Underwater acoustic target recognition method based on effective receptive field regulation

Abstract

Disclosed in the present invention is an underwater acoustic target recognition method based on effective receptive field regulation. A proposed AEU-Net model has a plurality of resolution branches, with each resolution branch having independent convolution kernels, which adaptively adjust their sizes during training while aligning with their respective resolutions. The AEU-Net model comprises an ERF-Server, which can regulate an effective receptive field during forward propagation, and has two operations of increasing or decreasing the effective receptive field of a feature map of any designated module. In the present invention, acoustic physical information of an underwater acoustic target sample can be captured in a plurality of interaction dimensions, and feature fusion of information from effective receptive fields across a plurality of scales is performed, so as to adapt to sonar image targets of different sizes, thereby improving the accuracy and recognition speed in the recognition of an underwater acoustic target.

Description

An underwater acoustic target recognition method based on effective receptive field regulation TECHNICAL FIELD

[0001] The present application relates to the field of marine underwater target recognition, and particularly relates to an underwater acoustic target recognition method based on effective receptive field regulation. BACKGROUND

[0002] Underwater acoustic target recognition essentially belongs to the field of image processing, and a semantic segmentation model is usually used to predict each pixel point of an image based on feature information extracted from the image. The semantic segmentation task on the current mainstream underwater acoustic dataset is mainly affected by problems such as complex marine environment (such as object occlusion, target posture change, multiple targets, uneven illumination), strong noise interference, etc. Moreover, the dataset samples of underwater acoustic target recognition contain a large amount of acoustic physical information, and are very sensitive to the global perception ability of the model in space. The current recognition method based on deep learning has shown great advantages in performance indicators and real-time performance compared with the time-frequency based method, and has become the mainstream method in the field.

[0003] The concept of effective receptive field is proposed in relation to the theoretical receptive field. The effective receptive field only accounts for a small part of the theoretical receptive field, and with the increase of the number of layers, the proportion of the effective receptive field of the model in the theoretical receptive field decreases, and the shrinkage speed of the ratio of the two will also become smaller. Therefore, only by stacking the number of layers, the improvement of the model ERF is not great when the number of layers is large. At present, there is a lack of research on deep learning models based on effective receptive field vision models in the related field.

[0004] Most of the existing convolutional visual models are based on the small kernel residual block proposed by ResNet, which has a size of 3*3, which means that even if the number of layers is very deep, the ERF cannot be obtained. In the underwater acoustic target recognition task, due to the uniform distribution of frequency domain information in the whole sample, such models used on general data sets cannot extract sufficient and global information. Even if the network using small kernel such as high-resolution network (High-Resolution Network, hereinafter referred to as HRNet) obtains better spatial sensitivity through multi-path interaction and global maintenance of high-resolution representation, but due to the lack of ERF resources, it cannot perform well on underwater acoustic target data sets. At the same time, in the existing model researches aiming at multi-resolution representation, different size convolution kernel, or directly at the effective receptive field, the main research ideas are based on the respective experiments of these high and low, large and small representations (feature semantic information) or convolution kernels (feature semantic extraction means) (for example, HRNet uses 3*3 residual blocks, and RepLKNet does not consider multi-resolution fusion), and often good results are obtained. However, the exploration of both multi-scale and matching and adaptive relationship is relatively rare. On the other hand, in visual models such as vision transformer, swin transformer (hereinafter referred to as ViTs), the effectiveness of the effective receptive field has been widely proven, and they have performed strongly in many computer vision downstream tasks, which is believed to be due to the large number of parameters in the model design, which brings part of the effective receptive field improvement.

[0005] SUMMARY

[0006] In view of the deficiencies of the prior art, the present application provides an underwater acoustic target recognition method based on effective receptive field regulation, which directly designs from the kernel size positively correlated with ERF, and extracts multi-scale ERF information and performs fusion and interaction.

[0007] The specific technical solutions are as follows:

[0008] An underwater acoustic target recognition method based on effective receptive field regulation, comprising the following steps:

[0009] S1: Assemble a pixel-by-pixel underwater sonar segmentation image data set with labels;

[0010] S2: Construct and train an AEU-Net model based on U-Net, which is realized by the following sub-steps:

[0011] S2.1: embed the information interaction module of the high-resolution network into each single resolution module of the U-Net, expand each module into two resolution branches of first and second, after the original image is pre-processed by a convolutional layer, down-sampling is performed, in the down-sampling stage, the first branch output of the previous module is input into the first branch of the corresponding module in the up-sampling stage through the skip connection, the judgment basis of the corresponding module is that the resolutions of the feature maps are the same, the second branch output of the previous module is divided into two paths, one path is input into the first branch of the next module after passing through the residual block convolution with a step of 1, the other path is input into the second branch of the next module after passing through the residual block down-sampling with a step of 2; in the up-sampling process, the first branch output of the previous module is input into the first branch of the next module after up-sampling, the first branch output of the previous module is input into the second branch of the next module after passing through the residual block convolution with a step of 1; the second branch output of each module in the up-sampling stage is spliced in the channel dimension, and then output classification labels in turn through one-dimensional convolution and fully connected layer;

[0012] S2.2: reconstruct the convolution kernel of each branch using channel-wise convolution, the channel-wise convolution has an adaptive size convolution kernel, the size of which matches the resolution of each branch feature map;

[0013] S2.3: construct an ERF-Server network branch for realizing the skip connection, and complete the construction of the AEU-Net model; the ERF-Server network branch includes U-Serve and D-Serve two operations, for the feature map representation output by each down-sampling module, new semantic information with ERF enhancement is obtained after passing through the large kernel convolution block of U-Serve; for the feature map representation output by each down-sampling module, new semantic information with ERF reduction is obtained after passing through the residual convolution block with a fixed size of 3*3 of D-Serve; the up-sampling output of the first branch of the previous module of the corresponding module in the up-sampling stage, the output of U-Serve and the output of D-Serve are spliced into the corresponding module in the up-sampling stage according to the channel distribution coefficient;

[0014] S2.4: input the underwater sonar segmentation image dataset into the AEU-Net for training, and obtain the trained AEU-Net model; during training, the kernel size of the adaptive size convolution kernel is adjusted within the preset kernel size interval of the feature map of the current module, and the training target is to minimize the loss function of the adaptive size convolution kernel;

[0015] S3: input the underwater sonar segmentation image to be tested into the trained AEU-Net model, and obtain the recognition result.

[0016] Further, in S2.2, one convolution kernel of the channel-by-channel convolution is responsible for one channel, one channel is only convolved by one convolution kernel, and the number of feature map channels after convolution is the same as the number of input channels; the adaptive size convolution kernel controls the size of the kernel size through the respective loss function.

[0017] Further, the update logic of the adaptive size convolution kernel loss function during training is that the adaptive size convolution kernel loss function guides the adaptive convolution kernel to select and update the kernel size according to the difference between the predicted pixel points and the real label pixel points of different categories of samples in each training.

[0018] Based on the update logic of the adaptive size convolution kernel loss function during training, the recognition target of the underwater sonar segmentation image dataset is artificially divided into two categories: a first target and a second target. If the predicted pixels of the first target have mislocation or insufficient filling compared with the real label, the corresponding loss function becomes larger, at which time the kernel size of the adaptive convolution kernel is increased to reduce the loss function. Conversely, if the predicted pixels of the second target have missed detection or misjudgment compared with the real label, the corresponding loss function becomes larger, at which time the kernel size of the adaptive convolution kernel is reduced to reduce the loss function. The expression of the loss function is as follows:

[0019] In the formula, k p is the value of the size of the current adaptive convolution kernel and the prediction deviation of the first target and the second target after normalization in dimension.

[0020] Further, the small target prediction deviation is represented by Hausdorff distance, and the specific formula is as follows: Hausdorff(A, B) = max{h(A, B), h(B, A)} h(A, B) = max a∈A min b∈B ||a-b||

[0021] h(B, A) = max b∈A min a∈B ||b-a||

[0022] The large target prediction deviation is represented by the Dice coefficient, and the specific formula is as follows:

[0023] In the formula, A = {a 1 , a 2 , …, a p}, B = {b 1 , b 2 , …, b p}, set A is a predicted pixel point set, set B is a real label pixel point set, and ||a-b|| represents the Euclidean distance between two pixel points.

[0024] Further, the kernel size of the adaptive convolution kernel is adjusted in a kernel size interval preset for the feature map of the current module, and the kernel size intervals corresponding to feature maps of different resolutions do not overlap with each other; the kernel size interval is determined by a size function, and the expression is as follows: k m =f(r)

[0025] In the formula, k m is the median value of the kernel size interval of the adaptive convolution kernel, f(r) is the size function, and r is the size represented by the feature map of the current module.

[0026] Further, the bilinear interpolation is selected for the up-sampling.

[0027] Further, in S2.3, the channel allocation is implemented by a channel allocation coefficient C k , which is a preset hyperparameter and is a 1*3 vector, each component of which is between 0 and 1 and the sum is 1, representing the proportions of the three groups of feature maps of the D-Serve output, the U-Serve output and the up-sampling output of the corresponding module of the previous module of the first branch of the up-sampling stage, and the channel number is rounded when allocated.

[0028] Further, in S2.4, the ERF distribution of the feature map of each module of the entire AEU-Net model is obtained, wherein the real ERF is calculated by the size of the adaptive size convolution kernel of each branch and the number of convolution layers it has experienced, and the expression is as follows:

[0029] That is

[0030] In the formula, is the standard deviation of the pixel value S n output by the nth convolution layer; (i, j) is the index of the i-th row and j-th column element in the k*k adaptive size convolution kernel; k is the size of the adaptive size convolution kernel for convolution of the feature map of each module of the underwater acoustic target signal, and n is the number of stacked layers accumulated from the input of the original image to the current convolution layer.

[0031] An electronic device comprising a memory and one or more processors, the memory having stored therein executable code which, when executed by the one or more processors, is configured to implement the underwater acoustic target recognition method based on effective receptive field regulation.

[0032] A computer-readable storage medium having stored thereon a program which, when executed by a processor, implements the underwater acoustic target recognition method based on effective receptive field regulation.

[0033] The beneficial effects of the present application are:

[0034] (1) The AEU-Net model constructed in the present application is based on the general model U-Net and HRNet in the field of computer vision, and through an independently designed network branch called ERF-Server, it is specialized for underwater acoustic data sets, can change the effective receptive field properties originally fixed with the network structure into controllable and trainable, and further can use this feature to make each module of the model better extract the physical features in the underwater acoustic data set, fuse the semantic information of each scale ERF from large to small, and realize the recognition accuracy and adaptability of sonar targets of different sizes.

[0035] (2) The structure of matching the size of the convolution kernel and the resolution of the feature map proposed in the present application can extract more global information and target shape information in high-resolution representation with large kernels, and extract more detailed local texture information in low-resolution representation with small kernels, instead of using a large group of convolution kernels or a small 3x3 residual fast throughout the previous network. Compared with the commonly used large kernel size group applied to the standard data set, it can significantly reduce the parameter amount and improve the training speed under the premise of improving the performance effect, so that the model can be applied to underwater and other scenes with high real-time requirements.

[0036] (3) The AEU-Net model proposed in the present application has a reconstructed adaptive multi-scale information interaction module, which uses matching convolution kernels to efficiently extract different information at each resolution module, and supplements the semantic information of the feature maps obtained by the two semantic branches U-Serve and D-Serve adjusted by the ERF-Server in the previous downsampling module, so that the current size resolution matching semantic information, global large effective receptive field information, and global small receptive field information are fused, and compared with the design of global large kernel, it can exchange information fusion from semantic resources and effective receptive field resources with smaller parameter amount, and further improve the feature extraction capability of underwater acoustic target images with high global perception requirements on physical information covering the entire sample image. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 is a flowchart of the underwater acoustic target recognition method based on effective receptive field regulation in the embodiment of the present application.

[0038] Figure 2 is a schematic diagram of the backbone network of the AEU-Net constructed in the embodiment of the present application.

[0039] Figure 3 is a matching schematic diagram of the variable kernel size median and the feature map representation size in the embodiment of the present application.

[0040] Fig. 4 is a comparison example of the effect of the ERF-Server of the present application and the prior art.

[0041] Fig. 5 is a schematic diagram of the fusion details of three-dimensional information of each resolution module in the up-sampling process in the embodiment of the present application.

[0042] Fig. 6 is a schematic diagram of an electronic device for implementing the underwater acoustic target recognition method based on effective receptive field regulation in the embodiment of the present application. DETAILED DESCRIPTION

[0043] The present application will be described in detail below according to the drawings and preferred embodiments, and the purposes and effects of the present application will become more apparent. The present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0044] An underwater acoustic target recognition method based on effective receptive field regulation comprises the following steps:

[0045] S1: Assemble an underwater sonar segmentation image dataset, wherein each pixel of each underwater sonar segmentation image in the dataset has a label.

[0046] At present, the number of datasets in the field of underwater acoustic target recognition is relatively small, and the underwater acoustic target recognition classification of underwater acoustic sonar image is a challenging problem. In the embodiment, the underwater sonar segmentation image is collected by a Tritech Tritech 1200i side scan sonar. The collected underwater sonar segmentation image is a two-dimensional matrix form of the original echo intensity information of the side scan sonar, and the image is stored in the format of bmp with a resolution of 12000x8200 pixels.

[0047] S2: Construct and train an AEU-Net (Adjustable Effective Receptive Field U-shape Network) model based on a U-shaped network (U-shape Network, hereinafter referred to as U-Net).

[0048] As shown in FIG. 1, the AEU-Net includes: a down-sampling stage, an up-sampling stage, an ERF-Server network branch; an information interaction module of the HRNet is used to expand each single resolution module of the U-Net, so that the original module has two resolution branches, wherein the first branch is the original resolution, and the second branch is 1 / 2 of the original resolution. After the original image is input into a pre-processing convolutional layer, the original image is input into the down-sampling stage, the output of each module of the down-sampling stage is connected to the corresponding up-sampling module through the ERF-Server skip connection while being transmitted to the next module; the ERF-Server specifically refers to the joint action of the U-Serve and the D-Serve to obtain the feature representation changed by the ERF; an adaptive size convolution kernel is used for compensation in the down-sampling process of each module of the down-sampling stage. The outputs of each module of the up-sampling stage are spliced in the channel dimension, so that each up-sampling stage has three groups of feature map sources (i.e., inputs), which are: the output of the up-sampling of the first branch of the previous module, the output of the U-Serve in the ERF-Server, and the output of the D-Serve in the ERF-Server. The three are spliced and fused in the channel according to the channel allocation coefficient C k The channel allocation coefficient gives each of the three groups of feature maps a proportion and the sum is 1. Finally, the outputs of each up-sampling module are spliced to obtain fused features without losing any channels, and the fused features contain multi-scale information; after one-dimensional convolution, a full connection layer is performed to output a classification label, and the recognition is completed.

[0049] S2 is specifically implemented through the following sub-steps:

[0050] S2.1: embed the information interaction module of the HRNet into the single resolution module of the U-Net (i.e., embed into each module) to obtain a U-shaped high-resolution feature extraction network with enhanced global information perception ability.

[0051] Specifically, the up-sampling stage and the down-sampling stage are taken as the backbone network, the up-sampling stage, the down-sampling stage, and the semantic information recovery of the skip connection are all derived from the structure of the U-Net itself, and the U-Net is used to extract the features of the image. Before entering the backbone network, the original image is first input into a pre-processing convolutional layer to down-sample the resolution to 1 / 4 of the original image size as the initial input of the backbone network. Each module of the down-sampling stage reduces the resolution by half through a convolution operation with a step of 2, and in this process, the semantic information becomes more abstract with the convolution, which can increase the robustness of some small disturbances of the input image, such as image translation and rotation, and can also reduce the risk of overfitting, reduce the amount of calculation, and increase the size of the receptive field. The skip connection takes the output of a certain module of the down-sampling stage as the input of the corresponding module (i.e., the module with the same resolution) of the up-sampling stage.

[0052] After the information interaction module of the HRNet is embedded in the single resolution module of the U-Net, each module is composed of two adjacent resolution branches (the resolution ratio differs by 1 / 2), referred to as a first branch and a second branch. The forward propagation module in each branch is reconstructed by a large-core convolution block, and the semantic information between the adjacent resolution branches is reconstructed in the information interaction manner of the HRNet, as shown in FIG. 2. The resolution reduction manner is 3*3 residual block down-sampling, and the resolution promotion manner is bilinear interpolation.

[0053] As shown in FIG. 2, in this embodiment, module 1 is taken as the first single resolution module, and modules 2 to 7 are sequentially referred to in order. The down-sampling stage includes modules 1 to 4, and the up-sampling stage includes modules 4 to 7. The feature map with a resolution of 1 / 4 after the pre-processing convolution layer is divided into two paths, one of which is input into the first branch of module 1 without any operation, and the other of which is input into the second branch of module 1 after being down-sampled to 1 / 8 (i.e., down-sampling to half of the original). The two branches are parallel. The two branches of each module are cross-scale interacted, and then output after several large-core convolutions.

[0054] In the down-sampling process, the output of the first branch of the previous module is input into the first branch of the corresponding resolution module in the up-sampling stage through skip connection. The output of the second branch of the previous module is divided into two paths, one of which is input into the first branch of the next module after being convolved by a residual block with a step of 1, and the other of which is input into the second branch of the next module after being down-sampled by a residual block with a step of 2.

[0055] In the up-sampling process, the output of the first branch of the previous module is input into the first branch of the next module after being up-sampled, and the output of the first branch of the previous module is input into the second branch of the next module after being convolved by a residual block with a step of 1. In this embodiment, bilinear interpolation is used for up-sampling. In this embodiment, the entire backbone network has five outputs, i.e., the second branch outputs of modules 4, 5, 6, and 7 and the first branch output of module 7. The second branch outputs of each module in the up-sampling stage are up-sampled to the highest resolution and then spliced, and finally one-dimensional convolution is performed, and then the classification result is output after the full connection layer.

[0056] S2.2: For the U-shaped high-resolution feature extraction network, the convolution kernel of each branch of the backbone network except the ERF-Server is further reconstructed by using depth-wise convolution (hereinafter referred to as DW conv). The depth-wise convolution has an adaptive size of the convolution kernel, and the size is in a matching relationship in positive correlation with the resolution of the feature map of each branch, so as to obtain new adaptive kernel size information interaction modules of each module.

[0057] As shown in FIG. 3, unlike the conventional convolution operation, one convolution kernel of the DW convolution is responsible for one channel, one channel is only convolved by one convolution kernel, and the number of feature map channels after the convolution is completed is the same as the number of input channels. The adaptive size convolution kernel controls the size of the kernel size through a respective loss function, which is independent of the loss function of the training module for the convolution parameters, and each adaptive size convolution kernel of the resolution feature map adjusts the kernel size thereof during training.

[0058] The specific update logic of the adaptive size convolution kernel loss function during training is that the adaptive size convolution kernel loss function guides the adaptive convolution kernel to select and update the kernel size according to the difference between the predicted pixel points of different categories of samples during each training and the pixel points in the real label.

[0059] Based on the update logic of the adaptive size convolution kernel loss function, based on the update logic of the adaptive size convolution kernel loss function during training, the recognition target of the sonar segmentation image dataset is divided into two categories of large target objects and small target objects according to the prior knowledge artificially added, if the predicted pixels of the large target objects have mispositioning or insufficient filling compared with the real label, the corresponding loss function becomes larger, tends to reduce the loss function by increasing the kernel size, and then iterates out a larger kernel size; on the contrary, if the predicted pixels of the small target objects have missed detection or misjudgment compared with the real label, the corresponding loss function becomes larger, tends to reduce the loss function by reducing the kernel size, and then iterates out a smaller kernel size; the expression of the loss function is as follows:

[0060] In the formula, k p is the size of the adaptive convolution kernel of the current module and the normalized value of the prediction deviation of the first target and the second target in dimension; ~ indicates positive correlation. It should be pointed out that both terms are only meaningful when the formula under the square root is positive, otherwise it is 0, and this operation logic is specifically implemented by using the relu function.

[0061] Wherein, the small target prediction deviation is represented by Hausdorff distance, and the specific formula is as follows: Hausdorff(A,B)=max{h(A,B),h(B,A)} h(A,B)=max a∈A min b∈B ||a-b|| h(B,A)=max p∈A min a∈B ||b-a||

[0062] The large target prediction deviation is represented by the Dice coefficient, and the specific formula is as follows:

[0063] In the formula, A={a 1 , a2 a p}, B = {b 1 b 2 , …, b p}, set A is a set of predicted pixel points, set B is a set of real label pixel points, and ||a-b|| represents the Euclidean distance between two pixel points.

[0064] The kernel size of the adaptive convolution kernel is an integer, which is positively correlated with the resolution of the feature map of the current module, and is adjusted and changed within a preset kernel size interval, and the median interval of the kernel size of the adaptive convolution kernel used by different resolution feature maps does not overlap. In addition, the median of these kernel size intervals is determined by a size function, and in the actual model training process, the change of the size is that after each update, one size in a group is selected for use. The correspondence between the median of the kernel size interval of the adaptive convolution kernel and the size of the feature map is shown in FIG. 4, and the size function expression is as follows: k m = f(r)

[0065] wherein k m is the median of the kernel size interval of the adaptive convolution kernel, f(r) is the size function, and r is the size represented by the current module feature map.

[0066] S2.3: For the reconstructed adaptive kernel size information interaction module, a network branch module ERF-Server is assembled to realize skip connection and complete the construction of the AEU-Net model. The main function of the ERF-Server is to take some ERF information out of the model, process it into the expected features, and then send it back to the model itself. As shown in FIG. 2, in this embodiment, the outputs of the module 1, the module 2 and the module 3 are input into the ERF-Server network branch, the output of the module 1 is input into the module 7, the output of the module 2 is input into the module 6, and the output of the module 3 is input into the module 5.

[0067] As shown in FIG. 3, the ERF-Server receives acoustic semantic information from different modules in the down-sampling stage of the model, and performs two different operations to obtain acoustic semantic information of effective receptive fields of different scales:

[0068] The semantic information of large effective receptive field is obtained through an operation called U-Serve and is spliced into the subsequent module. U-Serve refers to a network branch composed of multiple channel-by-channel convolutions with fixed kernel size and maximum size (the maximum value in the range of convolution kernel size of the highest resolution branch), with residual connection. It can receive the features of any module in the downsampling stage, and after the large kernel convolution of U-Serve itself, the semantic information of the feature map after the ERF of the relatively original 3*3 residual block is greatly improved. After adjusting the channel number with a 1*1 residual block, the result is spliced into the subsequent designated module according to the channel allocation, and the module obtains the new semantic information of the feature map after the ERF of this part of the improved information, and further improves the richness of the ERF information of all channels. When allocating by channel, there is a channel allocation coefficient C k The fusion efficiency of the new semantic information of this module is adjustable.

[0069] The semantic information of small effective receptive field is obtained through an operation called D-Serve and is spliced into the subsequent module. D-Serve refers to a network branch composed of a fixed number of residual blocks. It can receive the features of any module in the downsampling stage, and after a relatively small number of convolution layers compared to the backbone network, it retains the small features of early ERF in the acoustic frequency domain signal. After adjusting the channel number with a 1*1 residual block, the result is spliced into the subsequent designated module according to the channel allocation, and the module obtains the new semantic information of the feature map after the ERF of this part of the reduced information, and further improves the richness of the ERF information of all channels. Since there is a distribution coefficient when allocating by channel, the fusion efficiency of this part of the new semantic information is adjustable. In this embodiment, D-serve processes the received information through 3*3 residual blocks. Compared with the main branch using large kernel groups, D-serve controls the rapid improvement of ERF, enabling the model to capture the current scale of ERF information more deeply.

[0070] It should be noted that U-serve and D-serve at the output end of ERF-Server will be allocated with channels with the upsampled results of the first branch of the previous module of the corresponding module in the backbone network. The original channel is replaced with the new channel after the ERF-Server operation according to the channel allocation coefficient, obtaining the feature representation after the ERF change. Channel allocation is achieved through a channel allocation coefficient C k , which is a preset hyperparameter, is a 1*3 vector, each component is between 0 and 1 and the sum is 1, representing the proportion of the three groups of feature maps of D-Serve output, U-Serve output and the upsampled output of the first branch of the previous module of the corresponding module in the downsampling stage. When allocating, the channel number is rounded. The hyperparameter is a real number, which increases the variability of parameter tuning, Ck The greater the value assigned to D-Serve and U-Serve means that the corresponding Server will play a stronger role, and vice versa 0 means that the Server does not work.

[0071] S2.4: input the underwater sonar segmentation image dataset into the AEU-Net for training to obtain the trained AEU-Net. The kernel size of the adaptive size convolution kernel is adjusted within the preset kernel size interval of the feature map of the current module during training, and the training target is to minimize the loss function of the adaptive size convolution kernel. For the ERF changed feature representation obtained by the ERF-Server, the real ERF is calculated by the size of the adaptive size convolution kernel of each branch and the number of convolution layers it has experienced, and the expression is as follows:

[0072] That is

[0073] In the formula, is the standard deviation of the output pixel value S n of the nth convolution layer, S n is composed of the weighted sum of all input pixel values in its receptive field, and its variance is calculated by considering the contribution of each element in the adaptive size convolution kernel to the output of each pixel; (i, j) is the index of the i-th row and j-th column element in the k x k adaptive size convolution kernel; ERF in the formula refers to the effective receptive field of the feature map of each module, where k is the size of the adaptive size convolution kernel for convolution of the feature map of each module of the underwater acoustic target signal, and n refers to the number of stacked layers accumulated from the input of the original image to the current convolution layer.

[0074] During training, the ERF distribution of each module of the entire AEU-Net model is obtained while adjusting the kernel size, and it is used as the ERF distribution optimized for the information distribution of the sonar segmentation image dataset.

[0075] S3: input the underwater sonar segmentation image to be tested into the trained AEU-Net model to obtain the recognition result.

[0076] Fig. 5 is an effect diagram of the effect of the ERF-Server framework in the application on the change of the effective receptive field attribute of the feature map in the forward propagation by feature map splicing, compared with ResNet and RepLKNet, the first row is the change of the effective receptive field of the existing ResNet front and rear modules, the second row is the change of the effective receptive field of the existing RepKLNet front and rear modules, and the third row is the effective receptive field calculated by selecting all channel feature maps of a module in the AEU-Net model proposed in the application, and then inputting the ERF-Server for effective receptive field regulation test. The darker the color in the figure, the greater the influence on each pixel value from the center, that is, the larger the ERF, so as to reflect the size of the effective receptive field according to the lightness. It is illustrated that the ERF-Server involved in the application is effective in regulating the effective receptive field.

[0077] The underwater acoustic target recognition method based on effective receptive field regulation provided by the application has larger and more rapidly changing ERFs brought by large kernel convolution compared with the existing segmentation network method applied in the field, thereby having more ERF resources, so that the network has stronger global feature capturing capability for defect samples through the interaction of multi-scale ERF information, and more focuses on the overall shape of the defect when extracting information. On the other hand, it has an independent branch framework ERF-Server which can actively regulate the change of the network ERF. The function of this module can quickly improve the original effective receptive field due to the large kernel operation of U-serve, and can reduce the effective receptive field compared with the backbone network due to the size of the adaptive convolution kernel in the backbone network being larger than the traditional 3*3 convolution, so that the effective receptive field information of each scale from small to large in the underwater sonar segmentation image initially input into the AEU-Net can be fully utilized. The application can improve the overall recognition accuracy of the underwater sonar data set while reducing the model parameter amount by nearly half compared with the original U-Net and HRNet to meet the real-time requirement of underwater target recognition.

[0078] As shown in Fig. 6, the electronic device provided by the embodiment of the application includes a memory and one or more processors, the memory stores executable code, and the one or more processors execute the executable code to implement the underwater acoustic target recognition method based on effective receptive field regulation in the above-mentioned embodiments.

[0079] The device of the present application can be applied to any device with data processing capability, which can be a device or apparatus such as a computer. The device of the present application can be implemented by software, or by hardware or a combination of software and hardware. Taking the software implementation as an example, as a logical device, it is formed by reading the corresponding computer program instructions in the non-volatile memory into the memory for running by the processor of the device. From the hardware level, as shown in FIG. 6, it is a hardware structure diagram of the device of the present application in any device with data processing capability. In addition to the processor, memory, network interface and non-volatile memory shown in FIG. 6, the device of the present application in any device with data processing capability in the embodiment can also include other hardware according to the actual function of the device with data processing capability, which will not be described here.

[0080] The implementation process of the functions and roles of each unit in the above device is specifically described in the implementation process of the corresponding steps in the above method, which will not be described here.

[0081] For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can be referred to the part of the method embodiment. The above described device embodiment is only illustrative, and the units illustrated as separate components can be or can not be physically separated, and the components shown as units can be or can not be physical units, that is, they can be located in one place or distributed on multiple network units. According to the actual needs, part or all of the modules can be selected to achieve the purpose of the present application. Those skilled in the art can understand and implement it without creative labor.

[0082] The embodiment of the present application also provides a computer readable storage medium, which stores a program, and the program is executed by a processor to realize the underwater acoustic target recognition method based on effective receptive field regulation in the above embodiment.

[0083] The computer readable storage medium can be an internal storage unit of any of the aforementioned embodiments of the device with data processing capability, such as a hard disk or a memory. The computer readable storage medium can also be an external storage device of any of the aforementioned embodiments of the device with data processing capability, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. Further, the computer readable storage medium can include both an internal storage unit and an external storage device of any of the aforementioned embodiments of the device with data processing capability. The computer readable storage medium is used to store computer programs and other programs and data required by the device with data processing capability, and can also be used to temporarily store data that has been output or will be output.

[0084] Those skilled in the art can understand that the above description is only preferred examples of the application and is not used to limit the application. Although the application is described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions recorded in the foregoing examples or make equivalent replacements for part of the technical features. Any modification, equivalent replacement, etc. within the spirit and principle of the application shall be included in the protection scope of the application.

Claims

1. An underwater acoustic target recognition method based on effective receptive field regulation, characterized in that, The method comprises the following steps: S1: Assembling an underwater sonar segmentation image dataset with labels pixel by pixel; S2: Building and training an AEU-Net model based on U-Net, which is realized through the following sub-steps: S2.1: Embedding the information interaction module of the high-resolution network into each single resolution module of the U-Net, so that each module is expanded into first and second two resolution branches; after the original image is subjected to a pre-processing convolutional layer and down-sampling, the output of the first branch of the previous module is input into the first branch of the corresponding module in the up-sampling stage through a skip connection, and the judgment basis of the corresponding module is that the resolutions of the feature maps are the same; the output of the second branch of the previous module is divided into two paths, one of which is subjected to residual block convolution with a step of 1 and then input into the first branch of the next module as the first branch input, and the other is subjected to residual block down-sampling with a step of 2 and then input into the second branch of the next module as the second branch input; in the up-sampling process, the output of the first branch of the previous module is input into the first branch of the next module after up-sampling, and the output of the first branch of the previous module is subjected to residual block convolution with a step of 1 and then input into the second branch of the next module as the second branch input; after the output of the second branch of each module in the up-sampling stage is spliced in the channel dimension, it is sequentially subjected to one-dimensional convolution and a fully connected layer, and the classification label is output; S2.2: Reconstructing the convolution kernel of each branch using channel-wise convolution, which has an adaptive size convolution kernel with a size matching the resolution of each branch feature map; S2.3: Building an ERF-Server network branch for realizing the skip connection to complete the construction of the AEU-Net model; the ERF-Server network branch includes U-Serve and D-Serve two operations, for the feature map representation output by each down-sampling module, after the U-Serve large kernel convolution block, new semantic information with ERF promotion is obtained; for the feature map representation output by each down-sampling module, after the D-Serve fixed size 3*3 residual convolution block, new semantic information with ERF reduction is obtained; the up-sampling output of the first branch of the previous module of the corresponding module in the up-sampling stage, the output of the U-Serve and the output of the D-Serve are spliced into the corresponding module in the up-sampling stage according to the channel distribution coefficient; S2.4: Inputting the underwater sonar segmentation image dataset into the AEU-Net for training to obtain the trained AEU-Net model; during the training, the kernel size of the adaptive size convolution kernel is adjusted within the preset kernel size interval of the feature map of the current module, and the training target is to minimize the loss function of the adaptive size convolution kernel; S3: Inputting the underwater sonar segmentation image to be tested into the trained AEU-Net model to obtain the recognition result.

2. The method of claim 1, wherein, In S2.2, one convolution kernel of the channel-wise convolution is responsible for one channel, and one channel is only convolved by one convolution kernel, and the number of feature map channels after convolution is the same as the number of input channels; the adaptive size convolution kernel controls the size of the kernel size through the respective loss function.

3. The method of claim 2, wherein, The update logic of the loss function of the adaptive size convolution kernel during training is as follows: the loss function of the adaptive size convolution kernel guides the adaptive convolution kernel to select and update the kernel size according to the difference between the predicted pixel points of different categories of samples and the pixel points in the real label each time during training. Based on the update logic of the loss function of the adaptive size convolution kernel during training, the recognition target of the underwater sonar segmentation image dataset is artificially divided into two categories: a first target object and a second target object. If the predicted pixels of the first target object are mispositioned or insufficiently filled compared with the true labels, the corresponding loss function becomes larger. At this time, the kernel size of the adaptive convolution kernel is increased to reduce the loss function. Conversely, if the predicted pixels of the second target object are missed or misjudged compared with the true labels, the corresponding loss function becomes larger. At this time, the kernel size of the adaptive convolution kernel is reduced to reduce the loss function. The expression of the loss function is as follows: In the formula, k p is the size of the current module adaptive convolution kernel and the value of the first target, the second target prediction deviation after normalization in dimension.

4. The method of claim 3, wherein, The small target prediction bias is expressed by Hausdorff distance, and the specific formula is as follows: Hausdorff(A,B) = max{h(A,B),h(B,A)} h(A,B) = max a∈A min b∈B ||a-b|| h(B,A) = max b∈A min a∈B ||b-a|| The large target prediction bias is expressed by a Dice coefficient, and the specific formula is as follows: In the formula, A={a 1 , a 2 , …, a p}, B={b 1 , b 2 , …, b p}, set A is a predicted pixel point set, set B is a real label pixel point set, and ||a-b|| represents the Euclidean distance between two pixel points.

5. The method of claim 1, wherein, The kernel size of the adaptive convolution kernel is adjusted in a kernel size interval preset for the feature map of the current module, and the kernel size intervals corresponding to feature maps of different resolutions do not overlap with each other; the kernel size interval is determined by a size function, and the expression is as follows: k m =f(r) In the formula, k m is the median of the kernel size interval of the adaptive convolution kernel, f(r) is the size function, and r is the size of the current module feature map representation.

6. The method of claim 1, wherein, Bilinear interpolation is selected for upsampling.

7. The method of claim 1, wherein, In S2.3, the channel allocation is implemented through a channel allocation coefficient C k is a preset hyperparameter, is a 1*3 vector, each component is between 0 and 1 and the sum is 1, represents the proportion of the three groups of feature maps of the D-Serve output, the U-Serve output and the up-sampling output of the first branch of the previous module of the corresponding module of the up-sampling stage, and the channel number is rounded when allocated.

8. The method of claim 1, wherein, In the S2.4, the ERF distribution of the feature map of each module of the entire AEU-Net model is obtained, wherein the real ERF is calculated by the size of the adaptive size convolution kernel of each branch and the number of convolution layers it undergoes, and the expression is as follows: That is In the formulae, S is the output pixel value of the n-th convolutional layer n the standard deviation; (i, j) is the index of the element in the i-th row and j-th column of the adaptive size convolution kernel of k x k; k is the size of the adaptive size convolution kernel for convolution of the feature map of each module of the underwater acoustic target signal, and n is the number of stacking layers accumulated from input to the convolutional layer of the original image.

9. An electronic device, comprising: The device comprises a memory and one or more processors, the memory stores executable code, and the one or more processors execute the executable code to implement the underwater acoustic target recognition method based on effective receptive field regulation in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The device has a program stored thereon, and the program is executed by a processor to implement the underwater acoustic target recognition method based on effective receptive field regulation in any one of claims 1-8.

Images