Detection method and detection system for offshore wind power facility target

Abstract

The invention discloses a detection method and a detection system for an offshore wind power facility target. The detection method comprises the following steps: acquiring a satellite remote sensing image to be detected; performing block preprocessing on the satellite remote sensing image to obtain a plurality of block satellite remote sensing images; performing region-of-interest clustering on the plurality of block satellite remote sensing images in sequence; cutting the obtained clustering regions of the regions of interest in sequence, inputting the cut clustering regions into an offshore wind power facility detection network, and detecting independent wind power facility targets in the regions; and integrating the detected block satellite remote sensing images to obtain a final detection result of the wind power facility target. According to the invention, the target position of the offshore wind power facility is detected based on the satellite remote sensing technology, and long-time continuous control of ocean engineering related to a sea area is realized.

Description

Technical Field

[0001] The present invention relates to the field of marine engineering control related to sea areas, and in particular to a detection method and a detection system for offshore wind power facility targets. Background Art

[0002] In recent years, remote sensing technology has become increasingly closely related to economic development, ecological environment and marine resource supervision. It has been widely used in agriculture, forestry, ocean, geological and mineral exploration, meteorology, hydrology, environmental protection and other fields, and has gradually formed a multi-level, multi-angle and multi-field observation system for detecting the earth's environment and resources. Remote sensing technology has the advantages of wide coverage, real-time, objectivity and accuracy. It can be used to quickly obtain high-precision marine monitoring information in key sea areas, real-time and rapid tracking and marine dynamic supervision, timely formulate treatment measures, reduce losses, supervise the use of sea areas and islands, and protect the marine environment. Secondly, the complex and changeable marine environmental conditions make the application needs of the marine surveillance department highly professional and targeted.

[0003] In view of the long-term and continuous control requirements for relevant marine projects under the current situation, it is urgent to propose an effective method to detect offshore wind power facilities. By clustering the regions of interest of the satellite remote sensing images to be detected, a more accurate target location can be obtained, thereby strengthening the continuous control of the long-term series of relevant marine projects in the sea area, carrying out monitoring of key areas in the sea area, timely grasping the illegal and irregular behaviors within the jurisdiction of the city, regulating the order of sea area use, and comprehensively improving the technical support for the legal use of marine resources. Summary of the invention

[0004] (I) Purpose of the invention

[0005] The purpose of the present invention is to propose a detection technology solution for offshore wind power facility targets based on satellite remote sensing technology in view of the changeable marine environmental conditions and the need for monitoring of sea-related projects under the current situation.

[0006] (II) Technical solution

[0007] To solve the above problems, the first aspect of the present invention provides a method for detecting offshore wind power facility targets, including: acquiring a satellite remote sensing image to be detected; performing block preprocessing on the satellite remote sensing image to obtain a number of block satellite remote sensing images; clustering the several block satellite remote sensing images into regions of interest in turn; cropping the obtained region of interest cluster regions in turn and inputting them into an offshore wind power facility detection network to detect independent wind power facility targets in the region; integrating the detected block satellite remote sensing images to obtain the final wind power facility target detection result.

[0008] Preferably, in the detection method, the region of interest clustering region cropping includes: if the candidate frames of the two segmented satellite remote sensing images satisfy the following formula:

[0009] IoU(box1,box2)>0

[0010] The two candidate boxes are merged and cropped to obtain the region of interest; wherein box1 and box2 represent the candidate boxes of the two block satellite remote sensing images respectively.

[0011] Preferably, the detection method, the offshore wind power facility detection network includes: a feature encoder, a feature fusion module and a detection head.

[0012] Preferably, in the detection method, the feature encoder includes a convolution layer, a feature fusion layer and a fast spatial pyramid pooling layer, the convolution layer performs a convolution operation on each of the regions of interest; the feature fusion layer splits the region of interest into two parts, the first part is input into two bottleneck layers, the second part is spliced ​​with the outputs of the two bottleneck layers, and the information of the two parts is fused using a convolution layer with predetermined parameters; the fast spatial pyramid pooling layer performs a pooling operation on the regions of interest of different sizes and proportions of the input.

[0013] Preferably, in the detection method, the feature fusion module adopts a bidirectional design mode to fuse and splice from the highest-level feature map to the lower-level feature map, and then performs the same fusion operation from the lowest level to the higher level.

[0014] Preferably, in the detection method, the detection head uses Wasserstein distance to evaluate the position accuracy of the central target of the wind power facility to be detected, wherein the Wasserstein distance is expressed by the following formula:

[0015]

[0016] By normalizing the formula in exponential form, we get:

[0017]

[0018] Where C represents a constant coefficient, represents the Wasserstein distance, NWD(G1, G2) represents the normalized Wasserstein distance, G1 and G2 represent two different Gaussian distributions, respectively. represents the central abscissa of the inscribed ellipse of the Gaussian distribution G1 bounding box, represents the central ordinate of the inscribed ellipse of the Gaussian distribution G1 bounding box, w1 represents the width of the Gaussian distribution G1 bounding box, h1 represents the height of the Gaussian distribution G1 bounding box, The central abscissa of the inscribed ellipse of the Gaussian distribution G2 bounding box, represents the central ordinate of the inscribed ellipse of the Gaussian distribution G2 bounding box, w2 represents the width of the Gaussian distribution G2 bounding box, and h2 represents the height of the Gaussian distribution G2 bounding box.

[0019] Preferably, the detection method fits the position of the central target with the height and width of eight sampling points around the central target, and takes the average of multiple central target positions as the final independent wind power target.

[0020] Preferably, the detection method and the offshore wind power facility detection network further include a loss function calculation module, and the loss function adopts the following expression:

[0021] loss total = CrossLoss(p,q)+DFL(S i ,S i+1 )+L NWD (pred box ,gt box )+L NWD (pred center ,gt center )

[0022] Among them, loss total represents the total loss function, CrossLoss(p,q) represents the improved cross entropy loss function, DFL(S i ,S i+1 ) represents the distribution focus loss function, L NWD (pred box ,gt box ) represents the loss function of the target box regression based on the Wasserstein distance, L NWD (pred center ,gt center ) represents the loss function for regressing the center position of the target box based on the Wasserstein distance.

[0023] Preferably, the detection method integrates the detected block satellite remote sensing images, including: if there is a wind power facility target in the overlapping area of ​​two adjacent block satellite remote sensing images, a non-maximum suppression method is used to retain the detection result of one block satellite remote sensing image and delete the detection result of the other block satellite remote sensing image.

[0024] According to another aspect of the present invention, the present invention also proposes a detection system for offshore wind power facility targets, the detection system comprising: a remote sensing image acquisition module, which is used to acquire satellite remote sensing images to be detected; a remote sensing image blocking module, which is used to perform block preprocessing on the satellite remote sensing images to obtain a number of block satellite remote sensing images; an area of ​​interest clustering module, which is used to cluster the areas of interest of a number of block satellite remote sensing images in turn; a wind power target detection module, which is used to crop the obtained area of ​​interest clustering areas in turn and input them into an offshore wind power facility detection network to detect independent wind power facility targets in the area; a remote sensing image integration module, which is used to integrate the detected block satellite remote sensing images to obtain the final wind power facility target detection result.

[0025] (III) Beneficial effects

[0026] The above technical solution of the present invention has the following beneficial technical effects:

[0027] The present invention proposes a detection method and detection system for offshore wind power facility targets. The present invention solves the existing need for control of marine engineering related to the sea area. Based on remote sensing satellite technology, the satellite remote sensing images to be detected are clustered by regions of interest to lock the more accurate location of offshore wind power targets. By strengthening the long-term series monitoring of related marine engineering, the existing illegal behaviors at sea can be grasped in time and the order of use of the sea area can be regulated. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic diagram of the steps of a method for detecting an offshore wind power facility target according to an embodiment of the present invention;

[0029] Figure 2 It is a schematic diagram of the working process of the method for detecting an offshore wind power facility target according to an embodiment of the present invention;

[0030] Figure 3 It is a schematic structural diagram of a detection system for an offshore wind power facility target according to an embodiment of the present invention. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings. It should be understood that these descriptions are only exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, the description of well-known structures and technologies is omitted to avoid unnecessary confusion of the concept of the present invention.

[0032] The accompanying drawings show schematic diagrams of layer structures according to embodiments of the present invention. These figures are not drawn to scale, and some details are magnified and some details may be omitted for the purpose of clarity. The shapes of various regions and layers shown in the figures and the relative sizes and positional relationships therebetween are only exemplary, and may deviate in practice due to manufacturing tolerances or technical limitations, and those skilled in the art may additionally design regions / layers with different shapes, sizes, and relative positions according to actual needs.

[0033] Obviously, the described embodiments are only some embodiments of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0034] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0035] In recent years, remote sensing technology has become increasingly closely related to economic development, ecological environment and marine resource supervision. It has been widely used in agriculture, forestry, ocean, geological and mineral exploration, meteorology, hydrology, environmental protection and other fields, and has gradually formed a multi-level, multi-angle and multi-field observation system for detecting the earth's environment and resources. Remote sensing technology has the advantages of wide coverage, real-time, objectivity and accuracy. It can be used to quickly obtain high-precision marine monitoring information in key sea areas, real-time and rapid tracking and marine dynamic supervision, timely formulate treatment measures, reduce losses, supervise the use of sea areas and islands, and protect the marine environment. Secondly, the complex and changeable marine environmental conditions make the application needs of the marine surveillance department highly professional and targeted.

[0036] In response to the current requirement for long-term continuous monitoring of marine engineering projects in the sea area, the present invention proposes a detection method for offshore wind power facility targets based on remote sensing technology to achieve continuous control of related marine engineering projects.

[0037] Figure 1 Schematic diagram of the steps of the method for detecting an offshore wind power facility target according to an embodiment of the present invention. Figure 1 A schematic diagram illustrating steps of a method for detecting an offshore wind power facility target.

[0038] Step S110, obtaining a satellite remote sensing image to be detected;

[0039] Step S120, performing block preprocessing on the satellite remote sensing image to obtain a plurality of block satellite remote sensing images;

[0040] Step S130, clustering the regions of interest in sequence on a plurality of block satellite remote sensing images;

[0041] Step S140, sequentially clipping the obtained region of interest cluster regions and inputting them into the offshore wind power facility detection network to detect independent wind power facility targets in the region;

[0042] Step S150: Integrate the detected block satellite remote sensing images to obtain the final detection result of the wind power facility target.

[0043] According to some optional embodiments, in step S110, a satellite remote sensing image to be detected is obtained. First, a sample library is constructed using offshore wind power target remote sensing images, and then the offshore wind power target remote sensing images in the sample library are classified and labeled to obtain different types of satellite remote sensing images.

[0044] In one embodiment, the images in the sample library are annotated with the roLabelImg image annotation tool, and a .txt file is generated after the annotated image is completed. The file contains a variety of information about the image classification, including the category information of the target, the position of the center point of the target in the image, and the width, height, and angle information of the annotated target. By annotating the images in the sample library with the annotation tool, the location of the offshore wind power target can be intelligently retrieved.

[0045] In one embodiment, a data augmentation strategy is applied to the image to improve the accuracy of the target type detection. The use of the data augmentation strategy can increase the diversity of training samples and improve the robustness of detection. The sample diversity is increased by expanding the training samples by horizontal flipping, vertical flipping, random rotation, random scaling, random cropping or expansion. In addition, the image also uses color jitter, including contrast adjustment, brightness adjustment, saturation adjustment and hue adjustment. The above strategies are used to expand the richness of the target types to be detected, provide materials for the deep learning offshore wind power target detection model to detect and identify the target, and also provide evaluation data for the supervised learning algorithm.

[0046] After completing step S110 to obtain the satellite remote sensing image to be detected, the process proceeds to step S120 to perform block preprocessing on the satellite remote sensing image to obtain a plurality of block satellite remote sensing images.

[0047] In one embodiment, during the pre-processing of satellite remote sensing images, the image files in tiff format are first converted into jpg format, and then the original images are cropped using a sliding window of 4096×4096 size, so that each cropped image has 256 pixels of overlap around it. The window cropping method makes the upper, lower, left and right boundaries of the current window overlap with the images of adjacent windows respectively, and obtains a number of satellite remote sensing images in blocks.

[0048] Then, step S130 is performed to cluster the regions of interest in turn for the plurality of block satellite remote sensing images. The clustering of the regions of interest focuses on detecting the center point position of the region of interest and the width and height corresponding to the center point.

[0049] In one embodiment, in the process of clustering the region of interest, a target detector that can detect the approximate location of the target is first trained, and the target detector performs a meanshift clustering operation on the candidate boxes of the above-mentioned plurality of block satellite remote sensing images. If a satellite remote sensing image does not generate a candidate box, the detection process of the window is skipped. If a candidate box is generated, the two candidate boxes are merged and cropped to obtain the region of interest.

[0050] Among them, the clipping of the cluster area of ​​interest includes:

[0051] If the candidate frames of two block satellite remote sensing images satisfy the following formula:

[0052] IoU(box1,box2)>0

[0053] The two candidate boxes are merged and cropped to obtain the region of interest. Box1 and box2 represent the candidate boxes of two segmented satellite remote sensing images.

[0054] In one embodiment, ResNet50 is selected to extract feature maps at three scales, where the sizes of the output feature pyramids corresponding to the three feature maps are 128, 256, and 512, respectively. Then, a top-down feature fusion method is used to perform a bilinear interpolation operation on the previous layer features, and a bitwise addition operation is performed with the next layer features. Finally, the results of the top-down fusion of the pyramid are sequentially input into the prediction network to regress the clustering candidate box information at different scales. The candidate boxes with scores higher than the 0.5 threshold are then used as the input of the mean shift clustering algorithm to obtain the clustered areas of the region of interest.

[0055] Then, step S140 is entered, and the obtained region of interest cluster regions are clipped in turn and input into the offshore wind power facility detection network to detect independent wind power facility targets in the region. In one embodiment, the offshore wind power facility detection network includes: a feature encoder, a feature fusion module and a detection head. The offshore wind power facility detection network proposes a wind power facility detection network based on an improved YOLOv8 model structure for the characteristics of wind power targets such as small size, strong interference, and sensitivity to evaluation indicators, thereby overcoming the difficulties in detecting wind power targets.

[0056] Specifically, the feature encoder includes a convolution layer, a feature fusion layer, and a fast spatial pyramid pooling layer. The convolution layer performs convolution operations on each region of interest; the feature fusion layer splits the region of interest into two parts, the first part is input into the two bottleneck layers, and the second part is spliced ​​with the output of the two bottleneck layers, and the convolution layer uses a predetermined parameter to fuse the information of the two parts; the fast spatial pyramid pooling layer performs pooling operations on the input regions of interest of different sizes and proportions.

[0057] In one embodiment, each convolution layer is composed of a 2D convolution, a layer normalization function, and a SiLU activation function. The calculation formula of SiLU is:

[0058] SiLU = x × sigmoid(x),

[0059] Since the region of interest includes offshore wind power targets and potential interference targets in offshore wind power scenarios, the convolution operation can be used to extract high-dimensional features of the region of interest image, enhancing the distinction between wind power targets and interference targets. In addition, inputting the region of interest into the convolution layer can reduce the complexity of the operation, further reduce the image resolution of the region of interest, and extract the features of the image, which can facilitate the subsequent detection head to decode the image features. The convolution layer can also improve the feature fitting ability of the image, and learn the feature distribution of interference factors with similar characteristics to wind power targets in remote sensing images through convolution operations, enhance the distinction between wind power targets and interference items, and improve the detection ability and anti-interference ability of the offshore wind power facility detection network.

[0060] In one embodiment, when passing through the feature fusion layer, the feature map is first input into the convolution layer to obtain the features of the image, and then the channel splitting method is used to split the input feature map into two equal parts. The first part is input into the two bottleneck layers for channel dimension reduction and feature fusion, and the second part is spliced ​​with the output of the two bottleneck layers to form a new feature map, and then the predetermined parameter convolution layer is used to fuse the information of the two parts. Among them, the predetermined parameter convolution layer is a convolution layer with a convolution kernel and a step size of 1. This convolution layer can not share parameters with the convolution layer in the encoding stage, thereby improving the modeling capabilities of different features in the regional fusion stage and the encoding stage.

[0061] In one embodiment, the fast spatial pyramid pooling layer has a faster calculation speed and fewer parameters, and its structure is to concat three serial MaxPool inputs together, wherein the kernels of the three MaxPools are all 5×5.

[0062] The offshore wind power facility detection network also includes a feature fusion module, which adopts a bidirectional design mode to fuse and splice from the highest-level feature map to the lower-level feature map, and then performs the same fusion operation from the lowest level to the highest level.

[0063] In one embodiment, in the feature fusion stage, the feature fusion module adopts a bidirectional design mode, including top-down and bottom-up modes. The highest-level features are first upsampled to make their feature maps have the same size as the lower-level feature maps. Then, they are channel-joined with the lower-level feature maps and feature fusion layers are used to learn feature relationships at different scales. This process is repeated until the bottom layer of the pyramid used by the detection head. Then, after the network performs the same fusion operation from the bottom layer of the pyramid, the results of different layers are sequentially input into the detection head for target detection and regression of wind power facilities.

[0064] The offshore wind power facility detection network also includes a detection head, which uses the Wasserstein distance to evaluate the position accuracy of the central target of the wind power facility to be detected. The Wasserstein distance is expressed by the following formula:

[0065]

[0066] Normalizing the formula in exponential form, we get:

[0067]

[0068] Where C represents a constant coefficient, represents the Wasserstein distance, NWD(G1,G2) represents the normalized Wasserstein distance, G1 and G2 represent two different Gaussian distributions, represents the central abscissa of the inscribed ellipse of the Gaussian distribution G1 bounding box, represents the central ordinate of the inscribed ellipse of the Gaussian distribution G1 bounding box, w1 represents the width of the Gaussian distribution G1 bounding box, h1 represents the height of the Gaussian distribution G1 bounding box, The central abscissa of the inscribed ellipse of the Gaussian distribution G2 bounding box, represents the central ordinate of the inscribed ellipse of the Gaussian distribution G2 bounding box, w2 represents the width of the Gaussian distribution G2 bounding box, and h2 represents the height of the Gaussian distribution G2 bounding box.

[0069] In one embodiment, the height and width of eight sampling points around the central target are used to fit the position of the central target, and the sampling points at 8 positions around the central point are used as positive sample points, and the rest are negative sample points, so as to enhance the model position regression capability. And the average of multiple central target positions is used as the final independent wind power target. In the offshore wind power facility detection network, the target frame of the offshore wind power is usually characterized by the center point of the predicted target, the width and height of the target. However, this prediction method is difficult to accurately characterize the shape of the wind power target, which will cause invalid learning of features. Here, the prediction of the offshore wind power target sampling points is changed from a single center point prediction problem to a multi-position fitting problem, which enriches the feature information of the wind power target, thereby improving the position regression capability and anti-interference capability of the offshore wind power facility detection network.

[0070] In one embodiment, when evaluating the position of the central target of multiple block satellite remote sensing images, the Wasserstein distance is introduced to replace IoU as a new indicator to measure the performance of extremely small target detection, which can improve the robustness of small target detection. In addition, the embodiment of the present invention introduces a large number of data enhancement methods, including rotation, translation, random distortion, HSV random fluctuation, etc., to enhance the robustness of the detector to changes in different atmospheric conditions and lighting conditions.

[0071] In the embodiment of the present invention, the bounding boxes of the two segmented satellite remote sensing images are first modeled as a two-dimensional Gaussian distribution, where the weight of the central pixel is the largest and the weight decreases toward the boundary. If the current horizontal bounding box is:

[0072] R=(x,y,w,h)

[0073] Where x, y, w and h represent the center point, width and height of the bounding box respectively. The inscribed ellipse of the bounding box can be expressed as:

[0074]

[0075] Among them, c x , c y , r x and r y Represent the center coordinates of the ellipse and the lengths along the x-axis and y-axis respectively. The mean μ and covariance X of the Gaussian distribution of any two block satellite remote sensing images obey:

[0076]

[0077] For two Gaussian distributions G1 = N (μ1, X1) and G2 = N (μ2, X2), their Wasserstein distance can be defined as:

[0078]

[0079] Further, normalizing it in exponential form yields:

[0080]

[0081] Among them, C is a constant coefficient, and its value is 4 in this system. represents the Wasserstein distance, NWD(G1, G2) represents the normalized Wasserstein distance, G1 and G2 represent two different Gaussian distributions, respectively. represents the central abscissa of the inscribed ellipse of the Gaussian distribution G1 bounding box, represents the central ordinate of the inscribed ellipse of the Gaussian distribution G1 bounding box, w1 represents the width of the Gaussian distribution G1 bounding box, h1 represents the height of the Gaussian distribution G1 bounding box, The central abscissa of the inscribed ellipse of the Gaussian distribution G2 bounding box, represents the central ordinate of the inscribed ellipse of the Gaussian distribution G2 bounding box, w2 represents the width of the Gaussian distribution G2 bounding box, and h2 represents the height of the Gaussian distribution G2 bounding box.

[0082] The offshore wind power facility detection network also includes a loss function calculation module. The loss function based on Wasserstein distance is:

[0083] loss total = CrossLoss(p,q)+DFL(S i ,S i+1 )+L NWD (pred box ,gt box )+L NWD (pred center ,gt center )

[0084] Among them, loss total represents the total loss function, CrossLoss(p,q) represents the improved cross entropy loss function, DFL(S i ,S i+1 ) represents the distribution focus loss function, L NWD (pred box ,gt box ) represents the loss function of the target box regression based on the Wasserstein distance, L NWD (pred center ,gt center ) represents the loss function for regressing the center position of the target box based on the Wasserstein distance.

[0085] In one embodiment, the loss function includes classification loss, target frame regression loss, and target frame center position regression loss. The classification loss uses an improved cross entropy function, the frame regression uses a combination of distribution focus loss DFL and Wasserstein distance loss, and the center position loss uses the Wasserstein distance loss function.

[0086] The calculation formula of the improved cross entropy loss function is as follows:

[0087]

[0088] q is the IoU (intersection over union) of bbox (prediction box) and gt (real box), IoU is the intersection of the prediction box and the real box divided by the union of the two boxes, p is the probability. α and γ are hyperparameters, which are set to 2.0 and 0.1 in the embodiment of the present invention.

[0089] The calculation formula of DFL is:

[0090] DFL(S i ,S i+1 )=-((y i+1 -y)log(S i )+(yy i )log(S i+1 ))

[0091] Among them, y represents the true predicted value, y i and i+1 Represents two predicted values ​​near the true value, S i and S i+1 is the probability corresponding to the predicted value.

[0092] The overall loss function formula is:

[0093] loss total = CrossLoss(p,q)+DFL(S i ,S i+1 )+L NWD (pred box ,gt box )+L NWD (pred center ,gt center )

[0094] Finally, step S150 is entered to integrate the detected block satellite remote sensing images to obtain the final wind power facility target detection result. The detected block satellite remote sensing images are integrated, including: if there is a wind power facility target in the overlapping area of ​​two adjacent block satellite remote sensing images, a non-maximum suppression method is used to retain the detection result of one block satellite remote sensing image and delete the detection result of the other block satellite remote sensing image.

[0095] Figure 2 FIG. 1 is a schematic diagram of a work flow of a method for detecting an offshore wind power facility target according to an embodiment of the present invention. Figure 2 The workflow of the detection method for offshore wind power facility targets is described.

[0096] Firstly, the satellite remote sensing images to be detected are obtained, and a sample library is constructed using the remote sensing images of offshore wind power targets. The images in the sample library are classified and labeled to obtain different types of satellite remote sensing images.

[0097] Then, the acquired satellite remote sensing images are preprocessed by using the window clipping method to obtain several satellite remote sensing images, and then the several satellite remote sensing images are clustered into regions of interest. The clipped regions of interest include: if the candidate frames of the two satellite remote sensing images satisfy the following formula:

[0098] IoU(box1,box2)>0

[0099] The two candidate boxes are merged and cropped to obtain the region of interest, wherein box1 and box2 represent the candidate boxes of the two block satellite remote sensing images respectively.

[0100] Then, the obtained region of interest cluster regions are cut and input into the offshore wind power facility detection network to detect the independent wind power facility targets in the region. The offshore wind power facility detection network includes: feature encoder, feature fusion module and detection head.

[0101] Specifically, the feature encoder includes a convolution layer, a feature fusion layer and a fast spatial pyramid pooling layer. The convolution layer performs a convolution operation on each of the regions of interest. The feature fusion layer splits the region of interest into two parts, the first part is input into two bottleneck layers, the second part is spliced ​​with the output of the two bottleneck layers, and the information of the two parts is fused by the convolution layer using predetermined parameters. The fast spatial pyramid pooling layer performs a pooling operation on the regions of interest of different sizes and proportions of the input.

[0102] The feature fusion module adopts a bidirectional design mode to fuse and splice from the highest-level feature map to the lower-level feature map, and then performs the same fusion operation from the lowest level to the highest level.

[0103] The detection head uses Wasserstein distance to evaluate the position accuracy of the central target of the wind power facility to be detected, where the Wasserstein distance is expressed by the following formula:

[0104]

[0105] Normalizing the formula in exponential form, we get:

[0106]

[0107] Where C represents a constant coefficient, represents the Wasserstein distance, NWD(G1,G2) represents the normalized Wasserstein distance, G1 and G2 represent two different Gaussian distributions, represents the central abscissa of the inscribed ellipse of the Gaussian distribution G1 bounding box, represents the central ordinate of the inscribed ellipse of the Gaussian distribution G1 bounding box, w1 represents the width of the Gaussian distribution G1 bounding box, h1 represents the height of the Gaussian distribution G1 bounding box, The central abscissa of the inscribed ellipse of the Gaussian distribution G2 bounding box, represents the central ordinate of the inscribed ellipse of the Gaussian distribution G2 bounding box, w2 represents the width of the Gaussian distribution G2 bounding box, and h2 represents the height of the Gaussian distribution G2 bounding box. The height and width of the eight sampling points around the central target fit the position of the central target, and the average of the positions of multiple central targets is used as the final independent wind power target.

[0108] In addition, the offshore wind power facility detection network also includes a loss function calculation module, and the loss function uses the following expression:

[0109] loss total = CrossLoss(p,q)+DFL(S i ,S i+1 )+L NWD (pred box ,gt box )+L NWD (pred center ,gt center )

[0110] Among them, loss total represents the total loss function, CrossLoss(p,q) represents the improved cross entropy loss function, DFL(S i ,S i+1 ) represents the distribution focus loss function, L NWD (pred box ,gt box) represents the loss function of target box regression based on Wasserstein distance, L NWD (pred center ,gt center ) represents the loss function for target box center position regression based on Wasserstein distance.

[0111] Finally, the detected block satellite remote sensing images are integrated to obtain the final wind power facility target detection results. If there are wind power facility targets in the overlapping area of ​​two adjacent block satellite remote sensing images, the non-maximum suppression method is used to retain the detection result of one block satellite remote sensing image and delete the detection result of the other block satellite remote sensing image.

[0112] Based on the above detection method for offshore wind power facility targets, an embodiment of the present invention further provides a detection system for offshore wind power facility targets. The system uses the above detection method to detect offshore wind power facility targets.

[0113] Figure 3 Schematic diagram of the structure of the detection system of the offshore wind power facility target according to the embodiment of the present invention, Figure 3 To illustrate the various components and functions of the detection system for offshore wind power facility targets.

[0114] like Figure 3 As shown, the offshore wind power facility target detection system provided by the embodiment of the present invention includes a remote sensing image acquisition module 10, a remote sensing image segmentation module 20, an area of ​​interest clustering module 30, a wind power target detection module 40, and a remote sensing image integration module 50. Specifically, the remote sensing image acquisition module 10 is used to acquire the satellite remote sensing image to be detected. The remote sensing image segmentation module 20 is used to perform block preprocessing on the satellite remote sensing image to obtain a plurality of block satellite remote sensing images. The area of ​​interest clustering module 30 is used to cluster the plurality of block satellite remote sensing images into areas of interest in turn. The wind power target detection module 40 is used to crop the obtained area of ​​interest clustering areas in turn and input them into the offshore wind power facility detection network to detect independent wind power facility targets in the area. The remote sensing image integration module 50 is used to integrate the detected block satellite remote sensing images to obtain the final wind power facility target detection result.

[0115] The process of each module realizing its function in the offshore wind power facility target detection system provided in the embodiment of the present invention is the same as the steps in the offshore wind power facility target detection method provided in the above embodiment of the present invention, and its repeated description will be omitted here.

[0116] The present invention relates to a detection method and a detection system for offshore wind power facility targets. Based on satellite remote sensing technology, key monitoring of suspicious points and areas in the sea area is carried out, and the final detection results of offshore wind power facility targets are locked by clustering the regions of interest of several block satellite remote sensing images and detecting the offshore wind power facility detection network, so as to timely discover illegal and irregular behaviors in the jurisdiction of the sea area, realize long-term continuous control of relevant projects in the sea area, and standardize the use order of the sea area.

[0117] It should be understood that the above specific embodiments of the present invention are only used to illustrate or explain the principles of the present invention, and do not constitute a limitation of the present invention. Therefore, any modifications, equivalent substitutions, improvements, etc. made without departing from the spirit and scope of the present invention should be included in the protection scope of the present invention. In addition, the appended claims of the present invention are intended to cover all changes and modifications that fall within the scope and boundaries of the appended claims, or the equivalent forms of such scope and boundaries.

[0118] The present invention has been described above with reference to the embodiments of the present invention. However, these embodiments are for illustrative purposes only and are not intended to limit the scope of the present invention. The scope of the present invention is defined by the appended claims and their equivalents. Without departing from the scope of the present invention, a person skilled in the art may make various substitutions and modifications, which should all fall within the scope of the present invention.

[0119] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0120] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0121] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

Claims

1. A method for detecting an offshore wind power facility target, characterized in that: include: Acquire the satellite remote sensing image to be detected; Performing block preprocessing on the satellite remote sensing image to obtain a plurality of block satellite remote sensing images; Clustering several block satellite remote sensing images into regions of interest in sequence; The obtained clustered regions of interest are cut in turn and input into the offshore wind power facility detection network to detect independent wind power facility targets in the region; The detected block satellite remote sensing images are integrated to obtain the final detection results of the wind power facility target.

2. The detection method according to claim 1, characterized in that: The region of interest clustering region clipping includes: If the candidate frames of the two segmented satellite remote sensing images satisfy the following formula: IoU(box1,box2)>0 Then the two candidate frames are merged and cropped to obtain the region of interest; Among them, box1 and box2 respectively represent the candidate boxes of the two block satellite remote sensing images.

3. The detection method according to claim 1, characterized in that: The offshore wind power facility detection network includes: a feature encoder, a feature fusion module and a detection head.

4. The detection method according to claim 3, characterized in that: The feature encoder includes a convolution layer, a feature fusion layer and a fast spatial pyramid pooling layer. The convolution layer performs a convolution operation on each of the regions of interest; The feature fusion layer splits the region of interest into two parts, the first part is input into two bottleneck layers, the second part is concatenated with the outputs of the two bottleneck layers, and the information of the two parts is fused using a convolutional layer with predetermined parameters; The fast spatial pyramid pooling layer performs a pooling operation on the input regions of interest of different sizes and proportions.

5. The detection method according to claim 3, characterized in that: The feature fusion module adopts a bidirectional design mode to fuse and splice from the highest-level feature map to the lower-level feature map, and then performs the same fusion operation from the lowest level to the higher level.

6. The detection method according to claim 3, characterized in that: The detection head uses Wasserstein distance to evaluate the position accuracy of the central target of the wind power facility to be detected, wherein the Wasserstein distance is expressed by the following formula: By normalizing the formula in exponential form, we get: Where C represents a constant coefficient, represents the Wasserstein distance, NWD(G1, G2) represents the normalized Wasserstein distance, G1 and G2 represent two different Gaussian distributions, respectively. represents the central abscissa of the inscribed ellipse of the Gaussian distribution G1 bounding box, represents the central ordinate of the inscribed ellipse of the Gaussian distribution G1 bounding box, w1 represents the width of the Gaussian distribution G1 bounding box, h1 represents the height of the Gaussian distribution G1 bounding box, The central abscissa of the inscribed ellipse of the Gaussian distribution G2 bounding box, represents the central ordinate of the inscribed ellipse of the Gaussian distribution G2 bounding box, w2 represents the width of the Gaussian distribution G2 bounding box, and h2 represents the height of the Gaussian distribution G2 bounding box.

7. The detection method according to claim 6, characterized in that: The position of the central target is fitted by using the heights and widths of eight sampling points around the central target, and the average of multiple central target positions is used as the final independent wind power target.

8. The detection method according to claim 3, characterized in that: The offshore wind power facility detection network also includes a loss function calculation module, and the loss function adopts the following expression: loss total =CrossLoss(p,q)+DFL(S i ,S i+1 )+L NWD (pred box ,gt box )+L NWD (pred center ,gt center ) Among them, loss total represents the total loss function, CrossLoss(p,q) represents the improved cross entropy loss function, DFL(S i ,S i+1 ) represents the distribution focus loss function, L NWD (pred box ,gt box ) represents the loss function of the target box regression based on the Wasserstein distance, L NWD (pred center ,gt center ) represents the loss function for regressing the center position of the target box based on the Wasserstein distance.

9. The detection method according to claim 1, characterized in that: Integrate the detected block satellite remote sensing images, including: If there are wind power facility targets in the overlapping area of ​​two adjacent block satellite remote sensing images, the non-maximum suppression method is used to retain the detection result of one block satellite remote sensing image and delete the detection result of the other block satellite remote sensing image.

10. A detection system for an offshore wind power facility target, characterized in that: The detection system comprises: A remote sensing image acquisition module, which is used to acquire the satellite remote sensing image to be detected; A remote sensing image segmentation module is used to perform segmentation preprocessing on the satellite remote sensing image to obtain a plurality of segmented satellite remote sensing images; An area of ​​interest clustering module is used to cluster the areas of interest of a plurality of block satellite remote sensing images in sequence; A wind power target detection module is used to sequentially clip the obtained region of interest clustering regions and input them into an offshore wind power facility detection network to detect independent wind power facility targets in the region; The remote sensing image integration module is used to integrate the detected block satellite remote sensing images to obtain the final detection results of the wind power facility target.

Images