Obstacle automatic identification method based on improved inverse multivariate quadratic pooling function

Abstract

The invention discloses an automatic obstacle identification method based on an improved inverse multivariate quadratic pooling function, and relates to the technical field of seismic exploration obstacle identification, and the method comprises the following steps: S1, obtaining a ground feature marking map and a satellite picture in a work area range; s2, extracting obstacle information in the ground feature labeling map and corresponding satellite picture pixel values, establishing a network training label library according to the extracted information, and building a trained convolutional neural network; s3, training the constructed convolutional neural network to obtain an obstacle recognition model; in a training process, endowing each activation value in a pooling region of the convolutional neural network with a different weight; and S4, inputting the satellite picture in the work area into the trained obstacle recognition model for obstacle recognition, and outputting an obstacle recognition result. According to the pooling method, each activation value in the pooling area is endowed with a different weight, and the risk of losing important original information caused by losing most activation values is well reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of seismic exploration obstacle identification, and more specifically to an improved inverse multivariate quadratic pooling function obstacle automatic identification method. Background Art

[0002] At present, automatic obstacle avoidance technology has been applied in seismic exploration. The general practice is to combine satellite images in the work area, establish a deep learning sample library based on pixel values, and automatically identify satellite images through neural network training. When faced with satellite image information composed of massive data, in order to improve the accuracy and efficiency of automatic recognition, pooling functions will be added in deep learning. Currently, most of them use maximum pooling or average pooling. These two methods have problems in retaining the original image information, low recognition accuracy, and risks in the automatic obstacle recognition process.

[0003] In the prior art, the patent with publication number CN111142150A discloses a method for automatic intelligent obstacle avoidance design for seismic exploration, which includes: step 1, obtaining a map of ground object markings and satellite images within the scope of the work area; step 2, extracting obstacle information and corresponding satellite image pixel values ​​in the ground object marking map, and constructing a training sample library of different obstacles; step 3, training the sample library with a neural network to obtain neural network model parameters; step 4, automatically identifying satellite images within the entire work area; step 5, automatically avoiding obstacles for the shot points according to the identification results. The method for automatic intelligent obstacle avoidance design for seismic exploration trains the mapping relationship between ground object marking information and corresponding satellite image pixel values ​​through a neural network, and then automatically identifies and classifies obstacles in the satellite images of the entire work area through the trained neural network, which better solves the problem of identifying and classifying various obstacles in satellite images.

[0004] The above patents still have the above problems: pooling functions are added in deep learning, and most of them currently use maximum pooling or average pooling. These two methods have problems in retaining the original image information, low recognition accuracy, and risks in the automatic obstacle recognition process. Summary of the invention

[0005] In order to overcome the defects existing in the above-mentioned prior art, the present invention discloses an automatic obstacle recognition method of an improved inverse multivariate quadratic pooling function. The purpose of the present invention is to solve the problems existing in the prior art in maximum pooling or average pooling in retaining the original image information, low recognition accuracy, and risks in the automatic obstacle recognition process. The pooling method of the present invention assigns different weights to each activation value in the pooling area, which greatly reduces the risk of losing important original information caused by losing most of the activation values. Each activation value in the pooling area contributes to the final pooling output, which not only retains important information, but also does not distract attention too much. This method effectively retains the important information of the feature map and improves the final recognition accuracy of the network.

[0006] In order to achieve the above objectives, the technical solution adopted by the present invention is: An improved inverse multivariate quadratic pooling function obstacle automatic recognition method comprises the following steps: S1. Obtain the marked maps and satellite images of the objects within the work area; S2. Extract obstacle information and corresponding satellite image pixel values ​​from the object annotation map, establish a label library for network training based on the extracted information, and build a trained convolutional neural network; Preferably, in the convolutional neural network, the network includes 5 levels, each of levels 1 to 4 is divided into left and right sides, and there are two convolution operations on the left and right sides respectively, and each convolution layer is accompanied by an activation layer. In the present invention, the purpose of setting up the convolutional neural network in this way is to extract the feature information of the input data, complete the nonlinear transformation between the data, and enhance the ability of the network.

[0007] Preferably, in the convolutional neural network, the convolution on the left increases the number of channels of the feature map, and the convolution on the right reduces the number of channels of the feature map. In the present invention, the network on the left is a feature extraction network, which increases the number of channels while extracting feature information of the image; the network on the right is a feature fusion network, which concatenates the feature map on the left side of the same level and the feature map generated by upsampling on the right side, reduces the number of channels, and makes the final output feature map consistent with the specifications of the initial training input image.

[0008] Preferably, in the convolutional neural network, layer 5 has two convolution operations to increase the number of channels, and the left side between each layer is connected to the pooling layer, and the right side is connected to the transposed convolution. In the present invention, the pooling operation is equivalent to a dimensionality reduction in the spatial range, so that the model can extract a wider range of features, while reducing the input size of the next layer, thereby reducing the amount of calculation and preventing overfitting to a certain extent. Transposed convolution is also called upsampling, and its purpose is to make the size of the feature map of the final output consistent with the initial input size.

[0009] S3. Train the constructed convolutional neural network to obtain an obstacle recognition model. During the training process, assign different weights to each activation value in the pooling area of ​​the convolutional neural network. In the present invention, after starting training, the number of training times (epochs for short) for the sample set is set, and one epoch represents one forward calculation and back propagation. The training sample size and type input below are (256, 256, 3), where 256 represents the image size and 3 represents the number of channels. The convolution kernel size is 3×3, the step size is 1, and the padding type is set to be the same as the original image size (that is, the convolution layer input image size is consistent with the output image size). The output feature map type after the first convolution of layer 1 is (256, 256, 64), and the number of channels increases from 3 to 64. The output feature map type after the second convolution is (256, 256, 64), and the number of channels does not increase.

[0010] Preferably, in the step S3, training the constructed convolutional neural network includes the following steps: S31, initializing the weights of the constructed convolutional neural network; S32, inputting the sample images in the label library into the convolution layer of the convolutional neural network layer 1 in sequence, performing convolution calculation and outputting; In the present invention, the above steps are convolution calculations. Assume that a channel of the convolution input is , the convolution kernel is , the fill type is set to the same size as the original image, so the original image will be filled , and the final output after the convolution operation is .

[0011] S33, inputting the result of the convolution calculation output into the pooling layer of the convolutional neural network, performing pooling calculation and outputting; In the present invention, the above steps are pooling calculations. After the above convolution calculations, the data are directly input into the pooling layer for calculation. The present invention optimizes the traditional maximum pooling method into an improved inverse multi-quadratic function pooling method.

[0012] Preferably, in the pooling layer of the convolutional neural network, the pooling window size of the feature map of a channel is m×m, and the feature activation value is extracted as: in, Represents the feature activation value.

[0013] In the present invention, the pooling function is divided into two parts, namely a weighted calculation function and a pooling weight coefficient function.

[0014] Preferably, in the pooling layer of the convolutional neural network, the weighted calculation function is: in, for The average value of is the weight coefficient.

[0015] In the present invention, for the traditional maximum pooling function and average pooling function, there is a problem of excessive loss of original image feature information. The present invention proposes the following improved inverse multivariate quadratic basis function pooling weight coefficient method, which is as follows.

[0016] Preferably, in the pooling layer of the convolutional neural network, the pooling weight coefficient function is: in, is the center value of the inverse multi-quadratic basis function of the current pooling window, which is are vectors of the same dimension; is a vector The norm of and The distance between Indicates the corresponding feature activation value The weight of , which is the activation value of the pooling window Contribution to the final output; is a correction factor for Make The denominator of is zero, and the loss of original image information is corrected.

[0017] In the present invention, it can be seen from the above pooling weight coefficient function that the distance average activation value The farther the feature activation value is, The smaller the weight assigned, the less influential the feature activation values ​​that are farther away from the center will be while ensuring that all activation values ​​within the pooling window contribute.

[0018] Preferably, in the pooling layer of the convolutional neural network, after the pooling operation, the output result of the m×m pooling window is: .

[0019] In the present invention, if the pooling window size is selected as 2×2 and the pooling output type is (128, 128, 64), the width and height of the feature map can be reduced to 1 / 2 of the original, and the amount of data for subsequent recognition calculations can be reduced by half.

[0020] S34, in the convolutional neural network, repeat the convolution calculation of step S32 and the pooling calculation of step S33 for its level 2 to level 5; The above steps are repeated convolution and pooling calculations. Layers 2 to 5 in the network repeat the convolution calculation of step S32 and the pooling calculation of step S33. The output types of layers 2 to 5 are (128, 128, 128), (64, 64, 256), (32, 32, 512), and (16, 16, 1024).

[0021] S35, after reaching level 5, the output on the left is deconvolved, and each deconvolution increases the size of the feature map to twice its original size. Then the output result is propagated upward from the right side of the network. After reaching level 4, it is spliced ​​with the feature map on the left side of the corresponding level 4, and then two convolution operations are performed; The above steps are deconvolution calculations, the purpose of which is to reduce the number of channels. At this time, the output type after convolution is (32, 32, 512).

[0022] S36, repeat the deconvolution calculation and two convolution calculations of step S35 for level 4 to level 1; The above steps are upsampling calculations. In this stage, the deconvolution calculation of step S35 and two convolution calculations are repeated from level 4 to level 1. The output types from level 3 to level 1 are (64, 64, 256), (128, 128, 128), (256, 256, 64), and the final output is subject to a full convolution with a convolution kernel size of 1×1.

[0023] S37, calculate the error between the full convolution result and the expected result, then return the error layer by layer, calculate the error of each layer, and update the weights in step S31 according to the error, and then perform the next forward propagation.

[0024] In the present invention, the forward propagation ends and the backward propagation begins. The error between the full convolution result and the expected result is calculated, and then the error is returned layer by layer, the error of each layer is calculated, the weight update in step S31 is performed, and then the next forward propagation is performed. One forward propagation and one backward propagation represent one epoch, and iterative calculation is performed according to the given training epoch value.

[0025] Preferably, in step S37, the error is: in, Represents a sample, Indicates the actual label, represents the predicted output, Represents the total number of samples.

[0026] S4. Input the satellite images in the work area into the trained obstacle recognition model to perform obstacle recognition, and output the obstacle recognition results.

[0027] Beneficial effects of the present invention: The overall training accuracy of the neural network constructed by the present invention is significantly improved based on the advantages of the improved inverse multivariate quadratic function pooling, thereby obtaining more accurate results for identifying different types of obstacles on the surface. In the traditional average pooling method, all activation values ​​contribute equally, which reduces the feature strength of the overall area, while the pooling method of the present invention assigns different weights to each activation value in the pooling area, which greatly reduces the risk of losing important original information caused by discarding most of the activation values. Each activation value in the pooling area contributes to the final pooling output, which not only retains important information but also does not distract attention too much. This method effectively retains the important information of the feature map and improves the final recognition accuracy of the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a process diagram of the pooling method of the present invention; Figure 2 This is the actual work area satellite image; Figure 3 Obstacle recognition result (pooling method of the present invention); Figure 4 Obstacle recognition result (traditional maximum pooling method). DETAILED DESCRIPTION

[0029] The concept, specific structure and technical effects of the present invention will be clearly and completely described below in conjunction with the embodiments and drawings to fully understand the purpose, characteristics and effects of the present invention.

[0030] Example 1 An improved inverse multivariate quadratic pooling function obstacle automatic recognition method comprises the following steps: S1. Obtain the marked maps and satellite images of the objects within the work area; S2. Extract obstacle information and corresponding satellite image pixel values ​​from the object annotation map, establish a label library for network training based on the extracted information, and build a trained convolutional neural network; S3. Train the constructed convolutional neural network to obtain an obstacle recognition model. During the training process, assign different weights to each activation value in the pooling area of ​​the convolutional neural network. S4. Input the satellite images in the work area into the trained obstacle recognition model to perform obstacle recognition, and output the obstacle recognition results.

[0031] In the traditional average pooling method, all activation values ​​contribute equally, which reduces the feature strength of the overall area. The pooling method of this embodiment gives different weights to each activation value in the pooling area, which greatly reduces the risk of losing important original information caused by losing most of the activation values. Each activation value in the pooling area contributes to the final pooling output, which not only retains important information but also does not distract attention too much. This method effectively retains the important information of the feature map and improves the final recognition accuracy of the network. Example 2 Based on Example 1, this example further explains the convolutional neural network constructed in step S2.

[0032] In the convolutional neural network (U-Net is used as an example), the network is divided into 5 levels. Each level from level 1 to level 4 is divided into left and right sides. There are two convolution operations on each side, and each convolution layer is accompanied by an activation layer. The convolution on the left will increase the number of channels of the feature map, and the convolution on the right will reduce the number of channels of the feature map. Level 5 has only two convolution operations that increase the number of channels. The left side of each level is connected to the pooling layer, and the right side is connected to the transposed convolution.

[0033] Example 3 Based on Example 2, this example further explains the convolutional neural network training in step S3.

[0034] In this embodiment, after starting training, the number of training times (epochs for short) for the sample set is set, and one epoch represents one forward calculation and back propagation. The training sample size and type input below are (256, 256, 3), where 256 represents the image size and 3 represents the number of channels. The convolution kernel size is 3×3, the step size is 1, and the padding type is set to be the same as the original image size (that is, the convolution layer input image size is consistent with the output image size). The output feature map type after the first convolution of layer 1 is (256, 256, 64), and the number of channels increases from 3 to 64. The output feature map type after the second convolution is (256, 256, 64), and the number of channels does not increase.

[0035] In the step S3, training the constructed convolutional neural network includes the following steps: S31, initializing the weights of the constructed convolutional neural network; S32, inputting the sample images in the label library into the convolution layer of the convolutional neural network layer 1 in sequence, performing convolution calculation and outputting; In this embodiment, the above steps are convolution calculations. Assume that a channel of the convolution input is , the convolution kernel is , the fill type is set to the same size as the original image, so the original image will be filled , and the final output after the convolution operation is .

[0036] S33, inputting the result of the convolution calculation output into the pooling layer of the convolutional neural network, performing pooling calculation and outputting; In this embodiment, the above steps are pooling calculations. After the above convolution calculations, the convolution layer is directly input for calculation (see the process). Figure 1 ), this embodiment optimizes the traditional maximum pooling method to an improved inverse multi-quadratic function pooling method. Taking the feature map of one channel as an example, assuming that the pooling window size is m×m, after the feature activation value is extracted, it can be expressed as the following set X: in, Represents the feature activation value.

[0037] In this embodiment, the pooling function is divided into two parts, namely a weighted calculation function and a pooling weight coefficient function, which is referred to as the weighted calculation function below.

[0038] in, for The average value of is the weight coefficient.

[0039] In this embodiment, for the traditional maximum pooling function and average pooling function, there is a problem of excessive loss of original image feature information. This embodiment proposes the following improved inverse multivariate quadratic basis function pooling weight coefficient method, which is as follows.

[0040] in, is the center value of the inverse multi-quadratic basis function of the current pooling window, which is are vectors of the same dimension; is a vector The norm of and The distance between Indicates the corresponding feature activation value The weight of , which is the activation value of the pooling window Contribution to the final output; is a correction factor for Make The denominator of is zero, and the loss of original image information is corrected.

[0041] In this embodiment, it can be seen from the above pooling weight coefficient function that the distance average activation value The farther the feature activation value is, The smaller the weight assigned, the smaller the weight assigned to the feature activation value that is farther away from the center and has less influence, while ensuring that all activation values ​​in the pooling window contribute. So after the pooling operation, the output result of the m×m pooling window is: .

[0042] In this embodiment, if the pooling window size is selected as 2×2 and the pooling output type is (128, 128, 64), the width and height of the feature map can be reduced to 1 / 2 of the original, and the amount of data for subsequent recognition calculations can be reduced by half.

[0043] S34, repeat convolution and pooling calculation. Layers 2 to 5 in the network repeat the convolution calculation of step S32 and the pooling calculation of step S33. The output types of layers 2 to 5 are (128, 128, 128), (64, 64, 256), (32, 32, 512), and (16, 16, 1024).

[0044] S35, deconvolution calculation. After reaching level 5, the output on the left is deconvolved. Each deconvolution increases the size of the feature map to twice its original size. Then the output result is propagated upward from the right side of the network. After reaching level 4, it is concatenated with the feature map on the left side of the corresponding level 4 (jump connection), and then two more convolution operations are performed to reduce the number of channels. At this time, the output type after convolution is (32, 32, 512).

[0045] S36, sampling calculation. In this stage, the deconvolution calculation of step S35 and two convolution calculations are repeated from level 4 to level 1. The output types from level 3 to level 1 are (64, 64, 256), (128, 128, 128), (256, 256, 64), and the final output is subject to a full convolution with a convolution kernel size of 1×1.

[0046] S37, at this point, the forward propagation ends and the backward propagation begins. Calculate the error between the full convolution result and the expected result, then return the error layer by layer, calculate the error of each layer, perform the weight update in step S31, and then perform the next forward propagation. One forward propagation and one backward propagation represent one epoch, and iterate the calculation according to the given training epoch value. The above error calculation function is as follows: in, Represents a sample, Indicates the actual label, represents the predicted output, Represents the total number of samples.

[0047] Using the above steps, the actual collection of work area obstacle recognition results is as follows: Figure 2-Figure 4 . Satellite images of seismic acquisition areas such as Figure 2 The obstacle recognition result of this embodiment is as follows: Figure 3 , the results of the traditional maximum pooling method are as follows Figure 4 , as can be seen from the figure, Figure 3 The recognition accuracy is much higher than Figure 4 .

[0048] The above is a specific description of the implementation mode of the present invention, but the present invention is not limited to the described embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present invention, and these equivalents or substitutions are all included in the scope defined by the claims of the present invention.

Claims

1. An improved inverse multivariate quadratic pooling function obstacle automatic recognition method, characterized in that: The following steps are involved: S1. Obtain the marked maps and satellite images of the objects within the work area; S2. Extract obstacle information and corresponding satellite image pixel values ​​from the object annotation map, establish a label library for network training based on the extracted information, and build a trained convolutional neural network; S3. Train the constructed convolutional neural network to obtain an obstacle recognition model. During the training process, assign different weights to each activation value in the pooling area of ​​the convolutional neural network. S4. Input the satellite images in the work area into the trained obstacle recognition model to perform obstacle recognition, and output the obstacle recognition results.

2. The automatic obstacle recognition method according to claim 1, characterized in that: In the pooling layer of the convolutional neural network, the pooling window size of the feature map of a channel is m×m, and the feature activation value is extracted as: ， in, Represents the feature activation value.

3. The automatic obstacle recognition method according to claim 2, characterized in that: In the pooling layer of the convolutional neural network, the weighted calculation function is: ， in, for The average value of is the weight coefficient.

4. The automatic obstacle recognition method according to claim 3, characterized in that: In the pooling layer of the convolutional neural network, the pooling weight coefficient function is: ， in, is the center value of the inverse multi-quadratic basis function of the current pooling window, which is are vectors of the same dimension; is a vector The norm of and The distance between Indicates the corresponding feature activation value The weight of , which is the activation value of the pooling window Contribution to the final output; is a correction factor.

5. The automatic obstacle recognition method according to claim 4, characterized in that: In the pooling layer of the convolutional neural network, after the pooling operation, the output result of the m×m pooling window is: 。 6. The automatic obstacle recognition method according to claim 1, characterized in that: In the convolutional neural network, the network includes 5 levels, and each level from level 1 to level 4 is divided into a left side and a right side, and there are two convolution operations on the left and right sides respectively, and each convolution layer is accompanied by an activation layer.

7. The automatic obstacle recognition method according to claim 6, characterized in that: In the convolutional neural network, the convolution on the left increases the number of channels of the feature map, and the convolution on the right reduces the number of channels of the feature map.

8. The automatic obstacle recognition method according to claim 7, characterized in that: In the convolutional neural network, layer 5 has two convolution operations to increase the number of channels, and the left side of each layer is connected to the pooling layer, and the right side is connected to the transposed convolution.

9. The automatic obstacle recognition method according to claim 8, characterized in that: In the step S3, training the constructed convolutional neural network includes the following steps: S31, initializing the weights of the constructed convolutional neural network; S32, inputting the sample images in the label library into the convolution layer of the convolutional neural network layer 1 in sequence, performing convolution calculation and outputting; S33, inputting the result of the convolution calculation output into the pooling layer of the convolutional neural network, performing pooling calculation and outputting; S34, in the convolutional neural network, repeat the convolution calculation of step S32 and the pooling calculation of step S33 for its level 2 to level 5; S35, after reaching level 5, the output on the left is deconvolved, and each deconvolution increases the size of the feature map to twice its original size. Then the output result is propagated upward from the right side of the network. After reaching level 4, it is spliced ​​with the feature map on the left side of the corresponding level 4, and then two convolution operations are performed; S36, repeat the deconvolution calculation and two convolution calculations of step S35 for level 4 to level 1; S37, calculate the error between the full convolution result and the expected result, then return the error layer by layer, calculate the error of each layer, and update the weights in step S31 according to the error, and then perform the next forward propagation.

10. The automatic obstacle recognition method according to claim 9, characterized in that: In the step S37, the error is: ， in, Represents a sample, Indicates the actual label, represents the predicted output, Represents the total number of samples.

Images