Speed spectrum pickup method and device based on target detection

Abstract

The invention relates to the technical field of seismic data processing, in particular to a velocity spectrum pickup method and device based on target detection. The method comprises the following steps: converting pre-acquired velocity spectrum data into picture data; pre-picked tv pairs and corresponding velocity spectrum picture data are adopted as sample labels, and model training is performed on the YOLOv7 network to extract picture features; performing model processing on the velocity spectrum picture data to be processed by using the trained YOLOv7 network to output pickup result data obtained by reasoning; and correcting the pickup result data to avoid pickup multiples so as to obtain final pickup result data. The method can better adapt to the speed spectrum pickup work of data with different signal-to-noise ratios, the pickup result is close to the manual label pickup style, manual intervention is not needed in the whole process, compared with manual pickup, the pickup accuracy is close to that of a manual pickup method, and the pickup efficiency of an automatic pickup method is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of seismic data processing, and in particular to a velocity spectrum picking method and device based on target detection. Background Art

[0002] Velocity spectrum picking refers to analyzing seismic records to obtain velocity information of strata or lithology, thereby providing a basis for geological interpretation, oil and gas exploration, etc. In seismic exploration, when seismic waves propagate in strata, their propagation velocity is affected by factors such as stratum thickness, density, porosity, etc. By analyzing the velocity of seismic records, the velocity distribution of strata can be determined, and then information such as the lithology and structure of the strata can be inferred. Through velocity spectrum picking, the velocity distribution of strata can be obtained, providing an important basis for subsequent geological interpretation and oil and gas exploration. At the same time, velocity spectrum analysis is also one of the commonly used data processing methods in seismic exploration.

[0003] Currently, the velocity spectrum picking work is achieved through a human-computer interaction method, which requires professional technical personnel with rich seismic processing experience and geological academic backgrounds to spend a lot of time to complete. With the processing requirements for massive data and high-density interpretation of velocity spectra, manual velocity spectrum picking is increasingly difficult to meet the actual production needs, and there is an urgent need to automate the velocity spectrum picking work. In recent years, with the rapid development of artificial intelligence technology and its combined application in various industries, it has provided a new research direction for this work.

[0004] A large number of studies have been carried out at home and abroad on automatic velocity spectrum picking, which can be roughly divided into two categories: First, it is based on non-linear inversion methods, such as conjugate gradient method, Monte Carlo method, etc., that is, by establishing an objective function and constraint conditions, globally optimizing to solve the position of the best t-v pair, but this type of method requires certain prior constraints as conditions, manually setting parameters and initial models, and the calculation efficiency and accuracy are relatively low. Second, methods based on artificial intelligence technology, mainly including two categories: One is to use machine learning methods to obtain stacking velocity by identifying the peak values of energy clusters in the velocity spectrum, such as using BP neural network, or combining BP neural network with binary sorting tree, etc., and some other unsupervised learning machine learning methods, such as K-means clustering and DBSCAN clustering algorithms, etc. However, due to problems such as the weak learning ability of early shallow neural networks, being easily trapped in local minima or overfitting in machine learning, and the limitation of computing resources at that time, the accuracy of the results of this method is relatively low and cannot achieve satisfactory results; finally, it is a method based on deep learning of artificial intelligence to train network models to simulate manual picking experience, such as convolutional neural network (CNN), recurrent neural network (RNN), and long short-term memory network (LSTM), etc. However, due to the limitations of the network models used above, the picking accuracy is relatively low and it is difficult to meet the actual production requirements. Summary of the Invention

[0005] To solve the technical problems existing in the above-mentioned prior art, the present invention provides a velocity spectrum picking method and device based on object detection.

[0006] To achieve the above object, the embodiments of the present invention provide the following technical solutions:

[0007] In a first aspect, in an embodiment provided by the present invention, a velocity spectrum picking method based on object detection is provided, and the method includes the following steps:

[0008] Convert the pre-acquired velocity spectrum data into picture data;

[0009] Use the pre-picked tv pairs and the corresponding velocity spectrum picture data as sample labels to train the YOLOv7 network to extract picture features;

[0010] Process the velocity spectrum picture data to be processed by the trained YOLOv7 network model to output the inferred picking result data;

[0011] To avoid picking multiple waves, correct the picking result data to obtain the final picking result data.

[0012] As a further solution of the present invention, the picture data format is RGB3 channels; wherein, the picture dimension is H×W×3, H is the picture height, W is the picture width, and 3 is the picture channel number.

[0013] As a further solution of the present invention, the extraction of picture features by the YOLOv7 network includes:

[0014] First, input the picture with a size of H×W×3 into the backbone network, and then output three feature maps of different sizes through the head layer network, and output the prediction result after passing through REP and CBM.

[0015] As a further solution of the present invention, the training loss function of the YOLOv7 network consists of three parts, namely, localization regression loss, classification loss, and object confidence loss.

[0016] As a further solution of the present invention, the model training of the YOLOv7 network includes:

[0017] Train the network by inputting the sample label data into the YOLOv7 network. After calculating the loss function for the network output, update the parameters in the network through backpropagation, and then perform forward calculation and output. Repeat this process until the model training loss function converges to complete the model training of the YOLOv7 network.

[0018] As a further solution of the present invention, the method for correcting the picked result data to avoid picking multiple waves to obtain the final picked result data includes:

[0019] Performing anomaly detection on the obtained picked result data through a constrained K-means module, and obtaining effective final picked result data through screening.

[0020] As a further solution of the present invention, the method for performing anomaly detection on the obtained picked result data through a constrained K-means module includes:

[0021] Setting the K value according to the data type, randomly selecting K picked points as the clustering center points, calculating the Euclidean distance between the picked points and each center, and classifying the category of the center point with the closest distance as the category of the picked point, and then recalculating the clustering center points, and repeating the above steps until all points in the data set are closest to their corresponding centroids.

[0022] As a further solution of the present invention, the method for performing anomaly detection on the obtained picked result data through a constrained K-means module further includes:

[0023] Constraining each type of center point in the time direction. If the velocity value of the center point is less than the velocity value of the previous point in the time direction, then delete the picked points of this type, so as to correct the picked result.

[0024] As a further solution of the present invention, the formula for calculating the Euclidean distance between the picked points and each center is as follows:

[0025]

[0026] Second aspect, in another embodiment provided by the present invention, a velocity spectrum picking device based on object detection is provided, and the device includes: a sample processing module, a model training module, a processing module, and a correction module.

[0027] The sample processing module is used to convert the pre-acquired velocity spectrum data into picture data.

[0028] The model training module is used to use the pre-picked tv pairs and the corresponding velocity spectrum picture data as sample labels to train the YOLOv network to extract picture features.

[0029] The processing module is used to process the velocity spectrum picture data to be processed by the trained YOLOv network model to output the inferred picked result data.

[0030] The correction module is used to correct the picked result data to avoid picking multiple waves to obtain the final picked result data.

[0031] The technical solution provided by the present invention has the following beneficial effects:

[0032] The method and device for velocity spectrum picking based on object detection provided by the present invention include the following steps: converting pre-acquired velocity spectrum data into picture data; using pre-picked tv pairs and corresponding velocity spectrum picture data as sample labels to train a YOLOv7 network to extract picture features; using the trained YOLOv7 network to process the velocity spectrum picture data to be processed to output the inferred picking result data; and correcting the picking result data to avoid picking multiple waves to obtain the final picking result data.

[0033] The present invention can better adapt to the velocity spectrum picking work of different signal-to-noise ratio data. The picking result is close to the manual label picking style. The whole process does not require manual intervention. Compared with manual picking, the picking accuracy is close to that of manual picking, and the picking efficiency of the automatic picking method is significantly improved.

[0034] These aspects or other aspects of the present invention will be more clearly understood in the following description of the embodiments. It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present invention. Description of the Drawings

[0035] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other embodiments can be obtained based on these drawings without creative efforts.

[0036] Figure 1 It is a flowchart of the method for velocity spectrum picking based on object detection according to an embodiment of the present invention.

[0037] Figure 2 It is a structural diagram of the YOLOv7 network in the method for velocity spectrum picking based on object detection according to an embodiment of the present invention.

[0038] Figure 3 It is a structural block diagram of the device for velocity spectrum picking based on object detection according to an embodiment of the present invention.

[0039] In the figure: sample processing module - 100, model training module - 200, processing module - 300, correction module - 400. Detailed Embodiments

[0040] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0041] The flowchart shown in the accompanying drawings is only an example, and does not necessarily include all the contents and operations / steps, nor does it necessarily need to be executed in the described order. For example, some operations / steps can be decomposed, combined, or partially merged, so the actual execution order may change according to the actual situation.

[0042] It should be understood that the terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an", and "the" are intended to include the plural forms.

[0043] Specifically, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0044] Please refer to Figure 1 , Figure 1 which is a flowchart of a velocity spectrum picking method based on object detection provided by an embodiment of the present invention. As Figure 1 shown, the velocity spectrum picking method based on object detection includes steps S10 to S40.

[0045] S10. Convert the pre-acquired velocity spectrum data into picture data for use as a sample label data set for network model training.

[0046] In the embodiment of the present invention, the picture data format is RGB3 channels; wherein, the picture dimension is H×W×3, H is the picture height, W is the picture width, and 3 is the number of picture channels.

[0047] S20. Use the pre-picked tv pairs (time-velocity pairs, that is, a sequence composed of a series of time-velocity pairs picked on each velocity spectrum picture data) and the corresponding velocity spectrum picture data as sample label data to train the YOLOv7 network to extract picture features.

[0048] It should be noted that the picking of tv pairs can be obtained manually.

[0049] It should be noted that the network structure diagram of the YOLOv7 network is as Figure 2 shown, and the YOLOv7 network extracts picture features, including:

[0050] First, an input image with a size of H×W×3 is fed into the backbone network. Then, through the head layer network, three feature maps of different sizes are output. After passing through REP and CBM, the prediction results are obtained. The standard output is 80 categories, and this method is only for one category. Then each output (x, y, w, h, o), namely the coordinate position and foreground / background, and 3 refers to the number of anchors. Therefore, the output of each layer is (80 + 5)×3 = 255, which is then multiplied by the size of the feature map to obtain the final output.

[0051] The present invention also provides the backbone structure of the YOLOV7 network, as Figure 2 shown in Backbone.

[0052] First, it passes through 4 CBS modules (a neural network module composed of a Conv layer, a BN layer, and a Silu layer), which is composed of a two-dimensional convolutional layer Conv + normalization layer BN + activation function SiLU. Among them, the convolutional kernel size of the CBS1 convolutional layer is 3, and the stride is 1. The convolutional kernel size of the CBS2 convolutional layer is 3, and the stride is 2. The output feature map size is 1 / 4(H×W)×128, and it is input into the ELAN module.

[0053] The ELAN module (a neural network module for processing sequential data. Its main task is to capture long-range dependencies, usually applied to natural language processing or time series data. The ELAN module enables the network to learn more features and has stronger robustness by controlling the shortest and longest gradient paths) is an efficient network structure that enables the network to learn more features by controlling the shortest and longest gradient paths. ELAN has two branches. The first branch passes through a CBS3 for channel number change. The second branch first passes through a CBS3 module for channel number change, then through four CBS1 modules for feature extraction, and finally the four features are superimposed to obtain the final feature extraction result. Finally, it passes through a CBS3 module to output a feature map with a size of 1 / 4(H×W)×256, and then is input into the MP module (referring to the max pooling layer, a common structure in convolutional neural networks, used to downsample the input feature map to reduce the dimension of the feature map, thereby reducing the computational amount and storage requirements while maintaining the richness of features).

[0054] The MP module has two branches for downsampling. The first branch first passes through a maxpool max pooling layer for downsampling, and then passes through a CBS3 module to change the number of channels. The second branch first passes through a CBS3 module to change the number of channels, and then passes through a CBS2 module, which is also used for downsampling. Finally, the results of the first branch and the second branch are added together to obtain the result of super downsampling. Here, it is used as the MP-1 module, and the number of input and output channels remains unchanged.

[0055] After the output of the MP module, it passes through an ELAN module. The output feature map size is 1 / 8(H×W)×512 to the input branch of the first head, and continues to pass through the ELAN and MP modules to output a feature map size of 1 / 6(H×W)×1024 to the input branch of the second head. Finally, it continues to pass through the ELAN and MP modules to output a feature map size of 1 / 32(H×W)×1024 to the input branch of the third head. The above is the backbone part of the network.

[0056] As Figure 2 shown, the present invention also provides the head part of the YOLOV7 network.

[0057] First is the 32x downsampling input branch, whose input scale is 1 / 32(H×W)×1024. After passing through the SPPCSP module, which is composed of the SPP module and the CSP module. The role of the SPP module is to increase the receptive field, enabling the algorithm to adapt to images of different resolutions, and obtaining different receptive fields through maxpooling. In the first branch of the SPP module, it first passes through 3 CBS modules, and then through four branches of maxpool, which are 5, 9, 13, and 1 respectively. Among them, the four different maxpools can process different objects, and the four different scales of maxpooling have four receptive fields. Through different maxpool branches, small targets and large targets can be better distinguished. After cat splicing (the command to concatenate files in Unix or Linux systems. By using the cat command, the contents of multiple files can be concatenated together to form a new file), it then passes through two CBS modules. And through the CSP module, the features are first divided into two parts. One part undergoes normal processing, that is, only passing through one CBS module, and the other part undergoes the processing of the SPP module as described above. Finally, the two processed parts are merged together, which can reduce the computational amount by half, making the speed faster while the accuracy will increase. Finally, it passes through 1 CBS module and then outputs. The output feature map passes through the CBS3 module and then inputs to the UPSample upsampling module. Using transposed convolution and the nearest interpolation method, the operation of enlarging the size of the feature map is realized. The output feature map size is 1 / 16(H×W)×256, and it is concatenated with the feature map of the same size output by the 16x downsampling of backbone after passing through 1 CBS3 module in a cat manner, and then inputs to the ELAN-W module. The ELAN-W module is similar in structure to the ELAN module, and the output results of more hierarchical CBS modules are concatenated, enabling the network to learn more features and having stronger robustness. The output results pass through one CBS3 module and one UPSample upsampling module, and then are concatenated with the feature map of the same size output by the 8x downsampling of backbone after passing through 1 CBS3 module in a cat manner, and then output to the ELAN-W module for feature extraction. The output results are input to the 8x downsampling output branch of the REP module on the one hand, and input to the MP-2 module for downsampling on the other hand. The difference between it and MP-1 is that the output channels are doubled. The feature map output by MP-2 is concatenated with the result output by the 16x downsampling ELAN-W and then input to the next ELAN-W module. The output of the module is input to the 16x downsampling output branch of the REP module on the one hand, and on the other hand, after passing through the MP-2 module, it is concatenated with the output of the SPPCSPC module, and then after passing through the ELAN-W module, it is input to the 32x downsampling output branch of the REP module.

[0058] The three output judgment branches all pass through a REP module (Re-parameterization, a method for changing the distribution of random variables, which can help optimization algorithms better handle problems such as random noise and gradient disappearance. In neural networks, the REP module is usually used to introduce random noise or resample data to improve the generalization ability and robustness of the model) first, and then are adjusted by the CBM module (Component Business Model, a business component module, a simple and effective convolutional neural network attention module. Given an intermediate feature map, the CBM module infers the attention of the picture along two independent dimensions of channels and space in turn, and then multiplies the attention map by the input feature map for adaptive feature refinement) to adjust the channels and finally output. The REP module is divided into two mode branches, as Figure 1 shown, one is the train training mode branch, and the other is the deploy inference mode branch. The training module has three branches. The top branch is a 3x3 convolution for feature extraction, the middle branch is a 1x1 convolution for smoothing features, and the last branch does not perform convolution operations but only normalization operations. Finally, they are added together. In the inference module branch, there is a 3x3 convolution with a stride of 1, which is converted from the training module through re-parameterization. Model re-parameterization requires converting the 1x1 convolution into a 3x3 convolution and adding a 3x3 convolution to the normalization branch as well, and then performing a matrix addition fusion process, that is, adding the weights, to obtain a 3x3 convolution, and the three branches are fused into one with only a 3x3 convolution inside, and the weight is the superposition result of the three branches. The CBM module is basically the same as the CBS module, including a Conv convolutional layer, a BN normalization layer, and a sigmoid activation function layer, with a convolution kernel of 1x1 and a stride of 1.

[0059] In the embodiment of the present invention, the output sizes of the three output judgment branches are 1 / 8(H×W)×255, 1 / 16(H×W)×255, 1 / 32(H×W)×255. The number of channels of each output, 255, can be decomposed into 3 85s, that is, each branch corresponds to 3 prior boxes, and each prior box has 85 parameters. The 85 parameters of the prior box can be split into 4+1+80, where 4 represents the regression parameters of each feature point, which are the predicted center coordinates (x, y) of the bounding box and the width and height w, h of the bounding box, 1 is used to judge whether the feature point contains an object, and 80 is used to judge the type of object contained in each feature point.

[0060] The network is trained through the input sample label data. After calculating the loss function for the network output, backpropagation is used to update the parameters in the network, and then forward calculation is performed for the output. This process is repeated until the model training loss function converges to complete the model training of the YOLOv7 network.

[0061] The loss function consists of three parts, namely the localization regression loss loss local , the classification loss loss class and the object confidence loss loss obj .

[0062] The localization regression loss loss local is as shown in the following formula:

[0063] loss local = loss CIOU = 1 – CIOU(1)

[0064]

[0065]

[0066] S1 = (min(x p2 , x l1 ) - max(x p1 , x l1 )) * (min(y p2 , y l1 ) - max(y p1 , y l1 ) (4)

[0067] S2 = (x p2 - x p1 ) * (y p2 - y p1 ) + (x l2 - x l1 ) * (y l2 - y l1 ) - S1 (5)

[0068]

[0069]

[0070] x p1 = x p - w p / 2 (8)

[0071] y p1 = y p - h p / 2 (9)

[0072] x p2 = x p + w p / 2 (10)

[0073] y p2 = yp +h p / 2 (11)

[0074] x l1 =x l -w l / 2 (12)

[0075] y l1 =y l -h l / 2 (13)

[0076] x l2 =x l +w l / 2 (14)

[0077] y l2 =y l +h l / 2 (15)

[0078] Using the CIOU loss, the overlapping area, the distance between the center points, and the aspect ratio of the two rectangular boxes of the label box and the predicted box are simultaneously added to the calculation to improve the stability and convergence speed of training. In the formula, IOU is the intersection-over-union of the areas of the two rectangular boxes, S1 is the intersection area of the two rectangles, S2 is the union area of the two rectangles, ρ is the distance between the center points of the two boxes, c is the diagonal length of the smallest enclosing rectangle of the two boxes, v is the aspect ratio similarity of the two boxes, α is the influence factor of v, x p 、y p are the coordinates of the center point of the predicted box, w p 、h p are the width and height of the predicted box, x l 、y l are the coordinates of the center point of the label box, w l 、h l are the width and height of the label box.

[0079] The classification loss loss class is shown as follows:

[0080] loss class =loss BCE (16)

[0081]

[0082]

[0083] In the formula, the loss BCE binary cross-entropy loss function is adopted, where N is 80 classifications, x ip is the class prediction value, y ip is the probability of the current class obtained after passing through the activation function, y ilIs the true value of the label.

[0084] Target confidence loss obj Also use loss BCE Binary cross-entropy loss function.

[0085] Add weights to different loss functions, then the overall loss is:

[0086] Loss = 0.05loss local + 0.125loss class + 0.1loss obj (19).

[0087] S30. Process the velocity spectrum picture data to be processed by the YOLOv7 network after training is completed to output the picked result data obtained by inference.

[0088] It should be noted that the picked result data is the "time - velocity" pair picked result.

[0089] S40. To avoid picking multiple waves, correct the picked result data to obtain the final picked result data.

[0090] In the embodiment of the present invention, the step of avoiding picking multiple waves to correct the picked result data to obtain the final picked result data includes:

[0091] Perform anomaly detection on the obtained picked result data through the constrained K-means module, and obtain effective final picked result data through screening.

[0092] In the embodiment of the present invention, the step of performing anomaly detection on the obtained picked result data through the constrained K-means module includes:

[0093] Set the K value according to the data type, randomly select K picked points as the clustering center points, calculate the Euclidean distance between the picked points and each center, and classify the picked points with the closest center point category as the picked point category. Then recalculate the clustering center points, and repeat the above steps until all points in the dataset are closest to their corresponding centroids. Constrain the center points in the time direction. If the velocity value of the center point is less than the velocity value of the previous point in the time direction, then delete the picked points of this category, thereby correcting the picked result.

[0094] In the embodiment of the present invention, the formula for calculating the Euclidean distance between the picked points and each center is as follows:

[0095]

[0096] where (x, y) is the coordinate position of the picking point, and (x i , y i ) is the coordinate of the i-th center point among K points.

[0097] The present invention is based on a velocity spectrum picking model of the YOLOv7 network, and the network is optimized to improve the picking accuracy. The network is trained through velocity spectrum picture data and manually picked label samples. After training, the model can perform inference picking on the unpicked velocity spectrum, and the result is post-processed by the constrained K-means anomaly detection method to screen and eliminate outliers, making the result closer to manual picking.

[0098] It should be understood that although the above is described in a certain order, these steps are not necessarily executed in the above order. Unless there is a clear indication in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, a part of the steps in this embodiment may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be executed alternately or alternately with at least a part of other steps or steps in other steps.

[0099] In one embodiment, as shown in Figure 3 , a velocity spectrum picking device based on object detection is further provided in the embodiment of the present invention. The device includes a sample processing module 100, a model training module 200, a processing module 300, and a correction module 400.

[0100] The sample processing module 100 is used to convert the pre-acquired velocity spectrum data into picture data.

[0101] The model training module 200 is used to use the pre-picked tv pairs and the corresponding velocity spectrum picture data as sample labels to train the YOLOv7 network to extract picture features.

[0102] The processing module 300 is used to process the velocity spectrum picture data to be processed by using the trained YOLOv7 network model to output the inferred picking result data.

[0103] The correction module 400 is used to correct the picking result data to avoid picking multiple waves to obtain the final picking result data.

[0104] It should be understood that, as used herein, unless the context clearly supports an exception, the singular form "a" is intended to also include the plural form. It should also be understood that the "and / or" used herein refers to any and all possible combinations of one or more of the associated listed items. The serial numbers of the disclosed embodiments of the present invention above are only for description and do not represent the superiority or inferiority of the embodiments.

[0105] Those of ordinary skill in the art should understand that: the discussion of any of the above embodiments is only exemplary and is not intended to imply that the scope of the disclosure of the embodiments of the present invention (including the claims) is limited to these examples; under the concept of the embodiments of the present invention, the technical features in the above embodiments or different embodiments can also be combined, and there are many other variations in different aspects of the embodiments of the present invention as above, which are not provided in detail for the sake of brevity. Therefore, any omission, modification, equivalent replacement, improvement, etc. made within the spirit and principle of the embodiments of the present invention shall be included in the protection scope of the embodiments of the present invention.

Claims

1. A velocity spectrum picking method based on object detection, characterized in that, The method includes: Converting the pre-acquired velocity spectrum data into picture data; Using the pre-picked tv pairs and the corresponding velocity spectrum picture data as sample labels to train the YOLOv7 network for model training to extract picture features; Processing the velocity spectrum picture data to be processed by the trained YOLOv7 network model to output the picked result data obtained by inference; To avoid picking multiple waves, correcting the picked result data to obtain the final picked result data.

2. The velocity spectrum picking method based on object detection according to claim 1, characterized in that, The picture data format is RGB3 channels; Among them, the picture dimension is H×W×3, where H is the picture height, W is the picture width, and 3 is the number of picture channels.

3. The velocity spectrum picking method based on object detection according to claim 2, wherein The training of the YOLOv7 network for model training to extract picture features includes: First, input the picture with a size of H×W×3 into the backbone network, then output three feature maps of different sizes through the head layer network, and output the prediction results after passing through REP and CBM.

4. The velocity spectrum picking method based on object detection according to claim 1, characterized in that, The training loss function of the YOLOv7 network consists of three parts, namely the localization regression loss, the classification loss, and the object confidence loss.

5. The velocity spectrum picking method based on object detection according to claim 4, wherein, The model training of the YOLOv7 network includes: Training the network by inputting the sample label data into the YOLOv7 network. After calculating the loss function for the network output, backpropagate to update the parameters in the network, and then calculate the output forward. Repeat this process until the model training loss function converges to complete the model training of the YOLOv7 network.

6. The method for picking up velocity spectrum based on object detection according to claim 1, wherein The step of avoiding picking multiple waves to correct the picked result data to obtain the final picked result data includes: Performing anomaly detection on the obtained picked result data through the constrained K-means module, and obtaining the effective final picked result data through screening.

7. The method for picking up velocity spectrum based on object detection according to claim 6, characterized in that, The performing anomaly detection on the obtained picked result data through the constrained K-means module includes: Setting the K value according to the data type, randomly selecting K picking points as the clustering center points, calculating the Euclidean distance between the picking points and each center, and classifying the picking points with the closest center point category as the picking point category. Then recalculate the clustering center points, and repeat the above steps until all points in the dataset are closest to their corresponding centroids.

8. The method for picking up velocity spectrum based on object detection according to claim 7, wherein The performing anomaly detection on the obtained picked result data through the constrained K-means module further includes: Constraining the center points in the time direction. If the center point velocity value is less than the velocity value of the previous point in the time direction, then delete the picking points of this category, thereby correcting the picking result.

9. The velocity spectrum picking method based on object detection according to claim 8, characterized in that, The formula for calculating the Euclidean distance between the picking points and each center is as follows: L2 = (∑ i K ＝1 [(x - x i ) 2 + (y - y i ) 2 ) 1 / 2 where (x, y) is the coordinate position of the picking point, and (x i , y i ) is the coordinate of the i-th center point among K points, and K is the number of clustering center points.

10. A velocity spectrum picking device based on object detection, characterized in that, The device includes: a sample processing module, a model training module, a processing module, and a correction module; The sample processing module is used to convert the pre-acquired velocity spectrum data into picture data; The model training module is used to use the pre-picked tv pairs and the corresponding velocity spectrum picture data as sample labels to train the YOLOv network for model training to extract picture features; The processing module is used to process the velocity spectrum picture data to be processed by using the YOLOv network after training is completed, so as to output the picked result data obtained by inference; The correction module is used to correct the picked result data to avoid picking multiple waves, so as to obtain the final picked result data.

Images