A progressive multi-level semantic segmentation method for remote sensing images for machine vision

Abstract

This invention provides a progressive multi-level semantic segmentation method for remote sensing images oriented towards machine vision. First, target objects in the original remote sensing image are identified and labeled, and training and testing sets are created. Then, a progressive multi-level semantic segmentation model for remote sensing images is built, consisting of several identical processing stages stacked together. Each stage includes the same preprocessing module, segmentation module, thinning module, and super-resolution module. As the image is progressively segmented from the coarsest to the finest level, progressively refined segmentation results are output. This invention achieves accurate segmentation of targets in remote sensing images while preserving the global information of the entire image. Its processing speed far exceeds traditional techniques, playing a crucial role in the rapid and accurate processing of remote sensing images.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology, specifically relating to a progressive multi-level semantic segmentation method for remote sensing images for machine vision. Background Technology

[0002] With the continuous increase in the number of satellites and the continuous development of remote sensing technology in my country, the data volume of high-resolution remote sensing images has multiplied. Faced with this massive amount of data, the demand for rapid and accurate processing of remote sensing images has become increasingly prominent. In recent years, with the vigorous development of artificial intelligence technology, machine vision technology has been increasingly widely applied in the field of remote sensing image processing. Semantic segmentation technology is one way to process remote sensing images using machine vision principles. Its purpose is to manually label each pixel in the remote sensing image with a specific species category, assigning each pixel a specific category meaning. This allows for the segmentation of specific regions or targets in the remote sensing image, making the processing speed of remote sensing images far exceed that of traditional technologies, playing a crucial role in the rapid and accurate processing of remote sensing images.

[0003] Current advanced remote sensing image semantic segmentation techniques are based on deep learning, utilizing convolutional neural networks (CNNs) to acquire input images and output segmentation results. Patent application CN202110530385.2 proposes an unsupervised remote sensing image semantic segmentation method based on super-resolution and domain adaptation. It replaces the ASPP module in the original super-resolution domain-adaptive unsupervised remote sensing image semantic segmentation method with a feature pyramid attention module, and applies the residual feature pyramid attention module to the discriminator to obtain accurate pixel-level attention for high-level semantic features; it also alleviates the class imbalance problem through the Dice coefficient loss function. Patent application CN202111530033.3 proposes a deep learning-based remote sensing image semantic segmentation method that modifies the SegNet network by adding two dilated spatial pyramid pooling (ASPP) modules in the first two stages of the encoder. This effectively extracts multi-scale features and preserves spatial location information. After training the two ASPP modules, corresponding feature fusion is performed in the decoder. Patent application CN202110623160.1 proposes an unsupervised domain-adaptive semantic segmentation method for remote sensing images with feature adaptive alignment. This method is a deep learning-based unsupervised domain-adaptive semantic segmentation method for remote sensing images based on an adversarial learning framework. The method generally consists of two main modules: a semantic segmentation model and a domain discriminator. The domain discriminator aligns the target domain features extracted by the segmentation model with the source domain features in the output space, thereby improving the segmentation performance of the segmentation model on the target domain image.

[0004] However, due to limitations in computing and memory resources, most existing technologies cannot process high-resolution input images. A common method for processing high-resolution remote sensing images is downsampling, but this can result in low-resolution output segmentation maps. Another approach is to cut the image into small patches, perform semantic segmentation on each patch separately, and then stitch them together. However, this method does not consider the global information of the entire remote sensing image. The significant difference between the size of the entire remote sensing image and the size of local image patches leads to starkly contrasting segmentation results, and it is difficult to reduce and adjust this difference through a single processing stage. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a progressive multi-level semantic segmentation method for remote sensing images oriented towards machine vision. Based on a specific semantic segmentation network, the method continuously refines the output segmentation map through the superposition of segmented images at multiple levels. This process refines the segmentation map, giving it both global and rich detail information. Combined with global contextual clues, it generates more accurate segmentation. The refined segmentation map is then subjected to super-resolution reconstruction with different magnification factors, outputting a high-resolution, clear semantic segmentation map while satisfying computational and memory resource constraints.

[0006] This invention proposes a progressive multi-level semantic segmentation method for remote sensing images oriented towards machine vision, comprising:

[0007] Step 1: Identify and label target objects in the original remote sensing images to create training and testing sets; specifically:

[0008] Step 1.1: Classify the target objects in the original remote sensing image;

[0009] Step 1.2: Assign RGB values ​​to each item category;

[0010] Step 1.3: Colorize the item type according to the assigned RGB values ​​to obtain the remote sensing image label set;

[0011] Step 1.4: Crop the obtained original images and label sets, flip the images vertically and horizontally, and then divide them into training and test sets according to a certain ratio.

[0012] Step 2: Build a progressive multi-level semantic segmentation model for remote sensing images for machine vision; the segmentation model consists of several identical processing stages stacked together, each stage containing the same preprocessing module, segmentation module, thinning module and super-resolution module. As the image is progressively segmented from the coarsest level to the finest level, the segmentation result image is output with progressive refinement.

[0013] The preprocessing module is used to perform windowing and downsampling processing on the original remote sensing image X to output remote sensing images X of different sizes required for different levels S. w S- ;

[0014] The segmentation module is used to segment the input remote sensing image X. w S- Feature extraction and feature fusion are performed to obtain segmented image A. w S- ;

[0015] The refinement module is used to refine the segmented image Y output from the previous level. w (S-1)- With segmented image A w S- We obtain an initial combined segmentation image B, and calculate the segmentation image Y output from the previous level. w (S-1)- The uncertainty score C between the segmented image B and the image Y is determined, and the image Y is evaluated based on the score C. w (S-1)- The image is thinned to obtain the thinned image Y. w S+ ;

[0016] The super-resolution module is used to refine the image Y. w S+ Super-resolution processing is performed to obtain the super-resolution image Y. S Image Y S As input to the next level of preprocessing module.

[0017] The process of building the preprocessing module is as follows:

[0018] Step A1: Set a group of rectangular windows W S ={w|w=(x,y,h) S ,w S The original remote sensing image X is divided into blocks, with each window consisting of the coordinates of its upper left corner (x, y) and its height h. S and width w S designated;

[0019] Step A2: Based on the rectangular window W used to divide the image blocks S The input image X is divided into segments of size h. S ×w S Image Patch X w S ;

[0020] Step A3: For image patch X w S The image patch X is obtained by performing a downsampling operation.w S- Image block X w S- The data are respectively input into the segmentation module and the refinement module in the network model;

[0021] The process of building the segmentation module is as follows:

[0022] The image patch X output by the preprocessing module at the current level S. w S- As input to the PSPNet network, features are extracted by the first pre-trained model, ResNet101, outputting feature map T1. The pyramid pooling module uses pooling operations of four different sizes (1×1, 2×2, 3×3, and 6×6) to obtain feature maps T2, T3, T4, and T5 of multiple sizes. These feature maps are then subjected to 1×1 convolutions to reduce the number of channels, resulting in feature maps T2′, T3′, T4′, and T5′. Bilinear interpolation is then used to upsample feature maps T2′, T3′, T4′, and T5′ to obtain feature maps T2″, T3″, T4″, and T5″. Feature map T1 is then concatenated with T2″, T3″, T4″, and T5″ on each channel to obtain feature map T6 containing overall information. A 1×1 convolutional layer is used to change the number of channels in feature map T6 to obtain the segmented image A. w S- .

[0023] The process of building the refinement module is as follows:

[0024] Step B1: Overlay the semantic segmentation maps of all levels preceding the current level S into an image Y. w (S-1)- and segmentation image A w S- As input to the second pre-trained model ResNet101, it outputs an initial combined segmentation map B;

[0025] Step B2: Calculate the image Y w (S-1)- And the prediction uncertainty score Y of B u and B u ;

[0026] Step B3: Use two uncertainty scores Y u and B u To select image Y w (S-1)- Thinning operations are performed at n positions to obtain the thinned image Y. w S+ .

[0027] The construction process of the super-resolution module is as follows:

[0028] Step C1: Extract the thinned image Y w S+ Features, output feature map T7;

[0029] Step C2: Combine several residual blocks together to share parameters at different levels. Feature map T7 is output as feature map T8 through the combined residual blocks.

[0030] Step C3: Upsample the feature map T8 and pass it through a convolutional layer to obtain the super-resolution image Y. S If multiple refinement processes are required, increment the number of layers S by 1 to improve the super-resolution image Y. S As input to the preprocessing module in the next processing stage;

[0031] Step 3: Set training parameters, construct loss function, train the progressive remote sensing image multi-level semantic segmentation model for machine vision, update network parameters, and obtain the trained progressive remote sensing image multi-level semantic segmentation model for machine vision.

[0032] Step 4: Input the remote sensing image test set from Step 1 into the trained progressive multi-level semantic segmentation model for machine vision remote sensing images, and output the accurate segmentation results of the remote sensing images.

[0033] The beneficial effects of this invention are:

[0034] (1) The present invention divides the input image into blocks and generates a high-resolution segmentation output without affecting the GPU memory usage.

[0035] (2) The segmentation module of the present invention aggregates contextual information from different regions, thereby improving the ability to obtain global information and outputting high-quality semantic segmentation result maps.

[0036] (3) In order to avoid ignoring local details in remote sensing images, the refinement module of the present invention includes multiple segmentation stages. The output of the previous stage will be used as the input of the next stage, and the segmentation output will be gradually refined, including multiple image blocks of various scales from the coarsest to the finest.

[0037] (4) The super-resolution module of the present invention is a multi-level super-resolution network. This network can reconstruct high-resolution images with different magnification factors in a single model. By utilizing the correlation between levels, information at different levels is shared, thereby realizing super-resolution reconstruction of remote sensing images of different sizes. Attached Figure Description

[0038] Figure 1 This is a flowchart of the progressive multi-level semantic segmentation method for remote sensing images for machine vision in this invention.

[0039] Figure 2 This is a schematic diagram of the progressive remote sensing image multi-level semantic segmentation model for machine vision in this invention.

[0040] Figure 3 This is a schematic diagram of the preprocessing module structure in this invention;

[0041] Figure 4 This is a schematic diagram of the segmentation module structure in this invention;

[0042] Figure 5 This is a schematic diagram of the detailed module structure in this invention;

[0043] Figure 6 This is a schematic diagram of the super-resolution module structure in this invention. Detailed Implementation

[0044] The invention will be further explained below with reference to the accompanying drawings and specific implementation examples.

[0045] like Figures 1-2 As shown, a progressive multi-level semantic segmentation method for remote sensing images for machine vision includes:

[0046] Step 1: Identify and label target objects in the original remote sensing images to create training and testing sets; specifically:

[0047] Step 1.1: Classify the target objects in the original remote sensing image;

[0048] Step 1.2: Assign RGB values ​​to each item category;

[0049] Step 1.3: Colorize the item type according to the assigned RGB values ​​to obtain the remote sensing image label set;

[0050] Step 1.4: Crop the obtained original images and label sets, flip the images vertically and horizontally, and then divide them into training and test sets according to a certain ratio.

[0051] Step 2: Build a progressive multi-level semantic segmentation model for remote sensing images for machine vision; the segmentation model consists of several identical processing stages stacked together, each stage containing the same preprocessing module, segmentation module, thinning module and super-resolution module. As the image is progressively segmented from the coarsest level to the finest level, the segmentation result image is output with progressive refinement.

[0052] The preprocessing module is used to perform windowing and downsampling processing on the original remote sensing image X to output remote sensing images X of different sizes required for different levels S. w S- ;

[0053] The segmentation module is used to segment the input remote sensing image X.w S- Feature extraction and feature fusion are performed to obtain segmented image A. w S- ;

[0054] The refinement module is used to refine the segmented image Y output from the previous level. w (S-1)- With segmented image A w S- We obtain an initial combined segmentation image B, and calculate the segmentation image Y output from the previous level. w (S-1)- The uncertainty score C between the segmented image B and the image Y is determined, and the image Y is evaluated based on the score C. w (S-1)- The image is thinned to obtain the thinned image Y. w S+ ;

[0055] The super-resolution module is used to refine the image Y. w S+ Super-resolution processing is performed to obtain the super-resolution image Y. S Image Y S As input to the next level of preprocessing module.

[0056] like Figure 3 As shown, the construction process of the preprocessing module is as follows:

[0057] Step A1: Set a group of rectangular windows W S ={w|w=(x,y,h) S ,w S The original remote sensing image X is divided into blocks, with each window consisting of the coordinates of its upper left corner (x, y) and its height h. S and width w S Specify; as the hierarchy level S increases, the width and height of the rectangular window decrease, but W s The base number increases.

[0058] Step A2: Based on the rectangular window W used to divide the image blocks S The input image X is divided into segments of size h. S ×w S Image Patch X w S ;

[0059] Step A3: For image patch X w S The image patch X is obtained by performing a downsampling operation. w S- Image block X w S-The data are respectively input into the segmentation module and the refinement module in the network model;

[0060] like Figure 4 As shown, the segmentation module uses the PSPNet network, including a pre-trained ResNet101 model, a pyramid pooling module, and a 1×1 convolutional layer. The pre-trained ResNet101 model is used to extract features and output a feature map T1. The feature map T1 is then processed by the pyramid pooling module to obtain a fused feature map T6 containing overall information. The 1×1 convolutional layer is used to change the number of channels in the feature map T6 to obtain the segmented image A. w S- The specific setup process is as follows:

[0061] The image patch X output by the preprocessing module at the current level S. w S- As input to the PSPNet network, features are extracted by the first pre-trained model, ResNet101, outputting feature map T1. The pooling layers in the Pyramid Pooling Module employ pooling operations of four different sizes (1×1, 2×2, 3×3, and 6×6) to obtain feature maps T2, T3, T4, and T5 of multiple sizes. These feature maps are then subjected to 1×1 convolutions to reduce the number of channels, resulting in feature maps T2′, T3′, T4′, and T5′. Bilinear interpolation is then used to upsample feature maps T2′, T3′, T4′, and T5′ to obtain feature maps T2″, T3″, T4″, and T5″. Finally, feature maps T1 are concatenated with T2″, T3″, T4″, and T5″ on each channel to obtain feature map T6 containing overall information. A 1×1 convolutional layer is used to change the number of channels in feature map T6 to obtain the segmented image A. w S- .

[0062] like Figure 5 As shown, the refinement module includes a pre-trained ResNet101 model, a prediction uncertainty scoring module, and a replacement module; the pre-trained ResNet101 model is used to output an initial combined segmentation map B; the prediction uncertainty scoring module calculates the input image Y respectively. w (S-1)- And the prediction uncertainty score Y of B u and B u Replace the module using Y u and B u Select image Y w (S-1)- The image Y is obtained by thinning at n positions. w S+ The detailed image is constructed as follows:

[0063] Step B1: Overlay the semantic segmentation maps of all levels preceding the current level S into an image Y. w (S-1)- and segmentation image A w S- As input to the second pre-trained model ResNet101, it outputs an initial combined segmentation map B;

[0064] Step B2: Calculate the image Y w (S-1)- And the prediction uncertainty score Y of B u and B u ;

[0065] Step B3: Use two uncertainty scores Y u and B u To select image Y w (S-1)- Thinning operations are performed at n positions to obtain the thinned image Y. w S+ .

[0066] For the input image Y of the thinning module w (S-1)- For each pixel, the prediction confidence at that location is defined as the absolute value of the difference between the highest and lowest probability values; then, an uncertainty score is calculated based on the prediction confidence score at that location (the sum of these two scores must equal 1). The uncertainty score of the initial combined segmentation map B is calculated using the same principle. The two uncertainty scores Y are used in the replacement module. u and B u To select image Y w (S-1)- The image Y is obtained by thinning at n positions. w S+ At these n positions, Y w (S-1)- The prediction is uncertain, however, the prediction of the initial combined segmentation map B is deterministic. This is achieved through the formula C = Г(Y). u ⊙(1-B u The pixel value sorting score image C is calculated using ⊙ (⊙ represents point-by-point multiplication, and Г represents the median filter function, which blurs the median and smooths the image). Then, based on the score image C, the input image Y is sorted... w (S-1)- Partial replacement yields a scale subdivision map Y. w S+ The final output of the refinement module is the scale subdivision map Y. w S+ .

[0067] like Figure 6As shown, the super-resolution module is a novel multi-level super-resolution network that can reconstruct high-resolution images with different magnification factors in a single model. By constructing a multi-level architecture that utilizes inter-level correlations, processing modules at specific levels are introduced to perform super-resolution of images at different levels. The super-resolution module includes a super-resolution preprocessing module, a residual combination module, and an upsampling module; the specific construction process is as follows:

[0068] Step C1: Extract the thinned image Y w S+ Features, output feature map T7;

[0069] Step C2: Combine several residual blocks together to share parameters at different levels. Feature map T7 is output as feature map T8 through the combined residual blocks.

[0070] In this embodiment, three parallel residual block combinations X1, X2, and X3 corresponding to different input image levels are designed. Only one of these three parallel residual block combinations works at a time, corresponding to the input image level; combinations corresponding to levels other than the selected level do not work. Following the three parallel residual block combinations X1, X2, and X3, a residual combination module consisting of eight residual blocks and a convolutional layer is connected as an independent main branch to share most parameters across different levels. Feature map T7 outputs feature map T8 via the residual combination module. At the end of this multi-level network, a parallel upsampling module corresponding to a specific level is set to process the super-resolution reconstruction of the multi-level images. After upsampling feature map T8, the upsampling module passes it through a convolutional layer to obtain the final super-resolution image Y. S .

[0071] Step C3: Upsample the feature map T8 and pass it through a convolutional layer to obtain the super-resolution image Y. S If multiple refinement processes are required, increment the number of layers S by 1 to improve the super-resolution image Y. S As input to the preprocessing module in the next processing stage;

[0072] The thinning module outputs the thinned image Y at the current level S. w S+ Then, the refined result image Y w S+ The input is fed into the final super-resolution module to obtain the final high-resolution semantic segmentation map Y. S This is the final output diagram for level S. (This model requires S levels of neural network learning, so the output diagram Y for this level is...) S This refers to the output image of the entire network; if the model requires an S+1-level neural network learning process, the output image Y at level S is...S The image will be used as the input for the next stage, S+1, and then fed into a progressive multi-level semantic segmentation model for remote sensing images aimed at machine vision.

[0073] Step 3: Set training parameters, construct the loss function, train the progressive multi-level semantic segmentation model for remote sensing images for machine vision, update the network parameters, and obtain the trained progressive multi-level semantic segmentation model for remote sensing images for machine vision; the specific training process is as follows:

[0074] S1: First, randomly initialize the parameters of the segmentation module in the model, and then set the output image block X of the preprocessing module. w S- The input is fed into the segmentation module PSPNet of a progressive multi-level semantic segmentation model for remote sensing images aimed at machine vision. This generates a semantic segmentation probability map of the remote sensing image, and then the semantic segmentation loss is calculated. The loss is backpropagated to achieve semantic segmentation of the remote sensing image. During training, the SGD optimizer is used with a momentum of 0.8 and a decay weight of 4 × 10⁻⁶. -4 The initial learning rate is 10. -3 The network was trained for 500 epochs. After training, the training weight parameters of the PSPNet network were saved to obtain the training parameters of the segmentation module.

[0075] S2: First, the segmentation module in the progressive remote sensing image multi-level semantic segmentation model for machine vision is initialized using the model parameters saved in step S1. Simultaneously, the parameters of the refinement module are randomly initialized. Next, the parameters of the refinement module are trained and optimized. This refinement module is trained using a backbone network. First, the PSPNet network is selected as the backbone network, and the training parameters of the PSPNet network saved in step S1 are used as the pre-training parameters for the refinement module. Then, the refinement module is trained based on the PSPNet backbone network, using the pre-training parameters from the backbone network for further training. During the training of the refinement module, image patches are randomly extracted from the input image, and the input image is rotated, horizontally flipped, and vertically flipped. Then, the SGD optimizer is used again, with a momentum of 0.9 and a decay weight of 5×10. -4 The initial learning rate is 10. -3 Meanwhile, the refinement module was trained for 100 epochs using cross-entropy as the loss function. Finally, the training parameters of the refinement module were obtained and saved.

[0076] S3: First, the segmentation and refinement modules in the progressive remote sensing image multi-level semantic segmentation model for machine vision are initialized using the model parameters saved in steps S1 and S2. Simultaneously, the parameters of the super-resolution module are randomized. Next, the super-resolution module is trained. The remote sensing image training set data, after preprocessing by the module, is used to obtain output image patches, which are then input into the super-resolution module. The super-resolution loss is calculated, backpropagated, and the parameters of the super-resolution module are optimized to achieve super-resolution reconstruction of low-resolution images. During the training of the super-resolution module, image patches are randomly extracted from the input image, and the input image is rotated, horizontally flipped, and vertically flipped. The Adam optimizer is used during training, with an initial learning rate of 10. -4 The training process lasted 200 epochs, with the learning rate halved every 50 epochs. Training began with the X1 block, which, after convergence, was used as a pre-trained network for other layers. During each update of the trained multi-layer network, a mini-batch was randomly selected from X1, X2, and X3 for training. Only residual block combinations corresponding to the selected layer were enabled and updated. Therefore, specific residual block combinations corresponding to layers other than the selected layer were not enabled or updated. After training, the training parameters of the super-resolution module were saved.

[0077] Step 4: Input the remote sensing image test set from Step 1 into the trained progressive multi-level semantic segmentation model for machine vision, and output the accurate segmentation results of the remote sensing images; specifically:

[0078] The remote sensing image test set from step 1 is input into the trained progressive multi-level semantic segmentation model for machine vision. The multi-level semantic segmentation model obtains the corresponding grayscale output image, which is the initial result of the model. The relationship between the RGB values ​​and grayscale values ​​for each item category is found. Based on the RGB values ​​corresponding to the grayscale values ​​in the image output by the progressive multi-level semantic segmentation model for machine vision, the semantic segmentation result image is colorized to obtain the final semantic segmentation RGB image. Then, the RGB values ​​previously assigned to each item category are found, and the item category corresponding to the remote sensing image in the test set is obtained based on the RGB values ​​in the final semantic segmentation result image.

Claims

1. A progressive multi-level semantic segmentation method for remote sensing images for machine vision, characterized in that, include: Step 1: Identify and label target objects in the original remote sensing images to create training and test sets; Step 2: Build a progressive multi-level semantic segmentation model for remote sensing images for machine vision; The segmentation model consists of several identical processing stages stacked together. Each stage contains the same preprocessing module, segmentation module, thinning module, and super-resolution module. As the image is progressively segmented from the coarsest level to the finest level, a progressively thinned segmentation result image is output. The preprocessing module is used to process the original remote sensing image X and the super-resolution image Y output from the previous level. (S-1) Remote sensing images X of different sizes are output using windowing functions and downsampling processing to achieve different levels of S. w S- Y w (S-1)- ; The segmentation module is used to segment the input remote sensing image X. w S- Feature extraction and feature fusion are performed to obtain segmented image A. w S- ; The refinement module is used to refine the super-resolution image Y output from the previous level. w (S-1)- With segmented image A w S- We obtain an initial combined segmentation map B, and calculate the super-resolution image Y output from the previous level. w (S-1)- The uncertainty score C between the segmented image B and the image Y is determined, and the image Y is evaluated based on the score C. w (S-1)- The image is thinned to obtain the thinned image Y. w S+ ; The super-resolution module is used to refine the image Y. w S+ Super-resolution processing is performed to obtain the super-resolution image Y. S Image Y S As input to the next level of preprocessing module; Step 3: Set training parameters, construct loss function, train the progressive remote sensing image multi-level semantic segmentation model for machine vision, update network parameters, and obtain the trained progressive remote sensing image multi-level semantic segmentation model for machine vision. Step 4: Input the remote sensing image test set from Step 1 into the trained progressive multi-level semantic segmentation model for machine vision remote sensing images, and output the accurate segmentation results of the remote sensing images.

2. The progressive multi-level semantic segmentation method for remote sensing images oriented towards machine vision according to claim 1, characterized in that, Step 1 includes: Step 1.1: Classify the target objects in the original remote sensing image; Step 1.2: Assign RGB values ​​to each item category; Step 1.3: Colorize the item type according to the assigned RGB values ​​to obtain the remote sensing image label set; Step 1.4: Crop the obtained original images and label sets, flip the images vertically and horizontally, and then divide them into training and test sets according to a certain ratio.

3. The progressive multi-level semantic segmentation method for remote sensing images oriented towards machine vision according to claim 2, characterized in that, The process of building the preprocessing module is as follows: Step A1: Set a group of rectangular windows W S ={w|w=(x, y, h S , w S The original remote sensing image X and the super-resolution image Y output from the previous level are used to analyze the data. (S-1) The process is divided into blocks, with each window consisting of the coordinates (x, y) of its top-left corner and its height h. S and width w S designated; Step A2: Based on the rectangular window W used to divide the image blocks S The input image is segmented into parts of size h. S ×w S Image Patch X w S Y w (S-1) ; Step A3: For image patch X w S Y w (S-1) The image patch X is obtained by performing a downsampling operation. w S- Y w (S-1)- Image block X w S- The input is given to the segmentation module in the network model, and Y... w (S-1)- The input is fed into the refinement module of the network model.

4. The progressive multi-level semantic segmentation method for remote sensing images oriented towards machine vision according to claim 1, characterized in that, The process of building the segmentation module is as follows: The image patch X output by the preprocessing module at the current level S. w S- As input to the PSPNet network, features are extracted by the first pre-trained model, ResNet101, outputting feature map T1. The pooling layers in the pyramid pooling module use four different pooling sizes—1×1, 2×2, 3×3, and 6×6—to obtain feature maps of multiple sizes, T2, T3, T4, and T5. These feature maps of different sizes are then subjected to 1×1 convolutions to reduce the number of channels and obtain the final feature map. , , , Then, bilinear interpolation is used to modify the feature map. , , , Upsampling is performed to obtain the feature map , , , and on the channel, feature map T1 is compared with... , , , The feature maps T6, containing overall information, are concatenated to obtain a feature map T6; a 1×1 convolutional layer is used to change the number of channels in feature map T6 to obtain the segmented image A. w S- .

5. A progressive multi-level semantic segmentation method for remote sensing images oriented towards machine vision, as described in claim 1, is characterized in that... The process of building the refinement module is as follows: Step B1: Overlay the semantic segmentation maps of all levels preceding the current level S into an image Y. w (S-1)- and segmentation image A w S- As input to the second pre-trained model ResNet101, it outputs an initial combined segmentation map B; Step B2: Calculate the image Y w (S-1)- And the uncertainty score Y of B u and B u ; Step B3: Use two uncertainty scores Y u and B u To select image Y w (S-1)- Thinning operations are performed at n positions to obtain the thinned image Y. w S+ .

6. A progressive multi-level semantic segmentation method for remote sensing images oriented towards machine vision, as described in claim 1, is characterized in that... The construction process of the super-resolution module is as follows: Step C1: Extract the thinned image Y w S+ Features, output feature map T7; Step C2: Combine several residual blocks together to share parameters at different levels. Feature map T7 is output as feature map T8 through the combined residual blocks. Step C3: Upsample the feature map T8 and pass it through a convolutional layer to obtain the super-resolution image Y. S Super-resolution image Y S It can be used as input to the preprocessing module in the next processing stage or as the final segmented image.

Images