Building extraction method based on multi-level perception and boundary joint constraint
Through the method of multi-level perception and boundary joint constraint, combined with feature fusion and hollow space pyramid technology, the building extraction network is optimized, which solves the problems of low building extraction accuracy and irregular boundary in remote sensing images, and achieves efficient and accurate buildings Extraction effect.
Patent Information
- Application Number
- CN202411405014.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-10
- Publication Date
- 2024-11-08
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The prior art is due to the similarity and complex morphology of the building and the environment when extracting buildings in remote sensing images, resulting in low extraction accuracy, irregular boundaries and adhesion problems. The deep learning method is not effective in low contrast areas, and requires manual labor. optimization.
Using a method based on multi-level perception and boundary joint constraints, a building target extraction network composed of a convolutional network and an encoder decoder is used to extract multi-scale features through a boundary loss function and semantics. Loss function optimizes the extraction results of building boundary.
It improves the accuracy of building extraction and boundary positioning accuracy, reduces manual intervention, reduces the amount of calculation, solves the problems of boundary irregularity and adhesion, and enhances the robustness of the extraction network.
Smart Images

Figure CN118918481A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of building detection, and in particular to a building extraction method based on multi-level perception and boundary joint constraint. Background Art
[0002] Buildings are one of the most important components of surface information. Extracting buildings using high-resolution remote sensing images is of great significance in natural resource survey and monitoring, land change survey and monitoring, and illegal occupation of cultivated land monitoring. However, buildings in remote sensing images often show similarities with the surrounding background in color, texture, and shape. In particular, the complex and diverse shapes and irregular contours of buildings make high-precision remote sensing building extraction challenging.
[0003] Due to the diversity of size, shape, and texture of buildings, traditional algorithms that manually design features can only solve certain types of problems under certain conditions and are not universal. With the development of deep learning technology, deep learning networks can greatly overcome the problems of large amounts of remote sensing image data and complex features. However, the existing network feature utilization capabilities are insufficient, small target building information is seriously lost, and due to the influence of tree shadows, accurate positioning of building boundaries is still difficult, especially in low-contrast areas. Deep learning-based methods will produce blurred probability values at the edges of buildings, resulting in irregular and sticky building boundaries. Due to these limitations, image post-processing algorithms are usually required to optimize building extraction results, but they require manual design, and additional algorithms also increase the amount of calculation. Therefore, an efficient and accurate building extraction method is needed to improve the efficiency and accuracy of building extraction in remote sensing images. Summary of the invention
[0004] In order to solve the technical problems existing in the background technology, the present invention proposes a building extraction method based on multi-level perception and boundary joint constraint.
[0005] The building extraction method based on multi-level perception and boundary joint constraint proposed in the present invention comprises the following steps: S1, obtaining a remote sensing image of a building to be detected, and preprocessing the remote sensing image to obtain a first type of remote sensing image; S2, inputting the first type of remote sensing image into the convolutional network to obtain a first feature map; S3, performing feature encoding on the first feature map by using the trained encoder to obtain feature maps of multiple different levels; S4, inputting the feature maps of multiple different levels into the decoder for resolution recovery to obtain a second feature map; S5. Processing the second feature map according to a preset strategy to obtain a building extraction result corresponding to the remote sensing image.
[0006] Preferably, the convolutional network, encoder, and decoder constitute a building target extraction network; the encoder adopts a multi-level perception network, combined with feature fusion and void space pyramid technology, to extract multi-scale building features to obtain feature maps of multiple different levels; the decoder is a feature pyramid network, which processes feature maps of multiple different levels from the encoder, gradually restores building features, and fuses deep high-level semantic information with shallow detail information in the feature maps of different levels output from the encoder in turn, gradually restores building features from deep to shallow, performs resolution recovery, and finally generates a second feature map.
[0007] Preferably, the preset strategy specifically includes: Generate the probability of each pixel of the second feature map belonging to the background and building categories through the softmax function; Compare the background probability and building probability corresponding to each pixel, and take the category with the larger probability value as the category corresponding to the pixel; The second feature map is labeled pixel by pixel with a value of 0 representing the background category and 1 representing the building, so as to obtain a binary image corresponding to the second feature map; Based on the binary image, a building extraction result corresponding to the remote sensing image is obtained.
[0008] Preferably, the building target extraction network training process is as follows: Construct a remote sensing image building extraction dataset, including a training set and a validation set, where both the training set and the validation set contain building sample images, building labels, and building boundary labels; Obtain building sample images, building labels and building boundary labels in a training set, input the building sample images into a preset building target extraction network, use the building labels and building boundary labels as supervision information of the building target extraction network and output the building extraction results; Iteratively training the building object extraction network with the goal of minimizing the loss values of the boundary loss function and the semantic loss function; When the loss values of the boundary loss function and the semantic loss function are minimized, the current network parameters of the building target extraction network are saved to obtain a preliminarily trained building target extraction network; The validation set is used to verify whether the initially trained building target extraction network is overfitting, and finally the trained building target extraction network is obtained.
[0009] Preferably, the boundary loss function is a combination of the Focal loss function and the Dice loss function; the boundary loss function formula is as follows: in, To predict the detail feature map; It is the real boundary information of the surface; is the Focal loss function; is the Dice loss function; The Focal loss function calculation formula is as follows: in, is a balancing factor used to adjust the weights of positive and negative samples; is an adjustment factor used to adjust the weight of difficult and easy samples; The Dice loss function calculation formula is as follows:
[0010] in, Indicates the feature map pixels; and Represent the length and width of the feature map respectively; Represents the smoothing coefficient.
[0011] Preferably, the semantic loss function adopts a cross entropy loss function, and the calculation formula is as follows:
[0012] in, The probability that the sample predicted by the model belongs to a building; is the true label, with a value of 0 or 1, 0 represents background and 1 represents building.
[0013] Preferably, the remote sensing image of the building to be detected is a multi-band image including three bands of RGB; step S1 specifically comprises: Acquire a remote sensing image of the building to be detected, and crop the remote sensing image into a 512×512 RGB three-band image according to a preset overlap ratio; The cropped RGB three-band image is normalized to obtain the first type of remote sensing image.
[0014] Preferably, step S2 specifically includes: The first type of remote sensing image is input into the convolutional network, and a convolution calculation is performed after a 3×3 convolution with a step size of 2 to obtain a feature map of size 512×512×16, and this feature map is used as the first feature map.
[0015] In the present invention, the proposed building extraction method based on multi-level perception and boundary joint constraint obtains the remote sensing image of the building to be detected, and pre-processes the remote sensing image to obtain a first type of remote sensing image; inputs the first type of remote sensing image into the convolutional network to obtain a first feature map; performs feature encoding on the first feature map through the trained encoder to obtain multiple feature maps at different levels; inputs the multiple feature maps at different levels into the decoder for resolution recovery to obtain a second feature map; processes the second feature map according to a preset strategy to obtain the building extraction result corresponding to the remote sensing image. The present invention constructs a boundary optimization module to suppress the building boundary during training, effectively optimizes the building boundary extraction result, and does not increase the additional parameter amount when predicting the building. The present invention uses a multi-level perception network as the feature extraction network of the building encoder stage, while taking into account the extraction of features of different scales, retaining shallow feature information while obtaining deep features. To a certain extent, the problem of irregularity and adhesion of the boundaries of the building extraction results is solved, and the building extraction accuracy is improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Figure 1 This is a schematic diagram of the workflow structure of the building extraction method based on multi-level perception and boundary joint constraint proposed in the present invention; Figure 2 A schematic diagram of a building extraction network training structure of a building extraction method based on multi-level perception and boundary joint constraint proposed in the present invention; Figure 3 A schematic diagram of the structure of densely connected blocks of the building extraction method based on multi-level perception and boundary joint constraint proposed in the present invention; Figure 4 This is a schematic diagram of the structure of the first feature fusion mode of the building extraction method based on multi-level perception and boundary joint constraint proposed by the present invention; Figure 5 This is a schematic diagram of the structure of the second feature fusion mode of the building extraction method based on multi-level perception and boundary joint constraint proposed by the present invention; Figure 6 This is a schematic diagram of the structure of the third feature fusion mode of the building extraction method based on multi-level perception and boundary joint constraint proposed by the present invention; Figure 7 A schematic diagram of the structure of a hollow space pyramid of a building extraction method based on multi-level perception and boundary joint constraint proposed in the present invention; Figure 8 This is a schematic diagram of the feature pyramid network structure of the building extraction method based on multi-level perception and boundary joint constraint proposed in the present invention; Fig. 9Schematic diagram of the boundary optimization network structure of the building extraction method based on multi-level perception and boundary joint constraints proposed in the present invention. DETAILED DESCRIPTION
[0017] Reference Figure 1-9 The building extraction method based on multi-level perception and boundary joint constraint proposed in the present invention comprises the following steps: S1. Acquire a remote sensing image of a building to be detected, and preprocess the remote sensing image to obtain a first type of remote sensing image.
[0018] In this embodiment, the remote sensing image of the building to be detected is a multi-band image including three RGB bands; step S1 specifically includes: obtaining the remote sensing image of the building to be detected, and cropping the remote sensing image into a 512×512 RGB three-band image according to a preset overlap rate; and normalizing the cropped RGB three-band image to obtain a first type of remote sensing image.
[0019] In this embodiment, the remote sensing image of the building to be detected is a multi-band image including three RGB bands. First, the remote sensing image is cropped into a 512×512 RGB three-band image according to a 50% overlap rate.
[0020] In this embodiment, the cropped RGB three-band image is normalized to adjust the image pixel values to the same distribution as that during model training to ensure that the model can accurately perform inference in the prediction stage. The normalization formula is as follows:
[0021] in, is the pixel value of each band; is the normalized mean, and the RGB bands are taken as (0.485, 0.456, 0.406) respectively. is the normalized standard deviation, and the RGB bands are taken as (0.229, 0.224, 0.225) respectively.
[0022] S2. Inputting the first type of remote sensing image into the convolutional network to obtain a first feature map.
[0023] S3. Perform feature encoding on the first feature map through the trained encoder to obtain feature maps of multiple different levels.
[0024] In this embodiment, the encoder uses a multi-level perception network, combined with feature fusion and hollow space pyramid technology, to extract multi-scale building features. The multi-level perception network adopts a multi-branch parallel structure, and each branch network uses densely connected blocks to enhance the information transfer between convolutional layers. The feature fusion module is used to fuse the feature information of the multi-branch network, effectively reducing the loss of spatial information and enhancing the feature learning ability of the network. In the last branch network, the hollow space pyramid structure is used to obtain deep multi-scale feature information, further enhancing the multi-level perception network's ability to extract information on objects of different scales; Reference Figure 2 As shown in the figure, the multi-level perception network first uses a 3×3 convolutional layer with a step size of 2 to perform downsampling operations to reduce the image size, and then extracts features at different levels through four branch networks. The four branch networks are the first, second, third, and fourth branches from top to bottom. Each branch network obtains feature information through densely connected blocks, and uses the feature fusion module to realize the multi-level feature combination between different branch networks. In addition, the network adds a hollow space pyramid structure after the fourth branch.
[0025] In this embodiment, the first branch network passes through 2 densely connected blocks, 1 feature fusion module, 1 densely connected block, 1 feature fusion module, 1 densely connected block, and 1 feature fusion module in sequence to obtain a feature map of size 256×256×256. The input of each feature fusion module is the output of the previous densely connected block and the output of the second branch network densely connected block after upsampling. The second branch network receives the output of the first densely connected block of the first branch network after downsampling, passes through 1 densely connected block and 1 feature fusion module for 3 consecutive times, and obtains a feature map of size 128×128×384. The input of each feature fusion module is the output of the previous densely connected block, the output of the first branch network densely connected block after downsampling, and the output of the third branch network densely connected block after upsampling. The third branch network receives the down-sampled output of the first densely connected block of the second branch network, passes through one densely connected block and one feature fusion module for two consecutive times, and obtains a feature map of size 64×64×512. The input of each feature fusion module is the output of the previous densely connected block, the down-sampled output of the densely connected block of the second branch network, and the up-sampled output of the densely connected block of the fourth branch network. The fourth branch network receives the down-sampled output of the first densely connected block of the third branch network, passes through two densely connected blocks and a dilated spatial pyramid structure in sequence, and obtains a feature map of size 32×32×256.
[0026] Reference Figure 3As shown in the figure, the dense connection block directly passes the feature map of the previous layer of the network as input to all subsequent layers, that is, there is a direct connection between any two layers, so as to achieve the effect of feature reuse. For a network with L layers, there are L(L+1) / 2 connections in total, and the formula for dense connection is as follows:
[0027] in express All feature maps before the layer, It is a nonlinear combination function, including batch normalization layer, nonlinear activation layer, convolution layer and dropout layer.
[0028] Reference Figure 4 , 5 As shown in Figures 6 and 7, the feature fusion module enhances the combination and utilization of shallow and deep features in the branch network through three different feature fusion modes. The first is that the shallow feature map remains unchanged, the deep feature map is successively subjected to 1×1 convolution and upsampling, and then the shallow and deep feature maps are channel-superimposed, and finally the Leaky ReLU activation function is used to output the fused feature map; the second is that the deep feature map remains unchanged, the shallow feature map is successively subjected to 1×1 convolution and downsampling, and then the shallow and deep feature maps are channel-superimposed, and finally the Leaky ReLU activation function is used to output the fused feature map; the third is that the middle-layer feature map remains unchanged, the shallow feature map is successively subjected to 1×1 convolution and downsampling to become the same size as the middle-layer feature map, the deep feature map is successively subjected to 1×1 convolution and upsampling to become the same size as the middle-layer feature map, and then the deep, middle-layer, and shallow feature maps are channel-superimposed, and the Leaky ReLU activation function is used to output the fused feature map.
[0029] Reference Figure 7 As shown in the figure, the atrous spatial pyramid structure captures multi-scale information by connecting atrous convolutions with dilation rates of 12, 24, and 36 in parallel. Each scale is used as an independent branch, and finally the features of each scale are merged at the end of the network to obtain a larger receptive field.
[0030] S4. Inputting multiple feature maps at different levels into a decoder for resolution recovery to obtain a second feature map.
[0031] In this embodiment, refer to Figure 8 As shown in the figure, the decoder uses a feature pyramid module (FPN) to process multi-scale information from the encoder, gradually restores building features, and through bottom-up paths and lateral connections, the feature maps of different levels output from the encoder are sequentially fused with deep high-level semantic information and shallow detail information, gradually restore building features from deep to shallow layers, and finally generate high-resolution feature maps.
[0032] S5. Process the second feature map according to a preset strategy to obtain a building extraction result corresponding to the remote sensing image.
[0033] In this embodiment, the preset strategy specifically includes: generating the probability of each pixel of the second feature map belonging to the background and building categories through the softmax function; comparing the background probability and the building probability corresponding to each pixel, and taking the category with a larger probability value as the category corresponding to the pixel; marking the second feature map pixel by pixel with a value of 0 representing the background category and 1 representing the building to obtain a binary image corresponding to the second feature map; based on the binary image, obtaining the building extraction result corresponding to the remote sensing image.
[0034] In this embodiment, a convolutional network, an encoder, and a decoder constitute a building target extraction network; the encoder adopts a multi-level perception network, combined with feature fusion and void space pyramid technology, to extract multi-scale building features to obtain feature maps of multiple different levels; the decoder is a feature pyramid network, which processes feature maps of multiple different levels from the encoder, gradually restores the building features, and fuses the deep high-level semantic information with the shallow detail information of the feature maps of different levels output from the encoder in turn, gradually restores the building features from deep to shallow, performs resolution recovery, and finally generates a second feature map.
[0035] Specifically, Figure 2 As shown in the figure, the training process of the building target extraction network is as follows: construct a remote sensing image building extraction data set, including a training set and a validation set, wherein both the training set and the validation set contain building sample images, building labels, and building boundary labels; obtain the building sample images, building labels, and building boundary labels in the training set, input the building sample images into a preset building target extraction network, use the building labels and building boundary labels as the supervision information of the building target extraction network, and use the building extraction results as the output; iteratively train the building target extraction network with the goal of minimizing the loss values of the boundary loss function and the semantic loss function; when the loss values of the boundary loss function and the semantic loss function are minimized, save the current network parameters of the building target extraction network to obtain a preliminarily trained building target extraction network; use the validation set to verify whether the preliminarily trained building target extraction network is overfitting, and finally obtain a trained building target extraction network.
[0036] In this embodiment, a boundary optimization module is introduced in the process of model training. During training, the results processed by the encoder are input into the decoder and the boundary optimization module respectively to calculate the boundary loss and semantic loss respectively.
[0037] The boundary optimization module introduces a building boundary alignment mechanism as an auxiliary task for building extraction, constructs a boundary optimization network to extract building boundary features, and designs a boundary loss function to enhance and optimize the extraction of building edge details, and uses building boundaries to constrain the extraction of building semantic information.
[0038] Reference Fig. 9 As shown in the figure, the boundary optimization network extracts the building edge information through the gated convolution layer (GCL), and the output result is calculated with the real boundary information to optimize the extraction of the building boundary. The gated convolution can filter out other information in the network and only retain the feature information related to the shape. It stops the activation of color and texture related features, so that the network can focus more on the extraction of shape features. The boundary optimization network processes the multi-scale information from the encoder and extracts the boundary features of the building through the gated convolution in sequence through the bottom-up path. The output of the third branch of the encoder is reduced by 1×1 convolution to reduce the number of channels, and then passes through the first gated convolution in sequence. It is concat spliced with the output of the second branch of the encoder after 1×1 convolution, and then passes through the second gated convolution. It is concat spliced with the output of the first branch of the encoder after 1×1 convolution, and outputs 256×256×128 building boundary features. Finally, it is restored to the size of the input image through a 3×3 deconvolution operation.
[0039] In this embodiment, during training, the boundary loss function is used to calculate the loss between the building boundary extraction result and the building boundary true value, and the semantic loss function is used to calculate the loss between the building extraction result and the building true value. Then, the parameters of the building target extraction network are updated according to the loss value, and the verification set is used to verify whether the building target extraction network is overfitting, and finally a trained building target extraction network is obtained.
[0040] In this embodiment, the sample image of the building is a 3-band RGB image with a size of 512×512, the semantic information of the building is a binary image, in which the pixel value 0 represents the background and the pixel value 1 represents the building, and the size is 512×512, and the semantic information of the building boundary is a binary image, in which the pixel value 0 represents the background and the pixel value 1 represents the building boundary, and the size is 512×512.
[0041] In this embodiment, the boundary loss function is a combination of the Focal loss function and the Dice loss function; the boundary loss function formula is as follows:
[0042] in, To predict the detail feature map; It is the real boundary information of the surface; is the Focal loss function; is the Dice loss function; The calculation formula of Focal loss function is as follows:
[0043] in, is a balancing factor used to adjust the weights of positive and negative samples; is an adjustment factor used to adjust the weight of difficult and easy samples; The Dice loss function calculation formula is as follows:
[0044] in, Indicates the feature map pixels; and Represent the length and width of the feature map respectively; Represents the smoothing coefficient.
[0045] In this embodiment, the boundary loss function uses the Focal loss function and the Dice loss function to optimize edge feature supervision. The Focal loss function is used to solve the problem of imbalance between building edge and background categories. The loss of background samples is reduced, the loss of building edge samples is increased, and the building edge samples are focused. The Dice loss function is used to calculate the similarity between the building segmentation result and the label. The best segmentation result is 1 and the worst is 0.
[0046] In this embodiment, the building target extraction network is trained. During training, the boundary loss function is used to calculate the loss between the building boundary extraction result and the building boundary true value, and the semantic loss function is used to calculate the loss between the building extraction result and the building true value. Then, the parameters of the building target extraction network are updated according to the loss value, and the verification set is used to verify whether the building target extraction network is overfitting.
[0047] In this embodiment, the semantic loss function adopts the cross entropy loss function, and the calculation formula is as follows:
[0048] in, The probability that the sample predicted by the model belongs to a building; is the true label, with a value of 0 or 1, 0 represents background and 1 represents building.
[0049] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.
Claims
1. A building extraction method based on multi-level perception and boundary joint constraints, characterized in that: The following steps are involved: S1, obtaining a remote sensing image of a building to be detected, and preprocessing the remote sensing image to obtain a first type of remote sensing image; S2, inputting the first type of remote sensing image into the convolutional network to obtain a first feature map; S3, performing feature encoding on the first feature map by using the trained encoder to obtain feature maps of multiple different levels; S4, inputting the feature maps of multiple different levels into the decoder for resolution recovery to obtain a second feature map; S5. Processing the second feature map according to a preset strategy to obtain a building extraction result corresponding to the remote sensing image.
2. The building extraction method based on multi-level perception and boundary joint constraint according to claim 1 is characterized in that: The convolutional network, encoder, and decoder constitute a building target extraction network; the encoder adopts a multi-level perception network, combined with feature fusion and void space pyramid technology, to extract multi-scale building features to obtain feature maps of multiple different levels; the decoder is a feature pyramid network, which processes feature maps of multiple different levels from the encoder, gradually restores building features, and fuses deep high-level semantic information with shallow detail information in the feature maps of different levels output from the encoder in turn, gradually restores building features from deep to shallow, performs resolution recovery, and finally generates a second feature map.
3. The building extraction method based on multi-level perception and boundary joint constraint according to claim 1 is characterized in that: The preset strategy specifically includes: Generate the probability of each pixel of the second feature map belonging to the background and building categories through the softmax function; Compare the background probability and building probability corresponding to each pixel, and take the category with the larger probability value as the category corresponding to the pixel; The second feature map is labeled pixel by pixel with a value of 0 representing the background category and 1 representing the building, so as to obtain a binary image corresponding to the second feature map; Based on the binary image, a building extraction result corresponding to the remote sensing image is obtained.
4. The building extraction method based on multi-level perception and boundary joint constraint according to claim 2 is characterized in that: The building target extraction network training process is as follows: Construct a remote sensing image building extraction dataset, including a training set and a validation set, where both the training set and the validation set contain building sample images, building labels, and building boundary labels; Obtain building sample images, building labels and building boundary labels in a training set, input the building sample images into a preset building target extraction network, use the building labels and building boundary labels as supervision information of the building target extraction network and output the building extraction results; Iteratively training the building object extraction network with the goal of minimizing the loss values of the boundary loss function and the semantic loss function; When the loss values of the boundary loss function and the semantic loss function are minimized, the current network parameters of the building target extraction network are saved to obtain a preliminarily trained building target extraction network; The validation set is used to verify whether the initially trained building target extraction network is overfitting, and finally the trained building target extraction network is obtained.
5. The building extraction method based on multi-level perception and boundary joint constraint according to claim 4 is characterized in that: The boundary loss function is a combination of the Focal loss function and the Dice loss function; the boundary loss function formula is as follows: in To predict the detail feature map; It is the real boundary information of the surface; is the Focal loss function; is the Dice loss function; The Focal loss function calculation formula is as follows: in, is a balancing factor used to adjust the weights of positive and negative samples; is an adjustment factor used to adjust the weight of difficult and easy samples; The Dice loss function calculation formula is as follows: in, Indicates the feature map pixels; and Represent the length and width of the feature map respectively; Represents the smoothing coefficient.
6. The building extraction method based on multi-level perception and boundary joint constraint according to claim 4 is characterized in that: The semantic loss function adopts the cross entropy loss function, and the calculation formula is as follows: in, The probability that the sample predicted by the model belongs to a building; is the true label, with a value of 0 or 1, 0 represents background and 1 represents building.
7. The building extraction method based on multi-level perception and boundary joint constraint according to claim 1 is characterized in that: The remote sensing image of the building to be detected is a multi-band image including three bands of RGB; step S1 specifically includes: Acquire a remote sensing image of the building to be detected, and crop the remote sensing image into a 512×512 RGB three-band image according to a preset overlap ratio; The cropped RGB three-band image is normalized to obtain the first type of remote sensing image.
8. The building extraction method based on multi-level perception and boundary joint constraint according to claim 1 is characterized in that: Step S2 specifically includes: The first type of remote sensing image is input into the convolutional network, and a convolution calculation is performed after a 3×3 convolution with a step size of 2 to obtain a feature map of size 512×512×16, and this feature map is used as the first feature map.
Citation Information
Patent Citations
Remote sensing image building extraction method and system based on U-Net network and electronic equipment
CN111460936A
Medical image automatic segmentation method based on deep learning
CN113516659A
Remote sensing image building extraction method based on attention mechanism and boundary loss
CN114387521A
Building extraction method based on point-line-plane multi-task remote sensing image
CN116682004A
Cited By
High-resolution remote sensing image small building extraction method and system fusing local feature perception and boundary constraint
CN118823581A
High-resolution remote sensing image small building extraction method and system fusing local feature perception and boundary constraint
CN118823581B
