Target object detection method, device and server

Abstract

This specification provides a method, device and server for detecting a target object. Based on this method, before specific detection and identification, the high-spectral and high-resolution characteristics of the remote sensing image can be used to perform a minimum noise separation transformation by using a preset high-pass filter based on spectral information to effectively remove background noise and reduce the interference of the background environment with similar colors on the detection and identification of the target object; at the same time, the principal component analysis matrix obtained after the minimum noise separation transformation is used to process the target remote sensing image, so as to achieve data dimension reduction for the target remote sensing image with a large amount of data, and obtain a processed target remote sensing image with a relatively small amount of data suitable for subsequent model processing; during specific detection and identification, the target remote sensing image after the above processing can be used to detect and determine the target object of interest from the target area more efficiently and accurately, thereby improving the detection accuracy of the target object in the image processing scenario with high recognition difficulty.

Description

Technical Field

[0001] This specification belongs to the field of artificial intelligence technology, and in particular, relates to a method, device and server for detecting a target object. Background Art

[0002] In some image processing scenarios, it is often necessary to identify the physical object that the user is interested in from the image.

[0003] However, when the color of the physical object to be identified is close to the background environment, the degree of integration with the background is high, and the image area occupied by the physical object in the image is relatively small compared to the background environment, it is often difficult to accurately identify the target object from the image.

[0004] For the above-mentioned image processing scenarios with high recognition difficulty, when performing detection and recognition based on existing methods, errors such as false detection or missed detection are often prone to occur, resulting in low detection accuracy.

[0005] To address the above problems, no effective solution has been proposed yet. Summary of the invention

[0006] This specification provides a method, device and server for detecting a target object, which can detect and identify the target object of interest from a target area based on a target remote sensing image more efficiently and accurately, thereby improving the detection accuracy of target objects in image processing scenarios with high recognition difficulty.

[0007] This specification provides a method for detecting a target object, including:

[0008] Obtain the target remote sensing image of the target area in the current time period;

[0009] Calling a preset high-pass filter to process the target remote sensing image to obtain a noise covariance matrix; wherein the preset high-pass filter is established based on the spectral information of the target object and the spectral information of the background environment in the remote sensing image;

[0010] According to the noise covariance matrix, a minimum noise separation transformation is performed to obtain a principal component analysis matrix of the target remote sensing image;

[0011] Processing the target remote sensing image using the principal component analysis matrix to obtain a processed target remote sensing image;

[0012] The target object in the target area of ​​the current time period is determined according to the processed target remote sensing image.

[0013] In one embodiment, the target object includes a physical object whose color difference between the target area and the background environment is less than a preset difference threshold.

[0014] In one embodiment, the preset high-pass filter is established in the following manner:

[0015] Acquire a remote sensing image targeting a target area and containing a target object as a test image;

[0016] Determining in the test image a first image region where the target object is located and a second image region where the background environment is located;

[0017] Performing spectral information statistics on the first image area to obtain spectral information statistics results of the target object; performing spectral information statistics on the second image area to obtain spectral information statistics results of the background environment;

[0018] The preset high-pass filter is constructed according to the statistical results of the spectral information of the target object and the statistical results of the spectral information of the background environment.

[0019] In one embodiment, a minimum noise separation transformation is performed according to the noise covariance matrix to obtain a principal component analysis matrix of the target remote sensing image, including:

[0020] Constructing a first transformation matrix according to the noise covariance matrix;

[0021] Performing a first transformation process on the noise covariance matrix using the first transformation matrix to obtain a transformed data matrix;

[0022] constructing a second transformation matrix according to the transformed data matrix;

[0023] constructing a minimum noise separation transformation matrix according to the first transformation matrix and the second transformation matrix;

[0024] The target remote sensing image is processed using the minimum noise separation transformation matrix to obtain a principal component analysis matrix of the target remote sensing image.

[0025] In one embodiment, the target remote sensing image is processed using the principal component analysis matrix to obtain a processed target remote sensing image, including:

[0026] Determining the principal component bands in the target remote sensing image according to the principal component analysis matrix;

[0027] According to the principal component bands, corresponding spectral information is extracted from the target remote sensing image to generate a corresponding RGB pseudo-color picture as the processed target remote sensing image.

[0028] In one embodiment, determining the target object in the target area in the current time period according to the processed target remote sensing image includes:

[0029] According to a preset cutting rule, the target remote sensing image is cut into a plurality of sub-image blocks; wherein there is an overlapping area of ​​a preset range between two adjacent sub-image blocks in the plurality of sub-image blocks;

[0030] Calling a preset processing model to process the multiple sub-image blocks to obtain an image processing result corresponding to the current time period;

[0031] According to the image processing result, the target objects in the target area in the current time period and the number of the target objects are determined.

[0032] In one embodiment, the preset processing model at least includes a preset feature extraction network; wherein the preset feature extraction network includes: a main network, a first sub-network, a second sub-network, and a fusion structure; wherein,

[0033] The main network is used to extract a high-resolution feature map; the first sub-network is used to extract a first type of low-resolution feature map; the second sub-network is used to extract a second type of low-resolution feature map; the fusion structure is used to fuse the high-resolution feature map, the first type of low-resolution feature map, and the second type of low-resolution feature map to obtain a fused feature map.

[0034] In one embodiment, the preset processing model also includes a preset regression and classification network; the preset regression and classification network is used to locate multiple anchor frames based on the fused feature map, and determine the predicted probability of the existence of the target object in each anchor frame of the multiple anchor frames.

[0035] In one embodiment, the preset regression and classification network includes: a first convolutional layer, and a second convolutional layer and a third convolutional layer connected to the first convolutional layer; wherein the first convolutional layer is used to process the fused feature map and output the corresponding intermediate data; the second convolutional layer is used to locate multiple anchor frames by finding a combination of multiple coordinate points in the target area according to the intermediate data; the third convolutional layer is used to calculate the predicted probability of the existence of the target object in the anchor frame according to the intermediate data.

[0036] In one embodiment, the target object includes a wind power station, and the target area includes a snow area.

[0037] In one embodiment, after determining the target objects in the target area in the current time period and the number of the target objects, the method further includes:

[0038] Obtaining target objects in a target area of ​​a previous time period determined based on a target remote sensing image of a previous time period, and the number of target objects;

[0039] The construction progress of the wind power station in the snow area is determined based on the target objects in the target area in the previous time period and the number of target objects; and the target objects in the target area in the current time period and the number of target objects.

[0040] This specification also provides a detection device for a target object, comprising:

[0041] An acquisition module is used to acquire a target remote sensing image of a target area in the current time period;

[0042] A first processing module is used to call a preset high-pass filter to process the target remote sensing image to obtain a noise covariance matrix; wherein the preset high-pass filter is established based on the spectral information of the target object and the spectral information of the background environment in the remote sensing image;

[0043] A transformation module, used for performing minimum noise separation transformation according to the noise covariance matrix to obtain a principal component analysis matrix of the target remote sensing image;

[0044] A second processing module is used to process the target remote sensing image using the principal component analysis matrix to obtain a processed target remote sensing image;

[0045] The determination module is used to determine the target object in the target area of ​​the current time period according to the processed target remote sensing image.

[0046] This specification also provides a server, including a processor and a memory for storing processor executable instructions, and the processor implements the relevant steps of the target object detection method when executing the instructions.

[0047] This specification also provides a computer-readable storage medium on which computer instructions are stored, and when the instructions are executed, the relevant steps of the target object detection method are implemented.

[0048] The present specification provides a method, device and server for detecting a target object. After obtaining a target remote sensing image of a target area in a current time period, a preset high-pass filter established based on the spectral information of the target object and the spectral information of the background environment in the remote sensing image can be first called to filter the target image to obtain a noise covariance matrix; based on the noise covariance matrix, a minimum noise separation transformation is performed to obtain a principal component analysis matrix of the target remote sensing image; then, the principal component analysis matrix can be used to process the target remote sensing image to obtain a processed target remote sensing image; and based on the processed target remote sensing image, the target object in the target area in the current time period is detected and determined. In this way, the high-spectral and high-resolution characteristics of remote sensing images can be fully utilized. By using a preset high-pass filter based on spectral information to perform a minimum noise separation transformation, background noise can be effectively removed, and the interference of background environments with similar colors on target object detection and recognition can be reduced; at the same time, the principal component analysis matrix is ​​used to process the target remote sensing image to achieve data dimensionality reduction for target remote sensing images with a large data volume, and a processed target remote sensing image with a relatively small data volume and suitable for model processing is obtained; and then, based on the above-mentioned processed target remote sensing image, the target object of concern can be detected and identified from the target area more efficiently and accurately, thereby improving the detection accuracy of target objects in image processing scenarios with high recognition difficulty. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of this specification, the drawings required for use in the embodiments will be briefly introduced below. The drawings described below are only some embodiments recorded in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0050] Figure 1 It is a schematic diagram of an embodiment of the structural composition of a system using the target object detection method provided by the embodiments of this specification;

[0051] Figure 2 is a flowchart of a method for detecting a target object provided by an embodiment of this specification;

[0052] Figure 3 is a schematic diagram of an embodiment of a method for detecting a target object provided by an embodiment of this specification, in a scenario example;

[0053] Figure 4 It is a schematic diagram of the structure of a server provided by an embodiment of this specification;

[0054] Figure 5 is a schematic diagram of the structure of a target object detection device provided by an embodiment of this specification;

[0055] Figure 6 It is a schematic diagram of an embodiment of a target object detection method provided by an embodiment of this specification, in a scenario example. DETAILED DESCRIPTION

[0056] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this specification.

[0057] The present invention provides a method for detecting a target object, which can be applied to a system including a server and a ground station device. Figure 1 The server and the ground station equipment can be connected by wire or wirelessly to perform specific data interaction.

[0058] In this embodiment, the server may specifically include a background server applied to a data processing platform side that can realize functions such as data transmission and data processing. Specifically, the server may be, for example, an electronic device with data calculation, storage and network interaction functions. Alternatively, the server may also be a software program running in the electronic device to provide support for data processing, storage and network interaction. In this embodiment, the number of servers included in the server is not specifically limited. The server may specifically be one server, or several servers, or a server cluster formed by several servers.

[0059] In this embodiment, the ground station equipment may specifically include a front-end electronic device deployed on the ground, capable of communicating with a satellite and realizing functions such as data collection and data transmission.

[0060] In this embodiment, it is currently planned to build multiple wind power stations in batches in the XX snow area. Among them, the XX snow area is relatively wide, and almost all areas are covered with snow. The wind power station to be built is also white, and the color is close to the background environment of the XX snow area, which is easy to blend into the background environment, making it difficult to identify.

[0061] Currently, users need to monitor the construction progress of wind power stations in order to better manage the construction of wind power stations.

[0062] The satellite can be set to regularly collect remote sensing images of the entire area including the XX snow area (target area) at preset time intervals (for example, 1 day, etc.) as target remote sensing images; and send the collected target remote sensing images to the ground station equipment.

[0063] The ground station device receives the target remote sensing image of the target area in the current time period from the satellite, and sends the target remote sensing image to the server for specific processing via wired or wireless means. Correspondingly, the server can obtain the target remote sensing image of the target area in the current time period.

[0064] Before receiving the target remote sensing image of the current time period, the server can first obtain a remote sensing image of the target area and containing at least one target object as a test image. Considering that the remote sensing image itself has the characteristics of hyperspectral, it is considered that the hyperspectral characteristics of the remote sensing image can be utilized to pre-build a preset high-pass filter that can effectively remove background noise in scenes with high recognition difficulty based on the above test image.

[0065] Specifically, the server can first determine and select in the test image a first image area where the target object is located and can more typically reflect the characteristics of the target object, and a second image area where the background environment is located and can more typically reflect the characteristics of the background environment. Then, spectral information statistics are performed on the first image area and the second image area respectively to obtain the spectral information statistics results of the target object and the spectral information statistics results of the background environment. Furthermore, a preset high-pass filter can be constructed based on the above two spectral information statistics results.

[0066] The server uses the above preset high-pass filter to filter the target remote sensing image first, so that the image pixel blocks with similar spectral properties to the background environment in the target remote sensing image are first filtered and the corresponding noise covariance matrix is ​​obtained. Specifically, it can be recorded as: C N .

[0067] Next, the server can use the above noise covariance matrix to perform minimum noise separation transformation on the target remote sensing image to remove background noise to a large extent, so that the elements in the vector are independent of each other, and the principal components in the target remote sensing image are retained, so that the data dimension of the image is greatly reduced, and the principal component analysis matrix of the target remote sensing image is arranged from large to small according to the signal-to-noise ratio. This can effectively reduce the impact of background noise on image quality and obtain a relatively pure principal component analysis matrix with relatively good use effect.

[0068] When performing the minimum noise separation transformation, first, the server can construct the corresponding first transformation matrix P according to the noise covariance matrix. Specifically, the server can first transform CN The eigenvalues ​​of are arranged in descending order to obtain the corresponding diagonal matrix D N ; Then determine the orthogonal matrix U constructed based on the eigenvector; and according to the diagonal matrix D N And the orthogonal matrix U constructs the corresponding first transformation matrix according to the following formula:

[0069] Next, the server may use the first transformation matrix to perform a first transformation on the noise covariance matrix to obtain a transformed data matrix. Specifically, the server may use the first transformation matrix to perform a first transformation on the noise covariance matrix C N Perform the first transformation processing, perform the standard principal component transformation on the noise data to project the noise data into a new feature space, and obtain the transformed data matrix C D-adj This will make the noise data have unit variance and no longer have correlation in the band, making it easier to remove it later. For details, please refer to the following formula: C D-adj =P T C N P.

[0070] Then, the server may construct a second transformation matrix V according to the transformed data matrix. Specifically, the server may first determine a diagonal matrix D with eigenvalues ​​arranged in descending order according to the transformed data matrix: D-adj ; Then, according to the transformed data matrix and the diagonal matrix, the second transformation matrix V is determined according to the following formula: D D-adj =V T C D-adj V.

[0071] Furthermore, the server may construct a corresponding minimum noise separation transformation matrix T according to the first transformation matrix and the second transformation matrix. Specifically, the minimum noise separation transformation matrix may be constructed according to the following formula: T=PV.

[0072] Finally, the server can use the minimum noise separation transformation matrix to process the target remote sensing image to obtain a principal component analysis matrix of the target remote sensing image that meets the requirements and has high quality and good use effect.

[0073] After obtaining the principal component analysis matrix, it is considered that remote sensing images are not suitable for model processing due to their high data dimension and large data volume. As a result, if the remote sensing images are directly input into the model for processing, there will often be problems such as long processing time and poor processing effect. It is precisely because of the above problems that the server first uses the principal component analysis matrix to process the target remote sensing image to achieve data dimensionality reduction, and obtains the processed target remote sensing image with relatively low dimension and relatively small data volume, which is more suitable for subsequent model processing.

[0074] Specifically, the server can first determine the principal component bands in the target remote sensing image according to the principal component analysis matrix; then extract the corresponding spectral information from the target remote sensing image according to the principal component bands to generate a corresponding RGB pseudo-color image as the processed target remote sensing image. In this way, the original high-spectral color target remote sensing image can be converted into a corresponding RGB pseudo-color image, thereby achieving data dimensionality reduction. The subsequent use of the above-mentioned RGB pseudo-color image after dimensionality reduction instead of the remote sensing image for subsequent processing can effectively improve the overall model processing efficiency.

[0075] After the server obtains the processed target remote sensing image in the above manner, considering that the remote sensing image after dimensionality reduction still has a high resolution and a large amount of data, the subsequent model processing is still relatively complex and cumbersome. Therefore, the processed target remote sensing image can be further cut into multiple smaller sub-image blocks, so that the model processing efficiency can be further improved when the preset processing model based on the neural network is used for specific processing.

[0076] Specifically, the server can cut the target remote sensing image into multiple sub-image blocks according to a preset cutting rule. In addition, in the process of cutting, in order to avoid cutting the target object into different sub-image blocks, which may cause errors in subsequent processing and identification, an overlapping area can be reserved between two adjacent sub-image blocks during cutting, and the range size of the overlapping area is set to a preset range. The preset cutting rule includes a parameter value of a preset range. The parameter value of the preset range can be determined specifically based on the step size of the cutting used and / or the size parameters of the active window, combined with the size parameters of the RGB pseudo-color image.

[0077] When performing cutting according to the preset cutting rules, if it is found that the remaining image area of ​​the RGB pseudo-color image is not enough to be cut into a sub-image block, the server can perform a zero padding operation (padding) on ​​the RGB pseudo-color image to fill the insufficient image area with a black image, and then perform cutting. In this way, multiple sub-image blocks that are more suitable for subsequent processing by the preset processing model can be cut.

[0078] After obtaining the above-mentioned multiple sub-image blocks based on the processed target remote sensing image, the server can input the above-mentioned multiple sub-image blocks into a pre-trained preset processing model based on a neural network, and run the preset processing model to obtain model outputs for the multiple sub-image blocks as image processing results corresponding to the current time period.

[0079] The server can determine, based on the image processing result, an anchor frame with a predicted probability of the existence of a target object greater than a preset probability threshold as a target anchor frame containing the detected and identified target object (i.e., the wind power station established in the current time period). Then, based on the coordinate information carried by the target anchor frame, the location information of the target object in the target area is determined. At the same time, the server can also determine the number of target objects in the current target area by counting the number of target anchor frames.

[0080] The server may store the location information of the target objects identified and determined during the current time period and the number of the target objects in a preset progress record table for preservation.

[0081] Furthermore, the server may also query the above-preset progress record table to obtain the target objects and the number of target objects in the target area of ​​the previous time period determined based on the target remote sensing image of the previous time period. Then, by comparing the number of target objects in the target area of ​​the previous time period with the number of target objects in the target area of ​​the current time period, the construction progress of the wind power station in the XX snowy area in the time period between the previous time period and the current time period can be determined.

[0082] According to the above method, the server can regularly detect and determine the construction progress of the wind power station in the XX snowy area in each time period. And timely determine whether the current actual construction situation meets the construction plan based on the determined construction progress of the wind power station in the XX snowy area. When it is found that the construction situation does not meet the construction plan, prompt information can be generated in time to prompt, so that the construction of the wind power station can be supervised in a timely and effective manner.

[0083] Through the above system, the target object of interest can be detected and identified from the target area more efficiently and accurately, improving the detection accuracy of target objects in image processing scenarios with high recognition difficulty, and accurately determining and tracking the changing trends of target objects in the target area.

[0084] See also Figure 2 As shown, the embodiment of this specification provides a method for detecting a target object. The method is specifically applied to the server side. When implemented specifically, the method may include the following contents:

[0085] S201: Acquire a target remote sensing image of a target area in the current time period;

[0086] S202: calling a preset high-pass filter to process the target remote sensing image to obtain a noise covariance matrix; wherein the preset high-pass filter is established based on the spectral information of the target object and the spectral information of the background environment in the remote sensing image;

[0087] S203: performing minimum noise separation transformation according to the noise covariance matrix to obtain a principal component analysis matrix of the target remote sensing image;

[0088] S204: Processing the target remote sensing image using the principal component analysis matrix to obtain a processed target remote sensing image;

[0089] S205: Determine the target object in the target area in the current time period according to the processed target remote sensing image.

[0090] Through the above-mentioned embodiments, the high-spectral and high-resolution characteristics of remote sensing images can be fully utilized. By using a preset high-pass filter established based on spectral information to perform a minimum noise separation transformation, background noise can be effectively removed, and the interference of background environments with similar colors on target object detection and recognition can be reduced; at the same time, by using a principal component analysis matrix to process the target remote sensing image, data dimension reduction can be achieved for target remote sensing images with a large data volume, and processed target remote sensing images with a relatively small data volume and suitable for subsequent model processing can be obtained; and then, based on the above-mentioned processed target remote sensing image, the target object of concern to the user can be detected and identified from the target area more efficiently and accurately, thereby improving the detection accuracy of target objects in image processing scenarios with high recognition difficulty.

[0091] In some embodiments, the target remote sensing image provided by the satellite for the target area can be acquired every preset time period (e.g., 1 week, etc.), wherein each target remote sensing image corresponds to a time period, and is used to reflect the specific situation of the target object in the target area in the corresponding time period.

[0092] In some embodiments, the target object may specifically include a physical object whose color difference between the target area and the background environment is less than a preset difference threshold.

[0093] Through the above embodiments, the target object detection method provided in this specification can be applied to more accurately detect and identify target objects in image processing scenarios where identification is more difficult.

[0094] In some embodiments, the specific target object may be a white wind power station in a white snowy area, a yellow building in a yellow desert area, or a green camouflage vehicle in a green forest area. For the above-mentioned physical objects in the above-mentioned areas, since the color of the physical objects is close to the color of the surrounding background environment, it is difficult to clearly distinguish the contours of the physical objects from the background environment, so that the physical objects are easily integrated into the background environment in the image, and the recognition is difficult. For the detection and identification of target objects in the above-mentioned various scenarios, the target object detection method provided in this specification can be well applied.

[0095] In some embodiments, the target remote sensing image can be specifically understood as a remote sensing image collected for a wide target area. Usually, the target remote sensing image has the characteristics of high spectrum, high resolution, high data dimension, and large data volume.

[0096] In some embodiments, the preset high-pass filter can be pre-established in the following manner:

[0097] S1: Acquire a remote sensing image of the target area and containing the target object as a test image;

[0098] S2: determining in the test image a first image region where the target object is located and a second image region where the background environment is located;

[0099] S3: performing spectral information statistics on the first image area to obtain spectral information statistics results of the target object; performing spectral information statistics on the second image area to obtain spectral information statistics results of the background environment;

[0100] S4: constructing the preset high-pass filter according to the statistical results of the spectral information of the target object and the statistical results of the spectral information of the background environment.

[0101] Through the above embodiments, the hyperspectral characteristics of the remote sensing image can be fully utilized, and a preset high-pass filter that can effectively filter background environmental noise and has a good effect can be constructed based on the test image.

[0102] In some embodiments, during specific implementation, the preset high-pass filter may be used to filter the target image, so as to filter out background noise using the hyperspectral characteristics of the remote sensing image and obtain a corresponding noise covariance matrix.

[0103] In some embodiments, the noise covariance matrix can be denoted as C N . In which, each matrix element of the covariance matrix is ​​the covariance between each vector element.

[0104] In some embodiments, the above-mentioned minimum noise separation transformation is performed according to the noise covariance matrix to obtain the principal component analysis matrix of the target remote sensing image. When implemented specifically, the following contents may be included:

[0105] S1: constructing a first transformation matrix according to the noise covariance matrix;

[0106] S2: Performing a first transformation process on the noise covariance matrix using the first transformation matrix to obtain a transformed data matrix;

[0107] S3: constructing a second transformation matrix according to the transformed data matrix;

[0108] S4: constructing a minimum noise separation transformation matrix according to the first transformation matrix and the second transformation matrix;

[0109] S5: Processing the target remote sensing image using the minimum noise separation transformation matrix to obtain a principal component analysis matrix of the target remote sensing image.

[0110] Through the above embodiments, the target remote sensing image can be subjected to minimum noise separation transformation by constructing and utilizing the noise covariance matrix, so as to remove background noise to a large extent, obtain that the elements in the vector are independent of each other, and retain the principal components in the target remote sensing image, so that the data dimension of the image is reduced, and the principal component analysis matrix of the target remote sensing image is arranged from large to small according to the signal-to-noise ratio. In this way, the influence of background noise on image quality can be effectively reduced, and a relatively pure principal component analysis matrix with relatively good effect can be obtained. Among them, the above principal component analysis matrix can be specifically understood as a matrix containing the band information of the principal component bands in the target remote sensing image.

[0111] In some embodiments, when constructing the first transformation matrix, the server may transform C N The eigenvalues ​​of are arranged in descending order to obtain the corresponding diagonal matrix D N ; Determine the orthogonal matrix U constructed based on the eigenvector; and according to the diagonal matrix D N And the orthogonal matrix U constructs the corresponding first transformation matrix according to the following formula:

[0112] In some implementation examples, when performing the first transformation process, the server may use the first transformation matrix to transform the noise covariance matrix C according to the following formula: N Perform the first transformation: C D-adj =P T C N P; perform standard principal component transformation on the noise data, project the noise data into a new feature space, and obtain the transformed data matrix C D-adj This makes the noise data have unit variance and no longer correlated in the band, making it easier to remove it later.

[0113] In some embodiments, when constructing the second transformation matrix, the server may determine the diagonal matrix D in which the eigenvalues ​​are arranged in descending order according to the transformed data matrix: D-adj ; Based on the transformed data matrix and the diagonal matrix, the second transformation matrix V is constructed according to the following formula: D D-adj =V T C D-adj V.

[0114] In some embodiments, when specifically constructing the minimum noise separation transformation matrix, the server may construct the corresponding minimum noise separation transformation matrix using the first transformation matrix and the second transformation matrix according to the following formula: T=PV.

[0115] In some embodiments, the above-mentioned use of the principal component analysis matrix to process the target remote sensing image to obtain the processed target remote sensing image may include the following contents during specific implementation:

[0116] S1: Determine the principal component bands in the target remote sensing image according to the principal component analysis matrix;

[0117] S2: extracting corresponding spectral information from the target remote sensing image according to the principal component bands to generate a corresponding RGB pseudo-color image as the processed target remote sensing image.

[0118] Through the above embodiments, the hyperspectral characteristics of remote sensing images can be utilized and the principal component analysis matrix can be used to perform corresponding processing on the target remote sensing image to achieve data dimensionality reduction of the image, thereby obtaining a processed target remote sensing image with relatively low dimensionality, relatively small data volume, and more suitable for subsequent model processing.

[0119] In some embodiments, the above-mentioned determining the target object in the target area of ​​the current time period according to the processed target remote sensing image may include the following contents when implemented:

[0120] S1: cutting the target remote sensing image into a plurality of sub-image blocks according to a preset cutting rule; wherein there is an overlapping area of ​​a preset range between two adjacent sub-image blocks among the plurality of sub-image blocks;

[0121] S2: calling a preset processing model to process the plurality of sub-image blocks to obtain an image processing result corresponding to the current time period;

[0122] S3: Determine the target objects and the number of the target objects in the target area in the current time period according to the image processing result.

[0123] Through the above embodiments, the processed target remote sensing image can be further divided into multiple sub-image blocks with relatively smaller data volume and more suitable for subsequent model processing, and then the above-mentioned multiple sub-image blocks can be input into the preset processing model instead of the complete processed target remote sensing image for specific model processing to improve the efficiency of subsequent model processing.

[0124] It should be noted that even after the dimensionality reduction processing, the image size and data volume of the target remote sensing image are still relatively large. At this time, if the above-mentioned processed target remote sensing image is directly input into the preset processing model for processing, the preset processing model will often first compress the size of the above-mentioned processed target remote sensing image, resulting in the loss of image data, and then perform recognition based on the compressed image. However, the target object itself is relatively small relative to the target area, and the color is close to the color of the background environment of the target area. In this case, recognition based on the compressed image is relatively more prone to errors. In this embodiment, the processed target remote sensing image with a larger image size is pre-divided into a plurality of sub-image blocks with smaller image sizes, and then input into the preset processing model for processing. In this way, the preset processing model does not need to perform size compression, avoiding the loss of image data caused by compression.

[0125] In some embodiments, during the specific segmentation, the target remote sensing image can be cut into multiple sub-image blocks according to a preset cutting rule. In addition, during the cutting process, in order to avoid cutting a target object into different sub-image blocks, which may cause errors in subsequent processing and identification, an overlapping area can be reserved between two adjacent sub-image blocks during cutting according to the preset cutting rule, and the range size of the overlapping area is set to a preset range. The preset cutting rule includes a parameter value of a preset range. The parameter value of the preset range can be determined specifically according to the step size of the cutting used and / or the size parameters of the active window, and in combination with a variety of data such as the size parameters of the RGB pseudo-color image and the size parameters of the target object.

[0126] In some embodiments, when performing cutting according to a preset cutting rule, if it is found that the remaining image area of ​​the RGB pseudo-color image is insufficient to be cut into a sub-image block, the server can also perform a zero padding operation (padding) on ​​the RGB pseudo-color image to fill the insufficient image area with a black image, and then continue cutting. Thus, multiple sub-image blocks that are more suitable for subsequent processing by a preset processing model can be cut.

[0127] In some embodiments, the above-mentioned calling of a preset processing model to process the multiple sub-image blocks to obtain image processing results corresponding to the current time period may include: the server may use the above-mentioned multiple sub-image blocks as model inputs, input them into a pre-trained preset processing model based on a neural network, and run the preset processing model to obtain model outputs for the above-mentioned multiple sub-image blocks as the corresponding image processing results for the current time period.

[0128] In some embodiments, the image processing result may specifically include an anchor frame of the target object that may be contained and identified by a preset recognition model, and a predicted probability of the target object existing in the anchor frame. The anchor frame may specifically be a frame structure composed of four coordinates. In addition, the anchor frame also carries location information corresponding to the target area (e.g., latitude and longitude coordinate information).

[0129] In some embodiments, the above-mentioned determination of the target objects and the number of target objects in the target area of ​​the current time period based on the image processing results may include: comparing the predicted probability of the existence of the target object in each anchor frame with a preset probability threshold; based on the comparison result, determining an anchor frame with a preset probability greater than a preset probability threshold from multiple anchor frames as a target anchor frame where the target object exists. Further, the position of the target object in the target area may be determined based on the position information carried by the target anchor frame; and the number of target objects in the target area may also be determined by counting the number of target anchor frames.

[0130] In some embodiments, the above-mentioned preset processing model can be specifically understood as a pre-trained neural network model that can determine and identify the anchor frame where the target object may exist in the target remote sensing image based on multiple input sub-image blocks, and determine the predicted probability of the target object existing in the anchor frame.

[0131] In some embodiments, the preset processing model may include at least a preset feature extraction network; wherein the preset feature extraction network may specifically include: a main network, a first sub-network, a second sub-network, and a fusion structure; for details, please refer to Figure 3 As shown. The main network is used to extract high-resolution feature maps; the first sub-network is used to extract the first type of low-resolution feature maps; the second sub-network is used to extract the second type of low-resolution feature maps; the fusion structure is used to fuse the high-resolution feature maps, the first type of low-resolution feature maps, and the second type of low-resolution feature maps to obtain a fused feature map.

[0132] Through the above embodiments, the preset processing model can effectively extract a fused feature map (or feature heat map) that simultaneously retains and fuses the high-resolution features and low-resolution features in the remote sensing image based on a preset feature extraction network, and then a more accurate image processing result can be obtained based on the above fused feature map.

[0133] In some embodiments, the main network may include multiple convolutional layers in series. The first sub-network and the second sub-network may be down-sampled by 2 times and 4 times, respectively, to obtain the corresponding first type low-resolution feature map and second type low-resolution feature map, respectively.

[0134] In some embodiments, in a preset feature extraction network, the main network, the first sub-network, and the second sub-network may be connected in parallel to each other, and each is connected to the fusion structure after passing through three convolution layers. In this way, the high-resolution feature map, the first type of low-resolution feature map, and the second type of low-resolution feature map output from the main network, the first sub-network, and the second sub-network, respectively, can flow into the fusion structure for fusion after three convolution operations to obtain a fused feature map.

[0135] In some embodiments, the fusion structure may be specifically configured to be fused using a preset fusion method, wherein the preset fusion method may specifically include: pixel-by-pixel addition fusion (eg, elementwise-add), and / or channel stacking fusion.

[0136] In some embodiments, the preset processing model may further include a preset regression and classification network. Figure 3 As shown, the preset regression and classification network is used to locate multiple anchor frames according to the fused feature map, and determine the predicted probability that the target object exists in each anchor frame of the multiple anchor frames.

[0137] Through the above embodiments, the preset processing model is based on the preset regression and classification network, and can accurately locate multiple anchor frames where target objects may exist according to the fused feature map output by the preset feature extraction network, and determine the predicted probability of the target object existing in each anchor frame, thereby obtaining a more accurate image processing result.

[0138] In some embodiments, the above-mentioned preset regression and classification network can be specifically understood as a pre-trained neural network structure that can locate the anchor frame where the target object exists in the image and determine the predicted probability of the target object existing in the anchor frame.

[0139] In some embodiments, the preset regression and classification network may specifically include: a first convolutional layer, and a second convolutional layer and a third convolutional layer connected to the first convolutional layer. Figure 3 As shown. The first convolution layer can be used to process the fused feature map output intermediate data; the second convolution layer can be used to locate multiple anchor frames by finding a combination of multiple coordinate points in the target area according to the intermediate data; the third convolution layer can be used to calculate the predicted probability of the target object existing in the anchor frame according to the intermediate data. The second convolution layer is connected to the first convolution layer, and the third convolution layer is connected to the first convolution layer.

[0140] Through the above embodiment, using the above preset regression and classification network, it is possible to simultaneously and efficiently determine multiple anchor frames where target objects may exist, and accurately determine the predicted probability of the target object existing in each anchor frame.

[0141] In some embodiments, specifically, the first convolution layer may include a convolution layer of size 3*3. The second convolution layer may include a convolution layer of size 1*1. The third convolution layer may include a convolution layer of size 1*1. The second convolution layer is used to find the anchor frame composed of 4 coordinate points and output N*4 data. The first convolution layer is used to predict the probability of the target object existing in the anchor frame and output N*2 data.

[0142] In some embodiments, before specific implementation, when training the above-mentioned preset regression and classification network, a box that actually contains the target object can be marked in the sample image, recorded as the Ground thuth box, to obtain the marked sample image. Then, the sample image after the above-mentioned marking is input into the initial regression and classification network to output the initial basic anchor box. The overlap (IOU value) between the basic anchor box and the Ground thuth box is calculated. Further, the marked sample image with an overlap greater than the first overlap (for example, 0.7) in the marked sample image is determined and identified as a positive sample; the marked sample image with an overlap less than the second overlap (for example, 0.3) is determined and identified as a negative sample. Then, based on the positive samples and negative samples, the initial regression and classification network is trained, and the corresponding network loss is calculated. According to the network loss, the network parameters are adjusted in a targeted manner to obtain a preset regression and classification network that meets the requirements.

[0143] In some embodiments, the server may specifically call corresponding GPU resources to run a preset processing model so as to obtain corresponding image processing results more efficiently.

[0144] In some embodiments, the target object may specifically include a wind power station, and the target area may specifically include a snow area.

[0145] Through the above embodiments, based on the method provided in this specification, remote sensing images can be used to accurately identify wind power stations in snowy areas that are difficult to identify.

[0146] In some embodiments, after the server determines the target objects in the target area and the number of target objects in a time period, the server can save the above data in a preset database in the form of a json file, and identify the time period corresponding to the json file when saving.

[0147] In some embodiments, after determining the target objects and the number of target objects in the target area of ​​the current time period, the method may also include the following contents when implemented: obtaining the target objects and the number of target objects in the target area of ​​the previous time period determined based on the target remote sensing image of the previous time period; determining the construction progress of the wind power station in the snow area based on the target objects and the number of target objects in the target area of ​​the previous time period; and the target objects and the number of target objects in the target area of ​​the current time period.

[0148] Through the above embodiment, the server can accurately determine the construction progress of the wind power station in the snow area by comparing the target objects in the target area of ​​the current time period with the target objects in the target area of ​​the previous time period, thereby providing a better reference basis for the remote construction management of the wind power station.

[0149] In some embodiments, after determining the target objects in the target area of ​​the current time period and the number of target objects, the target objects in the target area of ​​multiple historical times before the current time period and the number of target objects can be obtained by querying the json file saved in the preset database. Then, the above data can be combined to more thoroughly determine the change trend of the target objects in the target area over a longer time period. And based on the above change area, the change of the target objects in the next time period is predicted.

[0150] In some embodiments, a preset processing model may be pre-built and trained before implementation. The specific implementation may include the following: constructing an initial processing model; wherein the initial processing model at least includes an initial feature extraction network, an initial regression and classification network; obtaining a sample image, and annotating the sample image accordingly to obtain an annotated sample image; using the annotated sample image, training the initial processing model to obtain a preset processing model that meets the accuracy requirements.

[0151] As can be seen from the above, the target object detection method provided in the embodiment of the present specification can first call a preset high-pass filter established based on the spectral information of the target object and the spectral information of the background environment in the remote sensing image to filter the target image after obtaining the target remote sensing image of the target area in the current time period to obtain a noise covariance matrix; perform a minimum noise separation transformation based on the noise covariance matrix to obtain a principal component analysis matrix of the target remote sensing image; then use the principal component analysis matrix to process the target remote sensing image to obtain a processed target remote sensing image; and determine the target object in the target area of ​​the current time period based on the processed target remote sensing image. In this way, the high-spectral and high-resolution characteristics of remote sensing images can be fully utilized. By using a preset high-pass filter established based on spectral information to perform a minimum noise separation transformation, background noise can be effectively removed, and the interference of background environments with similar colors on target object detection and recognition can be reduced; at the same time, the principal component analysis matrix is ​​also used to process the target remote sensing image to achieve data dimensionality reduction for target remote sensing images with a large amount of data, and a processed target remote sensing image with a relatively small amount of data suitable for model processing is obtained; and then, based on the above-mentioned processed target remote sensing image, the target object of concern can be detected and identified from the target area more efficiently and accurately, thereby improving the detection accuracy of target objects in image processing scenarios with high recognition difficulty.

[0152] The embodiment of the present specification also provides a server, including a processor and a memory for storing processor executable instructions, and the processor can perform the following steps according to the instructions when implemented: obtaining a target remote sensing image of a target area in a current time period; calling a preset high-pass filter to process the target remote sensing image to obtain a noise covariance matrix; wherein the preset high-pass filter is established based on the spectral information of the target object in the remote sensing image and the spectral information of the background environment; performing a minimum noise separation transformation according to the noise covariance matrix to obtain a principal component analysis matrix of the target remote sensing image; processing the target remote sensing image using the principal component analysis matrix to obtain a processed target remote sensing image; and determining the target object in the target area in the current time period according to the processed target remote sensing image.

[0153] In order to complete the above instructions more accurately, refer to Figure 4 As shown, the embodiment of this specification also provides another specific server, wherein the server includes a network communication port 401, a processor 402 and a memory 403, and the above structures are connected through internal cables so that each structure can perform specific data interaction.

[0154] The network communication port 401 may be used to obtain a target remote sensing image of a target area in the current time period.

[0155] The processor 402 can be specifically used to call a preset high-pass filter to process the target remote sensing image to obtain a noise covariance matrix; wherein the preset high-pass filter is established based on the spectral information of the target object in the remote sensing image and the spectral information of the background environment; according to the noise covariance matrix, a minimum noise separation transformation is performed to obtain a principal component analysis matrix of the target remote sensing image; the target remote sensing image is processed using the principal component analysis matrix to obtain a processed target remote sensing image; according to the processed target remote sensing image, the target object in the target area of ​​the current time period is determined.

[0156] The memory 403 may be specifically used to store corresponding instruction programs.

[0157] In this embodiment, the network communication port 401 can be a virtual port that is bound to different communication protocols so that different data can be sent or received. For example, the network communication port can be a port responsible for web data communication, a port responsible for FTP data communication, or a port responsible for email data communication. In addition, the network communication port can also be a physical communication interface or communication chip. For example, it can be a wireless mobile network communication chip, such as GSM, CDMA, etc.; it can also be a Wifi chip; it can also be a Bluetooth chip.

[0158] In this embodiment, the processor 402 may be implemented in any appropriate manner. For example, the processor may take the form of a microprocessor or processor and a computer-readable medium storing a computer-readable program code (such as software or firmware) executable by the (micro)processor, a logic gate, a switch, an application specific integrated circuit (ASIC), a programmable logic controller, and an embedded microcontroller, etc. This specification does not limit this.

[0159] In this embodiment, the memory 403 may include multiple levels. In a digital system, anything that can store binary data can be a memory; in an integrated circuit, a circuit with a storage function that has no physical form is also called a memory, such as RAM, FIFO, etc.; in a system, a storage device with a physical form is also called a memory, such as a memory stick, TF card, etc.

[0160] The embodiment of the present specification also provides a computer storage medium based on the above-mentioned target object detection method, wherein the computer storage medium stores computer program instructions, which, when executed, implement the following steps: obtaining a target remote sensing image for a target area in a current time period; calling a preset high-pass filter to process the target remote sensing image to obtain a noise covariance matrix; wherein the preset high-pass filter is established based on the spectral information of the target object in the remote sensing image and the spectral information of the background environment; performing a minimum noise separation transformation according to the noise covariance matrix to obtain a principal component analysis matrix of the target remote sensing image; processing the target remote sensing image using the principal component analysis matrix to obtain a processed target remote sensing image; and determining the target object in the target area in the current time period according to the processed target remote sensing image.

[0161] In this embodiment, the storage medium includes, but is not limited to, a random access memory (RAM), a read-only memory (ROM), a cache, a hard disk (HDD), or a memory card. The memory may be used to store computer program instructions. The network communication unit may be an interface for network connection communication set in accordance with the standard specified by the communication protocol.

[0162] In this embodiment, the functions and effects specifically implemented by the program instructions stored in the computer storage medium can be explained in comparison with other implementations and will not be described in detail here.

[0163] See also Figure 5 As shown, at the software level, the embodiment of this specification also provides a detection device for a target object, which may specifically include the following structural modules:

[0164] The acquisition module 501 may be specifically used to acquire a target remote sensing image of a target area in a current time period;

[0165] The first processing module 502 may be specifically used to call a preset high-pass filter to process the target remote sensing image to obtain a noise covariance matrix; wherein the preset high-pass filter is established based on the spectral information of the target object and the spectral information of the background environment in the remote sensing image;

[0166] The transformation module 503 may be specifically used to perform minimum noise separation transformation according to the noise covariance matrix to obtain a principal component analysis matrix of the target remote sensing image;

[0167] The second processing module 504 may be specifically configured to process the target remote sensing image using the principal component analysis matrix to obtain a processed target remote sensing image;

[0168] The determination module 505 may be specifically configured to determine the target object in the target area in the current time period according to the processed target remote sensing image.

[0169] It should be noted that the units, devices or modules described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. For the convenience of description, the above devices are described separately by functions divided into various modules. Of course, when implementing this specification, the functions of each module can be implemented in the same or more software and / or hardware, or the modules that implement the same function can be implemented by a combination of multiple sub-modules or sub-units. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0170] It can be seen from the above that the target object detection device provided in the embodiments of this specification can detect and identify the target object of interest from the target area more efficiently and accurately, thereby improving the detection accuracy of the target object in image processing scenarios with high recognition difficulty.

[0171] In a specific scenario example, the target object detection method provided in this specification can be used to monitor the construction progress of a snowy wind power station. For the specific implementation process, please refer to the following content.

[0172] In this scenario example, considering that the existing methods often have the following problems when detecting and identifying remote sensing satellite images:

[0173] 1. Traditional remote sensing image recognition methods usually detect targets based on the differences in properties such as spectrum and color between the target and the background. However, in scenarios where the target is highly integrated with the complex background, such as detecting white windmills (wind power stations) and other target objects in snowy areas, the recognition effect is often poor.

[0174] 2. Deep learning-based target detection algorithms are often difficult to adapt to high-resolution, memory-intensive remote sensing satellite images. Training and recognition involve large amounts of data processing and high costs.

[0175] 3. The target detection algorithm based on deep learning is also unable to effectively utilize the spectral characteristics of remote sensing satellite images, which leads to false detection and missed detection under the interference of complex weather and environment.

[0176] In this scenario example, to address the above problems, a monitoring system that combines traditional large-scale satellite remote sensing image processing methods with deep learning target detection models is proposed and applied in the scenario of monitoring the construction progress of a snowy wind power station.

[0177] In this scenario example, the advantages of using the spectral properties of remote sensing satellite images for image preprocessing by traditional methods and extracting features from deep neural networks in complex backgrounds are fully combined, so that the detection accuracy of large remote sensing images in small target detection tasks in difficult backgrounds has been greatly improved. Specifically, the traditional remote sensing image processing method is first applied to perform minimum noise separation transformation on the image according to the preset target and background spectra, separate the noise, reduce the dimension and cut the high-resolution remote sensing satellite data, and then send it to the trained neural network for prediction and iterative optimization; finally, the business information such as the number of windmills, coordinate positions, and images with labeled boxes are output to provide an objective basis for construction progress detection.

[0178] For the specific implementation process, please refer to Figure 6 As shown, the following steps are included.

[0179] 1. The remote sensing images input into this system are images taken and uploaded regularly by professional satellite companies using satellites (for example, remote sensing images of targets in the current time period). They have the characteristics of large memory usage, high spectrum, and high resolution. They require a remote sensing image reading module to read and save the hyperspectral information.

[0180] 2. After reading and obtaining image information for the first time (for example, test image), users can select the area of ​​interest and background area in the image with polygons. After selecting the typical target and background area, the system will count the spectral information corresponding to the pixels in the area. Users can choose to save the statistical information of these target and background bands so that they can be automatically loaded during the next inspection.

[0181] 3. Based on the band information of typical targets and background areas, the system will generate a high-pass filter (e.g., a preset high-pass filter). Then, the entire image X can be filtered through the filter so that image pixel blocks with similar spectral properties can be filtered and the noise covariance matrix C can be generated. N , then diagonalize it: D N =U T C N U. Among them, U is an orthogonal matrix composed of eigenvectors, and the obtained matrix D N is the noise covariance matrix C N The diagonal matrix is ​​formed by arranging the eigenvalues ​​of in descending order. Further transformation according to the following formula can be obtained:

[0182]

[0183] I=P T C N P

[0184] Where I is the unit matrix. After the original image X is projected into the new feature space through the transformation matrix P, the noise of the generated projection data has unit variance, so it has band-independent correlation.

[0185] Next, the noise data is transformed by the standard principal component transformation to obtain the matrix C after the transformation matrix P D-adj , and further diagonalized according to the following formula:

[0186] C D-adj =P T C N P

[0187] D D-adj =V T C D-adj V

[0188] Among them, D D-adj C D-adj The diagonal matrix obtained by arranging the eigenvalues ​​of in descending order, V is C D-adj is an orthogonal matrix composed of the eigenvectors of .

[0189] Finally, the minimum noise separation transformation matrix T is the dot product of P and V: T = PV.

[0190] The original image is processed twice through overlapping to obtain the principal component analysis matrix. The elements in the vector obtained after the transformation are independent of each other. The retained principal components greatly reduce the dimension of the original image, and the information is arranged from large to small according to the signal-to-noise ratio, thus overcoming the influence of noise on image quality.

[0191] 4. Based on the retained principal component bands, the original hyperspectral image will be converted into an RGB pseudo-color image (for example, a processed target remote sensing image). The reduced-dimensional pseudo-color image is cut into small image blocks of equal size by the overlapping cutting script in the preprocessing module so that it can be input into the subsequent network for prediction. Adjacent images need to reserve the length of the overlapping part according to the target size to ensure that the target of interest, such as a windmill, will not be split in half during cutting. Specifically, the user can first enter or directly use the preset width and height of the small image after cutting, as well as the length of the overlapping part of adjacent images. The system will then pad the image with zeros according to the width, height and step size and perform sliding window cutting to ensure that the cut images are of the same size and that the target of interest can always appear completely in a small image.

[0192] 5. The cut satellite images will be packaged and sent to the feature extraction network (e.g., the preset feature extraction network) to generate a feature heat map (e.g., the fused feature map). The feature extraction main network connects multiple convolutional layers in series, and generates two low-resolution feature maps in parallel through 2x and 4x downsampling, also using multiple convolutional layers in series. In order to retain the high-resolution features of the satellite images, the feature maps of the main network and the two low-resolution sub-networks will perform an elementwise-add feature fusion operation after every three convolutions, so that the high- and low-resolution feature information in the shallow network can still be retained in the deep network.

[0193] 6. The anchor frame regression and classification network (e.g., the preset regression and classification network) first transforms the dimension of the feature map obtained in the previous step through a 3*3 convolution and two 1*1 convolutions to obtain N*2 prediction scores of whether the object (e.g., the target object) is contained, and N*4 anchor frame coordinate predictions. During specific training, 5 basic anchor frames of different sizes are preset in the network, and the IOU value is calculated with the ground truth frame. A number of anchor frames with an IOU greater than 0.7 are set as positive samples, and b numbers with an IOU less than 0.3 are set as negative samples, thereby obtaining (a+b)*2 sample classification labels and a*4 anchor frame regression tensor for calculating network loss.

[0194] Finally, the system can store the predicted number of windmills, coordinate positions, images with labeled boxes and other business information results into a json file output, and compare it with the data of the previous month (data of the previous time period) to reflect the progress of construction progress of the current month compared with the previous month.

[0195] Users can also use the system to transfer misidentified and missed images and use them to input into the preprocessing module to add preset target and background bands, or input into the feature extraction network, anchor box regression and classification network for iterative optimization to further improve the model recognition accuracy.

[0196] Through the above scenario examples, it is verified that the target object detection method provided in this specification has the following advantages over the existing remote sensing satellite image target detection method: the high-spectral and high-resolution features of remote sensing images are fully utilized to design a remote sensing image preprocessing module to disentangle the target of interest from the difficult-to-distinguish background in the projection space; through dimensionality reduction and overlapping cutting methods, large-capacity and large-size original satellite images are converted into multiple small images, and GPU devices are used for parallel prediction, which greatly improves the detection speed of remote sensing images; deep neural networks are used for feature extraction and regression prediction, which fully retains the high-resolution features of remote sensing images, and greatly improves the detection accuracy of small targets under complex backgrounds.

[0197] Although the present specification provides method operation steps as described in the embodiments or flow charts, more or less operation steps may be included based on conventional or non-creative means. The order of steps listed in the embodiments is only one way of executing the order of many steps, and does not represent a unique execution order. When the device or client product in practice is executed, it can be executed in sequence or in parallel according to the method shown in the embodiments or the drawings (for example, a parallel processor or a multi-threaded processing environment, or even a distributed data processing environment). The term "include", "include" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, product or device including a series of elements includes not only those elements, but also includes other elements that are not explicitly listed, or also includes elements inherent to such a process, method, product or device. In the absence of more restrictions, it is not excluded that there are other identical or equivalent elements in the process, method, product or device including the elements. The first, second, etc. words are used to represent the name, and do not represent any particular order.

[0198] Those skilled in the art also know that, in addition to implementing the controller in a purely computer-readable program code, the controller can be made to implement the same function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, such a controller can be considered as a hardware component, and the devices for implementing various functions included therein can also be considered as structures within the hardware component. Or even, the devices for implementing various functions can be considered as both software modules for implementing the method and structures within the hardware component.

[0199] This specification may be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, classes, etc. that perform specific tasks or implement specific abstract data types. This specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules may be located in local and remote computer storage media including storage devices.

[0200] Through the description of the above embodiments, it can be known that those skilled in the art can clearly understand that the present specification can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present specification can essentially be embodied in the form of a software product, which can be stored in a storage medium such as ROM / RAM, a disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a mobile terminal, a server, or a network device, etc.) to execute the methods described in each embodiment of the present specification or some parts of the embodiments.

[0201] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. This specification can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld devices or portable devices, tablet devices, multi-processor systems, microprocessor-based systems, set-top boxes, programmable electronic devices, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc.

[0202] Although the present specification is described through embodiments, those skilled in the art will appreciate that there are many modifications and changes to the present specification without departing from the spirit of the present specification, and it is intended that the appended claims include these modifications and changes without departing from the spirit of the present specification.

Claims

1. A method for detecting a target object, characterized in that: include: Obtain the target remote sensing image of the target area in the current time period; Calling a preset high-pass filter to process the target remote sensing image to obtain a noise covariance matrix; wherein the preset high-pass filter is established based on the spectral information of the target object and the spectral information of the background environment in the remote sensing image; According to the noise covariance matrix, a minimum noise separation transformation is performed to obtain a principal component analysis matrix of the target remote sensing image; Processing the target remote sensing image using the principal component analysis matrix to obtain a processed target remote sensing image; Determining the target object in the target area of ​​the current time period according to the processed target remote sensing image; Wherein, according to the processed target remote sensing image, determining the target object in the target area of ​​the current time period includes: according to a preset cutting rule, cutting the target remote sensing image into a plurality of sub-image blocks; wherein there is an overlapping area of ​​a preset range between two adjacent sub-image blocks of the plurality of sub-image blocks; calling a preset processing model, processing the plurality of sub-image blocks, and obtaining an image processing result corresponding to the current time period; and according to the image processing result, determining the target object in the target area of ​​the current time period, and the number of the target objects; The preset processing model at least includes a preset feature extraction network; wherein the preset feature extraction network includes: a main network, a first sub-network, a second sub-network, and a fusion structure; wherein the main network is used to extract a high-resolution feature map; the first sub-network is used to extract a first type of low-resolution feature map; the second sub-network is used to extract a second type of low-resolution feature map; the fusion structure is used to fuse the high-resolution feature map, the first type of low-resolution feature map, and the second type of low-resolution feature map to obtain a fused feature map.

2. The method according to claim 1, characterized in that The target object includes a physical object whose color difference between the target area and the background environment is less than a preset difference threshold.

3. The method according to claim 1, characterized in that The preset high-pass filter is established in the following manner: Acquire a remote sensing image targeting a target area and containing a target object as a test image; Determining in the test image a first image region where the target object is located and a second image region where the background environment is located; Performing spectral information statistics on the first image area to obtain spectral information statistics results of the target object; Performing spectral information statistics on the second image area to obtain spectral information statistics results of the background environment; The preset high-pass filter is constructed according to the statistical results of the spectral information of the target object and the statistical results of the spectral information of the background environment.

4. The method according to claim 1, characterized in that: According to the noise covariance matrix, a minimum noise separation transformation is performed to obtain a principal component analysis matrix of the target remote sensing image, including: Constructing a first transformation matrix according to the noise covariance matrix; Performing a first transformation process on the noise covariance matrix using the first transformation matrix to obtain a transformed data matrix; constructing a second transformation matrix according to the transformed data matrix; constructing a minimum noise separation transformation matrix according to the first transformation matrix and the second transformation matrix; The target remote sensing image is processed using the minimum noise separation transformation matrix to obtain a principal component analysis matrix of the target remote sensing image.

5. The method according to claim 1, characterized in that Processing the target remote sensing image using the principal component analysis matrix to obtain a processed target remote sensing image includes: Determining the principal component bands in the target remote sensing image according to the principal component analysis matrix; According to the principal component bands, corresponding spectral information is extracted from the target remote sensing image to generate a corresponding RGB pseudo-color picture as the processed target remote sensing image.

6. The method according to claim 1, characterized in that The preset processing model also includes a preset regression and classification network; the preset regression and classification network is used to locate multiple anchor frames based on the fused feature map, and determine the predicted probability of the target object existing in each anchor frame of the multiple anchor frames.

7. The method according to claim 6, characterized in that The preset regression and classification network includes: a first convolutional layer, and a second convolutional layer and a third convolutional layer connected to the first convolutional layer; wherein the first convolutional layer is used to process the fused feature map and output the corresponding intermediate data; the second convolutional layer is used to locate multiple anchor frames by finding a combination of multiple coordinate points in the target area according to the intermediate data; the third convolutional layer is used to calculate the predicted probability of the existence of the target object in the anchor frame according to the intermediate data.

8. The method according to claim 1, characterized in that The target object includes a wind power station, and the target area includes a snow area.

9. The method according to claim 8, characterized in that After determining the target objects in the target area in the current time period and the number of the target objects, the method further includes: Obtaining target objects in a target area of ​​a previous time period determined based on a target remote sensing image of a previous time period, and the number of target objects; The construction progress of the wind power station in the snow area is determined based on the target objects in the target area in the previous time period and the number of target objects; and the target objects in the target area in the current time period and the number of target objects.

10. A device for detecting a target object, characterized in that: include: An acquisition module is used to acquire a target remote sensing image of a target area in the current time period; A first processing module is used to call a preset high-pass filter to process the target remote sensing image to obtain a noise covariance matrix; wherein the preset high-pass filter is established based on the spectral information of the target object and the spectral information of the background environment in the remote sensing image; A transformation module, used for performing minimum noise separation transformation according to the noise covariance matrix to obtain a principal component analysis matrix of the target remote sensing image; A second processing module is used to process the target remote sensing image using the principal component analysis matrix to obtain a processed target remote sensing image; A determination module, used to determine the target object in the target area of ​​the current time period according to the processed target remote sensing image; The determination module is specifically used to cut the target remote sensing image into a plurality of sub-image blocks according to a preset cutting rule; wherein there is an overlapping area of ​​a preset range between two adjacent sub-image blocks in the plurality of sub-image blocks; calling a preset processing model to process the plurality of sub-image blocks to obtain an image processing result corresponding to the current time period; and determining the target objects in the target area of ​​the current time period and the number of target objects according to the image processing result; The preset processing model at least includes a preset feature extraction network; wherein the preset feature extraction network includes: a main network, a first sub-network, a second sub-network, and a fusion structure; wherein the main network is used to extract a high-resolution feature map; the first sub-network is used to extract a first type of low-resolution feature map; the second sub-network is used to extract a second type of low-resolution feature map; the fusion structure is used to fuse the high-resolution feature map, the first type of low-resolution feature map, and the second type of low-resolution feature map to obtain a fused feature map.

11. A server, characterized in that: The method comprises a processor and a memory for storing processor-executable instructions, wherein the processor implements the steps of the method according to any one of claims 1 to 9 when executing the instructions.

12. A computer-readable storage medium, characterized in that: Computer instructions are stored thereon, and when the instructions are executed, the steps of the method according to any one of claims 1 to 9 are implemented.

Images