Method and system for identifying polymetallic nodules in underwater sonar images

Abstract

The present application relates to the technical field of ocean engineering, and discloses a kind of underwater sonar image polymetallic nodule identification method to realize the efficient, accurate identification of polymetallic nodule, improve the generalization ability and anti-interference ability of model.The method comprises: sequentially carrying out noise suppression, image enhancement, geometric correction and distortion compensation on the original sonar image;Extract the reflection intensity features composed of gray mean, variance and peak intensity from the preprocessed image, and the texture characteristic features composed of energy, entropy, contrast and correlation;Two kinds of features are respectively mapped to dimension-unified high-dimensional feature maps by independent 1x1 convolution, and then the attention weights of each dimension and each pixel are learned by using multilayer perception after element-by-element addition;Then, based on the attention weights, the weighted two kinds of feature maps and the preprocessed image are spliced along the channel to generate the target fusion feature map, which is then input into the recognition model to obtain the polymetallic nodule identification result.

Description

Technical Field

[0001] This invention relates to the field of marine engineering technology, specifically to a method for identifying multimetallic nodules in underwater sonar images based on deep learning. Background Technology

[0002] Polymetallic nodules are rich in various metallic elements. The distribution density of polymetallic nodules varies considerably across seabed mining areas; only mining areas with a certain nodule density are economically viable. By obtaining and aggregating basic measurements of individual nodules, spatial statistical data over a wide range can be obtained, such as polymetallic nodule abundance (kg / m²) and seabed coverage (%). Analysis of this statistical data allows for the selection of high-abundance, low-cost areas for rational and environmentally friendly mining. Therefore, to mine seabed polymetallic nodules, a thorough assessment and analysis of the nodule distribution in each mining area is essential.

[0003] Currently, the detection of underwater polymetallic nodules mainly relies on sonar equipment to acquire seabed images, and the nodules are identified and located through the analysis of sonar images. However, due to the complex environment of the deep sea (water flow interference, seabed topographic relief, sound wave propagation attenuation, artificial gain setting deviation, beam pattern effect, etc.), underwater sonar images are prone to serious noise pollution (Gaussian noise, impulse noise), image blurring, geometric distortion (caused by scanning angle), and radiation distortion (caused by uneven gain, beam brightness, and propagation distance attenuation) during the acquisition process.

[0004] Radiation distortion is a core bottleneck affecting image quality and subsequent recognition performance. For example, radiation distortion can distort key features such as the reflection intensity and texture characteristics of polymetallic nodules, causing nodule features to be confused with seabed background features. This directly leads to the failure of subsequent feature extraction and a significant drop in recognition accuracy. Furthermore, common distortion compensation processes are completely disconnected from feature extraction and feature fusion. Distortion compensation is treated as an independent preprocessing step without considering its impact on subsequent feature extraction or utilizing the results of subsequent feature processing to optimize the distortion compensation effect. This results in a vicious cycle of "inaccurate distortion compensation → distorted feature extraction → ineffective feature fusion → low recognition accuracy," which is one of the core reasons why many existing technologies cannot achieve accurate recognition of polymetallic nodules.

[0005] Secondly, the key features of polymetallic nodules in sonar images, such as reflection intensity and texture characteristics, are not significantly different from the seabed background. Traditional identification methods (such as threshold segmentation, template matching, and traditional machine learning methods) use a single feature extraction approach, which makes it difficult to fully capture the core features of the nodules, resulting in low feature discrimination. Even with the introduction of deep learning (DL) technology (such as the U-Net network), the problem of "disconnect between distortion compensation and feature processing" has not been solved, leading to distortion of the features input to the model. This prevents the model from leveraging the feature extraction advantages of deep learning, and still results in low recognition accuracy and poor generalization ability.

[0006] Furthermore, underwater sonar image datasets suffer from small size and high annotation costs (requiring manual pixel-level annotation). Directly training deep learning models is prone to overfitting. Transfer learning techniques can leverage the general features of pre-trained models on large-scale datasets to alleviate the challenges of training with small samples. Existing research has not yet constructed an integrated preprocessing workflow encompassing systematic distortion compensation, feature extraction, and dynamic weighted feature fusion, nor has it combined this workflow with residual U-Net network optimization and transfer learning strategies to form a dedicated scheme for identifying multimetallic nodules in underwater sonar images. In other words, existing methods struggle to simultaneously balance recognition accuracy, computational efficiency, and model generalization performance, failing to meet the engineering application requirements of in-situ exploration of deep-sea multimetallic nodules. Summary of the Invention

[0007] The purpose of this invention is to disclose a method and system for identifying multimetallic nodules in underwater sonar images, so as to achieve efficient and accurate identification of multimetallic nodules and improve the generalization ability and anti-interference ability of the model.

[0008] To achieve the above objectives, the underwater sonar image polymetallic nodule identification method disclosed in this invention includes:

[0009] Step S1: Preprocess the input raw sonar image. The preprocessing includes noise suppression, image enhancement, geometric correction and distortion compensation in sequence. Step S2: Extract the reflection intensity features and texture characteristics of the preprocessed image; the reflection intensity features consist of gray-level mean, gray-level variance, and peak intensity; the texture characteristics consist of energy, entropy, contrast, and correlation. Step S3: The 3D reflection intensity feature and the 4D texture feature are respectively mapped through two independent 1×1 convolutional layers to obtain a dimensionally unified up-dimensional feature map. The convolutional kernel parameters used by the two convolutional layers are adaptively learned through model training. Step S4: After adding the dimension-aligned reflection intensity feature map and texture feature map element by element, learn the attention weights of each pixel in each dimension based on the multilayer perceptron. Step S5: Concatenate the reflection intensity feature map and texture feature map of each pixel in each dimension, which are weighted according to the corresponding attention weights, with the preprocessed image along the channel dimension to generate the target fusion feature map; Step S6: Input the target fusion feature map into the recognition model to obtain the recognition result of the polymetallic nodule.

[0010] Preferably, the noise suppression process in step S1 includes: Step S111: Select the Sym4 wavelet basis function to perform 3-level wavelet decomposition to transform the original distorted sonar image. Decomposed into low-frequency approximate components With high-frequency detail components ; ; In the formula: The wavelet decomposition level is denoted as . ; for After layer decomposition, the low-frequency contour components retain the main structural information of the nodules; For high-frequency noise and texture detail components of each layer; These are the pixel coordinates; Step S112: Apply an adaptive noise threshold to the high-frequency noise and texture detail components. Processing to obtain smoothed interference signals The specific calculation formula is as follows: ; Step S113: Adaptive filtering of the low-frequency profile components using the variable step-size NLMS algorithm is performed to balance convergence speed and steady-state performance; the specific calculation formula is as follows: ; ; In the formula: This is the filter weight vector; This is a variable step size factor; This is the filter residual; The input vector is the local image of the low-frequency contour component; The input energy term is used to adaptively scale the actual step size with the input intensity. Image coordinates The low-frequency components after filtering; This is the desired noise-free reference component; It is a constant; Step S114: Reconstruct the denoised image using inverse wavelet transform. :in, .

[0011] Preferably, the image enhancement in step S1 includes: Step S121: Define the gray-level histogram statistics of local sub-blocks in the image. In the formula: Grayscale; The gray level within a local sub-block is The number of pixels; The total number of pixels in a local sub-block; by limiting the contrast threshold. Histogram truncation restrictions: ; Step S122: Truncate histogram components exceeding the threshold and distribute them evenly across all gray levels, then calculate the local cumulative distribution function. ; Step S123: Obtain the enhanced image based on the grayscale mapping enhancement formula. The grayscale mapping enhancement formula is specifically as follows: .

[0012] Preferably, the geometric correction in step S1 includes: Step S131: Use the Canny edge detection operator to extract the target edges and seabed lines from the sonar image; based on the sonar imaging geometry, and according to the correspondence between slant range, depth, and horizontal seabed feature points, map the original slant range projection to the true horizontal projection to complete the correction of global geometric distortion. Canny edge gradient calculation to extract nodule and seabed contours:

[0013] Gradient magnitude and direction:

[0014] After non-maximum suppression and double thresholding, a sub-pixel level edge coordinate set is output; based on the control points obtained from edge detection, a sonar slant range-horizontal coordinate mapping correction model is constructed. ; In the formula: To detect slant range with sonar; The depth of the seabed; This represents the original slant distance projected horizontal pixel distance. This is the corrected actual geographical horizontal distance; Step S133, Global geometric coordinate transformation correction formula: ; In the formula: This represents the actual beam depression angle of the sonar. These are the standard pixel coordinates after geometric correction; the geometrically corrected image is output after mapping and resampling. .

[0015] Preferably, the distortion compensation in step S1 includes: Step S141: Perform radiation distortion correction using statistical methods. Within a certain window along the flight path, assuming that the echo intensity changes caused by the terrain are consistent and that the various losses of sound waves in the lateral direction are consistent, the energy change curve of the sound wave in the lateral direction is obtained by calculating the mean value of each signal within the window. Step S142: The correction coefficients for each column of the image can be obtained using the energy change curve. If the gain base value is selected as the average gray value within the window, the distortion compensation formula is: ; In the formula, For window height, Image width, Indicates the line number. Indicates the column number. Indicates the output channel index. For the first The correction factor of the column, These are sonar images after radiation distortion compensation, with subscripts. This represents the eigenvalue of the image after geometric correction in the space containing the subscript.

[0016] Preferably, the dimension of the upgraded feature map in step S3 is 16-dimensional, and the mapping process is as follows: ; ; in, It is a 3D reflection intensity feature. It is a 4-dimensional texture feature. The reflection intensity characteristics are in 16 dimensions after dimensionality upgrade. , This represents the texture characteristics after 16-dimensional upsizing. , , There are two respectively The weight matrix of the convolutional layer, It is a bias term; and the two mapping networks do not share weights or interfere with each other to ensure that the physical meaning of texture and reflection is completely decoupled.

[0017] Preferably, the recognition model employs a U-Ne network model that fuses residuals. Further, the U-Ne network model is iteratively updated based on a transfer learning method.

[0018] To achieve the above objectives, the present invention also discloses an underwater sonar image polymetallic nodule identification system, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method.

[0019] The present invention has the following beneficial effects: This paper combines deep learning technology with underwater sonar image analysis, employing a progressive, multi-level image preprocessing technique to address the quality issues of deep-sea sonar images. The reflection intensity features are then extracted from the preprocessed images, comprising the mean gray level, gray level variance, and peak intensity. Texture features are further extracted from the preprocessed images, comprising energy, entropy, contrast, and correlation. Finally, feature fusion is achieved through dynamic weights based on an attention mechanism. During the fusion process, the following multiple dimensions are synergistically employed to ensure rich and discriminative feature support for subsequent model recognition: I. To avoid semantic confusion, information cancellation, and indistinguishability of sources caused by directly adding elements of features with different physical meanings, this invention adopts a fusion strategy of first mapping independently, then weighting separately, and finally concatenating channels. This ensures that reflection features and texture features always maintain physical semantic independence, and the network can clearly distinguish the source of features and achieve selective trust.

[0020] Second, reflection intensity and texture characteristics are physically derived from the strong reflectivity of nodules, and both types of features exhibit the same trend of change in this scene. Considering this characteristic of sonar images and the characteristics of distortion consistency suppression, a strategy of uniformly learning attention weights after element-wise addition is adopted—that is, the greater the difference and the higher the discriminative power of the data after feature processing, the greater the corresponding weight.

[0021] Third, after weighting, high-discrimination features in the nodule region (such as high reflectance intensity and high texture entropy) are given high weights, while invalid features in the background region and interference features in the distortion residual region are given low weights, which significantly improves the discriminative power of the features.

[0022] Furthermore, in generating the target fusion feature map by concatenating the reflection intensity feature map and texture feature map of each pixel in each dimension with the preprocessed image along the channel dimension, the fusion feature map not only contains the pixel information of the original image after distortion compensation, but also integrates the dynamically weighted reflection intensity and texture feature, which fully reflects the coupling effect of distortion compensation and dynamic weighting, and can more effectively characterize the essential features of multimetallic nodules and avoid the recognition bias caused by distortion residue.

[0023] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0024] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the underwater sonar image polymetallic nodule identification method disclosed in an embodiment of the present invention.

[0025] Figure 2 This is a flowchart of the sonar image preprocessing process disclosed in an embodiment of the present invention.

[0026] Figure 3 This is a flowchart of the feature fusion process based on attention weights disclosed in an embodiment of the present invention.

[0027] Figure 4 This is a flowchart of the core steps of the underwater sonar image polymetallic nodule identification method disclosed in the embodiments of the present invention. Detailed Implementation

[0028] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0029] Example 1 As shown in Figure 1, a method for identifying multimetallic nodules in underwater sonar images based on deep learning specifically includes the following steps: Step 1: Underwater sonar image preprocessing. For details, please refer to [link / reference]. Figure 2 .

[0030] Raw sonar images of a deep-sea mining area were acquired using a side-scan sonar device. The resolution is Due to interference from deep-sea currents and seabed topography, the image exhibits significant Gaussian noise, blurring, and geometric distortion caused by the scanning angle. Noise suppression, image enhancement, geometric correction, and distortion compensation are sequentially performed to obtain a high-quality preprocessed sonar image. ).

[0031] 1.1 Noise Suppression Processing: A method combining wavelet multi-scale decomposition and adaptive filtering was adopted, using the Sym4 wavelet basis function for three-level wavelet decomposition to transform the original distorted sonar image. Decomposed into low-frequency approximate components With high-frequency detail components The process expression can be specifically defined as: In the formula: These are the pixel coordinates. The wavelet decomposition level is denoted as . ; for After layer decomposition, the low-frequency contour components retain the main structural information of the nodules; This represents the high-frequency noise and texture detail components of each layer.

[0032] Adaptive thresholding is used for high-frequency components to smooth interference signals. In the formula: The noise threshold is adaptive and dynamically updated according to the local noise intensity of the image.

[0033] A variable-step-size NLMS algorithm is used for adaptive filtering of low-frequency components to balance convergence speed and steady-state performance. Combined with the definition of filtering error The filtered output can be obtained. Mapping this output to the image space yields the filtered low-frequency components: = In the formula: This is the filter weight vector; This is a variable step size factor; This is the filter residual; The low-frequency component is the local image input vector; The input energy term is used to adaptively scale the actual step size with the input intensity. Image coordinates The low-frequency components after filtering; This is the desired noise-free reference component; It is a very small constant to prevent the denominator from being 0.

[0034] Finally, the denoised image is obtained by reconstructing it using inverse wavelet transform. : .

[0035] This method effectively suppresses Gaussian noise and salt-and-pepper noise in the deep-sea environment while preserving key shallow features such as the edges and textures of polymetallic nodules, thus retaining effective information for subsequent image enhancement and feature extraction.

[0036] 1.2 Image Enhancement Processing: An adaptive histogram equalization algorithm is used to adjust the grayscale distribution of the image, enhance the grayscale contrast between the multimetallic nodules and the seabed background, and highlight the contour features of the nodules. First, define the gray-level histogram statistics for local sub-blocks of the image: In the formula: Grayscale; The gray level within a local sub-block is The number of pixels; This represents the total number of pixels in the local sub-block.

[0037] By limiting the contrast threshold Histogram truncation restrictions: .

[0038] The histogram components exceeding the threshold are truncated and evenly distributed across all gray levels, and then the local cumulative distribution function is calculated. The specific calculation formula is as follows: .

[0039] Final grayscale mapping enhancement formula: .

[0040] 1.3 Geometric Correction Processing: The Canny edge detection operator can be used to extract the target edges and seabed lines of the sonar image to obtain high-precision depth information; based on the geometric relationship of sonar imaging, according to the correspondence between slant range, depth and horizontal seabed feature points, the original slant range projection is mapped to the real horizontal projection to complete the correction of global geometric distortion.

[0041] Canny edge gradient calculation accurately extracts nodule and seabed contours: Gradient magnitude and direction: After non-maximum suppression and double thresholding, a sub-pixel level set of edge coordinates is output. Specifically, by employing the Canny edge detection operator to extract the contour edges of the seabed line and polymetallic nodules in the sonar image, the contour information can be used to distinguish between the seabed nodule region and the water body region. Combined with the feature point pixel coordinates, the measured slant range, and the vertical depth, the slant range-horizontal coordinate mapping model parameters are calibrated. In other words, Canny edge detection provides the geometric model with "feature points with known geometric relationships," enabling parameters that previously required guesswork (such as the actual distance represented by each pixel) to be accurately solved. Conversely, without edge detection, the slant range-horizontal coordinate mapping model calibration loses its reliable method of obtaining control points from the image.

[0042] The sonar slant range–horizontal coordinate mapping correction model is as follows: In the formula: To detect slant range with sonar; The depth of the seabed; This represents the original slant distance projected horizontal pixel distance. This is the corrected actual geographical horizontal distance.

[0043] The global geometric coordinate transformation correction formula is: In the formula: This represents the actual beam depression angle of the sonar. These are the coordinates of the distorted pixels; These are the standard pixel coordinates after geometric correction.

[0044] Output geometrically corrected image after mapping resampling This thoroughly corrects the geometric deformations introduced by seabed topography and sonar scanning angles, ensuring that the nodule morphology, size, and spatial distribution characteristics are true and reliable.

[0045] 1.4 Distortion Compensation Processing: Polymetallic nodules are mainly distributed in the soft surface sediments of gently sloping deep-sea plains at depths of 4000 to 6000 meters in the global ocean. Statistical methods can be used to correct radiation distortion in sonar images caused by artificial gain settings, beam pattern effects, and propagation distance attenuation. This reduces image entropy, suppresses and corrects high-brightness anomalies near the seabed line, smooths the overall image grayscale, and improves image quality. Within a certain window along the flight path, assuming consistent echo intensity variations due to topography and consistent transverse acoustic losses, the mean value of each signal column is calculated within this window, yielding the transverse energy variation curve of the acoustic wave. A reasonable correction baseline is set, and this curve can be used to calculate the correction coefficients for each column of the image. If the gain baseline is chosen as the mean grayscale value within the window, the radiation distortion is as follows: In the formula, For window height, Image width, Indicates the line number. Indicates the column number. Indicates the output channel index. For the first The correction factor of the column, These are sonar images after radiation distortion compensation, with subscripts. This represents the eigenvalue of the image after geometric correction in the space containing the subscript.

[0046] Step 2: Processing and fusing key features of underwater sonar images. For details, please refer to [link / reference needed]. Figure 3 .

[0047] 2.1 Two core features are extracted from the preprocessed sonar image: reflection intensity and texture characteristics. Feature fusion is then performed using a dynamic weighted fusion network based on an attention mechanism, providing rich and discriminative feature support for model recognition. To avoid semantic confusion, information cancellation, and indistinguishability of origin caused by directly adding features with different physical meanings element-wise, this invention adopts a fusion strategy of independent mapping, separate weighting, and channel concatenation. This ensures that reflection and texture features remain physically and semantically independent, allowing the network to clearly distinguish feature sources and achieve selective trust.

[0048] Reflection intensity characteristics: These are obtained by extracting the mean gray level, gray level variance, and peak intensity statistics of the image, reflecting the differences in sonar reflection signals between the polymetallic nodules and the seabed background. Let the preprocessed sonar image be... (in , (These are the height and width of the image, respectively), then the reflection intensity characteristics It consists of the following statistics, the calculation process of which directly provides the basic feature basis for subsequent weight learning.

[0049] Mean grayscale value: This reflects the overall reflection intensity level of the image; the average gray value of the nodule region is usually higher than that of the background region.

[0050] Gray variance: The variance reflects the dispersion of the reflection intensity. The variance of the nodule region is usually larger, reflecting its difference from the background.

[0051] Peak intensity: The strong reflective points corresponding to the nodule region are one of the key features that distinguish the nodule from the background.

[0052] Finally, reflection intensity characteristics , dimension .

[0053] Texture characteristics: Extracted using the gray-level co-occurrence matrix algorithm, including energy, entropy, contrast, and correlation texture parameters, reflecting the roughness and texture distribution patterns of the polymetallic nodule surface; the extraction results complement the reflection intensity characteristics, jointly determining the core weighting logic of the dynamic weighting. Let the gray-level co-occurrence matrix be... ,in, For image gray levels, the selection of the two spatial parameters, distance d and direction θ, can be specifically as follows: Take... =2, take After calculating the features in the four directions ∈{0°,45°,90°,135°} and taking the average, the texture characteristic features are obtained. The calculation is as follows, and its parameter value directly affects the allocation ratio of attention weights.

[0054] energy: This reflects the uniformity of the texture; the energy value of the nodule region is usually lower than that of the background (the texture is more complex).

[0055] entropy: This reflects the complexity of the texture, with higher entropy values ​​in the nodule region.

[0056] Contrast Ratio: This reflects the clarity of the texture, with higher contrast in the nodule area.

[0057] Correlation: This reflects the correlation of textures, and the correlation of nodule regions differs significantly from that of the background.

[0058] Ultimately, texture characteristics Integrate into , dimension .

[0059] 2.2 Feature Mapping: Due to the characteristics of reflection intensity Texture characteristics Different dimensions Since direct weighted fusion is not possible, feature mapping is needed to achieve dimensionality unification, which is the key bridge connecting feature processing and dynamic weighting. The extracted reflection intensity features... and texture characteristics Feature mapping is performed using two independent 1×1 convolutional layers. The convolutional kernel parameters are adaptively learned during model training to ensure that the mapped features retain the discriminative power of the original features to the greatest extent. The mapping process is as follows: ; ;in, , There are two respectively The weight matrix of the convolutional layer, As a bias term, the resulting feature maps are all 16-dimensional. and This provides a unified dimension of feature input for subsequent dynamic weighted fusion. The two mapping networks do not share weights or interfere with each other, ensuring complete decoupling of the physical meaning of texture and reflection.

[0060] 1.3 Attention Weight Calculation: Based on the feature processing results, the importance weights of the two types of features are adaptively learned. The weight allocation directly depends on the representational capabilities of the reflection intensity feature and the texture feature. Reflection intensity and texture features are physically derived from the strong reflectivity of nodules, and the two types of features exhibit the same direction of change in this scene. Considering this characteristic of sonar images and the characteristics of distortion consistency suppression, a strategy of uniformly learning attention weights by adding elements one by one is adopted—that is, the greater the difference and the higher the discriminative power of the data after feature processing, the greater the corresponding weight.

[0061] First of all and Element-by-element addition is performed to obtain the fused feature map. ,in, Indicates spatial location All channels, that is, all 16 channel values ​​of the pixel in the i-th row and j-th column, are input into the Multi-Layer Perceptron (MLP) at once. This feature map integrates the preliminary information of the two types of features, providing a basis for weight calculation; then... The input is an attention module composed of an MLP, which generates a 16-dimensional attention weight map through nonlinear transformation. The specific calculation process is as follows: ; ;in, Indicates the first line, number Column, No. Unnormalized attention weights of the channel Indicates the first line, number Column, No. Normalized attention weights for the channel; For learnable parameters of MLP, This is the Sigmoid activation function.

[0062] In tuberculous areas, and All are high values, and the positive gradient generated by the loss function is propagated to [the next value] via the chain rule. This increases the weight, achieving "high reflectivity + high texture → high weight"; in the background area, and All are low values, negative gradient makes It is suppressed to achieve "low reflection + low texture → low weight".

[0063] Dynamic weighted fusion of features: using attention weight maps For feature maps and Weighted fusion is performed, and the weighting process combines the effect of distortion compensation to enhance effective features, suppress ineffective features and residual distortion interference, forming a process flow of distortion compensation, feature processing and dynamic weighting. The specific formula is as follows: ; ;in, , These are the weighted reflection intensity feature map and texture feature map, respectively. After weighting, high-discriminative features in the nodule region (such as high reflection intensity and high texture entropy) are given high weights, while invalid features in the background region and interference features in the distortion residual region are given low weights, significantly improving the discriminative power of the features. Finally, the weighted feature map is compared with the preprocessed image data obtained after distortion compensation in step 1. By concatenating the channels, the final fused features are obtained. Its dimensions are The dimension serves as the input data for subsequent models. This fused feature not only includes the pixel information of the original image after distortion compensation, but also integrates dynamically weighted reflection intensity and texture characteristics, fully demonstrating the coupling effect of distortion compensation and dynamic weighting. It can more effectively characterize the essential features of multimetallic nodules and avoid recognition bias caused by distortion residues.

[0064] Step 3: Optimize the U-Net model based on deep residual networks.

[0065] This embodiment constructs a U-Net optimization model that integrates deep residual networks to solve the problems of gradient vanishing and insufficient feature extraction in traditional U-Net, thereby improving the ability to identify small target nodules.

[0066] The basic architecture of the U-Net network model in this embodiment includes an encoder and a decoder. The encoder performs image downsampling and shallow / mid-level feature extraction, while the decoder performs upsampling and deep feature extraction. A deep residual network module is introduced at key locations in both the encoder and decoder. This deep residual network module uses a residual block structure and achieves deep and shallow feature fusion through skip connections, thereby effectively alleviating the gradient vanishing problem during model training.

[0067] Optionally, the deep residual network module in this embodiment can be set at the end of each convolutional stage of the U-Net encoder and at the beginning of each deconvolutional stage of the decoder. Each residual block can consist of two conventional convolutional layers, a batch normalization layer, and a ReLU activation function, which enhances the nonlinear fitting ability of the model and enables the model to extract the deep features of polymetallic nodules more fully.

[0068] Step 4: Introduction of transfer learning and model training.

[0069] In this embodiment, the introduction of transfer learning technology can solve the problem of model overfitting in small sample scenarios, thereby improving model training efficiency and generalization ability, and completing the training of the recognition model.

[0070] Pre-trained model selection: The ResNet50 model, which was pre-trained on the large-scale image dataset ImageNet, was selected as the pre-trained model. Its underlying feature extractor has learned common features such as the edge, shape, and texture of the image, and can be directly transferred to the multi-metal nodule recognition task.

[0071] Network parameter replacement and freezing: Replace some convolutional layers of the U-Net network in step 3 with the corresponding convolutional layers of the ResNet50 model, and freeze the underlying parameters of the replaced convolutional layers. Only train the parameters of the unreplaced convolutional layers of the U-Net network and the deep residual network modules.

[0072] Dataset construction: 1000 underwater sonar polymetallic nodule images with manual pixel-level annotation (label 1 for nodule, 0 for background) were selected, of which 800 images were used as the training set and 200 images were used as the validation set. All images were preprocessed in step 1 and feature fusion processed in step 2. Model training: using fused features The input and labeled images are used as labels. The cross-entropy loss function is used as the loss function of the model. The initial learning rate is set to 0.001, the batch size is 16, and the number of iterations is 100. The Adam gradient descent algorithm is used for model training. During the training process, when the recognition accuracy on the validation set no longer improves for 10 consecutive rounds, the training is stopped to avoid overfitting.

[0073] Model Performance Validation: As shown in Table 1, the trained multi-metal nodule recognition model achieved a recognition accuracy of 96.2%, an F1 score of 94.1%, and a recall of 93.5% on the validation set. Compared to the traditional method (recognition accuracy 92.5%, F1 score 90.1%, recall 88.7%) without using the attention fusion mechanism but with the same image preprocessing as this embodiment, the accuracy was significantly improved by 3.7%, demonstrating that the attention mechanism effectively focuses on the key feature regions of the nodule, suppresses background noise interference, and greatly improves the accuracy of target recognition. The improvement in F1 score (+4.0%) and recall (+4.8%) indicates that the algorithm performs better in the detection task of small-sized, low-contrast nodules.

[0074] Table 1: Performance Comparison

[0075] Step 5: Identification and output of polymetallic nodules.

[0076] The underwater sonar image to be identified (resolution) The pixels are processed according to the preprocessing method in step 1 to obtain the preprocessed image to be identified; then, according to the method in step 2, the reflection intensity features and texture characteristics of the image to be identified are extracted, and feature cross-fertilization and data fusion are performed to obtain 33-dimensional fused data to be identified; the fused data to be identified is input into the recognition model trained in step 4, and the model outputs the recognition result of polymetallic nodules, including the location, outline and quantity information of the nodules, so as to achieve accurate identification of polymetallic nodules in underwater sonar images.

[0077] In summary, the core of the underwater sonar image polymetallic nodule identification method disclosed in this embodiment is as follows: Figure 4 As shown, it includes: Step S1: Preprocess the input raw sonar image. The preprocessing includes noise suppression, image enhancement, geometric correction, and distortion compensation.

[0078] Step S2: Extract the reflection intensity features and texture characteristics of the preprocessed image; the reflection intensity features consist of gray-level mean, gray-level variance, and peak intensity; the texture characteristics consist of energy, entropy, contrast, and correlation.

[0079] Step S3: The 3D reflection intensity feature and the 4D texture feature are respectively mapped through two independent 1×1 convolutional layers to obtain a dimensionally unified up-dimensional feature map. The convolutional kernel parameters used in the two convolutional layers are adaptively learned through model training.

[0080] Step S4: The dimension-aligned reflection intensity feature map and texture feature map are respectively added element-wise, and then the attention weights of each pixel in each dimension are learned based on the multilayer perceptron.

[0081] Step S5: Concatenate the reflection intensity feature map and texture feature map of each pixel in each dimension, which are weighted according to the corresponding attention weights, with the preprocessed image along the channel dimension to generate the target fusion feature map.

[0082] Step S6: Input the target fusion feature map into the recognition model to obtain the recognition result of the polymetallic nodule.

[0083] Example 2 This embodiment discloses an underwater sonar image polymetallic nodule identification system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a series of steps of the method in Embodiment 1 above.

[0084] In summary, the underwater sonar image polymetallic nodule identification method and system disclosed in the above embodiments of the present invention have the following beneficial effects: This paper combines deep learning technology with underwater sonar image analysis, employing a progressive, multi-level image preprocessing technique to address the quality issues of deep-sea sonar images. The reflection intensity features are then extracted from the preprocessed images, comprising the mean gray level, gray level variance, and peak intensity. Texture features are further extracted from the preprocessed images, comprising energy, entropy, contrast, and correlation. Finally, feature fusion is achieved through dynamic weights based on an attention mechanism. During the fusion process, the following multiple dimensions are synergistically employed to ensure rich and discriminative feature support for subsequent model recognition: I. To avoid semantic confusion, information cancellation, and indistinguishability of sources caused by directly adding elements of features with different physical meanings, this invention adopts a fusion strategy of first mapping independently, then weighting separately, and finally concatenating channels. This ensures that reflection features and texture features always maintain physical semantic independence, and the network can clearly distinguish the source of features and achieve selective trust.

[0085] Second, reflection intensity and texture characteristics are physically derived from the strong reflectivity of nodules, and both types of features exhibit the same trend of change in this scene. Considering this characteristic of sonar images and the characteristics of distortion consistency suppression, a strategy of uniformly learning attention weights after element-wise addition is adopted—that is, the greater the difference and the higher the discriminative power of the data after feature processing, the greater the corresponding weight.

[0086] Third, after weighting, high-discrimination features in the nodule region (such as high reflectance intensity and high texture entropy) are given high weights, while invalid features in the background region and interference features in the distortion residual region are given low weights, which significantly improves the discriminative power of the features.

[0087] Furthermore, in generating the target fusion feature map by concatenating the reflection intensity feature map and texture feature map of each pixel in each dimension with the preprocessed image along the channel dimension, the fusion feature map not only contains the pixel information of the original image after distortion compensation, but also integrates the dynamically weighted reflection intensity and texture feature, which fully reflects the coupling effect of distortion compensation and dynamic weighting, and can more effectively characterize the essential features of multimetallic nodules and avoid the recognition bias caused by distortion residue.

[0088] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying polymetallic nodules in underwater sonar images, characterized in that, include: Step S1: Preprocess the input raw sonar image. The preprocessing includes noise suppression, image enhancement, geometric correction and distortion compensation in sequence. Step S2: Extract the reflection intensity features and texture characteristics of the preprocessed image; the reflection intensity features consist of gray-level mean, gray-level variance, and peak intensity; the texture characteristics consist of energy, entropy, contrast, and correlation. Step S3: The 3D reflection intensity feature and the 4D texture feature are respectively mapped through two independent 1×1 convolutional layers to obtain a dimensionally unified up-dimensional feature map. The convolutional kernel parameters used by the two convolutional layers are adaptively learned through model training. Step S4: After adding the dimension-aligned reflection intensity feature map and texture feature map element by element, learn the attention weights of each pixel in each dimension based on the multilayer perceptron. Step S5: Concatenate the reflection intensity feature map and texture feature map of each pixel in each dimension, which are weighted according to the corresponding attention weights, with the preprocessed image along the channel dimension to generate the target fusion feature map; Step S6: Input the target fusion feature map into the recognition model to obtain the recognition result of the polymetallic nodule.

2. The method for identifying polymetallic nodules in underwater sonar images according to claim 1, characterized in that, The noise suppression process in step S1 includes: Step S111: Select the Sym4 wavelet basis function to perform 3-level wavelet decomposition to transform the original distorted sonar image. Decomposed into low-frequency approximate components With high-frequency detail components ; ； In the formula: The wavelet decomposition level is denoted as . ; for After layer decomposition, the low-frequency contour components retain the main structural information of the nodules; For high-frequency noise and texture detail components of each layer; These are the pixel coordinates; Step S112: Apply an adaptive noise threshold to the high-frequency noise and texture detail components. Processing to obtain smoothed interference signals The specific calculation formula is as follows: ； Step S113: Adaptive filtering of the low-frequency profile components using the variable step-size NLMS algorithm is performed to balance convergence speed and steady-state performance; the specific calculation formula is as follows: ； ； In the formula: This is the filter weight vector; This is a variable step size factor; This is the filter residual; The input vector is the local image of the low-frequency contour component; The input energy term is used to adaptively scale the actual step size with the input intensity. Image coordinates The low-frequency components after filtering; This is the desired noise-free reference component; It is a constant; Step S114: Reconstruct the denoised image using inverse wavelet transform. :in, .

3. The method for identifying polymetallic nodules in underwater sonar images according to claim 2, characterized in that, The image enhancement in step S1 includes: Step S121: Define the gray-level histogram statistics of local sub-blocks in the image. In the formula: Grayscale; The gray level within a local sub-block is The number of pixels; The total number of pixels in a local sub-block; by limiting the contrast threshold. Histogram truncation restrictions: ； Step S122: Truncate histogram components exceeding the threshold and distribute them evenly across all gray levels, then calculate the local cumulative distribution function. ; Step S123: Obtain the enhanced image based on the grayscale mapping enhancement formula. The grayscale mapping enhancement formula is specifically as follows: .

4. The method for identifying polymetallic nodules in underwater sonar images according to claim 3, characterized in that, The geometric correction in step S1 includes: Step S131: Use the Canny edge detection operator to extract the target edges and seabed lines from the sonar image; based on the sonar imaging geometry, and according to the correspondence between slant range, depth, and horizontal seabed feature points, map the original slant range projection to the true horizontal projection to complete the correction of global geometric distortion. Canny edge gradient calculation to extract nodule and seabed contours: Gradient magnitude and direction: After non-maximum suppression and double thresholding, a sub-pixel level edge coordinate set is output; based on the control points obtained from edge detection, a sonar slant range-horizontal coordinate mapping correction model is constructed. ； In the formula: To detect slant range with sonar; The depth of the seabed; This represents the original slant distance projected horizontal pixel distance. This is the corrected actual geographical horizontal distance; Step S133, Global geometric coordinate transformation correction formula: ； In the formula: This represents the actual beam depression angle of the sonar. These are the standard pixel coordinates after geometric correction; the geometrically corrected image is output after mapping and resampling. .

5. The method for identifying polymetallic nodules in underwater sonar images according to claim 4, characterized in that, The distortion compensation in step S1 includes: Step S141: Perform radiation distortion correction using statistical methods. Within a certain window along the flight path, assuming that the echo intensity changes caused by the terrain are consistent and that the various losses of sound waves in the lateral direction are consistent, the energy change curve of the sound wave in the lateral direction is obtained by calculating the mean value of each signal within the window. Step S142: The correction coefficients for each column of the image can be obtained using the energy change curve. If the gain base value is selected as the average gray value within the window, the distortion compensation formula is: ； In the formula, For window height, Image width, Indicates the line number. Indicates the column number. Indicates the output channel index. For the first The correction factor of the column, These are sonar images after radiation distortion compensation, with subscripts. This represents the eigenvalue of the image after geometric correction in the space containing the subscript.

6. The method for identifying polymetallic nodules in underwater sonar images according to any one of claims 1 to 5, characterized in that, The dimension of the upgraded feature map in step S3 is 16-dimensional, and the mapping process is as follows: ； ； in, It is a 3D reflection intensity feature. It is a 4-dimensional texture feature. The reflection intensity characteristics are in 16 dimensions after dimensionality upgrade. , This represents the texture characteristics after 16-dimensional upsizing. , , There are two respectively The weight matrix of the convolutional layer, It is a bias term; and the two mapping networks do not share weights or interfere with each other to ensure that the physical meaning of texture and reflection is completely decoupled.

7. The method for identifying polymetallic nodules in underwater sonar images according to claim 6, characterized in that, The identification model adopts a U-Ne network model that fuses residuals.

8. The method for identifying polymetallic nodules in underwater sonar images according to claim 7, characterized in that, The U-Ne network model is iteratively updated based on the transfer learning method.

9. A system for identifying multimetallic nodules in underwater sonar images, characterized in that, The invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the computer program, implements the method described in any one of claims 1 to 8.

Images