An extraction method for a building extraction network based on high-resolution remote sensing images and a bias cosine network

Abstract

The present invention belongs to the technical field of optical remote sensing image processing, and particularly relates to an extraction method of a building extraction network based on high-resolution remote sensing images and a bias cosine network. Step 1: Collect high-resolution optical remote sensing images; Step 2: Extract the buildings in the remote sensing images collected in Step 1 and make them into a data set; Step 3: Build a bias cosine network; Step 4: Use the building data set in Step 2 to train the bias cosine network; Step 5: Verify the network trained in Step 4; Step 6: Use the network verified in Step 5 to perform semantic segmentation and identify the building outlines. The present invention is used to solve the problems of the utilization of transformation features between different scales by existing traditional semantic segmentation methods and the limitations in the extraction of buildings of different sizes.

Description

Technical Field

[0001] The present invention belongs to the technical field of optical remote sensing image processing, and particularly relates to an extraction method of a building extraction network based on high-resolution remote sensing images and a bias cosine network. Background Art

[0002] Extracting buildings from remote sensing images is a field of urban planning and urban information research. With the rapid development of the urbanization process, obtaining urban building information is of great significance for the formulation of urban planning and development strategies. At the present stage, the field of building extraction usually includes two parts: feature extraction and modeling. Among them, feature extraction is to select effective features from the input image to form a feature space; and the modeling process is to establish a mapping model between the image and the segmentation target based on the feature space through artificial or machine learning methods. Therefore, how to construct the feature space of the segmentation model and then adopt an effective method to establish the segmentation model is the key to achieving high-precision segmentation of remote sensing image targets. Traditional remote sensing image segmentation methods are based on image features such as gray values, gradients, and texture information, and use artificial or machine learning to establish a mapping model from the feature space to the segmentation target. This method is completed manually from feature extraction to model establishment. For this reason, machine learning methods such as neural networks and support vector machines are used to establish a mapping model between the feature space and the segmentation target. This type of method is based on manually extracted features and automatically establishes a mapping relationship between the features and the segmentation target through sample training. Due to the complex types and distribution states of ground objects, the true mapping model from the image to the segmentation target often exhibits high-dimensionality, strong coupling, and strong nonlinearity. If the dimension of the feature space is low, the characteristics of the targets in high-resolution remote sensing images cannot be effectively described. In recent years, deep neural networks have achieved fruitful results in the field of image segmentation because they can automatically extract features and establish a high-dimensional strong nonlinear model from input to output through the training of a large number of samples, realizing end-to-end classification or prediction. In 2015, Long Jonathan first proposed the fully convolutional neural network image semantic segmentation architecture, which uses VGG as the basic network and realizes the semantic segmentation of images through FCN (Fully Convolutional Netwoks) and deconvolution. In addition to VGG, common basic networks also include the ResNet network, the Inception network, etc. At the same time, the frameworks of semantic segmentation also include the U-net structure, the Deeplab structure, and the Hrnet structure, etc.

[0003] However, although the above methods make certain use of features at high-resolution scales, they have limited utilization of transformation features between different scales and limited adaptability to buildings of different sizes. Summary of the Invention

[0004] The present invention provides an extraction method for a building extraction network based on high-resolution remote sensing images, aiming to solve the problems of the utilization of transformation features between different scales by existing traditional semantic segmentation methods and the limitations in the extraction of buildings of different sizes.

[0005] The present invention provides a bias cosine network based on high-resolution remote sensing images, aiming to solve the problem of limited segmentation accuracy of buildings in existing high-resolution remote sensing images.

[0006] The present invention is realized through the following technical solutions:

[0007] An extraction method for a building extraction network based on high-resolution remote sensing images, the building network extraction method comprising the following steps:

[0008] Step 1: Collect high-resolution optical remote sensing images;

[0009] Step 2: Extract the buildings in the remote sensing images collected in Step 1 and make them into a data set;

[0010] Step 3: Build a bias cosine network;

[0011] Step 4: Use the building data set in Step 2 to train the bias cosine network;

[0012] Step 5: Verify the network trained in Step 4;

[0013] Step 6: Use the network verified in Step 5 to perform semantic segmentation and identify the building outlines.

[0014] An extraction method for a building extraction network based on high-resolution remote sensing images, specifically in Step 1, the optical remote sensing images are collected from remote sensing data with a resolution better than one meter in the Jilin-1 satellite constellation.

[0015] An extraction method for a building extraction network based on high-resolution remote sensing images, specifically in Step 2,

[0016] Step 2.1: Select the remote sensing images for training the building extraction model;

[0017] Step 2.2: Crop the original remote sensing images for making the data set based on the remote sensing images in Step 2.1; set the cropping size to 512×512 pixels, and for the parts with a size less than 512×512 pixels after cutting, fill them with 0 values;

[0018] Step 2.3: Divide the dataset in Step 2.2 into a training set, a validation set, and a test set; divide them according to the ratio of 7:2:1 for the training set, validation set, and test set.

[0019] An extraction method of a building extraction network based on high-resolution remote sensing images. The specific construction of the bias cosine network in Step 3 is as follows.

[0020] Step 3.1: Perform downsampling convolutions on the input image at three scales.

[0021] Pass each sized image through a batch normalization layer BN, an activation layer ReLU, and a max pooling layer MP in sequence; thus forming a complete downsampling process.

[0022] Step 3.2: Perform strided convolutions and transposed convolutions for different scales to achieve cyclic upsampling and downsampling.

[0023] All downsamplings are implemented through ConvD convolution operations. The convolution kernel size of ConvD is 3×3, the convolution stride is 2, and the number of convolution kernels is 64.

[0024] All upsamplings are implemented through ConvU transposed convolution operations. The convolution kernel size of ConvU is 3×3, the convolution stride is 2, and the number of convolution kernels is 64.

[0025] An extraction method of a building extraction network based on high-resolution remote sensing images. Step 3.2 also includes adding a batch normalization layer BN, an activation layer ReLU, and a max pooling layer MP respectively after each convolution or transposed convolution in the cyclic upsampling and downsampling.

[0026] An extraction method of a building extraction network based on high-resolution remote sensing images. Accumulate the cyclic upsampling and downsampling networks at three scales according to the cyclic depth and corresponding feature scales, and then merge the features of different stages.

[0027] An extraction method of a building extraction network based on high-resolution remote sensing images. Perform upsampling on the ends of the cyclic upsampling and downsampling at different scales through a transposed convolution operation with a convolution kernel size of 3×3, a convolution stride of m, and a number of convolution kernels of 64.

[0028] An extraction method of a building extraction network based on high-resolution remote sensing images. Perform an accumulation operation on the features of different scales and add a batch binary classification convolution operation with a convolution kernel size of 3×3, a convolution stride of 1, and a number of convolution kernels of 2.

[0029] Finally, add a batch normalization layer BN, an activation layer ReLU, and a max pooling layer MP, a prediction result normalization layer Softmax, and a binary classification output layer PixelClassification to the network.

[0030] An extraction method for a building extraction network based on high-resolution remote sensing images. The process of using the offset cosine network to extract buildings includes the following steps:

[0031] Step 6.1: The program will run on a machine with an Intel(R) Xeon(R) Gold 6254 CPU @ 3.10 GHz for the CPU, four NVIDIA GeForce RTX 3090 for the GPU, 132 GB of memory, and Ubuntu 18.04 as the operating system, using MCR2021b;

[0032] Step 6.2: The hyperparameters during dataset training are as follows: the initial learning rate learning_rate = 0.001, the batch size batch_size = 32, the number of training epochs epochs = 100, and the number of segmentation classes n_classes = 2;

[0033] Step 6.3: Input the cropped image into the trained network model to extract the buildings in the image.

[0034] An offset cosine network based on high-resolution remote sensing images, where the offset cosine network includes the first set of downsampling units Ⅰ, Ⅲ, Ⅳ, Ⅴ, Ⅵ; the second set of downsampling units Ⅰ, Ⅱ, Ⅲ, Ⅳ; the third set of downsampling units Ⅰ, Ⅱ, Ⅲ, Ⅳ; the fourth set of downsampling units Ⅰ, Ⅱ, Ⅲ, and Ⅳ;

[0035] The input image is respectively input into the first set of downsampling unit Ⅰ, the second set of downsampling unit Ⅰ, and the third set of downsampling unit Ⅰ after convolution. The first set of downsampling unit Ⅰ transmits the convolved image to the second set of downsampling unit Ⅱ.

[0036] The second set of downsampling unit Ⅰ transmits the convolved image to the third set of downsampling unit Ⅱ.

[0037] The third set of downsampling unit Ⅰ transmits the deconvolved image to the second set of downsampling unit Ⅱ.

[0038] The second set of downsampling unit Ⅱ transmits the deconvolved image to the first set of downsampling unit Ⅲ.

[0039] The second set of downsampling unit Ⅱ transmits the convolved image to the third set of downsampling unit Ⅲ.

[0040] The third set of downsampling unit Ⅱ transmits the deconvolved image to the second set of downsampling unit Ⅲ.

[0041] The first set of downsampling units Ⅲ transmits the convolved image to the second set of downsampling units Ⅳ.

[0042] The third set of downsampling units Ⅲ transmits the deconvolved image to the second set of downsampling units Ⅳ.

[0043] The second set of downsampling units Ⅲ transmits the deconvolved image to the first set of downsampling units Ⅳ.

[0044] The first set of downsampling units Ⅳ transmits the convolved image to the second set of downsampling units Ⅴ.

[0045] The second set of downsampling units Ⅳ transmits the deconvolved image to the first set of downsampling units Ⅴ.

[0046] The second set of downsampling units Ⅳ transmits the convolved image to the third set of downsampling units Ⅴ.

[0047] The first set of downsampling units Ⅴ transmits the convolved image to the first set of downsampling units Ⅵ.

[0048] Both the second set of downsampling units Ⅴ and the third set of downsampling units Ⅴ transmit the deconvolved image to the first set of downsampling units Ⅵ.

[0049] The first set of downsampling units Ⅵ outputs the image after normalization.

[0050] The beneficial effects of the present invention are as follows:

[0051] The offset cosine network of the present invention is more suitable for building extraction in remote sensing images.

[0052] The present invention realizes more accurate building extraction.

[0053] The research on such problems through deep learning is mainly completed by extracting features at different scales, and its effect has limitations, bringing great pressure to the post-processing of semantic segmentation. Therefore, the present invention innovatively proposes an offset cosine network, which can solve the above problems from two aspects. One is to take the features generated when the input image changes between different scales as the main body, and the other is to ensure that different-sized segmentation targets have the same scale features by breaking through the upsampling limit.

[0054] The present invention proves through experiments that the offset cosine network can realize the semantic segmentation task, and has consistent segmentation ability for the same type of targets with different sizes, and the recognition contour is more accurate, which is convenient for extracting the accurate contour of the target through post-processing. At the same time, the offset cosine network takes the features generated when changing between different scales as the main body, so it can also be extended to the fields of image classification and target recognition with different target sizes. Description of the Drawings

[0055] Figure 1It is a structural schematic diagram of the present invention.

[0056] Figure 2 It is the initial downsampling network structure diagram of the present invention.

[0057] Figure 3 It is the network structure diagram of cyclic upsampling and downsampling at each scale of the present invention.

[0058] Figure 4 It is the complete network structure diagram of cyclic upsampling and downsampling at each scale of the present invention.

[0059] Figure 5 It is the combined network structure diagram of cyclic upsampling and downsampling at each scale of the present invention.

[0060] Figure 6 It is the network structure diagram of offset cosine of the present invention. Detailed implementation manners

[0061] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0062] An extraction method of a building extraction network based on high-resolution remote sensing images, the building network extraction method includes the following steps:

[0063] Step 1: Collect high-resolution optical remote sensing images;

[0064] Step 2: Extract the buildings in the remote sensing images collected in Step 1 and make them into a data set;

[0065] Step 3: Build an offset cosine network;

[0066] Step 4: Use the building data set in Step 2 to train the offset cosine network;

[0067] Step 5: Verify the network trained in Step 4;

[0068] Step 6: Use the network verified in Step 5 to perform semantic segmentation and identify the building contours.

[0069] An extraction method of a building extraction network based on high-resolution remote sensing images, specifically for Step 1, the optical remote sensing images are collected from remote sensing data with a resolution better than one meter in the Jilin-1 satellite constellation. The shooting time is January and September.

[0070] An extraction method for a building extraction network based on high-resolution remote sensing images. Specifically, step 2 is as follows:

[0071] Step 2.1: Select the remote sensing images for training the building extraction model;

[0072] Step 2.2: Crop the original remote sensing images used for making the dataset based on the remote sensing images in step 2.1; Set the cropping size to 512×512 pixels. For the parts with sizes less than 512×512 pixels after cutting, fill them with 0 values;

[0073] Step 2.3: Divide the dataset in step 2.2 into a training set, a validation set, and a test set; Divide them according to the ratio of 7:2:1 for the training set, validation set, and test set.

[0074] The total number of cropped images is 97,565. Finally, the training set contains 68,296 cropped images, the validation set contains 19,513 cropped images, and the test set contains 9,756 cropped images.

[0075] An extraction method for a building extraction network based on high-resolution remote sensing images. Specifically, building the bias cosine network in step 3 is as follows:

[0076] Step 3.1: Perform downsampling convolutions on the input image at three scales; The convolution kernel size is 3×3, the convolution stride is n, where n = 1, 2, 4; The number of convolution kernels is 64;

[0077] The image sizes corresponding to the three scales are 512×512 (first level), 256×256 (second level), and 128×128 (third level);

[0078] Pass each size image through the batch normalization layer BN, the activation layer ReLU, and the max pooling layer MP in sequence; Thus, a complete downsampling process is formed, and the network structure is as Figure 2 shown;

[0079] Step 3.2: Perform strided convolution and transposed convolution for different scales to achieve cyclic upsampling and downsampling. Set the cycle depth to 5. Then the order of the intermediate layer sizes for the cyclic upsampling and downsampling at the first level is 512×512, 256×256, 128×128, 256×256, 512×512; The order of the intermediate layer sizes for the cyclic upsampling and downsampling at the second level is 256×256, 128×128, 256×256, 512×512, 256×256; The order of the intermediate layer sizes for the cyclic upsampling and downsampling at the first level is 128×128, 256×256, 512×512, 256×256, 128×128;

[0080] The downsampling is all implemented through ConvD convolution operation. The convolution kernel size of ConvD is 3×3, the convolution stride is 2, and the number of convolution kernels is 64;

[0081] The upsampling is all implemented through ConvU transposed convolution operation. The convolution kernel size of ConvU is 3×3, the convolution stride is 2, and the number of convolution kernels is 64. The network structure is as Figure 3 shown.

[0082] Steps 3.1 - 3.2 are in series.

[0083] An extraction method of a building extraction network based on high - resolution remote sensing images. Step 3.2 also includes adding a batch normalization layer BN, an activation layer ReLU, and a max - pooling layer MP respectively after each convolution or transposed convolution in the cyclic up - and - downsampling. The network structure is as Figure 4 shown.

[0084] An extraction method of a building extraction network based on high - resolution remote sensing images. The cyclic up - and - downsampling networks at three scales are subjected to an accumulation operation according to the cyclic depth and the corresponding feature scales, thereby merging the features at different stages. The merged network structure is as Figure 5 shown.

[0085] An extraction method of a building extraction network based on high - resolution remote sensing images. The end of the cyclic up - and - downsampling at different scales is upsampled through a transposed convolution operation with a convolution kernel size of 3×3, a convolution stride of m, and a number of convolution kernels of 64, where m = 1, 2, 4, and its size is restored to 512×512.

[0086] An extraction method of a building extraction network based on high - resolution remote sensing images. The features at different scales are subjected to an accumulation operation, and a batch binary classification convolution operation is added. Its convolution kernel size is 3×3, the convolution stride is 1, and the number of convolution kernels is 2;

[0087] Finally, a normalization layer BN, an activation layer ReLU, and a max - pooling layer MP, a prediction result normalization layer Softmax, and a binary classification output layer PixelClassification are added to the network. Its complete network structure is as Figure 6 shown.

[0088] A usage method of a building extraction network "biased cosine network" based on high - resolution remote sensing images includes the following steps,

[0089] Step 6.1: The program will run on a machine with an Intel(R) Xeon(R) Gold 6254 CPU @ 3.10 GHz, four NVIDIA GeForce RTX 3090 GPUs, 132 GB of memory, and the Ubuntu 18.04 operating system, using MCR 2021b;

[0090] When training the dataset, the hyperparameters are as follows: the initial learning rate learning_rate = 0.001, the batch size batch_size = 32, the number of training epochs epochs = 100, and the number of segmentation classes n_classes = 2;

[0091] Input the cropped images into the trained network model to extract buildings in the images.

[0092] A bias cosine network based on high-resolution remote sensing images uses the transformation between different scales as the main feature to form a cosine effect. At the initial stage, the input image is downsampled at multiple levels to form a bias effect. The bias cosine network includes the first group of downsampling units Ⅰ, Ⅲ, Ⅳ, Ⅴ, Ⅵ; the second group of downsampling units Ⅰ, Ⅱ, Ⅲ, Ⅳ; the third group of downsampling units Ⅰ, Ⅱ, Ⅲ, Ⅳ; and the fourth group of downsampling units Ⅰ, Ⅱ, Ⅲ, Ⅳ;

[0093] The input image is respectively input into the first group of downsampling unit Ⅰ, the second group of downsampling unit Ⅰ, and the third group of downsampling unit Ⅰ after convolution. The first group of downsampling unit Ⅰ transmits the convolved image to the second group of downsampling unit Ⅱ,

[0094] The second group of downsampling unit Ⅰ transmits the convolved image to the third group of downsampling unit Ⅱ,

[0095] The third group of downsampling unit Ⅰ transmits the deconvolved image to the second group of downsampling unit Ⅱ,

[0096] The second group of downsampling unit Ⅱ transmits the deconvolved image to the first group of downsampling unit Ⅲ,

[0097] The second group of downsampling unit Ⅱ transmits the convolved image to the third group of downsampling unit Ⅲ,

[0098] The third group of downsampling unit Ⅱ transmits the deconvolved image to the second group of downsampling unit Ⅲ,

[0099] The first group of downsampling unit Ⅲ transmits the convolved image to the second group of downsampling unit Ⅳ,

[0100] The third set of downsampling units Ⅲ transmits the deconvolved image to the second set of downsampling units Ⅳ.

[0101] The second set of downsampling units Ⅲ transmits the deconvolved image to the first set of downsampling units Ⅳ.

[0102] The first set of downsampling units Ⅳ transmits the convolved image to the second set of downsampling units Ⅴ.

[0103] The second set of downsampling units Ⅳ transmits the deconvolved image to the first set of downsampling units Ⅴ.

[0104] The second set of downsampling units Ⅳ transmits the convolved image to the third set of downsampling units Ⅴ.

[0105] The first set of downsampling units Ⅴ transmits the convolved image to the first set of downsampling units Ⅵ.

[0106] Both the second set of downsampling units Ⅴ and the third set of downsampling units Ⅴ transmit the deconvolved image to the first set of downsampling units Ⅵ.

[0107] The first set of downsampling units Ⅵ outputs the image after normalization.

[0108] Each downsampling unit includes normalization processing, convolution processing, and deconvolution processing.

Claims

1. An extraction method for a building extraction network based on high-resolution remote sensing images, characterized in that, The extraction method of the building extraction network includes the following steps: Step 1: Collect high-resolution optical remote sensing images; Step 2: Extract the buildings in the remote sensing images collected in Step 1 and make them into a dataset; Step 3: Build a bias cosine network; The bias cosine network includes the first group of downsampling units Ⅰ, the first group of downsampling units Ⅲ, the first group of downsampling units Ⅳ, the first group of downsampling units Ⅴ, the first group of downsampling units Ⅵ; the second group of downsampling units Ⅰ, the second group of downsampling units Ⅱ, the second group of downsampling units Ⅲ, the second group of downsampling units Ⅳ; the third group of downsampling units Ⅰ, the third group of downsampling units Ⅱ, the third group of downsampling units Ⅲ, the third group of downsampling units Ⅳ; the fourth group of downsampling units Ⅰ, the fourth group of downsampling units Ⅱ, the fourth group of downsampling units Ⅲ, and the fourth group of downsampling units Ⅳ; The input image is respectively input into the first group of downsampling units Ⅰ, the second group of downsampling units Ⅰ, and the third group of downsampling units Ⅰ after convolution, and the first group of downsampling units Ⅰ transmits the convolved image to the second group of downsampling units Ⅱ. The second group of downsampling units Ⅰ transmits the convolved image to the third group of downsampling units Ⅱ. The third group of downsampling units Ⅰ transmits the deconvolved image to the second group of downsampling units Ⅱ. The second group of downsampling units Ⅱ transmits the deconvolved image to the first group of downsampling units Ⅲ. The second group of downsampling units Ⅱ transmits the convolved image to the third group of downsampling units Ⅲ. The third group of downsampling units Ⅱ transmits the deconvolved image to the second group of downsampling units Ⅲ. The first group of downsampling units Ⅲ transmits the convolved image to the second group of downsampling units Ⅳ. The third group of downsampling units Ⅲ transmits the deconvolved image to the second group of downsampling units Ⅳ. The second group of downsampling units Ⅲ transmits the deconvolved image to the first group of downsampling units Ⅳ. The first group of downsampling units Ⅳ transmits the convolved image to the second group of downsampling units Ⅴ. The second group of downsampling units Ⅳ transmits the deconvolved image to the first group of downsampling units Ⅴ. The second group of downsampling units Ⅳ transmits the convolved image to the third group of downsampling units Ⅴ. The first group of downsampling units Ⅴ transmits the convolved image to the first group of downsampling units Ⅵ. Both the second group of downsampling units Ⅴ and the third group of downsampling units Ⅴ transmit the deconvolved image to the first group of downsampling units Ⅵ. The first group of downsampling units Ⅵ outputs an image after normalization; Step 4: Use the building dataset in Step 2 to train the bias cosine network; Step 5: Verify the network trained in Step 4; Step 6: Use the network verified in Step 5 for semantic segmentation to identify the building outlines.

2. The extraction method of the building extraction network according to claim 1, characterized in that Specifically, in Step 1, the optical remote sensing images are collected from the remote sensing data with a resolution better than one meter in the Jilin-1 satellite constellation.

3. The extraction method of the building extraction network according to claim 1, characterized in that, Specifically, in Step 2, Step 2.1: Select the remote sensing images for training the building extraction model; Step 2.2: Crop the original remote sensing images for making the dataset based on the remote sensing images in Step 2.1; set the cropping size to 512×512 pixels, and for the parts with a size less than 512×512 pixels after cutting, fill them with 0 values. Step 2.3: Divide the dataset in Step 2.2 into a training set, a validation set, and a test set; divide them according to the ratio of 7:2:1 for the training set, validation set, and test set.

4. The extraction method of the building extraction network according to claim 1, characterized in that, The specific construction of the bias cosine network in Step 3 is as follows. Step 3.1: Perform downsampling convolutions on the input image at three scales. Each sized image is successively passed through a batch normalization layer BN, an activation layer ReLU, and a max pooling layer MP; thus forming a complete downsampling process. Step 3.2: Perform strided convolutions and transposed convolutions for different scales to achieve cyclic upsampling and downsampling. All downsamplings are implemented through ConvD convolution operations. The convolution kernel size of ConvD is 3×3, the convolution stride is 2, and the number of convolution kernels is 64. All upsamplings are implemented through ConvU transposed convolution operations. The convolution kernel size of ConvU is 3×3, the convolution stride is 2, and the number of convolution kernels is 64.

5. The extraction method of the building extraction network according to claim 4, characterized in that, Step 3.2 also includes adding a batch normalization layer BN, an activation layer ReLU, and a max pooling layer MP respectively after each convolution or transposed convolution in the cyclic upsampling and downsampling.

6. The extraction method of the building extraction network according to claim 4, characterized in that, Accumulate the cyclic upsampling and downsampling networks at three scales according to the cyclic depth and corresponding feature scales, and then merge the features at different stages.

7. The extraction method of the building extraction network according to claim 6, characterized in that Perform upsampling on the ends of the cyclic upsampling and downsampling at different scales through a transposed convolution operation with a convolution kernel size of 3×3, a convolution stride of m, and a number of convolution kernels of 64.

8. The extraction method of the building extraction network according to claim 7, wherein Accumulate the features at different scales and add a batch binary classification convolution operation with a convolution kernel size of 3×3, a convolution stride of 1, and a number of convolution kernels of 2. Finally, add a batch normalization layer BN, an activation layer ReLU, and a max pooling layer MP, a prediction result normalization layer Softmax, and a binary classification output layer PixelClassification to the network.

9. The extraction method of the building extraction network according to claim 1, characterized in that The process of using the bias cosine network to extract buildings includes the following steps. Step 6.1: The program will run on a machine with an Intel(R) Xeon(R) Gold 6254 CPU @ 3.10 GHz for the CPU, four NVIDIA GeForce RTX 3090 for the GPU, 132 GB of memory, and an Ubuntu 18.04 operating system, using MCR2021b. Step 6.2: The hyperparameters for training the dataset are as follows: the initial learning rate learning_rate = 0.001, the batch size batch_size = 32, the number of training epochs epochs = 100, and the number of segmentation classes n_classes = 2. Step 6.3: Input the cropped image into the trained network model to extract buildings in the image.

Images