follow the detailed description are hereby expressly incorporated into the detailed description, with each claim itself serving as a separate embodiment of the present invention.

[0104] It should be noted that the above-mentioned embodiments illustrate rather than limit the invention and that those skilled in the art will be able to design alternative embodiments without departing from the scope of the appended claims.

Claims

1. A small sample remote sensing image semantic segmentation system based on multi-level feature aggregation and loss weighting, characterized in that: The small sample remote sensing image semantic segmentation system based on multi-level feature aggregation and loss weighting includes: A multi-level feature extraction module, used to extract feature information of a support data set and a query data set, wherein the support data set is a set of support image data with labels, and the query data set is a set of query image data, and the feature information includes support image features and query image features; A multi-level feature aggregation module is used to aggregate the feature information extracted by the multi-level feature extraction module from top to bottom to obtain a feature aggregation result; The accuracy loss weighting module is used to calculate the corresponding segmentation accuracy of the feature aggregation result output by the multi-level feature aggregation module, and calculate the loss weight corresponding to the feature aggregation result according to the segmentation accuracy to determine the overall training loss of the model in the system.

2. The small sample remote sensing image semantic segmentation system based on multi-level feature aggregation and loss weighting according to claim 1, characterized in that: The multi-level feature extraction module is composed of 5 feature extraction layers connected in sequence, and each of the feature extraction layers is composed of basic neurons including a stacked convolution layer, a pooling layer, and a normalization layer.

3. The small sample remote sensing image semantic segmentation system based on multi-level feature aggregation and loss weighting according to claim 2, characterized in that: The supporting image data is sequentially passed through five sequentially connected feature extraction layers to extract supporting image features of the supporting data set, and except for the feature extraction layer of the first layer, the input of each feature extraction layer is the output of the feature extraction layer of the previous layer.

4. The small sample remote sensing image semantic segmentation system based on multi-level feature aggregation and loss weighting according to claim 2, characterized in that: The query image data is sequentially passed through five sequentially connected feature extraction layers to extract query image features of the query data set, and except for the first feature extraction layer, the input of each feature extraction layer is the output of the previous feature extraction layer.

5. The small sample remote sensing image semantic segmentation system based on multi-level feature aggregation and loss weighting according to claim 2, characterized in that: The multi-level feature aggregation module is specifically used for: Extracting a supporting prototype for semantic segmentation of input features according to supporting picture features, the supporting prototype comprising an upsampler, a convolver, and a segmenter; The resolution of the query image features of the fourth layer corresponding to the query image data is enlarged by an upsampler; The query image features of the third layer corresponding to the query image data and the query image features of the fourth layer after the resolution is enlarged are concatenated, and the feature dimensions are aligned through a convolution device to obtain a first aggregation result, and a segmentation result of the first aggregation result is outputted using a segmenter; After concatenating the query image features of the second layer and the first aggregation result, the feature dimensions are aligned through a convolution device to obtain a second aggregation result, and a segmentation result of the second aggregation result is output using a segmenter; After concatenating the query image features of the first layer and the second aggregation result, the feature dimensions are aligned through a convolution device to obtain a third aggregation result, and a segmentation result of the third aggregation result is output using a segmenter; The segmentation result of the first aggregation result, the segmentation result of the second aggregation result, and the set of the segmentation result of the third aggregation result are used as the feature aggregation result.

6. The small sample remote sensing image semantic segmentation system based on multi-level feature aggregation and loss weighting according to claim 5, characterized in that: The accuracy loss weighting module is specifically used for: The feature aggregation result is enlarged in resolution by an upsampler so that each segmentation result corresponding to the feature aggregation result is consistent with the label size of the query image data; Calculate the segmentation accuracy of each segmentation result and express it as intersection-over-union ratio; The loss weight corresponding to the feature aggregation result is calculated according to the segmentation accuracy to determine the overall training loss of the model in the system.

Images