A method and system for automatically identifying buildings in a seismic survey area

Abstract

The application provides a seismic exploration work area building automatic identification method and system, and belongs to the technical field of intelligent identification of remote sensing images of seismic exploration work areas. The method combines an original Unet network model, a multi-scale convolution feature fusion structure and a hollow spatial pyramid structure to obtain an optimized network model, and uses the optimized network model to automatically identify buildings in a seismic exploration work area. The application uses a deep learning method to realize automatic identification of buildings in remote sensing images of seismic exploration work areas, introduces a multi-scale convolution feature fusion and an ASPP (Atrous Spatial Pyramid Pooling) module based on an original Unet network structure, enhances the network feature expression capability, and obtains semantic information of multiple sizes of targets. The application can more accurately locate building targets and has better segmentation effect on small targets, and has good universality.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent identification of seismic exploration work area remote sensing images, and particularly relates to a seismic exploration work area building automatic identification method and system. BACKGROUND

[0002] Three-dimensional seismic exploration has high requirements for seismic data acquisition accuracy, and the coverage area of the work area is wide. The distribution of obstacles represented by buildings in the work area is complex, which directly affects the quality of the seismic data acquisition results. Therefore, it is necessary to mark the buildings and other obstacles before construction to perform observation system re-observation and other operations. At present, in the field of seismic exploration, the building area information is generally manually marked based on the work area remote sensing image to quickly understand the building distribution in the work area. However, the manual marking method is inefficient, time-consuming and labor-intensive, and cannot meet the actual demand.

[0003] In recent years, with the rapid development of remote sensing technology, the imaging resolution of satellites has reached sub-meter level, and the amount of high-resolution remote sensing image data has increased sharply. Using high-resolution remote sensing images combined with image processing technology for target recognition has become a common method. In the field of field seismic exploration, unmanned aerial vehicles have been used for reconnaissance and inspection, which provides a convenient condition for obtaining remote sensing images of the exploration work area. It is of great significance to quickly and accurately identify buildings in seismic exploration work area remote sensing images for physical point layout, obstacle avoidance and acquisition in the field.

[0004] Chinese patent publication CN111652892A discloses a remote sensing image building vector extraction and optimization method based on deep learning, which includes using a full convolution form image segmentation model to generate a building target probability map from a remote sensing image, and generating an initial building vector contour through binaryzation, clustering and edge contour tracking processing. It also includes changing the convolution kernel size and applying a recursive padding method to improve the PointNet model for building contour geometry optimization. It also includes providing a new loss function for PointNet model training, which can estimate the similarity of two polygons with a large difference in node number. However, this patent mainly extracts the boundary of the building target through the boundary features of the building target, the target features are relatively simple, and the calculation steps are complex.

[0005] The buildings in the seismic exploration work area remote sensing image are widely distributed and numerous, and automatic identification is difficult. At present, the traditional machine learning method is mainly used for building target extraction in remote sensing images. This method needs to manually set the extraction features, and the subjective consciousness is strong, so the efficiency and accuracy of identification are very limited. In recent years, deep learning methods have become a powerful tool for image processing, so some research has begun to apply deep learning methods to building segmentation in remote sensing images, but there are problems such as incomplete target boundary segmentation and inaccurate small target segmentation. SUMMARY

[0006] The present application aims to solve the problems existing in the prior art, and provides a seismic exploration work area building automatic identification method and system, which is used for fast and high-precision automatic identification of building targets, and provides a basis for physical point layout, obstacle avoidance and collection in the field.

[0007] The present application is implemented by the following technical solutions:

[0008] In a first aspect, the present application provides a seismic exploration work area building automatic identification method, which combines an original Unet network model and a multi-scale convolution feature fusion structure and a hollow spatial pyramid structure to obtain an optimized network model, and uses the optimized network model to automatically identify the buildings in the seismic exploration work area.

[0009] Further improvements of the present application are as follows:

[0010] The method comprises:

[0011] (1) Collecting a city remote sensing image dataset related to building target segmentation, and performing image preprocessing and dataset division;

[0012] (2) Constructing a convolutional neural network;

[0013] (3) Constructing a convolutional neural network loss function;

[0014] (4) Training the convolutional neural network to obtain an optimized network model;

[0015] (5) Using the optimized network model to automatically identify the buildings in the seismic exploration work area.

[0016] Further improvements of the present application are as follows:

[0017] The preprocessing operation in step (1) comprises:

[0018] (11) Cutting all original images and corresponding label images into small images of a fixed size;

[0019] (12) Enhancing the original images and label images;

[0020] (13) Compressing the pixel values of the enhanced original images and label images to a preset value range space to obtain preprocessed original images and label images.

[0021] The dataset division operation in step (1) comprises:

[0022] The preprocessed original image and the label image are randomly divided into three data sets according to a set proportion, and the three data sets are respectively an experimental training set, a verification set and a test set.

[0023] The further improvement of the present application is that:

[0024] The step (2) operation includes:

[0025] The convolutional neural network is constructed as follows: the left side is the encoding structure of the original Unet network model, the right side is the decoding structure of the original Unet network model, the empty space pyramid module ASPP is arranged between the encoding structure and the decoding structure, and the multi-scale convolution feature fusion structure is arranged below the encoding structure;

[0026] The data are input into the encoding structure and the multi-scale convolution feature fusion structure, then the output of each pooling layer of the decoding structure and the output of each convolution layer of the multi-scale convolution feature fusion structure are subjected to feature splicing operation respectively, and the results obtained by the feature splicing operation are input into the decoding structure respectively;

[0027] Meanwhile, the output result of the last convolution layer of the encoding structure is input into the empty space pyramid module ASPP, and the output result of the empty space pyramid module ASPP is input into the decoding structure.

[0028] The further improvement of the present application is that:

[0029] The multi-scale convolution feature fusion structure adopts three convolution kernel sizes, which are 1*1, 3*3 and 5*5.

[0030] The further improvement of the present application is that:

[0031] The empty space pyramid structure ASPP includes four empty convolution layers, and the empty parameters are selected as 1, 2, 4 and 8.

[0032] The step (5) operation includes:

[0033] The collected data are preprocessed to obtain preprocessed data;

[0034] The preprocessed data are input into the optimized network model, and the optimized network model outputs a matrix, and the elements in the matrix are the classification probabilities of each pixel point, and the pixel point is classified according to the probability value of the same pixel point, that is, the automatic identification of the building in the seismic exploration area is realized.

[0035] The second aspect of the present application provides a building automatic identification system in a seismic exploration area.

[0036] The system includes:

[0037] The acquisition unit is used for acquiring the urban remote sensing image dataset about building target segmentation, and performing image preprocessing and dataset division;

[0038] The network construction unit is used for constructing the convolutional neural network;

[0039] The function construction unit is used for constructing the loss function of the convolutional neural network;

[0040] The training unit is connected with the network construction unit and the function construction unit respectively, and is used for training the convolutional neural network to obtain an optimized network model;

[0041] The recognition unit is connected with the training unit, and is used for automatically recognizing the buildings in the seismic exploration work area by using the optimized network model.

[0042] In a third aspect, the present application provides a computer readable storage medium, which stores at least one program executable by a computer, and the at least one program causes the computer to execute the steps of the automatic building recognition method for seismic exploration work area when executed by the computer.

[0043] Compared with the prior art, the present application has the following beneficial effects:

[0044] The present application realizes automatic building recognition of remote sensing images in seismic exploration work area based on deep learning method, introduces multi-scale convolutional feature fusion and an empty space pyramid module ASPP on the basis of the original Unet network structure, enhances the network feature expression ability and obtains semantic information of target multi-size. The present application can more accurately locate the building target and has better segmentation effect on small targets, and has good universality. The present application has very important significance for realizing physical point layout, obstacle avoidance and collection in the field. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 is a step block diagram of the method of the present application;

[0046] Figure 2 is a network structure diagram of the optimized Unet network model provided by the present application;

[0047] Figure 3 is a structure diagram of multi-scale convolutional feature fusion in the present application;

[0048] Fig. 4(a) is an original image in the embodiment of the present application;

[0049] Fig. 4(b) is a label image corresponding to the model in the embodiment of the present application;

[0050] Fig. 4(c) is a building segmentation result obtained by applying the method of the present application in an embodiment of the present application. DETAILED DESCRIPTION

[0051] The present application will be further described in detail below with reference to the accompanying drawings:

[0052] In view of the problems of incomplete target boundary segmentation and inaccurate small target segmentation in building target segmentation in remote sensing images in the prior art, the present application provides a building automatic identification method for remote sensing images of a seismic exploration work area based on deep learning, which is used for fast and high-precision automatic identification of building targets and provides a basis for physical point layout, obstacle avoidance and collection and the like in the field.

[0053] The method of the present application comprises the following steps: collecting remote sensing image data for image preprocessing and dataset division, constructing an optimized Unet network structure for building segmentation of remote sensing images, constructing a network loss function, training the optimized Unet network model using remote sensing image data and optimizing network hyperparameters, testing the network model using a test dataset and calculating evaluation indicators.

[0054] The network structure introduces multi-scale convolution feature fusion and a hollow spatial pyramid module based on the Unet network, enhances the network feature expression capability and obtains more semantic information of the target. By training the network model using remote sensing image data related to city building segmentation, the purpose of more complete building target segmentation and more accurate small target segmentation is achieved.

[0055] The implementation of the method of the present application is as follows:

[0056]

Example One

[0057] As shown in Figure 1 , the method of the present application specifically comprises the following steps:

[0058] S1, collect city remote sensing image data related to building target segmentation, and perform image preprocessing and dataset division;

[0059] Further, the specific steps of image preprocessing and dataset division in step S1 are as follows:

[0060] S11, crop all original images and corresponding label images (both of which are open source public datasets, and the label image is a black and white binary image, wherein black pixel points represent the background and white pixel points represent the building, which is equivalent to labeling each pixel point of the original image. The label image here comes from an open source public dataset.) into small images of a fixed size in a sliding window manner, preferably, the sliding window size is 1024*1024 and the step size is 1024;

[0061] S12, enhancing the original image: the original image is subjected to image enhancement by using existing random hue saturation brightness value transformation, random rotation by 90 degrees, random scaling by 15%-25%, random horizontal flipping and random vertical flipping; if the data enhancement method causes deformation to the original image, the label image corresponding to the original image also needs to be subjected to data enhancement at the same time, because the label image needs to be one-to-one corresponding to the pixel points of the original image.

[0062] S13, the original image and the label image subjected to minimum-maximum normalization processing, so that the pixel values of the original image and the label image are compressed into a preset value range space, to obtain preprocessed original image and label image;

[0063] S14, the preprocessed original image and label image are divided into three data sets, respectively, and the division is randomly divided according to the ratio of 8:1:1, and the three data sets are respectively: experimental training set, validation set and test set.

[0064] S2, constructing a convolutional neural network for remote sensing image building target segmentation;

[0065] The present application introduces the existing multi-scale convolutional feature fusion structure and the empty space pyramid structure ASPP into the Unet neural network to form a new convolutional neural network, as follows:

[0066] The existing Unet network is used to form a basic network framework, including an encoding module and a decoding module. The specific steps of constructing the network structure are as follows:

[0067] S21, in the encoding module, an original Unet encoding structure, a multi-scale convolutional feature fusion structure and an empty space pyramid module ASPP are set.

[0068] The structure organization and data transmission process are as follows:

[0069] As shown in Figure 2 The network model of the optimized Unet of the application is centered on ASPP, the left side is the encoding structure, the right side is the decoding structure, and the lower side is the multi-scale convolutional feature fusion structure.

[0070] The original Unet encoding structure includes five convolutional blocks: Conv_Block1, Conv_Block2, Conv_Block3, Conv_Block4 and Conv_Block5, and the data transmission is sequentially through the five convolutional blocks.

[0071] The multi-scale convolutional feature fusion structure mainly includes four multi-scale convolutional feature fusion modules, which are xx1, xx2, xx3 and xx4.

[0072] The ASPP mainly includes four cavity convolutional layers.

[0073] The decoding structure mainly includes four deconvolution blocks and four feature fusion modules (the four "feature fusion modules" have nothing to do with the aforementioned "multi-scale convolutional feature fusion module", and the "feature fusion module" here is to concatenate the symmetric encoding results, multi-scale convolutional feature fusion results and the output of the corresponding decoding stage), and the deconvolution blocks are Deconv_Block1, Deconv_Block2, Deconv_Block3 and Deconv_Block4.

[0074] The above are all existing structures, but the number of multi-scale convolutional feature fusion modules, the convolution size and the size of the ASPP need to be optimized according to specific data and experimental results.

[0075] The data is input into the encoding structure and the multi-scale convolutional feature fusion structure at the same time. In the encoding structure, the data passes through Conv_Block1, Maxpooling1, Conv_Block2, Maxpooling2, Conv_Block3, Maxpooling3, Conv_Block4, Maxpooling4 and Conv_Block5.

[0076] In the multi-scale convolutional feature fusion structure, the data passes through xx1, xx2, xx3 and xx4 in turn. At the same time, Maxpooling1 and xx1, Maxpooling2 and xx2, Maxpooling3 and xx3, and Maxpooling4 and xx4 are subjected to feature splicing operations (xx1-xx4 are the output results of the multi-scale convolutional feature fusion module) so as to be input into the subsequent symmetric decoding structure for feature fusion.

[0077] The intermediate part is an ASPP structure, the output of the Conv_Block5 is input into the ASPP structure, and the output result is input into the Conv_Block6. The splicing result of Maxpooling4 and xx4 and the output of the Conv_Block6 are spliced in the decoding structure and input into the Deconv_Block1. The splicing result of Maxpooling3 and xx3 and the output of the Deconv_Block1 are spliced and input into the Deconv_Block2. The splicing result of Maxpooling2 and xx2 and the output of the Deconv_Block2 are spliced and input into the Deconv_Block3. The splicing result of Maxpooling1 and xx1 and the output of the Deconv_Block3 are spliced and input into the Deconv_Block4. Then, the classification result is output through a softmax layer.

[0078] S22, as shown in Figure 3 Each multi-scale convolution feature fusion structure adopts three convolution kernel sizes: 1*1, 3*3 and 5*5, performs convolution operation in combination in sequence, selects four parallel operations of down-sampling branches (four down-sampling modules in the encoding stage, multi-scale convolution feature fusion is performed after each down-sampling operation, and then the fusion result and the feature map in the decoding stage are subjected to concatenation operation, and the down-sampling operation and the multi-scale convolution feature fusion operation are parallel.), and then the parallel results are subjected to concatenation operation.

[0079] S23, the hollow space pyramid structure ASPP includes four hollow convolution layers, the hollow parameters are selected as 1, 2, 4 and 8, and finally the last output of the original Unet encoding structure is added to the output feature map of each hollow convolution layer, that is, the feature matrices of the outputs selected by different hollow parameters are added.

[0080] S24, the data processed in step S1 are simultaneously input into the original Unet encoding structure and the multi-scale feature fusion structure, the output matrix of each pooling layer of the original Unet network is subjected to concatenation operation (matrix splicing operation) with the corresponding multi-scale convolution feature fusion result matrix, the output of the Conv_Block5 is input into the hollow space pyramid module ASPP for operation, and the output result is subjected to up-sampling. The feature fusion result of the encoding module Conv_Block4 and xx4 and the output feature map of the Deconv_Block1 are subjected to concatenation feature fusion operation, the fusion result is subjected to subsequent decoding operation, and after four times of up-sampling operation, the building segmentation result is finally output.

[0081] S3, constructing a loss function of the convolutional neural network;

[0082] Further, the specific steps of constructing the loss function in step S3 are as follows:

[0083] The present application adopts the existing loss function DiceLoss, and the specific formula is as follows:

[0084]

[0085] In the formula, K represents the total number of categories in the image, and two categories of building pixel points and non-building pixel points need to be classified, and the value is 2. N represents the total number of pixel points. kn and g kn are probability values, the value range is [0, 1], the former represents the output of the softmax layer (belongs to the decoding structure, the output layer of the network), and the latter represents the label value of the pixel point n predicted to belong to the category k.

[0086] S4, using the training set and the validation set to train the convolutional neural network, and obtaining an optimized network model;

[0087] The training set and the validation set processed by step S1 are combined with the network loss function to learn the parameters of the network model. Through the training result, the learning rate and the training iteration number and other hyperparameters are adjusted according to the optimization strategy, so that the network loss reaches the minimum, and the network loss minimum model is selected as the final building segmentation model under the premise that the network does not overfit and underfit. This step adopts the conventional network parameter adjustment process, which will not be described here;

[0088] S5, using the test data set to test the network model effect, outputting the building target extraction segmentation result, and calculating the model evaluation IOU index;

[0089] Further, for step S5, the specific steps of network model testing and precision evaluation index calculation are as follows:

[0090] S51, writing test code, loading the model, and testing using the test set to obtain the building segmentation result of the test set;

[0091] S52, comparing and calculating the test result image and the label image, using the intersection over union IOU as the evaluation index, and the calculation formula is as follows:

[0092]

[0093] In the formula, k+1 represents the pixel point category (including a background category), P ij represents the total number of pixel points of category i predicted by the model to be pixel points of category j, P jiThe total number of pixel points whose class is j predicted by the model as pixel points whose class is i, P ii The total number of pixel points whose class is i predicted by the model as pixel points whose class is i.

[0094] In actual use, the collected data needs to be preprocessed, specifically including: if the input image size is large, the image is cut into 1024*1024 size using the sliding window method, then the image is subjected to data enhancement, and then input into the optimized network model, the optimized network model outputs a matrix with a size of 1024*1024*2, the elements in the matrix are the classification probabilities of each pixel point, and the pixel point with a larger class probability value is the class of the pixel point.

[0095] The introduction of the multi-scale convolution feature fusion module enables the network to extract feature information of different scales of the target, including small targets, containing more target boundary detail features and global information, and being able to eliminate the influence of factors such as occlusion and light on the segmentation result, enhance the anti-interference ability of the network, and further improve the building segmentation precision. The introduction of the empty space pyramid module enables the network to obtain a larger receptive field, while the image resolution is not lost too much, so that the network extracts features containing more spatial position information. Therefore, the network involved in the present application can achieve the purposes of more complete building target segmentation and more accurate small target segmentation.

[0096] The implementation of the method of the present application is as follows:

[0097]

Example Two

[0098] A remote sensing image data in the test set is selected for illustration, and the method implementation process is as shown in Figure 1 The specific steps are as follows:

[0099] S1, collect city remote sensing image data sets related to building target segmentation, and perform image preprocessing and data set division;

[0100] S2, construct a convolutional neural network model for remote sensing image building target segmentation;

[0101] S3, construct a convolutional neural network loss function;

[0102] S4, use the training set and the validation set to train the network model, and optimize the network hyperparameters;

[0103] S5, use the test data set to test the network model effect, output the building target extraction segmentation result, and calculate the model evaluation IOU index;

[0104] Further, the specific steps of image preprocessing and data set division in step S1 are as follows:

[0105] S11, crop all original images and corresponding label images into small images with fixed size using sliding window method, sliding window size is 1024*1024, step size is 1024;

[0106] S12, perform image enhancement on the original remote sensing image data using random hue saturation lightness value transformation, random rotation by 90 degrees, random scaling by 15%-25%, random horizontal flipping and random vertical flipping, etc.

[0107] S13, perform minimum-maximum normalization processing on the enhanced data, and compress the pixel values of the original image and the label into a preset value range space;

[0108] S14, randomly divide the preprocessed image and label data into an experimental training set, a validation set and a test set according to a ratio of 8:1:1.

[0109] Further, for step S2, the specific steps of constructing a convolutional neural network model for building target segmentation of remote sensing images are as follows:

[0110] The basic network framework is selected to be composed of Unet, and the network contains an encoding module and a decoding module. The network structure is built as shown in Figure 2 The specific steps are as follows:

[0111] S21, in the encoding module, an original Unet encoding structure, a multi-scale convolution feature fusion structure and an empty space pyramid module ASPP are contained. The original Unet encoding structure mainly contains five convolution blocks: Conv_Block1, Conv_Block 2, Conv_Block 3, Conv_Block 4, Conv_Block 5. The multi-scale convolution feature fusion structure mainly contains four multi-scale convolution feature fusion modules: xx1, xx2, xx3, xx4 (the four multi-scale convolution feature fusion modules, Figure 2 from left to right in the multi-scale convolution feature fusion structure are xx1, xx2, xx3, xx4). The empty space pyramid structure ASPP mainly contains four empty convolution layers. The decoder mainly contains four deconvolution blocks and four feature fusion modules, and the deconvolution blocks are: Deconv_Block1, Deconv_Block2, Deconv_Block3, Deconv_Block4.

[0112] S22, the multi-scale convolution feature fusion structure selects three convolution kernel sizes: 1*1, 3*3, 5*5, performs convolution operation in combination, selects four parallel down-sampling branches, and then performs concatenation operation on the parallel results. The multi-scale convolution feature fusion network structure is as shown in Figure 3 .

[0113] S23, the cavity space pyramid structure ASPP includes four cavity convolution layers, and the cavity parameters are 1, 2, 4 and 8 respectively, and finally the last output of the original Unet coding structure is added to the output feature map of each cavity convolution layer.

[0114] S24, the data processed in step S1 are input into the original Unet coding structure and the multi-scale feature fusion structure at the same time, the input results of each pooling layer are subjected to concatenation operation with the corresponding multi-scale convolution feature fusion results, the output of Conv_Block5 is input into the cavity space pyramid module ASPP for operation, and the output result is up-sampled. The feature fusion results of the coding module Conv_Block4 and xx4 and the output feature map of Deconv_Block1 are subjected to concatenation feature fusion operation, and the fusion results are subjected to subsequent decoding operation, and after 4 times of up-sampling operation, the building segmentation result is finally output.

[0115] Further, the specific steps of constructing the loss function in step S3 are as follows:

[0116] The loss function used in the application is DiceLoss, and the specific formula is as follows:

[0117]

[0118] In the formula, K represents the total number of categories in the image, and two target categories of building pixel points and non-building pixel points need to be classified, and the value is 2. N represents the total number of pixel points. kn And g kn are probability values, the value range is [0, 1], the former represents the output of the softmax layer, and the latter represents the label value of the pixel point n predicted to belong to the category k.

[0119] Further, the specific steps of network model training in step S4 are as follows:

[0120] The training set and the validation set processed in step S1 are used in combination with the network loss function to learn the parameters of the network model. Through the training result, the learning rate and the training iteration times and other hyperparameters are adjusted according to the optimization strategy, so that the network loss reaches the minimum, and the network loss minimum model is selected as the final building segmentation model under the premise that the network is not over-fitted and under-fitted;

[0121] Further, for the specific steps of network model testing and precision evaluation index calculation in step S5, they are as follows:

[0122] S51, write test code, load model, test using test set, here select the original image as shown in Figure 4(a) for model test, the corresponding label image as shown in Figure 4(b), obtain the building segmentation result, as shown in Figure 4(c);

[0123] S52, compare the segmentation result image of the test set with the label image, use the intersection over union IOU as the evaluation index, the calculation formula is as follows:

[0124]

[0125] In the formula, k+1 represents the pixel point category (including a background category), P ij represents the total number of pixel points of the category i predicted by the model as the pixel point of the category j, P ji represents the total number of pixel points of the category j predicted by the model as the pixel point of the category i, P ii represents the total number of pixel points of the category i predicted by the model as the pixel point of the category i. The effect comparison of the two models is shown in Table 1, and from Table 1, after adopting the application, the features of the whole image can be more completely extracted.

[0126] Detection method IoU / % Unet 68.14 The method of the invention 75.69

[0127] Table 1

[0128] The application also provides a building automatic identification system for seismic exploration work area, and the implementation of the system is as follows:

[0129]

Example three

[0130] The system comprises:

[0131] The acquisition unit is used for acquiring the urban remote sensing image dataset related to building target segmentation, and performing image preprocessing and dataset division;

[0132] The network construction unit is used for constructing the convolutional neural network;

[0133] The function construction unit is used for constructing the loss function of the convolutional neural network;

[0134] The training unit is connected with the network construction unit and the function construction unit respectively, and is used for training the convolutional neural network to obtain an optimized network model;

[0135] The identification unit is connected with the training unit, and is used for automatically identifying the building in the seismic exploration work area by using the optimized network model.

[0136] The application also provides a computer readable storage medium, and the implementation of the computer readable storage medium is as follows:

[0137] Example 4

[0138] The computer readable storage medium stores at least one program executable by the computer, and the at least one program causes the computer to execute the steps in the automatic building identification method for seismic exploration work area.

[0139] The application can extract the features of the whole image, and segment the target boundary through optimizing Unet, and the application introduces multi-scale convolution feature fusion on the basis of increasing the empty space pyramid module ASPP, which expands the network width, so that the network can extract different size features of the target, thereby enhancing the feature extraction capability of the network.

[0140] In summary, the application introduces multi-scale convolution feature fusion and empty space pyramid module based on Unet network structure, enhances the network feature expression capability and obtains more semantic information of the target. Through training the network model by using the remote sensing image data related to city building segmentation, the application achieves the purpose of more complete building target segmentation and more accurate small target segmentation. The application provides an effective method for fast and high-precision automatic identification of buildings in seismic exploration work area remote sensing images, greatly improves the work efficiency, and has important significance for subsequent seismic exploration work.

[0141] In the description of the application, unless otherwise specified and limited, the terms "connected", "connected" should be understood broadly, for example, it can be fixed connection, or detachable connection, or integral connection, it can be mechanical connection, or electrical connection, it can be direct connection, or indirect connection through intermediate medium.

[0142] In the description of the application, unless otherwise specified, the terms "up", "down", "left", "right", "in", "out" and the like indicate the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore it cannot be understood as a limitation on the application.

[0143] Finally, it should be pointed out that the above technical solutions are only one embodiment of the application, for those skilled in the art, on the basis of the application disclosed application method and principle, various types of improvements or modifications can be easily made, and are not limited to the method described in the above embodiment, therefore the above description is only preferred, and does not have limiting significance.

Claims

1. A method for automatic identification of buildings in a seismic exploration area, characterized in that: The method combines the original Unet network model with a multi-scale convolutional feature fusion structure and a hollow spatial pyramid structure to obtain an optimized network model, and uses the optimized network model to automatically identify buildings in the seismic exploration area. The method includes: Step (1): Collect urban remote sensing image datasets for building target segmentation, and perform image preprocessing and dataset partitioning; Step (2): Construct a convolutional neural network; Step (3): Construct the loss function for the convolutional neural network; Step (4): Train the convolutional neural network to obtain an optimized network model; Step (5): Automatic identification of buildings in the seismic exploration area using an optimized network model; Step (2) includes the following operations: The convolutional neural network is constructed as follows: the left side is the encoding structure of the original Unet network model, the right side is the decoding structure of the original Unet network model, a hollow spatial pyramid module (ASPP) is set between the encoding structure and the decoding structure, and a multi-scale convolutional feature fusion structure is set below the encoding structure. The data is simultaneously input into the encoding structure and the multi-scale convolutional feature fusion structure. Then, the outputs of each pooling layer of the decoding structure and the outputs of each convolutional layer of the multi-scale convolutional feature fusion structure are respectively subjected to feature concatenation operation, and the results of each feature concatenation operation are respectively input into the decoding structure. Simultaneously, the output of the last convolutional layer of the encoding structure is input into the Spatial Pyramid Module (ASPP), and the output of the Spatial Pyramid Module (ASPP) is input into the decoding structure.

2. The method for automatic identification of buildings in seismic exploration areas according to claim 1, characterized in that: The preprocessing operations in step (1) include: (11) Cropping all original images and their corresponding label images into small images of a fixed size; (12) Enhance the original image and the label image; (13) Compress the pixel values ​​of the enhanced original image and label image into the preset value range space to obtain the preprocessed original image and label image.

3. The method for automatic identification of buildings in seismic exploration areas according to claim 1, characterized in that: The dataset partitioning operation in step (1) includes: The preprocessed original images and labeled images are randomly divided into three datasets according to a set ratio: experimental training set, validation set, and test set.

4. The method for automatic identification of buildings in seismic exploration areas according to claim 1, characterized in that: The multi-scale convolutional feature fusion structure employs three convolutional kernel sizes: 1 1, 3 3, 5 5.

5. The method for automatic identification of buildings in seismic exploration areas according to claim 1, characterized in that: The hollow spatial pyramid structure ASPP includes four hollow convolutional layers with hollow parameters of 1, 2, 4 and 8 respectively.

6. The method for automatic identification of buildings in seismic exploration areas according to claim 1, characterized in that: The operation of step (5) includes: The collected data is preprocessed to obtain preprocessed data; The preprocessed data is input into the optimized network model, which outputs a matrix. The elements in the matrix are the probability of each pixel being classified. The pixel with the larger probability value belongs to the class of that pixel, thus realizing the automatic identification of buildings in the seismic exploration area.

7. An automatic building identification system for seismic exploration areas, based on the automatic building identification method for seismic exploration areas as described in any one of claims 1-6, characterized in that: The system includes: Acquisition Unit: Used to acquire urban remote sensing image datasets for building target segmentation, and to perform image preprocessing and dataset partitioning; Network building blocks: used to build convolutional neural networks; Function building unit: used to construct the loss function of a convolutional neural network; Training unit: Connected to the network building unit and function building unit respectively, used to train the convolutional neural network to obtain an optimized network model; Identification unit: Connected to the training unit, used for automatic identification of buildings in the seismic exploration area using an optimized network model.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores at least one computer-executable program, which, when executed by the computer, causes the computer to perform the steps in the automatic building identification method for seismic exploration areas as described in any one of claims 1-6.

Images