Sound field region segmentation method based on improved ResNet-34

Abstract

The invention belongs to the technical field of crossing of marine acoustics and machine vision, and particularly relates to a sound field region segmentation method based on improved ResNet-34, which comprises the following steps of: 1, performing sound propagation loss image data set and sound field simulation setting; 2, designing a segmentation network and improving the structure; constructing a segmentation backbone network based on ResNet-34, extracting multi-layer semantic and fine-grained texture features by adopting residual connection and convolution stacking, and outputting a multi-level feature map required by a segmentation head as input of a subsequent improvement module; a deformable convolution self-adaptive sampling sound linear density mutation region is introduced into the Stage3, multi-scale feature fusion is carried out across to improve the sensing capability of convergence / shadow region structures with different sizes, and a global statistical feature module is added behind the Stage4 to carry out remarking and enhancement on multi-scale statistics, so that local boundaries and global distribution are considered; and step 3, evaluating standard definition based on a result of gray calculation. According to the invention, the accuracy and generalization ability of the segmentation result in different sea areas, different seasons and various sound channel structures are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of marine acoustics and machine vision, and particularly relates to a sound field region segmentation method based on an improved ResNet-34. BACKGROUND

[0002] The shadow zone and convergence zone in the marine sound field are typical energy distribution phenomena formed by the joint action of the seawater environment (temperature, salinity, depth / static pressure). The sound speed of seawater is jointly determined by temperature, salinity and depth, among which temperature has the most significant influence. Empirical formulas (such as the Mackenzie model) are commonly used in engineering to calculate the sound speed. The spatiotemporal changes of these environmental factors will directly change the sound propagation path and energy dispersion pattern, thereby affecting the spatial distribution and boundary morphology of the shadow zone and convergence zone. In engineering applications (ships, unmanned underwater vehicles, offshore engineering, etc.), high-precision, pixel-level region segmentation of the shadow zone and convergence zone is a key pre-step in acoustic environment recognition and subsequent tasks (such as long-range propagation evaluation, communication and detection), and its accuracy directly affects the subsequent feature classification, strategy formulation and system performance evaluation.

[0003] Traditional discrimination relies on numerical acoustic models and statistical thresholds: the typical approach is to calculate the propagation loss matrix and sound line coverage density based on ray tracing, and to combine thresholds and clustering to determine the high-energy "convergence" and high-loss "shadow" regions; or to use empirical indicators such as depth allowance to give a probabilistic judgment of the presence of the convergence zone. This approach has clear physical meaning, but is highly sensitive to sound speed profile accuracy and is better at coarse-grained diagnosis of "whether it appears" and is difficult to quantify the spatial position and geometric morphology of complex, non-stationary boundaries in real sea areas. In reproducible experiments, the overall accuracy of the traditional method based on sound line density threshold and clustering for large-scale sea area convergence zone recognition is generally low (the comparison results are generally around 40%, and the highest is about 65.1%), which is difficult to meet the current demand for fine characterization of marine acoustic environment.

[0004] To bypass complex environment inversion and improve robustness, the academia and engineering practice gradually attempt to directly segment the region based on the sound propagation loss image (first generate a high-resolution image from the ray model, then do image layer enhancement, edge detection / region growing, gray difference analysis), but this method has two types of bottlenecks: first, there is a lack of physical labels that can be used for supervision, and indirect gray consistency / contrast and other unsupervised evaluation indicators are often relied on to side with the segmentation rationality; second, engineering images often present weak texture and multi-scale fragmented structure, leading to misclassification and missed classification at the boundary for segmentation algorithms based on threshold or region growing, and it is difficult to establish global statistical features.

[0005] In recent years, convolutional neural networks (CNN) have been introduced to the region segmentation of sound propagation loss images. However, direct application of general networks also has obvious problems: 1. In the original ResNet-34 network model, the convolutional block is biased towards local texture extraction, the last layer feature map size is small, and the global statistics and cross-scale context are easy to lose, which leads to inaccurate delineation of convergence / acoustic shadow boundary and area ratio quantification. Compared with traditional methods, there is still room for improvement in the segmentation accuracy of convergence and acoustic shadow (the reported segmentation accuracy example is 74.2% / 87.5%).

[0006] 2. There are problems in cross-temporal and cross-regional generalization. The data is mainly from reanalysis / simulation combination, and there is still statistical deviation from the real marine environment. In real sea conditions, internal waves, eddies, fronts, and bottom reflection factors are superimposed, making the boundary non-stationary, multi-scale, and fuzzy, further increasing the difficulty of segmentation algorithm migration in different sea areas / seasons and different sound channel structures.

[0007] In summary, the existing technology has common shortcomings in four aspects: physical annotation scarcity, fuzzy and multi-scale boundary, difficulty in modeling global statistics, and insufficient cross-domain adaptation, which directly affects the pixel-level accurate segmentation of acoustic shadow area and convergence area, and limits the reliable classification and application of subsequent sound field features. Therefore, the present application provides a sound field region segmentation method based on improved ResNet-34; which can simultaneously capture global statistics and local details, and is friendly to the segmentation of marine acoustic data domain characteristics, taking into account the consistency of complex boundaries and the quantization accuracy of convergence / acoustic shadow area ratio, thereby meeting the fine needs of engineering scenarios. SUMMARY

[0008] The purpose of the present application is to provide a sound field region segmentation method based on improved ResNet-34, which aims to overcome the common shortcomings of existing acoustic shadow area and convergence area segmentation methods in terms of physical annotation scarcity, fuzzy and multi-scale boundary, difficulty in modeling global statistical features, and insufficient cross-domain adaptation. A marine sound field feature region segmentation method based on improved ResNet-34 is proposed. This method introduces deformable convolution (Deformable Convolution) into the residual structure to enhance the network's adaptive modeling ability for irregular sound line patterns and complex geometric boundaries, and combines with a global statistical feature module to depict the overall distribution trend of the sound field while maintaining local detail resolution, thereby achieving high-precision pixel-level segmentation of acoustic shadow area and convergence area.

[0009] The present application aims to overcome the misclassification and missed classification problems of traditional methods in weak texture, boundary transition, and multi-convergence area / multi-acoustic shadow area scenarios, improve the accuracy and generalization ability of the segmentation results in different sea areas, different seasons, and various sound channel structures, and provide reliable prior data support for subsequent sound field feature classification, marine acoustic environment modeling, and engineering applications (such as long-distance communication, target detection, anti-submarine, etc.).

[0010] The technical scheme adopted by the present application is specifically as follows: The sound field region segmentation method based on the improved ResNet-34 specifically includes the following steps: Step 1: Sound propagation loss image dataset and sound field simulation setting. Based on the multi-temporal and spatial three-dimensional temperature and salinity field data sample mainly in the northwest Pacific, the coherent propagation (CG) mode in the Bellhop model is used for sound ray tracing and propagation loss calculation to generate the sound propagation loss image in the range-depth domain; ideal Munk profile is introduced to generate a control sample, the same image resolution and dynamic range, complete training / validation / test set division; normalization and image enhancement are performed on all samples to improve the generalization ability of the method.

[0011] Step 2: Segmentation network design and structure improvement. Based on ResNet-34, a segmentation backbone network is constructed, residual connection and 3x3 convolution stacking are used to extract multi-layer semantic and fine-grained texture features, global average pooling is used to replace the fully connected layer to keep the consistency of spatial representation, and the multi-level feature map required by the segmentation head is output as the input of the subsequent improvement module. In Stage 3, deformable convolution is introduced to adaptively sample the sound ray density mutation area, and cross Multi-scale feature fusion is performed to improve the perception ability of converging / shadow area structures of different sizes, and a global statistical feature module is added after Stage 4 to re-label and enhance the multi-scale statistics, so as to consider local boundaries and global distribution. The joint loss function uses weighted cross-entropy and Dice to alleviate class imbalance and strengthen the fitting ability of boundaries and small target regions Step 3: Definition of result evaluation standard based on gray scale calculation. A kind of unsupervised evaluation index is designed to make a preliminary judgment on the gray scale value of the sound propagation loss image, and then the correctness of the segmentation region is verified according to the gray scale value difference.

[0012] The technical effects achieved by the present application are: In the present application, the segmentation accuracy is significantly improved. Under the best learning rate, the original ResNet-34 network has a segmentation accuracy of 74.2% for shadow area and 87.5% for convergence area, and the improved ResNet-34 network has a segmentation accuracy of 76.9% for shadow area and 90.6% for convergence area after introducing global statistics and joint loss, which significantly improves the area ratio quantization accuracy.

[0013] Compared with the traditional method, the highest accuracy of the traditional method is 63.5%, and the area accuracy of the improved ResNet-34 network is 86.9%, which improves the area segmentation effect by 23.5%, and can accurately divide the sound field feature region in the complex sound propagation loss image.

[0014] The improved ResNet-34 network can be used as a pretreatment module for sound field feature classification and marine acoustic environment modeling, and improves the accuracy and efficiency of the overall acoustic detection and analysis system. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a whole flow chart Figure 2 is a ResNet-34 network structure schematic diagram Figure 3 is a sound field feature region segmentation model based on the improved ResNet-34 Figure 4 is a typical sound propagation loss diagram Figure 5 is a sound propagation loss image of measured data DETAILED DESCRIPTION

[0016] In order to make the purpose and advantages of the present application clearer and more apparent, the present application will be specifically described below in combination with embodiments. It should be understood that the following text is only used to describe one or several specific embodiments of the present application, and does not strictly limit the specific protection scope requested by the present application.

[0017] As shown in Figures 1-5 the sound field region segmentation method based on the improved ResNet-34 specifically includes the following steps: Step 1: sound propagation loss image dataset and sound field simulation setting Based on the multi-temporal and spatial three-dimensional temperature and salinity field data samples mainly in the Northwest Pacific, the coherent propagation CG mode in the Bellhop model is used for sound ray tracing and propagation loss calculation to generate the sound propagation loss image in the range-depth domain; the ideal Munk profile is introduced to generate a control sample, the same image resolution and dynamic range, complete the training / validation / test set division; all samples are normalized and image enhanced to improve the generalization ability of the method; The step 1 includes the following steps: Step 101: In the process of depicting the characteristics of deep-sea sound field, the sound propagation loss image can intuitively represent the propagation law of the sound field and the loss area of the sound ray. The Bellhop model is designed to perform two-dimensional ray tracing, which is based on ray theory and uses Gaussian beam tracking to handle sound field problems in horizontally inhomogeneous environments. By associating a Gaussian intensity distribution with each sound ray, the sound ray can smoothly transition from the convergence zone to the shadow zone, which is highly consistent with the actual propagation of sound energy. The output options of this method mainly include ray coordinates, propagation time, amplitude, eigenrays, sound pressure, or propagation loss, and the final received point sound field is the result of the superposition of all eigenrays. When determining the ray coordinates, Bellhop calculates the ray equation to obtain the amplitude and sound pressure. For a general cylindrical symmetric system, the ray equation is represented as: where r(s) and z(s) represent the ray coordinates in the cylindrical symmetric system, s represents the arc length along the ray, and c(r, z) is the sound speed distribution. This equation solves the curvature change of the ray trajectory by numerical integration, thereby reflecting the influence of the sound speed gradient on the propagation path.

[0018] The calculation of the propagation loss is based on the superposition of the sound pressure field. Bellhop calculates the sound pressure p(r, z) by superimposing Gaussian beams and converts it to propagation loss: where p0 is the sound pressure intensity at a typical reference distance. For non-coherent propagation loss, the model superimposes the energy of each sound ray and ignores the phase information, generating an image that is more suitable for complex multi-path scenarios. In addition, the output propagation loss image can generate sound ray trajectory maps and propagation loss cloud maps after point cloud or digitization processing, providing multi-dimensional visualization support for deep-sea sound field characteristic analysis.

[0019] Step 102: Bellhop model environment file setting. The environment file of the Bellhop model defines the physical boundary conditions and medium parameters of sound propagation. Reasonable parameter settings can more accurately represent the sound propagation characteristics of deep-sea sound fields. The structure and specific content of the environment file are shown in Table 1.

[0020]

[0021] The Bellhop environment file adopts a hierarchical format to describe the calculation region of the sound field, where the frequency corresponding to the sound source frequency is set to 50 Hz, which can support long-distance propagation of sound waves, and the sound source frequency is set to 50 Hz, which has strong penetration and propagation distance in deep sea environment, and can effectively capture the periodic characteristics of the far-field convergence zone. The water body medium is set as a single layer; nmedia=1, assuming that the sound speed profile changes continuously in the vertical direction, avoiding the interference of complex layered media on the sound ray trajectory, and facilitating the analysis of the influence of sound speed gradient on the boundary of the shadow zone.

[0022] In the sound speed profile part, the sound speed value at the corresponding position is calculated by using the sound speed formula, and piecewise linear interpolation (`SSPOPT='SVW'`) is performed to describe the depth-dependent sound speed variation, covering a range from the sea surface to 5000 meters deep sea; the minimum value of the sound speed appears in the middle depth; about 1200 meters deep; forming a deep sea sound channel axis, which will dominate the horizontal convergence effect of sound energy. The seabed parameter is set as sandy sediment layer; density 1.8 g / cm³, longitudinal wave velocity 1600 m / s, which has high acoustic impedance characteristics and can enhance the seabed reflection loss, thereby highlighting the relevance of the shadow zone (low signal-to-noise ratio area) and the seabed multiple reflection effect.

[0023] The sound source depth is set to 1000 meters, located above the deep sea sound channel axis, ensuring that the sound energy propagates horizontally along the sound channel. The receiver covers the full water depth of 0~5000 meters and the horizontal distance of 0~100km, and the spatial sampling interval meets the Nyquist criterion of the sound field interference fringes, which can completely capture the spatial distribution law of the sound ray convergence zone and the shadow zone.

[0024] The coherent propagation loss calculation mode (`'CG'`) is adopted, which takes into account the phase interference effect of the sound field and the energy accumulation characteristics of the convergence zone. The range of the exit angle is set to -20.3°~20.3°, covering the effective propagation path from the sound source to the receiver, avoiding misjudgment of the boundary of the shadow zone due to insufficient sound ray beam. The setting of the sound ray tracking step and the calculation domain range (`STEP=50m, ZBOX=5500m, RBOX=100km`) ensures the accurate calculation of the sound ray curvature.

[0025] Step 103: generate an ideal Munk profile sound propagation loss image based on the above parameters, as shown in Figure 4 and Figure 5 , the horizontal axis is the propagation distance (0-100km), the vertical axis is the water depth (0-5000m), and the color scale represents the sound energy loss intensity. The difference in the location of the shadow zone / convergence zone between the ideal Munk profile sound propagation loss image and the sound propagation loss image generated from the actual ocean data is mainly due to the deviation of the sound channel axis depth and the deviation of the measured value and the theoretical absorption value of the seabed sediment layer absorption coefficient.

[0026] Step 104: The high-loss features of the convergence zone and the low-loss features of the acoustic shadow zone are highlighted by image enhancement techniques, and the geometric boundaries of the energy anomaly area are extracted using edge detection algorithms or region growing methods to determine the differences in acoustic energy between the acoustic shadow zone and the convergence zone.

[0027] Step 2: Network design and structure improvement; Based on ResNet-34, the segmentation backbone network is constructed, residual connection and 3x3 convolution stacking are used to extract multi-layer semantic and fine-grained texture features, global average pooling is used to replace the fully connected layer to maintain the consistency of spatial representation, and the multi-level feature maps required by the segmentation head are output as the input of the subsequent improvement module; at Stage3, deformable convolution is introduced to adaptively sample the acoustic line density mutation area, and the cross Multi-scale feature fusion is performed to improve the structure perception ability of convergence / shadow zones of different sizes, and a global statistical feature module is added after Stage4 to re-label and enhance multi-scale statistics, taking into account local boundaries and global distribution; the joint loss function uses weighted cross-entropy and Dice to alleviate class imbalance and enhance boundary and small target region fitting ability; The step 2 specifically comprises the following steps: Step 201: Select ResNet network as the backbone network, and the network structure diagram is as shown in Figure 2 The core structure of the ResNet network is the design of the residual module, and each residual block of the ResNet network contains two 3x3 convolution layers and a skip link structure. This structure can add the input image features and the output data after convolution processing step by step to construct the residual mapping relationship of F(x)+x. The design of the residual block can make the network automatically learn the difference between the input and the output, reduce the model parameters, and improve the network performance.

[0028] Differentiated feature map processing strategies are used at different depth stages. Standard residual blocks are used in shallow layers to extract basic features, and deep layers combine down-sampling convolution to adjust the feature map size. The network uses global average pooling instead of traditional fully connected layers to reduce parameter quantity while improving the integrity of feature space information.

[0029] Step 202: Residual structure optimization based on deformable convolution. The recognition object of the deep residual network is mainly for natural images, and the details and global information of the simulation image are not optimized. To solve this problem, two shallow feature preservation and adaptive sampling improvement schemes are added to the backbone network.

[0030] In the sound field segmentation task, the features of the image are less than those of a natural image. In the convolution process, the multi-scale feature maps of Stage 2 and Stage 4 are retained, and the perception intensity of the sound ray density and the regional gray scale is enhanced through cross-stage feature fusion.

[0031] To ensure the sensitivity of the network model to the sound ray density change, a deformable convolution is introduced in the residual block of Stage 3. Through the adaptive offset in the deformable convolution, the sampling position of the convolution kernel is dynamically adjusted, which can improve the recognition rate of the network in the area where the sound ray density changes greatly and improve the segmentation accuracy of the proportion change.

[0032] Step 203: Global statistical feature module design. In the regional segmentation process of the sound propagation loss image, the uneven distribution of the sound ray density will cause the problem of fuzzy regional segmentation. A multi-scale density statistical feature enhancement module is added to the ResNet-34 network to improve the network model's ability to distinguish the convergence area of the shadow area in the image through a three-level pooling and feature re-labeling mechanism.

[0033] If the feature map output by the last residual block Stage 4 of the network is , the module uses a three-level pooling branch to extract the density features, where the global pooling, 4x4 adaptive pooling and 16x16 maximum pooling are respectively as shown in the formula: After three-level pooling is performed in succession, the pooling results are flattened and spliced, and a density description vector is generated through a fully connected layer: The global statistical feature module composed of three-level pooling and feature fusion can not only realize the overall perception of the global density benchmark, but also finely identify the local high-density area, effectively improving the network model's boundary detection ability for the shadow area and the convergence area.

[0034] Step 204: Joint loss function design. The global proportion matching loss can directly constrain the pixel proportion difference between the predicted region and the real label, as shown in formula (1.7): where is the predicted probability, is the real label, is the total number of pixels. It can effectively suppress the overall proportion deviation caused by local misjudgment. On the basis of the traditional loss function, a spatial weight is introduced to fit the loss of the sound ray density gradient area, as shown in formula (1.8): where the weight coefficient The gradient amplitude of the real label is calculated as: The gradient is calculated using the Sobel operator, which is a tunable parameter (experimentally set to 0.6). By combining the global proportion matching loss with the density gradient loss, the final joint loss function is obtained as: The optimized overall network architecture is shown in Figure 3 .

[0035] Step 3: Define the evaluation criteria based on the results of the gray scale calculation; A kind of unsupervised evaluation index is designed to make an early judgment on the gray scale value of the sound propagation loss image, and then verify the correctness of the segmentation region according to the difference in gray scale value.

[0036] In step 3, a kind of unsupervised evaluation index is designed to make an early judgment on the gray scale value of the sound propagation loss image, and then verify the correctness of the segmentation region according to the difference in gray scale value. The segmentation results are verified using the method of regional gray scale contrast and regional internal consistency. The preprocessed sound propagation loss image will have obvious differences in the gray scale value distribution of the shadow zone and the convergence zone. According to the setting of the image gray scale value in the data preprocessing process, the gray scale value range of the shadow zone and the convergence zone is described in advance, and the gray scale difference between the shadow zone and the convergence zone is calculated, which can provide reliable basis for the subsequent segmentation results.

[0037] The step 3 specifically includes the following steps: Step 301: Formula (1.11) defines the gray scale difference between the shadow zone and the convergence zone: In the formula, reflects the average gray scale value of the shadow region in the segmentation result, and the average gray scale value of the convergence zone in the segmentation result. The gray scale difference index is obtained, which is larger, indicating that the gray scale difference between the two segmentation results is significant, and the segmentation result is reasonable, otherwise it represents that the segmentation region has been confused.

[0038] Step 302: While determining the difference between the segmentation regions, the consistency within the region is also judged and identified. Formula (1.12) measures the uniformity of the gray scale within a single class region: Wherein, represents the gray scale variance of the shadow zone, represents the gray scale variance of the convergence zone. By evaluating the consistency index , it can be reflected whether the segmentation region is consistent. This index can judge whether there is noise or missegmentation phenomenon in the result of region segmentation.

[0039] The unsupervised statistics of the gray scale contrast of the region and the consistency of the region can judge the segmentation difference of different regions, and evaluate the integrity of each region, and the index system can comprehensively reflect the segmentation ability of the model for different regions.

[0040] The unsupervised indicators CR (inter-class gray scale contrast) and IRC (intra-class consistency) are used to evaluate the segmentation quality; in order to be comparable with the baseline method, if the label is provided, the supervised indicators such as PA / mIoU are also given.

[0041] In the present example, the northwest Pacific Ocean is selected For fixed region, the sound propagation loss image obtained based on the temperature and salinity data of the northwest Pacific Ocean is taken as the input sample to carry out the region segmentation experiment of the convergence zone and the shadow zone, a contrast experiment is designed for verification, and the evaluation index is set to the accuracy of the shadow zone and the convergence zone under the same data set. The experimental environment configuration is as follows: the hardware platform is NVIDIA RTX 3060 GPU, and the software framework is PyTorch1.12.1. In order to avoid overfitting, the early stopping strategy (when the validation set accuracy does not improve for 10 consecutive epochs, the training is terminated) is adopted in the training process, and the maximum training round is limited to 100.

[0042] The region segmentation strategies of various methods are shown in Table 2.

[0043]

[0044] In order to verify the advancement of the method and obtain reliable region segmentation results, four groups of contrast experiments are set, which respectively use the traditional threshold method of Otsu, region growing method, original ResNet-34 network and improved ResNet-34 network for experiment, and in order to ensure the reliability of the experiment, the original ResNet-34 method and the improved method adopt the same training strategy.

[0045] The input image participates in the experiment in the original resolution, and the batch size and the initial learning rate are compared as shown in Table 3, the batch size is selected as 8, 16 and 32, and the initial learning rate is selected as 1×10 -4 , 1×10 -5 , 1×10 -6 , and the overall accuracy (formula 1.11) of the shadow zone and the convergence zone is compared to determine the best batch size and learning rate participating in the experiment.

[0046]

[0047] According to the comparison experiment result, the batch size is set to 32, and the initial learning rate is set to 1×10 -5, and the other methods. The improved network integrates a global statistical feature module in the residual block, which essentially retains the topology and parameter size of the original ResNet-34 network. Therefore, this experiment maintains the batch size and learning rate settings of the original network.

[0048] The offset of deformable convolution is generated by an additional 3x3 convolution layer, which takes the feature map of the current layer as input. The output channel number of the 9 sampling points corresponding to the x and y axes is 18, the standard deviation of the offset initialization is 0.1, the bilinear interpolation sampling is used, and the gradient threshold is 1 for clipping to stabilize the training. Sub-pixel level position awareness is achieved through bilinear interpolation sampling. In addition to improving the position, the consistency of other parameter settings effectively excludes irrelevant variables that interfere with the experimental results, ensuring that the improvement in the accuracy of the network model results only comes from the contribution of the new module rather than potential influences from changes in other parameters, thereby fully verifying the effectiveness of the improvement strategy. The comparison experiments on batch size and initial learning rate are shown in Table 4.

[0049]

[0050] The traditional threshold method, region growing method, original ResNet-34 network, and improved ResNet-34 network were compared in the experiment, and the accuracy of the convergence area proportion of the acoustic shadow area after region segmentation of the sound propagation loss image was obtained. The experimental results are shown in Table 5.

[0051]

[0052] The Otsu threshold segmentation method will have errors in judging the gray value when facing the relatively gentle density gradient area in the sound propagation loss image. The Otsu method uses a single global threshold for calculation, which is prone to obvious errors in boundary calculation. The region growing method uses an automatic seed point selection method to promote the algorithm process in the identification and diagnosis of the sound propagation loss image. However, due to the complex distribution of the convergence area in some sound field images, the automatically selected seed points may deviate and miss some areas. At the same time, due to the complex boundary texture of the sound propagation loss image, the seed points may lead to inaccurate boundary segmentation when growing, making it difficult to establish a global statistical feature distribution in complex patterns. In the comparison experiment, the original ResNet-34 network has an accuracy of 74.2% and 87.5% for the segmentation of the acoustic shadow area and the convergence area at the best learning rate, which is higher than the traditional method, but still has room for improvement.

[0053] The improved ResNet-34 network introduces a global statistical module after Stage4, extracts the global statistical features of the shadow area and convergence area in the sound propagation loss image through multi-scale pooling and feature fusion, and optimizes the loss function by using the joint optimization strategy of weighted cross-entropy loss and Dice loss. The implementation of the two optimization schemes can make the network capture the global distribution law and local details, and give higher weight to the low-frequency category, increase the sensitivity of the model to the minority category image, and significantly improve the quantization accuracy of the area proportion of the shadow area and the convergence area.

[0054] The results of the comparative experiment show that under the optimal learning rate, the improved ResNet-34 network has a judgment accuracy of 76.9% and 90.6% for the proportion of the shadow area and the convergence area, which is significantly higher than the traditional method and the original network, and the improved method can improve the accuracy and precision of the region segmentation.

[0055] The above is only the preferred embodiment of the present application, it should be pointed out that for the ordinary skilled in the art, without departing from the principles of the present application, can make a number of improvements and refinements, these improvements and refinements should also be considered as the protection scope of the present application. The structures, devices and operation methods not specifically described and explained in the present application, such as no special description and limitation, are implemented according to the conventional means in the art.

Claims

1. A sound field region segmentation method based on an improved ResNet-34, characterized in that: Specifically, the following steps are included: Step 1: Setting up the sound propagation loss image dataset and sound field simulation; Based on multi-temporal three-dimensional temperature and salinity field data samples, the coherent propagation CG mode in the Bellhop model is used to perform ray tracking and propagation loss calculation, generating a range-depth domain sound propagation loss image; an ideal Munk profile is introduced to generate control samples with the same image resolution and dynamic range, completing the division of training / validation / test sets. Normalize and enhance the images of all samples; Step 2: Design and structural improvement of the segmentation network; A segmentation backbone network is constructed based on ResNet-34. Residual connections and stacked 3×3 convolutions are used to extract multi-layer semantic and fine-grained texture features. Global average pooling replaces fully connected layers to maintain the consistency of spatial representation, and the output multi-level feature maps required by the segmentation head serve as input for subsequent improvement modules. Deformable convolutions are introduced in Stage 3 to adaptively sample regions of abrupt changes in ray density, spanning... Multi-scale feature fusion is performed to improve the ability to perceive convergence / shadow area structures of different sizes. A global statistical feature module is added after Stage 4 to rescale and enhance multi-scale statistics, thereby taking into account both local boundaries and global distribution. The joint loss function adopts weighted cross-entropy and Dice to alleviate class imbalance and enhance the fitting ability of boundaries and small target regions. Step 3: Define the evaluation criteria for the results based on grayscale calculation; Design an unsupervised evaluation index to pre-judge the gray values ​​of an image with sound propagation loss, and then verify the correctness of the segmented region based on the difference in gray values.

2. The sound field region segmentation method based on improved ResNet-34 according to claim 1, characterized in that: Step 1 includes the following steps: Step 101: In the process of depicting the characteristics of the deep-sea sound field, the sound propagation loss image can intuitively represent the propagation law of the sound field and the loss region of the sound rays; the Bellhop model is designed to perform two-dimensional sound ray tracing. When determining the ray coordinates, Bellhop obtains the amplitude and sound pressure by calculating the ray equation. For a general cylindrical symmetric system, the ray equation is expressed as: Where r(s) and z(s) represent the ray coordinates in the cylindrical symmetric system, s represents the arc length along the ray, c(r, z) is the sound velocity distribution, and the propagation loss is calculated based on the superposition of sound pressure fields. Bellhop calculates the sound pressure p(r, z) through the superposition of Gaussian beams and converts it into propagation loss: Where p0 is the sound pressure level at a typical reference distance; Step 102: Bellhop Model Environment File Settings; The Bellhop model's environment file defines the physical boundary conditions and medium parameters for sound propagation. Appropriate parameter settings can more accurately reflect the sound propagation characteristics of the deep-sea sound field. The Bellhop environment file uses a layered format to describe the computational region of the sound field. The frequency corresponding to the sound source frequency is set to 50Hz, a frequency band that supports long-distance sound wave propagation. The water medium layer is set to a single layer; nmedia=1. In the sound velocity profile, the sound velocity value at the corresponding location is calculated using the sound velocity formula, and piecewise linear interpolation is performed to describe the depth-related sound velocity changes, covering the range from the sea surface to 5000 meters deep. The sound velocity minimum occurs in the middle and deep layers, forming a deep-sea acoustic channel axis, which will dominate the horizontal convergence effect of sound energy. The seabed parameters are uniformly set as sandy sedimentary layers with a density of 1.8 g / cm³ and a P-wave velocity of 1600 m / s. The sound source depth is set to 1000 meters and located above the deep-sea acoustic channel axis to ensure that the sound wave energy propagates horizontally along the acoustic channel; the receiver covers the entire water depth from 0 to 5000 meters and the horizontal distance from 0 to 100 km, and the spatial sampling interval satisfies the Nyquist criterion of the sound field interference fringes, which can completely capture the spatial distribution pattern of the sound ray convergence area and the sound shadow area. A coherent propagation loss calculation model is adopted, taking into account both the sound field phase interference effect and the energy accumulation characteristics of the convergence region; the emission angle range is set to -20.3°~20.3°, covering the effective propagation path from the sound source to the receiver; Step 103: Generate an ideal Munk profile acoustic propagation loss image based on the above parameters. The positional differences between the acoustic shadow area / convergence area in the ideal Munk profile acoustic propagation loss image and the acoustic propagation loss image generated from actual ocean data are mainly due to the deviation of the acoustic duct axis depth and the deviation between the measured value and the theoretical absorption value of the absorption coefficient of the seabed sediment layer. Step 104: Highlight the high-loss features of the convergence region and the low-loss features of the sound shadow region through image enhancement technology, and use edge detection algorithm or region growing method to extract the geometric boundaries of the energy anomaly region to determine the difference in sound ray energy between the sound shadow region and the convergence region.

3. The sound field region segmentation method based on improved ResNet-34 according to claim 2, characterized in that: Step 2 specifically includes the following steps: Step 201: Select ResNet network as the backbone network. The core structure of ResNet network lies in the design of residual modules. Each residual block of ResNet network contains two 3×3 convolutional layers and a skip connection structure. This structure can add the input image features to the output data after convolution processing step by step to construct the residual mapping relationship F(x)+x. Step 202: Based on deformable convolution, residual structure optimization is performed by adding two shallow feature retention and adaptive sampling improvements to the backbone network. In the sound field segmentation task, images have fewer features than natural images. Therefore, we choose to retain the multi-scale feature maps of Stage 2 and Stage 4 during the image convolution process, and enhance the perception intensity of sound ray density and regional grayscale through cross-stage feature fusion. We choose to introduce deformable convolution in the residual block of Stage 3, and dynamically adjust the sampling position of the convolution kernel through adaptive offset in the deformable convolution. Step 203: Design of the global statistical feature module. A multi-scale density statistics feature enhancement module is chosen to be added to the ResNet-34 network. If the feature map output by the last residual block Stage4 of the network is The module employs a three-level pooling branch for density feature extraction, where global pooling, 4×4 adaptive pooling, and 16×16 max pooling are respectively shown in the following equations: After repeated three-stage pooling, the pooling results are flattened, concatenated, and then a density description vector is generated through a fully connected layer. ; Step 204: Joint loss function design; the global proportion matching loss can directly constrain the difference in pixel proportion between the predicted region and the true label, as shown in equation (1.7): in To predict probabilities, For real labels, This represents the total number of pixels. It can effectively suppress the overall proportion shift caused by local misjudgment; based on the traditional loss function, spatial weights are introduced to fit the loss in the acoustic density gradient region, as shown in Equation (1.8): Among them, the weighting coefficient Calculated from the gradient magnitude of the real label: The parameter is adjustable (0.6 in the experiment), and the gradient is calculated using the Sobel operator. The joint loss function is obtained by combining the global proportion matching loss and the density gradient loss: 。 4. The sound field region segmentation method based on improved ResNet-34 according to claim 3, characterized in that: In step 3, an unsupervised evaluation index is designed to make an initial judgment on the gray value of the sound propagation loss image, and then the correctness of the segmentation region is verified based on the difference in gray value. The segmentation results are verified by using regional grayscale contrast and regional consistency methods. After preprocessing, the sound propagation loss image will show significant differences in grayscale value distribution between the sound shadow area and the convergence area. Based on the image grayscale value settings during data preprocessing, the grayscale value range of the sound shadow area and the convergence area can be described in advance, and the grayscale difference between the sound shadow area and the convergence area can be calculated, which can provide a reliable basis for the subsequent segmentation results.

5. The sound field region segmentation method based on improved ResNet-34 according to claim 4, characterized in that: Step 3 specifically includes the following steps: Step 301: Formula (1.11) defines the degree of grayscale difference between the sound-shadow region and the convergence region: In the formula This reflects the average gray value of the sound-shadow region as shown in the segmentation result. This reflects the average gray value of the convergent region in the segmentation result; the resulting gray value difference index The larger the value, the more significant the difference in grayscale values ​​between the two segmentation results, indicating that the segmentation result is reasonable; conversely, the smaller the value, the more likely the segmented regions have been confused. Step 302: While determining the differences in segmented regions, the consistency within each region is judged and identified. Formula (1.12) measures the uniformity of gray levels within a single type of region: in, The grayscale variance representing the sound and shadow area. The gray variance representing the convergence region is used to evaluate the consistency index. This indicator can reflect whether the segmented region is consistent, and can determine whether noise or missegmentation occurs in the region segmentation result.

Images