An anti-aliasing acoustic wave image interpolation method based on self-adjusting closed-loop network

Abstract

The present invention proposes an anti-aliasing acoustic wave image interpolation method based on a self-adjusting closed-loop network, and builds a closed-loop CNN interpolation model, which includes two interpolation sub-models, one interpolation sub-model is used to interpolate missing data, and the other interpolation sub-model is used to re-interpolate the interpolation result of the missing channel. The two sub-models have the same structure and share weights, and on this basis, a closed-loop reconstruction error is defined to evaluate the interpolation result; the closed-loop interpolation model is trained using label data to obtain a model that converges on the label data set; a self-adjusting module is built to perform shearing transformation on the test data at different angles, and the test data are respectively input into the closed-loop interpolation model, and the optimal shearing angle is selected according to the closed-loop reconstruction error, and the interpolation result under the optimal shearing angle is inversely sheared to obtain the final interpolation result. The present invention can improve the application effect of the interpolation model on the test data and achieve the anti-aliasing result.

Description

Technical Field

[0001] The invention belongs to the field of image processing, and in particular relates to an anti-aliasing sound wave image interpolation method based on a self-adjusting closed-loop network. Background Art

[0002] Acoustic imaging is an imaging technology based on sensor arrays. The sensor array collects sound wave signals in space and images the target object. Active acoustic imaging excites sound wave signals through the source. The sound wave signals propagate in space, reflect when encountering the reflective interface, and are received by the sensor array. Passive acoustic imaging collects existing sound signals in space through the sensor array for positioning and imaging. Seismic exploration imaging and medical ultrasound imaging both belong to the field of acoustic imaging. In actual application, the cost and speed of acoustic imaging are affected by the number of array sensors. Due to the continuity of sound field propagation, the sound wave signals received by adjacent sensors often have a certain degree of redundancy. The number of sensors can be reduced, and the received signals of missing sensors can be interpolated and reconstructed through image interpolation methods to reduce the acquisition cost.

[0003] In recent years, deep learning has been widely used in the field of acoustic image interpolation [1-4]. Compared with traditional interpolation methods, interpolation methods based on deep learning have stronger learning ability and feature expression ability. When there are sufficient training samples, interpolation methods based on deep learning can achieve better interpolation effects. Wang et al. [1] first proposed an interpolation method based on deep learning. This method trains a convolutional neural network (CNN) on a synthetic dataset and effectively applies the trained CNN model to the actual dataset. Hu et al. [2] proposed a CNN training method based on self-supervision. This method only uses missing seismic data for training and can achieve better interpolation effects on the test dataset. Xiao et al. [3] applied the deep learning interpolation algorithm to ultrasonic data, effectively reducing the number of sensor channels and improving the ultrasonic imaging effect after beamforming. Vyas et al. [4] proposed an anti-aliasing interpolation method based on the frequency domain segmentation network.

[0004] Although deep learning algorithms have achieved good application results in the field of acoustic image interpolation, their generalization capabilities are limited. Due to the complex structure of the CNN model and poor interpretability, the effective application range of the CNN model is difficult to determine. Therefore, automatically evaluating the generalization effect of the CNN interpolation model on a new test set can help engineers evaluate the application effect of CNN, which is conducive to promoting the promotion and application of the CNN interpolation model in production and life, and has important application value. Summary of the invention

[0005] The present invention proposes an anti-aliasing acoustic wave image interpolation method based on a self-adjusting closed-loop network. The method adopts a closed-loop interpolation structure, which can automatically evaluate the interpolation results of the interpolation model. On this basis, the present invention further constructs a self-adjusting module, which adjusts the input image according to the evaluation results of the closed-loop network, finds the optimal miscutting parameters, and converts difficult samples into simple samples, which can effectively suppress aliasing, improve the interpolation effect, and achieve anti-aliasing acoustic wave image interpolation. The method proposed in the present invention can effectively improve the interpolation effect while reducing the number of sensors, thereby improving the quality of acoustic wave imaging.

[0006] An anti-aliasing acoustic wave image interpolation method based on a self-adjusting closed-loop network comprises the following steps:

[0007] Step 1: Build a closed-loop interpolation model;

[0008] Step 2: Train the closed-loop interpolation model using labeled data;

[0009] Step 3: Build the self-adjusting module.

[0010] Furthermore, the closed-loop interpolation model in step 1 includes two interpolation sub-models, one interpolation sub-model is used to interpolate the missing data, and the other interpolation sub-model is used to re-interpolate the interpolation results of the missing channels. The two sub-models have the same structure and share weights.

[0011] Preferably, the sub-model is a CNN model or a traditional interpolation model.

[0012] Preferably, in step 2, the interpolation loss function L is used interp And the closed-loop loss function L cyc Train the model.

[0013] Furthermore, in step 2, the overall function of the training process is:

[0014] L=λ1L interp +λ2L cyc

[0015] λ1,λ2>0, used to adjust the weights of the two loss functions;

[0016] Interpolation loss function L interp :

[0017] L interp =||zf W (x)||2

[0018] f W (·) represents the interpolation model, W represents the weight parameter of the model, z represents the original non-missing seismic data, x represents the missing data, x=z⊙m, m represents the non-missing trace mask, ⊙ represents the dot product;

[0019] Closed-loop loss function L cyc :

[0020]

[0021] f W′ (·) is another interpolation model, is the new missing data formed by the interpolation results of the missing channels,

[0022]

[0023] Furthermore, in step 3, a closed-loop reconstruction error of the test data is defined based on the closed-loop model to evaluate the application effect of the interpolation model on the new data set. A small reconstruction error indicates that the application effect of the interpolation model on the test set is good, otherwise, the application effect is poor.

[0024] Preferably, the closed-loop reconstruction error is:

[0025]

[0026] x * is the unlabeled test data, m * For x * The corresponding mask, is the closed loop output, Represents unlabeled test data x * The corresponding missing track interpolation results satisfy

[0027] Furthermore, in step 3, the test data is self-adjusted according to the closed-loop reconstruction error, the test data is sheared at different angles, and input into the closed-loop interpolation model respectively. The optimal shearing angle is selected according to the closed-loop reconstruction error, and the interpolation result at the optimal shearing angle is inversely sheared to obtain the final interpolation result.

[0028] The advantages of the present invention are that a closed-loop interpolation structure is adopted, a closed-loop reconstruction error is defined, and the application effect of the interpolation model on the test data is automatically evaluated; based on the application effect, combined with the shear transformation, the application effect of the interpolation model on the test data is improved to achieve the result of anti-aliasing. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a closed-loop network structure diagram of the present invention;

[0030] Figure 2 is the interpolation network structure diagram of the present invention;

[0031] Figure 3 It is a structural diagram of the self-adjusting module of the present invention;

[0032] Figure 4 is the test result diagram of actual seismic data, where: Figure 4 (a) is missing earthquake data, Figure 4 (b)-4(d) are the interpolation results of Bicubic interpolation, traditional CNN algorithm and the method of the present invention, respectively. Figure 4 (f)-4(h) are the interpolation errors corresponding to the three methods. Figure 4 (e) is the target earthquake data without missing data;

[0033] Figure 5 is the amplitude spectrum result diagram of actual seismic data, where: Figure 5 (a) is the target data amplitude spectrum, Figure 5 (b) is the amplitude spectrum of the downsampled data, Figure 5 (c)-5(e) are the amplitude spectrum of the Bicubic interpolation result, the amplitude spectrum of the traditional CNN interpolation result and the amplitude spectrum of the interpolation result of the method of the present invention respectively;

[0034] Figure 6 This is the actual ultrasonic data test result diagram; among them, Figure 6 (a) is the downsampled ultrasound data, Figure 6 (b)-6(d) are the Bicubic interpolation results, the traditional CNN interpolation results and the interpolation results of the method of the present invention respectively;

[0035] Figure 7 This is the amplitude spectrum result diagram of actual ultrasonic data. Figure 7 (a) is the amplitude spectrum of downsampled ultrasound data, Figure 7 (b)-7(d) are respectively the amplitude spectrum of the Bicubic interpolation result, the amplitude spectrum of the traditional CNN interpolation result and the amplitude spectrum of the interpolation result of the method of the present invention. DETAILED DESCRIPTION

[0036] The specific implementation of the present invention will be further described below in conjunction with the accompanying drawings.

[0037] (1) Building a closed-loop interpolation model

[0038] The closed-loop network structure diagram proposed by the present invention is as follows: Figure 1 As shown. For the label data, the original non-missing data is represented as z, the observed missing data is represented as x, m represents the non-missing channel mask, and the m value of 1 represents non-missing data. The missing data satisfies: x = z⊙m, where the symbol ⊙ represents the dot product. Figure 1 As shown, first, an interpolation model is used to learn the mapping from missing data x to non-missing data z. In this embodiment, the ResNets structure in the CNN model is used [5], and the corresponding interpolation loss function is as follows:

[0039] L interp=||zf W (x)||2

[0040] Among them, f W (·) represents the CNN interpolation model, and W represents the weight parameter of the CNN model.

[0041] According to the CNN model, the interpolated data f of the missing earthquake data can be obtained. W (x). Use the missing track mask (1-m) to extract the interpolation result of the missing track separately. The formula is as follows:

[0042]

[0043] Thus, the new missing data consisting of the interpolation results of the missing channels is obtained.

[0044] Use another CNN interpolation model f W′ (·) For new missing data Interpolation is performed, where W′ represents the weight parameter of the CNN interpolation model, and the closed-loop loss function is defined as follows:

[0045]

[0046] Among them, since the interpolation tasks of the two interpolation models are similar, the present invention constrains the two CNN interpolation sub-models to have the same structure and share weights, that is, constrains W and W′ to have the same value.

[0047] (2) Use labeled data to train the closed-loop interpolation model

[0048] The overall function of the training process is:

[0049] L=λ1L interp +λ2L cyc

[0050] Wherein, λ1, λ2>0, which is used to adjust the weights of the two loss functions. The CNN interpolation model used in the embodiment of the present invention is as follows: Figure 2 As shown, the ResNets structure [5] is adopted, including three residual modules, and residual connections are also used in the network output layer. In addition, other end-to-end network models such as U-Net [6] can also be used. During the training process, the weight W is adjusted to minimize the loss function L through gradient descent algorithms such as Adam [7] to achieve network optimization. The above-mentioned ResNets interpolation model structure is a type of CNN model, which can be replaced by other CNN models or traditional interpolation models.

[0051] (3) Build a self-adjusting module

[0052] Similarly, the unlabeled test data x* and its corresponding mask m * Input to the closed-loop network to get the closed-loop output The closed-loop reconstruction error of the test data is further defined as:

[0053]

[0054] in, Represents unlabeled test data x * The corresponding missing track interpolation results satisfy

[0055] Closed-loop reconstruction error It is used to evaluate the interpolation effect. If the reconstruction error is small, it means that the interpolation model has a good application effect on the test set. Otherwise, it means that the application effect is poor.

[0056] The present invention further performs self-adjustment on the test data according to the closed-loop reconstruction error to improve the interpolation effect. Figure 3 As shown. First, the seismic data is sheared according to the angle θ to obtain the missing data x after shearing. * (θ), and at the same time for the mask m * According to the same angle, the offset cut is performed to obtain the mask m after offset cutting. * (θ). After inputting the miscut data into the closed-loop network, the closed-loop reconstruction error L corresponding to the angle θ is obtained. cyc (θ). Select the θ value that minimizes the reconstruction error from a series of candidate θ values:

[0057]

[0058] Using θ * The corresponding interpolation result is taken as the optimal interpolation result, and after the inverse shear transformation, the final interpolation result is obtained. * The corresponding shear transformation can transform the test data from difficult samples to simple samples and obtain better interpolation results.

[0059] In order to verify the effectiveness and superiority of the present invention, the method proposed in the present invention is applied to actual seismic data and actual ultrasonic data. Figure 4-7 , Table 1 shows the interpolation effect of the present invention.

[0060] Figure 4 , 5 Experimental results on real seismic data are shown. Figure 4 (a) is the downsampled earthquake data, i.e., missing earthquake data. Figure 4 (e) is the target earthquake data without missing data, Figure 4(b)-(d) are the interpolation results of Bicubic interpolation, traditional CNN algorithm [1] and the method of the present invention, respectively. Figure 4 (f)-(h) are the interpolation errors corresponding to the three methods. Figure 4 The results show that the interpolation result of the method of the present invention is closer to the real non-missing data, and the interpolation error is lower. Figure 4 As shown by the black arrow, the method of the present invention can effectively reduce the aliasing phenomenon. Table 1 shows the quantitative results of the interpolation results of the three methods.

[0061] Table 1 Comparison of interpolation SNR of actual seismic data

[0062]

[0063] As can be seen from Table 1, the signal-to-noise ratio (SNR) of the Bicubic interpolation result is 12.46 dB, the signal-to-noise ratio of the traditional CNN interpolation result is 18.94 dB, and the signal-to-noise ratio of the interpolation result of the method of the present invention is 22.31 dB. It can be seen that the interpolation result of the method of the present invention has a significant improvement in quantitative indicators. Figure 5 The amplitude spectrum results are shown. Figure 5 (a) is the target data amplitude spectrum, Figure 5 (b) is the amplitude spectrum of the downsampled data, Figure 5 (c) is the amplitude spectrum of the Bicubic interpolation result. Figure 5 (d) is the amplitude spectrum of the traditional CNN interpolation result. Figure 5 (e) is the amplitude spectrum of the interpolation result of the method of the present invention. As shown by the white arrow in the figure, the interpolation results of Bicubic interpolation and traditional CNN cannot effectively suppress aliasing, while the anti-aliasing effect of the method of the present invention is better, and the amplitude spectrum of the interpolation result of the method of the present invention is closer to the target amplitude spectrum.

[0064] Figure 6 , 7 Experimental results on real ultrasound data are shown. Figure 6 (a) is the downsampled ultrasound data, Figure 6 (b)-6(d) are the Bicubic interpolation results, the traditional CNN interpolation results and the interpolation results of the method of the present invention. Figure 6 It can be seen from the black circle part that, compared with other methods, the method of the present invention has a better anti-aliasing effect. Figure 7 The corresponding amplitude spectrum is shown, Figure 7 (a) is the amplitude spectrum of downsampled ultrasound data, Figure 7(b)-7(d) are the amplitude spectra of the Bicubic interpolation result, the traditional CNN interpolation result and the interpolation result of the method of the present invention. The white arrows in the figure indicate the false frequency part, from which it can be seen that the method of the present invention can effectively suppress the false frequency and improve the interpolation effect.

[0065] References

[0066] [1]Wang, B., Zhang, N., Lu, W., and Wang, J.

[2018] Deep-learning-based seismic data interpolation: A preliminary result. Geophysics, 84(1), v11-v20.

[0067] [2]Hu, L., Zheng, X., Duan, Y., and Yan,

[0068] [3]Xiao,D.,Yiu,BY,Chee,AJand Alfred,CH

[2019] Channel CountReduction for Plane Wave Ultrasound Through Convolutional Neural NetworkInterpolation.In International Conference on Image Analysis and Recognition(pp.442-451).Springer,Cham.

[0069] [4]Vyas,M.and Liao,Q.

[2020] De-aliasing using the U-Net imagesegmentation algorithm.In SEG Technical Program Expanded Abstracts 2020(pp.1476-1480).Society of Exploration Geophysicists.

[0070] [5]He, K., Zhang, X., Ren, S. and Sun, J.

[2016] Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp.770-778).

[0071] [6] Ronneberger, O., Fischer, P., and Brox, T.

[2015] U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pp. 234-241.

[0072] [7] Kingma, DP, and Ba, J.

[2014] Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.

[0073] The above embodiments are only preferred specific implementations of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. An anti-aliasing acoustic wave image interpolation method based on a self-adjusting closed-loop network, characterized in that: The following steps are involved: Step 1: Build a closed-loop interpolation model; Step 2: Use the interpolation loss function L interp And the closed-loop loss function L cyc The closed-loop interpolation model is trained; the overall function of the training process is: L=λ1L interp +λ2L cyc λ1,λ2>0, used to adjust the weights of the two loss functions; Interpolation loss function L interp : L interp =‖z-f W (x)‖2 f W (·) represents the interpolation model, W represents the weight parameter of the model, z represents the original non-missing seismic data, x represents the missing data, x=z⊙m, M represents the non-missing trace mask, ⊙ represents the dot product; Closed-loop loss function L cyc : f W′ (·) is another interpolation model, is the new missing data formed by the interpolation results of the missing tracks, Step 3: Build a self-adjustment module; specifically, define a closed-loop reconstruction error of the test data based on the closed-loop interpolation model to evaluate the application effect of the interpolation model on the new data set. A small reconstruction error indicates that the application effect of the interpolation model on the new data set is good, otherwise, the application effect is poor; self-adjust the test data according to the closed-loop reconstruction error, perform staggered transformation on the test data at different angles, input the closed-loop interpolation model respectively, select the optimal staggered angle according to the closed-loop reconstruction error, perform inverse staggered transformation on the interpolation result under the optimal staggered angle, and obtain the final interpolation result.

2. The method for anti-aliasing acoustic wave image interpolation based on a self-adjusting closed-loop network as claimed in claim 1, characterized in that: The closed-loop interpolation model in step 1 includes two interpolation sub-models, one interpolation sub-model is used to interpolate missing data, and the other interpolation sub-model is used to re-interpolate the interpolation results of the missing channels. The two sub-models have the same structure and share weights.

3. The method for anti-aliasing acoustic wave image interpolation based on a self-adjusting closed-loop network as claimed in claim 2, characterized in that: The sub-model is a CNN model or a traditional interpolation model.

4. The method for anti-aliasing acoustic wave image interpolation based on a self-adjusting closed-loop network as claimed in claim 1, characterized in that: The closed-loop reconstruction error is: x * is the unlabeled test data, m * For x * The corresponding mask, is the closed loop output, Represents unlabeled test data x * The corresponding missing track interpolation results satisfy

Images