Remote sensing image segmentation method combined with vector prior data

Abstract

The invention discloses a remote sensing image segmentation method combined with vector prior data, and the method comprises the steps: 1) sequentially inputting satellite remote sensing segmentation image vectors Ub into a remote sensing image classification model Vb, determining a recognition target image, marking a recognition target Hb in a satellite remote sensing image Kb, and replacing the gray values of four boundary points of a recognition target marking box with a gray value Tb; 2) adjusting the resolution of a satellite remote sensing image K = {K1, K2,..., Ka} with an identification target label to D0, and determining the pixel point coordinates of a gray value T = {T1, T2,..., Ta}; and 3) in the original satellite remote sensing image, connecting pixel points where the gray value Tb is located into a closed rectangle, and marking an identification target label. According to the method, recognition sizes and resolutions of different targets are fully considered, and different segmentation sizes and resolutions are adopted to process the remote sensing image, so that the input image most suitable for target recognition is obtained. According to the method, the marking box is positioned by adopting the gray value, so that the multi-target recognition results are combined to the same layer.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and specifically to a remote sensing image segmentation method combining vector prior data. Background Art

[0002] High spatial resolution images, especially high spatial resolution remote sensing images, provide rich information on aspects such as the geometric structure of ground objects, texture details, and the spectra of ground objects, making it possible to observe the detailed changes on the earth's surface at a smaller spatial scale, conduct large-scale remote sensing mapping, and monitor the impact of human activities on the environment, with broad application prospects.

[0003] With the development of satellite remote sensing technology, the rapid query of a large amount of remote sensing images has become an urgent issue to be solved. The content-based remote sensing image retrieval method makes it possible to quickly and accurately query remote sensing images.

[0004] Image segmentation is a key part in region-based image retrieval. The segmentation methods of remote sensing images mainly include watershed-based segmentation, mathematical morphology-based segmentation, edge-based segmentation, statistics-based segmentation, etc.

[0005] After existing image segmentation performs multi-target recognition, it is difficult to quickly merge the recognized targets into the same layer. Summary of the Invention

[0006] The purpose of the present invention is to provide a remote sensing image segmentation method combining vector prior data, including the following steps:

[0007] 1) Construct a recognition target type vector A = {A 1 , A 2 ,..., A n}, and a remote sensing image classification model C = {C 1 , C 2 ,..., C n} for each recognition target; n is the number of recognition target types;

[0008] The remote sensing image classification model includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer;

[0009] The input layer acquires a remote sensing image and uses the remote sensing image as input data;

[0010] The convolutional layer performs a convolution operation on the input data through a filter to extract features at different positions, thereby outputting a feature map;

[0011] Among them, the feature map E f output by the convolutional layer f is as follows:

[0012]

[0013] Where: E f-1 represents the feature map of the convolutional layer f-1, represents the convolution operation, W f , B f represent the weights and biases of the convolutional layer f; σ() represents the activation function; f = 1, 2,..., F; F is the number of convolutional layers;

[0014] The pooling layer reduces the data dimension of the output feature map of the convolutional layer through downsampling operation, and outputs the downsampled feature map;

[0015] Among them, the feature map G output by the pooling layer r r is as follows:

[0016] G r = σ(W' r down(G r-1 ) + B' r ) (2)

[0017] Where: G r-1 represents the feature map output by the pooling layer r-1, W' r , B' r represent the weights and biases of the pooling layer r; down() represents the pooling function; r = 1, 2,..., R; R is the number of pooling layers;

[0018] The fully connected layer maps the extracted feature map to the recognized target image;

[0019]

[0020] Where: is the output of the j-th neuron in the l-th layer; are the weights and biases; is the output of the i-th neuron in the (l-1)-th layer;

[0021] The output layer is used to output the recognized target image;

[0022] 2) Construct the first mapping database of the recognition target and the resolution, and the steps include:

[0023] 2.1) Obtain the remote sensing image with the recognition target A i , extract the recognition target image, and record the resolution D i of the current recognition target image; the initial value of i is 1;

[0024] 2.2) Input the recognition target image into the remote sensing image classification model C i . If the output result is unrecognizable, go to step 2.3), otherwise, go to step 2.4);

[0025] 2.3) Update the resolution D of the recognition target image i = D i + ΔD, and return to step 2.2); ΔD is the resolution increase step size;

[0026] 2.4) Determine whether i > n holds. If not, set i = i + 1 and return to step 2.1). If so, proceed to step 2.5);

[0027] 2.5) Construct the first mapping database O between the recognition target and the resolution 1 , that is:

[0028]

[0029] 3) Construct the second mapping database between the recognition target and the segmentation scale;

[0030] 3.1) Obtain m remote sensing images with the recognition target A i , and adjust the resolution of the remote sensing images to D i to obtain the remote sensing images to be processed;

[0031] 3.2) Extract the recognition target image from the j-th remote sensing image to be processed, and record the size M Ai,j × N Ai,j ; The initial value of j is 1;

[0032] 3.3) Determine whether j > m holds. If not, set j = j + 1 and return to step 3.2). If so, proceed to step 3.4);

[0033] 3.4) Determine the segmentation scale M i of the recognition target A Ai × N Ai ; M Ai = max{M Ai,1 , M Ai,2 ,..., M Ai,m}; N Ai = max{N Ai,1 , N Ai,2 ,..., N Ai,m}

[0034] 3.5) Determine whether i > n holds. If not, set i = i + 1 and return to step 3.1). If so, proceed to step 3.6);

[0035] 3.6) Construct the second mapping database O between the recognition target and the segmentation scale 2 , that is:

[0036]

[0037] 4) Select multiple recognition targets from the recognition target type vector A = {A 1 , A 2 ,..., A n} as the objects to be recognized, denoted as H = {H 1 , H 2 ,..., H a}; a is the number of objects to be recognized; Based on the recognition target names, determine the remote sensing image classification models corresponding to each object to be recognized, denoted as V = {V 1 , V 2 ,..., V a};

[0038] 5) Based on the first mapping database, determine the resolutions corresponding to the objects to be recognized H = {H 1 , H 2 ,..., H a}, denoted as I = {I 1 , I 2 ,..., I a};

[0039] Based on the second mapping database, determine the segmentation sizes corresponding to the objects to be recognized H = {H 1 , H 2 ,..., H a}, denoted as J = {J 1 , J 2 ,..., J a} = {M H1 × N H1 , M H2 × N H2 ,..., M Ha × N Ha};

[0040] 6) Obtain the original satellite remote sensing image and preprocess the original satellite remote sensing image, record the resolution D of the original satellite remote sensing image 0 ;

[0041] 7) Based on the resolution I, adjust the resolution of the satellite remote sensing image to obtain a satellite remote sensing images with resolutions of I 1 , I 2 ,..., I a , denoted as K = {K 1 , K 2 ,..., K a} = {(K 1 , M K1 × N K1 ), (K 2 , M K2 × N K2),...,(K a ,M Ka ×N Ka )};M Kb ×N Kb is the size of the satellite remote sensing image K b ; b = 1, 2,..., a;

[0042] 8) Based on the segmentation size J, segment the satellite remote sensing image K = {K 1 , K 2 ,..., K a} to obtain the satellite remote sensing segmentation image matrix; among them, the segmentation size of the satellite remote sensing image K b is J b = M Hb ×N Hb ; b = 1, 2,..., a;

[0043] The satellite remote sensing segmentation image matrix U is as follows:

[0044]

[0045] In the formula, represents rounding up; the satellite remote sensing segmentation image vector U b is the segmented sub-image of the satellite remote sensing image K b ;

[0046] 9) Convert the original satellite remote sensing image and the satellite remote sensing image K = {K 1 , K 2 ,..., K a} into grayscale images, record the grayscale values of each pixel point, and construct a grayscale matrix;

[0047] 10) Randomly generate a grayscale values that are all different from all elements in the grayscale matrix, denoted as T = {T 1 , T 2 ,..., T a}, and construct the third mapping matrix O 3 of the grayscale value and the recognition target;

[0048] The third mapping matrix O 3 is as follows:

[0049]

[0050] 11) Input the satellite remote sensing segmentation image vector U b into the remote sensing image classification model V b in turn to determine the recognition target image, and mark the recognition target H b in the satellite remote sensing image K b , and use the grayscale value Tb Replace the gray values of the four boundary points of the recognition target annotation box; the initial value of b is 1;

[0051] 12) Determine whether b > a holds. If so, proceed to step 13). Otherwise, set b = b + 1 and return to step 11);

[0052] 13) Adjust the resolution of the satellite remote sensing image K = {K 1 , K 2 ,..., K a} with the recognition target annotation to D 0 , and determine the pixel coordinates where the gray values T = {T 1 , T 2 ,..., T a} are located;

[0053] 14) In the original satellite remote sensing image, connect the pixel points where the gray value T b is located into a closed rectangle and label it with the recognition target label H b ; the initial value of b is 1;

[0054] 15) Determine whether b > a holds. If not, set b = b + 1 and return to step 14).

[0055] Furthermore, the types of recognition targets include cultivated land, park, fruit forest, grassland, wetland, agricultural building facilities, residential houses, commercial service facilities, mining areas, warehouses, railways, rail transit, highways, airports, port terminals, water areas, and sea areas.

[0056] Furthermore, the steps for preprocessing the satellite remote sensing image include: denoising and dimensionality reduction.

[0057] Furthermore, the steps for denoising the satellite remote sensing image include:

[0058] a1) Move the neighborhood window Ω k pixel by pixel in the satellite remote sensing image, calculate the entropy F(x, y) of each pixel point, and construct the entropy image F;

[0059] The entropy F(x, y) of the pixel point is as follows:

[0060]

[0061] In the formula, p l is the probability that the gray level l appears in the neighborhood window Ω k ; L is the maximum gray level; l is the gray level;

[0062] Among them, the probability p k that the gray level l appears in the neighborhood window Ω l is as follows:

[0063]

[0064] In the formula, n l represents the number of pixels with gray level j; M k ×R k is the size of the neighborhood window Ω k size;

[0065] a2) Based on the entropy image, establish a denoising model, that is:

[0066]

[0067] In the formula, I represents the satellite remote sensing image; div and ▽ are the divergence operator and the gradient operator; B is the satellite remote sensing image after denoising; Q is the Gaussian kernel function; F is the entropy image; g(▽B,F) and g(▽I,F) are diffusion coefficient functions;

[0068] Among them, the diffusion coefficient functions g(▽B,F) and g(▽I,F) at the pixel point (x,y) are respectively shown as follows:

[0069]

[0070] In the formula, |▽I(x,y)| and |▽B(x,y)| are the gradient magnitudes of the satellite remote sensing image before and after denoising at the pixel point (x,y); K is the preset edge intensity threshold; f(F(x,y) is the entropy function; F max 、F min respectively represent the maximum and minimum values of the entropy; T 1 is the maximum value of the gradient magnitude |▽B(x,y)|; T 2 is the maximum value of the gradient magnitude |▽I(x,y)|;

[0071] a3) Use the denoising model to denoise the satellite remote sensing image.

[0072] Furthermore, the steps for dimensionality reduction of the satellite remote sensing image include: performing dimensionality reduction processing on the satellite remote sensing image using the gradient operator.

[0073] Furthermore, the steps for constructing a remote sensing image classification model include:

[0074] b1) Obtain multiple satellite remote sensing images with recognition targets;

[0075] b2) Expand the satellite remote sensing images with recognition targets, mark the recognition target images, and construct a satellite remote sensing image sample set;

[0076] b3) Use the satellite remote sensing image sample set to train the convolutional neural network to obtain a remote sensing image classification model.

[0077] Furthermore, when training the convolutional neural network, the loss function adopted is as follows:

[0078]

[0079] In the formula, ω is the weight; p(x) is the expected output; q(x) is the actual output; λ is the regularization parameter.

[0080] Furthermore, the steps for augmenting the satellite remote sensing image with the recognition target include flipping, rotating, translating, scaling, color enhancement, adding random noise, and image blending.

[0081] Furthermore, in step 8), if is not an integer, the satellite remote sensing image is augmented by padding with zeros.

[0082] Furthermore, in steps 2.1) and 3.2), the recognition target image is extracted manually.

[0083] The technical effect of the present invention is beyond doubt, and the beneficial effects of the present invention are as follows:

[0084] 1) Fully consider the recognition sizes and resolutions of different targets, and process the remote sensing image with different segmentation sizes and resolutions to obtain the input image most suitable for the recognition target.

[0085] 2) Use the gray value to locate the annotation box, so as to merge the multi-target recognition results into the same layer. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 It is a flowchart of the method. DETAILED DESCRIPTION OF THE INVENTION

[0087] The present invention will be further described below with reference to the embodiments, but it should not be understood that the above-mentioned subject scope of the present invention is limited to the following embodiments. Without departing from the above-mentioned technical idea of the present invention, various substitutions and changes should be included in the protection scope of the present invention according to the common general knowledge and customary means in the art.

[0088] Embodiment 1:

[0089] Refer to Figure 1 , a remote sensing image segmentation method combining vector prior data, including the following steps:

[0090] 1) Construct a recognition target type vector A = {A 1 , A 2 ,..., A n}, and a remote sensing image classification model C = {C 1 , C2 ,..., C n}; n is the number of recognized target types;

[0091] The remote sensing image classification model includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer;

[0092] The input layer obtains the remote sensing image and uses the remote sensing image as input data;

[0093] The convolutional layer performs a convolution operation on the input data through a filter to extract features at different positions, and thus outputs a feature map;

[0094] Among them, the feature map E output by the convolutional layer f f is as follows:

[0095]

[0096] In the formula: E f-1 represents the feature map of the convolutional layer f - 1, represents the convolution operation, W f , B f represents the weight and bias of the convolutional layer f; σ() represents the activation function; f = 1, 2,..., F; F is the number of convolutional layers;

[0097] The pooling layer reduces the data dimension of the feature map output by the convolutional layer through a downsampling operation and outputs the downsampled feature map;

[0098] Among them, the feature map G output by the pooling layer r r is as follows:

[0099] G r = σ(W' r down(G r-1 ) + B' r ) (2)

[0100] In the formula: G r-1 represents the feature map output by the pooling layer r - 1, W' r , B' r represents the weight and bias of the pooling layer r; down() represents the pooling function; r = 1, 2,..., R; R is the number of pooling layers;

[0101] The fully connected layer maps the extracted feature map to the recognized target image;

[0102]

[0103] In the formula: is the output of the j-th neuron in the l-th layer; is the weight and bias; is the output of the i-th neuron in the (l - 1)-th layer; the input of the fully connected layer is the output of the pooling layer;

[0104] The output layer is used to output the recognized target image;

[0105] 2) Construct the first mapping database of the recognition target and the resolution, and the steps include:

[0106] 2.1) Obtain the remote sensing image with the recognition target A i and extract the recognition target image, and record the resolution D of the current recognition target image i ; the initial value of i is 1;

[0107] 2.2) Input the recognition target image into the remote sensing image classification model C i . If the output result is unrecognizable, go to step 2.3), otherwise, go to step 2.4);

[0108] 2.3) Update the resolution D of the recognition target image i = D i + ΔD, and return to step 2.2); ΔD is the resolution increase step;

[0109] 2.4) Judge whether i > n holds. If not, let i = i + 1 and return to step 2.1). If so, go to step 2.5);

[0110] 2.5) Construct the first mapping database O of the recognition target and the resolution, that is: 1 namely:

[0111]

[0112] 3) Construct the second mapping database of the recognition target and the segmentation scale;

[0113] 3.1) Obtain m remote sensing images with the recognition target A i and adjust the resolution of the remote sensing images to D i to obtain the remote sensing images to be processed;

[0114] 3.2) Extract the recognition target image in the j-th remote sensing image to be processed, and record the size M Ai,j × N Ai,j ; the initial value of j is 1;

[0115] 3.3) Judge whether j > m holds. If not, let j = j + 1 and return to step 3.2). If so, go to step 3.4);

[0116] 3.4) Determine the segmentation scale M i of the recognition target A Ai × NAi ; M Ai = max{M Ai,1 , M Ai,2 ,..., M Ai,m}; N Ai = max{N Ai,1 , N Ai,2 ,..., N Ai,m}

[0117] 3.5) Determine whether i > n holds. If not, set i = i + 1 and return to step 3.1). If so, proceed to step 3.6);

[0118] 3.6) Construct the second mapping database O of the recognition target and the segmentation scale 2 , that is:

[0119]

[0120] 4) Select multiple recognition targets from the recognition target type vector A = {A 1 , A 2 ,..., A n} as the objects to be recognized, denoted as H = {H 1 , H 2 ,..., H a}; a is the number of objects to be recognized; Based on the recognition target name, determine the remote sensing image classification model corresponding to each object to be recognized, denoted as V = {V 1 , V 2 ,..., V a};

[0121] 5) Based on the first mapping database, determine the resolution corresponding to the objects to be recognized H = {H 1 , H 2 ,..., H a}, denoted as I = {I 1 , I 2 ,..., I a};

[0122] Based on the second mapping database, determine the segmentation size corresponding to the objects to be recognized H = {H 1 , H 2 ,..., H a}, denoted as J = {J 1 , J 2 ,..., J a} = {M H1 × N H1 , M H2 × N H2 ,..., M Ha × N Ha};

[0123] 6) Obtain the original satellite remote sensing image, preprocess the original satellite remote sensing image, and record the resolution D of the original satellite remote sensing image 0 ;

[0124] 7) Based on the resolution I, adjust the resolution of the satellite remote sensing image to obtain a satellite remote sensing images with resolutions of I 1 , I 2 ,..., I a a, denoted as K = {K 1 , K 2 ,..., K a} = {(K 1 , M K1 ×N K1 ), (K 2 , M K2 ×N K2 ),...,(K a , M Ka ×N Ka )}; M Kb ×N Kb is the size of the satellite remote sensing image K b ; b = 1, 2,..., a;

[0125] 8) Based on the segmentation size J, perform size segmentation on the satellite remote sensing image K = {K 1 , K 2 ,..., K a} to obtain the satellite remote sensing segmentation image matrix; where the segmentation size of the satellite remote sensing image K b is J b = M Hb ×N Hb ; b = 1, 2,..., a;

[0126] The satellite remote sensing segmentation image matrix U is as follows:

[0127]

[0128] In the formula, represents rounding up; the satellite remote sensing segmentation image vector U b is the segmented sub-image of the satellite remote sensing image K b ;

[0129] 9) Convert the original satellite remote sensing image, the satellite remote sensing image K = {K 1 , K 2 ,..., K a} into grayscale images, record the grayscale values of each pixel point, and construct a grayscale matrix;

[0130] 10) Randomly generate a grayscale values that are all different from all elements in the grayscale matrix, denoted as T = {T 1 , T 2 ,..., T a}, and construct the third mapping matrix O 3 of the grayscale value and the recognition target;

[0131] The third mapping matrix O 3 is as follows:

[0132]

[0133] 11) Input the satellite remote sensing segmentation image vector U b into the remote sensing image classification model V b in sequence to determine the recognition target image, and label the recognition target H b in the satellite remote sensing image K b , and replace the grayscale values of the four boundary points of the recognition target annotation box with the grayscale value T b ; The initial value of b is 1;

[0134] 12) Judge whether b > a holds. If so, go to step 13). Otherwise, let b = b + 1 and return to step 11);

[0135] 13) Adjust the resolution of the satellite remote sensing image K = {K 1 , K 2 ,..., K a} with the recognition target annotation to D 0 , and determine the pixel point coordinates where the grayscale value T = {T 1 , T 2 ,..., T a} is located;

[0136] 14) In the original satellite remote sensing image, connect the pixel points where the grayscale value T b is located into a closed rectangle and label it with the recognition target label H b ; The initial value of b is 1;

[0137] 15) Judge whether b > a holds. If not, let b = b + 1 and return to step 14).

[0138] The types of the recognition targets include cultivated land, park, fruit forest, grassland, wetland, agricultural building facilities, residential houses, commercial service facilities, mining areas, warehouses, railways, rail transit, highways, airports, port terminals, water areas, sea areas.

[0139] The steps for preprocessing the satellite remote sensing image include: denoising and dimensionality reduction.

[0140] The steps for denoising the satellite remote sensing image include:

[0141] a1) Move the neighborhood window Ω pixel by pixel in the satellite remote sensing image k , calculate the entropy F(x,y) of each pixel, and construct the entropy image F;

[0142] The entropy F(x,y) of the pixel is as follows:

[0143]

[0144] where p l is the probability that the gray level l appears in the neighborhood window Ω k ; L is the maximum gray level; l is the gray level;

[0145] Among them, the probability p k that the gray level l appears in the neighborhood window Ω l is as follows:

[0146]

[0147] where n l represents the number of pixels with gray level j; M k ×R k is the size of the neighborhood window Ω k ;

[0148] a2) Based on the entropy image, establish a denoising model, that is:

[0149]

[0150] where I represents the satellite remote sensing image; div and ▽ are the divergence operator and the gradient operator; B is the denoised satellite remote sensing image; Q is the Gaussian kernel function; F is the entropy image; g(▽B,F), g(▽I,F) are the diffusion coefficient functions;

[0151] Among them, the diffusion coefficient functions g(▽B,F), g(▽I,F) at the pixel point (x,y) are respectively as follows:

[0152]

[0153] where |▽I(x,y)|, |▽B(x,y)| are the gradient norms of the satellite remote sensing image before and after denoising at the pixel point (x,y); K is the preset edge intensity threshold; f(F(x,y) is the entropy function; F max , F min respectively represent the maximum and minimum values of the entropy; T 1 is the maximum value of the gradient norm |▽B(x,y)|; T 2 is the maximum value of the gradient norm |▽I(x,y)|;

[0154] a3) Denoise the satellite remote sensing image using a denoising model.

[0155] The steps for dimensionality reduction of the satellite remote sensing image include: performing dimensionality reduction processing on the satellite remote sensing image using a gradient operator.

[0156] The steps for constructing a remote sensing image classification model include:

[0157] b1) Obtain multiple satellite remote sensing images with recognition targets;

[0158] b2) Augment the satellite remote sensing images with recognition targets, label the recognition target images, and construct a satellite remote sensing image sample set;

[0159] b3) Use the satellite remote sensing image sample set to train a convolutional neural network to obtain a remote sensing image classification model.

[0160] When training the convolutional neural network, the loss function used is as follows:

[0161]

[0162] In the formula, ω is the weight; p(x) is the expected output; q(x) is the actual output; λ is the regularization parameter.

[0163] The steps for augmenting the satellite remote sensing images with recognition targets include flipping, rotating, translating, scaling, color enhancement, adding random noise, and image blending.

[0164] In step 8), if is not an integer, then pad the satellite remote sensing image with zeros for augmentation.

[0165] In steps 2.1) and 3.2), the recognition target images are extracted manually.

[0166] Example 2:

[0167] A remote sensing image segmentation method combining vector prior data, comprising the following steps:

[0168] 1) Construct a recognition target type vector A = {A 1 , A 2 ,..., A n}, and a remote sensing image classification model C = {C 1 , C 2 ,..., C n} for each recognition target; n is the number of recognition target types;

[0169] The remote sensing image classification model includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer;

[0170] The input layer obtains a remote sensing image and uses the remote sensing image as input data;

[0171] The convolutional layer performs a convolution operation on the input data through a filter to extract features at different positions, thereby outputting a feature map;

[0172] Among them, the feature map E output by the convolutional layer f f is as follows:

[0173]

[0174] In the formula: E f-1 represents the feature map of the convolutional layer f-1, represents the convolution operation, W f , B f represent the weights and biases of the convolutional layer f; σ() represents the activation function; f = 1, 2,..., F; F is the number of convolutional layers;

[0175] The pooling layer reduces the data dimension of the feature map output by the convolutional layer through a downsampling operation and outputs the downsampled feature map;

[0176] Among them, the feature map G output by the pooling layer r r is as follows:

[0177] G r = σ(W' r down(G r-1 ) + B' r ) (2)

[0178] In the formula: G r-1 represents the feature map output by the pooling layer r-1, W' r , B' r represent the weights and biases of the pooling layer r; down() represents the pooling function; r = 1, 2,..., R; R is the number of pooling layers;

[0179] The fully connected layer maps the extracted feature map to an identified target image;

[0180]

[0181] In the formula: is the output of the j-th neuron in the l-th layer; are the weights and biases; is the output of the i-th neuron in the (l-1)-th layer; the input of the fully connected layer is the output of the pooling layer;

[0182] The output layer is used to output the identified target image;

[0183] 2) Construct the first mapping database of the recognition target and the resolution, and the steps include:

[0184] 2.1) Obtain the remote sensing image with the recognition target A i and extract the recognition target image, and record the resolution D of the current recognition target image i ; The initial value of i is 1;

[0185] 2.2) Input the recognition target image into the remote sensing image classification model C i . If the output result is unrecognizable, go to step 2.3), otherwise, go to step 2.4);

[0186] 2.3) Update the resolution D of the recognition target image i = D i + ΔD, and return to step 2.2); ΔD is the resolution increase step size;

[0187] 2.4) Judge whether i > n holds. If not, let i = i + 1 and return to step 2.1). If so, go to step 2.5);

[0188] 2.5) Construct the first mapping database of the recognition target and the resolution, that is:

[0189]

[0190] 3) Construct the second mapping database of the recognition target and the segmentation scale;

[0191] 3.1) Obtain m remote sensing images with the recognition target A i and adjust the resolution of the remote sensing image to D i to obtain the remote sensing image to be processed;

[0192] 3.2) Extract the recognition target image in the jth remote sensing image to be processed, and record the size M Ai,j × N Ai,j ; The initial value of j is 1;

[0193] 3.3) Judge whether j > m holds. If not, let j = j + 1 and return to step 3.2). If so, go to step 3.4);

[0194] 3.4) Determine the segmentation scale M i × N Ai of the recognition target A Ai ; M Ai = max{M Ai,1 , M Ai,2 ,..., M Ai,m}; N Ai = max{N Ai,1 , NAi,2 ,...,N Ai,m}

[0195] 3.5) Determine whether i > n holds. If not, let i = i + 1 and return to step 3.1). If so, proceed to step 3.6);

[0196] 3.6) Construct the second mapping database of the recognition target and the segmentation scale, i.e.:

[0197]

[0198] 4) Select multiple recognition targets from the recognition target type vector A = {A 1 , A 2 ,..., A n} as the objects to be recognized, denoted as H = {H 1 , H 2 ,..., H a}; a is the number of objects to be recognized. Based on the recognition target name, determine the remote sensing image classification model corresponding to each object to be recognized, denoted as V = {V 1 , V 2 ,..., V a};

[0199] 5) Based on the first mapping database, determine the resolution corresponding to the object to be recognized H = {H 1 , H 2 ,..., H a}, denoted as I = {I 1 , I 2 ,..., I a};

[0200] Based on the second mapping database, determine the segmentation size corresponding to the object to be recognized H = {H 1 , H 2 ,..., H a}, denoted as J = {J 1 , J 2 ,..., J a} = {M H1 ×N H1 , M H2 ×N H2 ,..., M Ha ×N Ha};

[0201] 6) Obtain the original satellite remote sensing image and preprocess the original satellite remote sensing image, and record the resolution D of the original satellite remote sensing image 0 ;

[0202] 7) Based on resolution I, adjust the resolution of the satellite remote sensing image to obtain a satellite remote sensing images with resolutions of I 1 , I 2 ,..., I a of a satellite remote sensing images, denoted as K = {K 1 , K 2 ,..., K a} = {(K 1 , M K1 ×N K1 ), (K 2 , M K2 ×N K2 ),...,(K a , M Ka ×N Ka )}; M Kb ×N Kb is the size of the satellite remote sensing image K b ; b = 1, 2,..., a;

[0203] 8) Based on the segmentation size J, perform size segmentation on the satellite remote sensing image K = {K 1 , K 2 ,..., K a} to obtain a satellite remote sensing segmentation image matrix; where the segmentation size of the satellite remote sensing image K b is J b = M Hb ×N Hb ; b = 1, 2,..., a;

[0204] The satellite remote sensing segmentation image matrix is shown as follows:

[0205]

[0206] In the formula, represents rounding up; the satellite remote sensing segmentation image vector U b is the segmented sub-image of the satellite remote sensing image K b ;

[0207] 9) Convert the original satellite remote sensing image and the satellite remote sensing image K = {K 1 , K 2 ,..., K a} into grayscale images, record the grayscale values of each pixel point, and construct a grayscale matrix;

[0208] 10) Randomly generate a grayscale values that are all different from all elements in the grayscale matrix, denoted as T = {T 1 , T 2 ,..., T a}, and construct a third mapping matrix O 3;

[0209] The third mapping matrix O 3 is as follows:

[0210]

[0211] 11) Input the satellite remote sensing segmentation image vector U b sequentially into the remote sensing image classification model V b to determine the recognized target image, and label the recognized target H b in the satellite remote sensing image K b , and replace the gray values of the four boundary points of the recognized target annotation box with the gray value T b ; The initial value of b is 1;

[0212] 12) Judge whether b > a holds. If so, go to step 13). Otherwise, let b = b + 1 and return to step 11);

[0213] 13) Adjust the resolution of the satellite remote sensing image K = {K 1 , K 2 ,..., K a} with the recognized target annotation to D 0 , and determine the pixel point coordinates where the gray value T = {T 1 , T 2 ,..., T a} is located;

[0214] 14) In the original satellite remote sensing image, connect the pixel points where the gray value T b is located into a closed rectangle and label it with the recognized target label H b ; The initial value of b is 1;

[0215] 15) Judge whether b > a holds. If not, let b = b + 1 and return to step 14).

[0216] Example 3:

[0217] A remote sensing image segmentation method combining vector prior data, the technical content is the same as that of Example 2. Further, the recognized target types include cultivated land, park, fruit forest, grassland, wetland, agricultural building facilities, residential houses, commercial service facilities, mining areas, warehouses, railways, rail transit, highways, airports, port terminals, waters, sea areas.

[0218] Example 4:

[0219] A remote sensing image segmentation method combining vector prior data, the technical content is the same as any one of Examples 2 - 3. Further, the steps for preprocessing the satellite remote sensing image include: denoising, dimensionality reduction.

[0220] Example 5:

[0221] A remote sensing image segmentation method combining vector prior data, the technical content is the same as any one of Embodiments 2-4. Further, the steps of denoising the satellite remote sensing image include:

[0222] 1) Move the neighborhood window Ω pixel by pixel in the satellite remote sensing image k , calculate the entropy F(x,y) of each pixel point, and construct the entropy image F;

[0223] The entropy F(x,y) of the pixel point is as follows:

[0224]

[0225] In the formula, p l is the probability that the gray level l appears in the neighborhood window Ω k ; L is the maximum gray level; l is the gray level;

[0226] Among them, the probability p k that the gray level l appears in the neighborhood window Ω l is as follows:

[0227]

[0228] In the formula, n l represents the number of pixels with gray level j; M k ×R k is the size of the neighborhood window Ω k ;

[0229] 2) Based on the entropy image, establish a denoising model, that is:

[0230]

[0231] In the formula, I represents the satellite remote sensing image; div and ▽ are the divergence operator and the gradient operator; B is the denoised satellite remote sensing image; Q is the Gaussian kernel function; F is the entropy image; g(▽B,F), g(▽I,F) are diffusion coefficient functions;

[0232] Among them, the diffusion coefficient functions g(▽B,F), g(▽I,F) at the pixel point (x,y) are respectively as follows:

[0233]

[0234] In the formula, |▽I(x,y)|, |▽B(x,y)| are the gradient moduli of the satellite remote sensing image before and after denoising at the pixel point (x,y); K is the preset edge intensity threshold; f(F(x,y) is the entropy function; F max 、F minrespectively represent the maximum and minimum values of entropy; T 1 is the maximum value of the gradient modulus |▽B(x,y)|; T 2 is the maximum value of the gradient modulus |▽I(x,y)|;

[0235] 3) Use the denoising model to denoise the satellite remote sensing image.

[0236] Example 6:

[0237] A remote sensing image segmentation method combining vector prior data, the technical content is the same as any one of Examples 2-5. Further, the steps of dimensionality reduction of the satellite remote sensing image include: performing dimensionality reduction processing on the satellite remote sensing image using a gradient operator.

[0238] Example 7:

[0239] A remote sensing image segmentation method combining vector prior data, the technical content is the same as any one of Examples 2-6. Further, the steps of constructing a remote sensing image classification model include:

[0240] 1) Obtain multiple satellite remote sensing images with recognition targets;

[0241] 2) Expand the satellite remote sensing images with recognition targets, label the recognition target images, and construct a satellite remote sensing image sample set;

[0242] 3) Use the satellite remote sensing image sample set to train a convolutional neural network to obtain a remote sensing image classification model.

[0243] Example 8:

[0244] A remote sensing image segmentation method combining vector prior data, the technical content is the same as any one of Examples 2-7. Further, when training the convolutional neural network, the loss function used is as follows:

[0245]

[0246] In the formula, ω is the weight; p(x) is the expected output; q(x) is the actual output; λ is the regularization parameter.

[0247] Example 9:

[0248] A remote sensing image segmentation method combining vector prior data, the technical content is the same as any one of Examples 2-8. Further, the steps of expanding the satellite remote sensing images with recognition targets include flipping, rotating, translating, scaling, color enhancement, adding random noise, and image mixing.

[0249] Example 10:

[0250] A remote sensing image segmentation method combining vector prior data, the technical content is the same as any one of Embodiments 2-9. Further, in step 8), if is a non-integer, the satellite remote sensing image is expanded by padding with zeros.

[0251] Embodiment 11:

[0252] A remote sensing image segmentation method combining vector prior data, the technical content is the same as any one of Embodiments 2-10. Further, in steps 2.1) and 3.2), the target image is identified by manual extraction.

Claims

1. A remote sensing image segmentation method combined with vector prior data, characterized in that: The following steps are involved: 1) Construct the recognition target type vector A = {A1, A2, ..., A n }, and the remote sensing image classification model C = {C1, C2, ..., C n }; n is the number of identified target types. The remote sensing image classification model includes an input layer, a convolution layer, a pooling layer, a fully connected layer and an output layer; The input layer obtains remote sensing images and uses the remote sensing images as input data; The convolution layer performs a convolution operation on the input data through a filter to extract features at different positions, thereby outputting a feature map; Among them, the feature map E output by the convolution layer f f As shown below: Where: E f-1 represents the feature map of convolutional layer f-1, represents the convolution operation, W f , B f represents the weight and bias of the convolutional layer f; σ() represents the activation function; f = 1, 2, ..., F; F is the number of convolutional layers; The pooling layer reduces the data dimension of the feature map output by the convolution layer through a downsampling operation, and outputs the feature map after dimensionality reduction; Among them, the feature map G output by the pooling layer r r As shown below: G r =σ(W r 'down(G r-1 )+B r ')(2) Where: G r-1 Represents the feature map output by pooling layer r-1, W' r , B' r Represents the weight and bias of pooling layer r; down() represents the pooling function; r = 1, 2, ..., R; R is the number of pooling layers; The fully connected layer maps the extracted feature map to a recognition target image; The output layer is used to output the recognition target image; 2) Constructing a first mapping database of recognition targets and resolutions, the steps include: 2.1) Get the target A with identification i remote sensing image, and extract the recognition target image, and record the resolution D of the current recognition target image i ;The initial value of i is 1; 2.2) Input the identified target image into the remote sensing image classification model C i If the output result is unrecognizable, go to step 2.3), otherwise, go to step 2.4); 2.3) Update the resolution D of the recognition target image i =D i +ΔD, and return to step 2.2); ΔD is the resolution increase step size; 2.4) Determine whether i>n is true. If not, set i=i+1 and return to step 2.1). If yes, proceed to step 2.5); 2.5) Construct the first mapping database O1 of recognition target and resolution, namely: 3) constructing a second mapping database of recognition targets and segmentation scales; 3.1) Get the target A with identification i m remote sensing images, and adjust the remote sensing image resolution to D i , get the remote sensing image to be processed; 3.2) Extract the identified target image in the jth remote sensing image to be processed and record the size M of the identified target image Ai,j ×N Ai,j ; The initial value of j is 1; 3.3) Determine whether j>m is true. If not, set j=j+1 and return to step 3.2). If yes, proceed to step 3.4); 3.4) Determine the identification target A i The segmentation scale M Ai ×N Ai ;M Ai =max{M Ai,1 ,M Ai,2 ,...,M Ai,m }; N Ai =max{N Ai,1 ,N Ai,2 ,...,N Ai,m } 3.5) Determine whether i>n is true. If not, set i=i+1 and return to step 3.1). If yes, proceed to step 3.6); 3.6) Construct a second mapping database O2 of recognition targets and segmentation scales, namely: 4) From the identification target type vector A = {A1, A2, ..., A n }, select multiple recognition targets as the objects to be recognized, denoted as H = {H1, H2, ..., H a }; a is the number of objects to be identified; based on the name of the identified target, the remote sensing image classification model corresponding to each object to be identified is determined, denoted as V = {V1, V2, ..., V a }; 5) Based on the first mapping database, determine the object to be identified H = {H1, H2, ..., H a }, denoted as I={I1,I2,...,I a }; Based on the second mapping database, determine the object to be identified H = {H1, H2, ..., H a }, denoted as J = {J1, J2, ..., J a }={M H1 ×N H1 ,M H2 ×N H2 ,...,M Ha ×N Ha }; 6) Obtaining original satellite remote sensing images, preprocessing the original satellite remote sensing images, and recording the resolution D0 of the original satellite remote sensing images; 7) Based on the resolution I, the resolution of the satellite remote sensing image is adjusted to obtain resolutions I1, I2, ..., I a a satellite remote sensing images, denoted by K = {K1, K2, ..., K a }={(K1,M K1 ×N K1 ),(K2,M K2 ×N K2 ),...,(K a ,M Ka ×N Ka )};M Kb ×N Kb K is the satellite remote sensing image b The size of ; b = 1, 2, ..., a; 8) Based on the segmentation size J, the satellite remote sensing image K = {K1, K2, ..., K a } size segmentation to obtain the satellite remote sensing segmentation image matrix; among them, the satellite remote sensing image K b The segmentation size is J b =M Hb ×N Hb ; b = 1, 2, ..., a; The satellite remote sensing segmentation image matrix U is as follows: In the formula, Indicates rounding up; Satellite remote sensing segmentation image vector U b K is the satellite remote sensing image b The segmented sub-image of 9) The original satellite remote sensing image, satellite remote sensing image K = {K1, K2, ..., K a }Convert to grayscale image, record the grayscale value of each pixel, and construct a grayscale matrix; 10) Randomly generate a grayscale value that is different from all elements in the grayscale matrix, denoted by T = {T1, T2, ..., T a }, and construct the third mapping matrix O3 between the gray value and the recognition target; The third mapping matrix O3 is as follows: 11) Segment the satellite remote sensing image vector U b Input into the remote sensing image classification model V b In the satellite remote sensing image K b Mark the target H b , and the gray value T b Replace the grayscale values ​​of the four boundary points of the target annotation box; the initial value of b is 1; 12) Determine whether b>a is true, if so, proceed to step 13), otherwise, set b=b+1 and return to step 11); 13) The satellite remote sensing image K with the target identification label is K = {K1, K2, ..., K a } is adjusted to D0, and the gray value T={T1,T2,...,T a }The pixel coordinates where the 14) In the original satellite remote sensing image, the gray value T b The pixels are connected into a closed rectangle and marked with the recognition target label H b ; The initial value of b is 1; 15) Determine whether b>a is established. If not, set b=b+1 and return to step 14).

2. The remote sensing image segmentation method combined with vector prior data according to claim 1, characterized in that: The identified target types include cultivated land, parks, orchards, grasslands, wetlands, agricultural buildings and facilities, residential houses, commercial service facilities, mining areas, warehouses, railways, rail transit, highways, airports, ports and terminals, waters, and sea areas.

3. The remote sensing image segmentation method combined with vector prior data according to claim 1, characterized in that: The steps for preprocessing satellite remote sensing images include: denoising and dimensionality reduction.

4. The remote sensing image segmentation method combined with vector prior data according to claim 3, characterized in that: The steps for denoising satellite remote sensing images include: 1) Move the neighborhood window Ω pixel by pixel in satellite remote sensing images k , calculate the entropy F(x,y) of each pixel and construct the entropy image F; The entropy of a pixel point F(x,y) is as follows: In the formula, p l is the gray level l in the neighborhood window Ω k The probability of occurrence in; L is the maximum gray level; l is the gray level; Among them, the gray level l is in the neighborhood window Ω k The probability of occurrence p l As shown below: Where n l Represents the number of pixels with gray level j; M k ×R k is the neighborhood window Ω k size; 2) Based on the entropy image, a denoising model is established, namely: Where, I represents satellite remote sensing image; div and ▽ are divergence operator and gradient operator; B is the denoised satellite remote sensing image; Q is the Gaussian kernel function; F is the entropy image; g(▽B,F) and g(▽I,F) are diffusion coefficient functions; Among them, the diffusion coefficient functions g(▽B,F) and g(▽I,F) at the pixel point (x,y) are as follows: In the formula, is the gradient modulus of the satellite remote sensing image at the pixel point (x, y) before and after denoising; K is the preset edge intensity threshold; f(F(x, y) is the entropy function; F max 、F min They represent the maximum and minimum values ​​of entropy respectively; T1 is the gradient modulus The maximum value of T2 is the gradient mode. The maximum value of 3) Use the denoising model to denoise satellite remote sensing images.

5. The remote sensing image segmentation method combined with vector prior data according to claim 3, characterized in that: The steps of reducing the dimension of satellite remote sensing images include: using a gradient operator to reduce the dimension of satellite remote sensing images.

6. The remote sensing image segmentation method combined with vector prior data according to claim 1, characterized in that: The steps to build a remote sensing image classification model include: 1) Obtain multiple satellite remote sensing images with identified targets; 2) Expand the satellite remote sensing images with identified targets, mark the identified target images, and construct a satellite remote sensing image sample set; 3) Use the satellite remote sensing image sample set to train the convolutional neural network to obtain the remote sensing image classification model.

7. A remote sensing image segmentation method combined with vector prior data according to claim 6, characterized in that: When training a convolutional neural network, the loss function L used is as follows: Where ω is the weight; p(x) is the expected output; q(x) is the actual output; and λ is the regularization parameter.

8. The remote sensing image segmentation method combined with vector prior data according to claim 6, characterized in that: The steps of expanding satellite remote sensing images with identified targets include flipping, rotating, translating, scaling, color enhancement, adding random noise, and image mixing.

9. The remote sensing image segmentation method combined with vector prior data according to claim 1, characterized in that: In step 8), if If it is a non-integer, the satellite remote sensing image is expanded by filling with zeros.

10. The remote sensing image segmentation method combined with vector prior data according to claim 1, characterized in that: In step 2.1) and step 3.2), the target image is identified by manual extraction.

Images