Underwater Target Detection Method and System Based on Dual-Complementary Dynamic Convolution

Abstract

The present application provides an underwater target detection method and system based on dual complementary dynamic convolution. The method includes: obtaining light field distribution data of an underwater target through a photonic crystal sensing array, where the light field distribution data includes non-uniform scattering characteristics and spatial frequency domain distortion characteristics; using a dual complementary dynamic convolution operator to perform dual complementary coupling on the non-uniform scattering characteristics and the frequency domain response characteristics of the photonic crystal sensing to generate a dynamic convolution kernel weight distribution; using the dynamic convolution kernel weight distribution to perform spatio-temporal degradation suppression on the spatial frequency domain distortion characteristics, and combining the temporal correlation characteristics of the pulse network to generate an anti-interference enhanced feature sequence; combining the anti-interference enhanced feature sequence to perform positioning and classification on the underwater target, and eliminating the error of the underwater light transmission path during the positioning and classification process to generate a positioning and classification result of the underwater target. The present application improves the accuracy and robustness of underwater target detection.

Description

Technical Field

[0001] Embodiments of the present application relate to the technical field of dual complementary dynamic convolution, and in particular, to an underwater target detection method and system based on dual complementary dynamic convolution. Background Art

[0002] Underwater target detection has important application values in fields such as marine resource exploration, underwater robot navigation, and military target recognition. However, the underwater environment is complex and changeable, and factors such as suspended particle scattering, turbulent disturbance, and medium inhomogeneity in the optical transmission path will cause serious degradation of the optical field distribution data, manifested as spatial frequency domain distortion and temporal motion distortion.

[0003] Currently, for the degradation problem of underwater target detection, a typical solution is the multi-scale convolutional neural network method based on deep learning. This method extracts the spatial features of underwater images by constructing multi-scale convolutional layers and combines a temporal recurrent network to capture the motion information of the target to achieve the detection and classification of underwater targets.

[0004] Although the existing solutions can extract the spatial and temporal features of underwater targets to a certain extent, their ability to suppress the degradation effect of the underwater optical transmission path is limited. Specifically, it is difficult for the multi-scale convolutional neural network to effectively distinguish the target signal from the artifact noise caused by path scattering, and the temporal recurrent network is sensitive to the phase distortion caused by turbulent disturbance, resulting in insufficient target detection accuracy and robustness, especially poor performance in complex underwater environments. Summary of the Invention

[0005] Embodiments of the present application provide an underwater target detection method and system based on dual complementary dynamic convolution to solve the problems of low accuracy and poor robustness in underwater target detection in the prior art.

[0006] In a first aspect, embodiments of the present application provide an underwater target detection method based on dual complementary dynamic convolution, including:

[0007] Obtaining the optical field distribution data of an underwater target through a photonic crystal sensing array, where the optical field distribution data includes non-uniform scattering characteristics and spatial frequency domain distortion characteristics;

[0008] Using a dual complementary dynamic convolution operator to perform dual complementary coupling on the non-uniform scattering characteristics and the frequency domain response characteristics of the photonic crystal sensing, generating a dynamic convolution kernel weight distribution, and adaptively matching the attenuation characteristics of the underwater optical transmission path during the generation of the dynamic convolution kernel weight distribution;

[0009] Using the dynamic convolution kernel weight distribution to perform spatio-temporal degradation suppression on the spatial frequency domain distortion characteristics, and combining the temporal correlation characteristics of the pulse network to generate an anti-interference enhanced feature sequence;

[0010] Combined with the anti-interference enhanced feature sequence, the underwater target is positioned and classified, and during the positioning and classification process, the error of the underwater optical transmission path is eliminated to generate the positioning and classification result of the underwater target.

[0011] Optionally, the utilization of the dynamic convolution kernel weight distribution to perform spatio-temporal degradation suppression on the spatio-frequency domain distortion features, combined with the temporal correlation characteristics of the pulse network, to generate the anti-interference enhanced feature sequence includes:

[0012] Adaptive correction of the frequency-domain response characteristics of the dynamic convolution kernel weight distribution through the attenuation coefficient, where the attenuation coefficient is a subset of the attenuation characteristics;

[0013] Based on the corrected dynamic convolution kernel weight distribution, spatio-temporal degradation suppression is performed on the spatio-frequency domain distortion features, and the corrected dynamic convolution kernel weight distribution and the frequency-domain response characteristics of the photonic crystal sensing form a dual complementary constraint;

[0014] Decompose the spatio-frequency domain distortion features after spatio-temporal degradation suppression into multi-scale frequency band components;

[0015] Each scale of the frequency band components respectively performs cross-domain interaction with the temporal correlation characteristics of the pulse network. During the cross-domain interaction process, according to the attenuation coefficient, the frequency-domain response characteristics of the photonic crystal sensing and the temporal correlation characteristics of the pulse network are dynamically allocated respectively to generate the frequency-domain temporal weight ratio parameter;

[0016] Based on the frequency-domain temporal weight ratio parameter, weighted fusion is performed on the multi-scale frequency band components to generate the anti-interference enhanced feature sequence.

[0017] Optionally, the adaptive correction of the frequency-domain response characteristics of the dynamic convolution kernel weight distribution through the attenuation coefficient includes:

[0018] Decompose the dynamic convolution kernel weight distribution into a high-frequency suppression kernel component and a low-frequency compensation kernel component according to the attenuation characteristics of the underwater optical transmission path;

[0019] Based on the attenuation characteristics, non-linear attenuation compensation is performed on the amplitude-frequency response of the high-frequency suppression kernel component to generate the gain compensation coefficient;

[0020] Based on the attenuation characteristics, time-varying correction is performed on the phase-frequency response of the low-frequency compensation kernel component to obtain the phase compensation offset;

[0021] Fuse the gain compensation coefficient and the phase compensation offset to generate the corrected dynamic convolution kernel weight distribution.

[0022] Optionally, according to the attenuation coefficient, the frequency-domain response characteristics of the photonic crystal sensing and the timing correlation characteristics of the pulse network are dynamically allocated respectively to generate a frequency-domain timing weight ratio parameter, including:

[0023] Dividing the multi-scale frequency band components into a high-frequency scattering-dominated frequency band and a low-frequency turbulence distortion-dominated frequency band according to the attenuation coefficient;

[0024] Extracting local detail features corresponding to the high-frequency scattering-dominated frequency band from the frequency-domain response characteristics of the photonic crystal sensing;

[0025] Performing non-linear gain compensation on the local detail features based on the attenuation coefficient;

[0026] Using a pre-trained multi-band attenuation compensation model to map the compensated local detail features to the weight allocation value of the frequency-domain response characteristics;

[0027] Extracting the motion trajectory features of the underwater target corresponding to the low-frequency turbulence distortion-dominated frequency band from the timing correlation characteristics of the pulse network;

[0028] Performing phase distortion suppression and correction on the motion trajectory features of the underwater target based on the attenuation coefficient;

[0029] Based on the corrected motion trajectory features of the underwater target, calculating the weight allocation value of the timing correlation characteristics through a preset time-varying attenuation correction function;

[0030] Fusing the weight allocation value of the frequency-domain response characteristics and the weight allocation value of the timing correlation characteristics to generate a frequency-domain timing weight ratio parameter, and the frequency-domain timing weight ratio parameter is used to control the contribution ratio of the photonic crystal sensing and the pulse network in the multi-scale frequency band components.

[0031] Optionally, using the dual complementary dynamic convolution operator to perform dual complementary coupling on the non-uniform scattering characteristics and the frequency-domain response characteristics of the photonic crystal sensing to generate a dynamic convolution kernel weight distribution, including:

[0032] Based on the spatial distribution characteristics of the non-uniform scattering characteristics and the frequency-domain response characteristics of the photonic crystal sensing, constructing a dual complementary coupling model;

[0033] In the dual complementary coupling model, extracting spatial high-frequency detail information and timing perturbation information from the spatial distribution characteristics;

[0034] Performing dual complementary fusion on the spatial high-frequency detail information and the timing perturbation information through a dynamic weight allocation mechanism;

[0035] Input the dual complementary fusion result into a preset dynamic convolution kernel generation module to output the weight distribution of the dynamic convolution kernel. The preset dynamic convolution kernel generation module incorporates a dual complementary dynamic convolution operator.

[0036] Optionally, when combining the anti-interference enhanced feature sequence to locate and classify the underwater target, and eliminating the error of the underwater light transmission path during the location and classification process to generate the location and classification result of the underwater target, it includes:

[0037] Calculate the spatial coordinates of the underwater target through the non-linear mapping relationship between the spatial distribution characteristics of the anti-interference enhanced feature sequence and the attenuation coefficient of the underwater light transmission path, and eliminate the spatial positioning deviation caused by the path scattering of the underwater light transmission path through iterative optimization during the calculation process;

[0038] Calculate the category probability distribution of the underwater target through the non-linear mapping relationship between the temporal motion characteristics of the anti-interference enhanced feature sequence and the attenuation coefficient of the underwater light transmission path, and eliminate the classification error caused by the medium inhomogeneity of the underwater light transmission path through iterative optimization during the calculation process;

[0039] Generate the location and classification result of the underwater target based on the spatial coordinates and the category probability distribution.

[0040] Optionally, after combining the anti-interference enhanced feature sequence to locate and classify the underwater target, and eliminating the error of the underwater light transmission path during the location and classification process to generate the location and classification result of the underwater target, the method further includes:

[0041] Perform collaborative feedback on the frequency domain response characteristics and the weight distribution of the dynamic convolution kernel according to the location and classification result and the attenuation coefficient;

[0042] Adjust the weight ratio parameter of the frequency domain response characteristics and the temporal correlation characteristics according to the result of the collaborative feedback to generate an optimized control parameter in a cross-domain suppression complex degradation scenario, and the optimized control parameter is used to correct the capture and processing process of the light field distribution data to form a closed-loop suppression mechanism.

[0043] In a second aspect, an embodiment of the present application provides an underwater target detection system based on dual complementary dynamic convolution, including:

[0044] An acquisition module for acquiring the light field distribution data of the underwater target through a photonic crystal sensing array, and the light field distribution data includes non-uniform scattering characteristics and spatial frequency domain distortion characteristics;

[0045] A coupling module, configured to perform dual complementary coupling on the non-uniform scattering features and the frequency-domain response characteristics of the photonic crystal sensing by using a dual complementary dynamic convolution operator, generate a dynamic convolution kernel weight distribution, and adaptively match the attenuation characteristics of the underwater optical transmission path during the generation process of the dynamic convolution kernel weight distribution;

[0046] A generation module, configured to perform spatio-temporal degradation suppression on the spatial frequency-domain distortion features by using the dynamic convolution kernel weight distribution, and generate an anti-interference enhanced feature sequence in combination with the timing correlation characteristics of the pulse network;

[0047] A positioning module, configured to perform positioning and classification on the underwater target in combination with the anti-interference enhanced feature sequence, and eliminate the error of the underwater optical transmission path during the positioning and classification process, and generate a positioning and classification result of the underwater target.

[0048] In a third aspect, an embodiment of the present application provides a computing device, including a processor and a memory, where a computer program is stored in the memory, and the processor is configured to run the computer program to execute any one of the methods for underwater target detection based on dual complementary dynamic convolution in the first aspect.

[0049] In a fourth aspect, an embodiment of the present application provides a computer storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, they implement any one of the methods for underwater target detection based on dual complementary dynamic convolution in the first aspect.

[0050] In an embodiment of the present application, a method for underwater target detection based on dual complementary dynamic convolution is provided. The method includes: obtaining light field distribution data of an underwater target through a photonic crystal sensing array, where the light field distribution data includes non-uniform scattering features and spatial frequency-domain distortion features; performing dual complementary coupling on the non-uniform scattering features and the frequency-domain response characteristics of the photonic crystal sensing by using a dual complementary dynamic convolution operator, generating a dynamic convolution kernel weight distribution, and adaptively matching the attenuation characteristics of the underwater optical transmission path during the generation process of the dynamic convolution kernel weight distribution; performing spatio-temporal degradation suppression on the spatial frequency-domain distortion features by using the dynamic convolution kernel weight distribution, and generating an anti-interference enhanced feature sequence in combination with the timing correlation characteristics of the pulse network; performing positioning and classification on the underwater target in combination with the anti-interference enhanced feature sequence, and eliminating the error of the underwater optical transmission path during the positioning and classification process, and generating a positioning and classification result of the underwater target.

[0051] The technical solution of the present application has the following beneficial effects:

[0052] This application acquires optical field distribution data containing non-uniform scattering characteristics and spatial frequency domain distortion characteristics, providing high-quality input for subsequent processing. Through the synergistic effect of the frequency domain response characteristics and the dynamic convolution kernel, a dynamic convolution kernel weight distribution that adaptively matches the attenuation characteristics of the underwater light transmission path is generated, enhancing the ability to suppress non-uniform scattering characteristics. The spatial frequency domain distortion characteristics are suppressed by the dynamic convolution kernel weight distribution, and combined with the temporal correlation characteristics of the pulse network, an anti-interference enhanced feature sequence is generated, effectively reducing the influence of path scattering and turbulence disturbance. Based on the anti-interference enhanced feature sequence, high-precision underwater target positioning and classification are achieved, while eliminating errors in the underwater light transmission path, improving the robustness and accuracy of the detection results.

[0053] Furthermore, in the embodiments of this application, the frequency domain response characteristics of the dynamic convolution kernel weight distribution are adaptively corrected by the attenuation coefficient to generate a corrected dynamic convolution kernel weight distribution, which forms a dual complementary constraint with the frequency domain response characteristics of the photonic crystal sensing; the spatial frequency domain distortion characteristics are suppressed spatiotemporally based on the corrected dynamic convolution kernel weight distribution, and the suppressed characteristics are decomposed into multi-scale frequency band components; each scale of the frequency band components respectively performs cross-domain interaction with the temporal correlation characteristics of the pulse network, and the weights of the frequency domain response characteristics and the temporal correlation characteristics are dynamically allocated according to the attenuation coefficient to generate a frequency domain temporal weight ratio parameter; based on the frequency domain temporal weight ratio parameter, the multi-scale frequency band components are weighted and fused to generate an anti-interference enhanced feature sequence.

[0054] Moreover, by adaptively correcting the dynamic convolution kernel weight distribution and combining the frequency domain response characteristics of the photonic crystal sensing, the effective suppression of the spatial frequency domain distortion characteristics is achieved; through cross-domain interaction and dynamic weight allocation, the frequency domain response characteristics and the temporal correlation characteristics are synergistically optimized to generate an anti-interference enhanced feature sequence; ultimately, the accuracy and robustness of underwater target detection are improved, and the degradation effect caused by path scattering and turbulence disturbance is effectively suppressed.

[0055] These aspects or other aspects of this application will be more clearly understood in the following description of the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are some embodiments of this application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0057] Figure 1 It is a flowchart of a method for underwater target detection based on dual complementary dynamic convolution provided by the embodiments of this application;

[0058] Figure 2 Schematic diagram of a structure of an underwater target detection system based on dual complementary dynamic convolution provided by an embodiment of the present application;

[0059] Figure 3 Schematic diagram of a structure of a computing device provided by an embodiment of the present application. Detailed implementation manners

[0060] In order to enable those skilled in the art of the present technology to better understand the solution of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application.

[0061] In some processes described in the specification, claims and above-mentioned accompanying drawings of the present application, a plurality of operations that appear in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order in which they appear herein or may be executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions such as "first" and "second" in this article are used to distinguish different messages, devices, modules, etc., and do not represent a sequence, nor do they limit that "first" and "second" are of different types.

[0062] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative efforts belong to the scope of protection of the present application.

[0063] The research and development idea of the present application is to obtain the light field distribution data of the underwater target through a photonic crystal sensing array, use the dual complementary dynamic convolution operator to perform dual complementary coupling on the non-uniform scattering characteristics and the frequency domain response characteristics of the photonic crystal sensing, and generate a dynamic convolution kernel weight distribution that adaptively matches the attenuation characteristics of the underwater light transmission path; suppress the spatio-temporal degradation of the spatial frequency domain distortion characteristics based on the dynamic convolution kernel weight distribution, and generate an anti-interference enhanced feature sequence in combination with the timing correlation characteristics of the pulse network; finally, perform positioning and classification on the underwater target in combination with the anti-interference enhanced feature sequence, eliminate the error of the underwater light transmission path during the positioning and classification process, and generate a high-precision positioning and classification result.

[0064] Figure 1 A flowchart of a method for underwater target detection based on dual complementary dynamic convolution provided by an embodiment of the present application is as Figure 1 shown, and the method includes:

[0065] Step 101: Obtain the optical field distribution data of the underwater target through a photonic crystal sensing array.

[0066] In this step, the photonic crystal sensing array refers to an optical sensor array based on photonic crystal materials. Specifically, it can refer to a network composed of multiple photonic crystal sensing units, which can accurately capture the underwater optical field distribution data, and has excellent spatial resolution and anti-interference performance. The optical field distribution data includes non-uniform scattering characteristics and spatial frequency domain distortion characteristics. The non-uniform scattering characteristic refers to the non-uniform noise in the optical field distribution data caused by the scattering of underwater suspended particles, which is manifested as random fluctuations in space. The spatial frequency domain distortion characteristic refers to the distortion of the optical field distribution data in the frequency domain caused by the non-uniformity of the underwater medium and turbulent perturbation, which is manifested as high-frequency noise and low-frequency ghosting.

[0067] In the embodiment of the present application, the photonic crystal sensing array generates optical field distribution data by receiving the optical field signal reflected by the underwater target. This data includes non-uniform scattering characteristics and spatial frequency domain distortion characteristics. The specific implementation process is as follows: The photonic crystal sensing array performs spatial sampling and frequency domain analysis on the optical field signal, extracts the non-uniform scattering characteristics and spatial frequency domain distortion characteristics in the optical field distribution data, and provides input for subsequent processing.

[0068] For example, in a shallow sea environment, the research team deployed a network composed of multiple photonic crystal sensing units to monitor the marine biological activities near the coral reef. Through continuous data collection and analysis for several weeks, the team successfully captured the optical field changes caused by the swimming of fish, and extracted the non-uniform scattering characteristics and spatial frequency domain distortion characteristics from them, providing basic data support for further research.

[0069] Step 102: Use the dual complementary dynamic convolution operator to perform dual complementary coupling on the non-uniform scattering characteristics and the frequency domain response characteristics of the photonic crystal sensing, generate the dynamic convolution kernel weight distribution, and adaptively match the attenuation characteristics of the underwater light transmission path during the generation process of the dynamic convolution kernel weight distribution.

[0070] In this step, the dual complementary dynamic convolution operator refers to a convolution operation method that combines spatial frequency domain characteristics and temporal characteristics. Photonic crystal sensing is a functional description, and a photonic crystal sensing array is a specific hardware form. The dynamic convolution kernel weight distribution is used to suppress the spatial frequency domain distortion features. The underwater light transmission path refers to the complete path where an optical signal is emitted from a light source (such as a laser emitter), propagates through the underwater medium to an underwater target, and then reflects back to the detection device (including the photonic crystal sensing array) from the underwater target. It is related to the underwater medium, rather than directly related to the underwater target, detection device, or photonic crystal sensing array. The attenuation characteristic refers to the overall characteristic of the signal attenuation of the optical signal due to physical phenomena such as absorption and scattering in the above path, and usually includes parameters such as attenuation coefficient, scattering coefficient, and absorption coefficient. The underwater target refers to the object that affects the reflection characteristic of the optical signal but does not directly determine the attenuation characteristic of the light transmission path. The detection device receives and processes the optical signal but does not directly determine the attenuation characteristic of the light transmission path. The photonic crystal sensing array is part of the detection device and is used to capture the optical signal but does not directly determine the attenuation characteristic of the light transmission path. The underwater medium determines the attenuation characteristic of the light transmission path, including water, suspended particles, turbulence, etc.

[0071] In the embodiment of the present application, first, a dual complementary constraint model is constructed based on the spatial distribution of non-uniform scattering characteristics and the frequency domain distribution of frequency domain response characteristics. Then, by adaptively matching the attenuation characteristics of the underwater light transmission path, the convolution kernel weight distribution is dynamically adjusted to generate a dynamic convolution kernel weight distribution applicable to the current underwater environment.

[0072] For example, continuing with the example of the previous step, after analyzing the light field changes around the coral reef, the researchers applied the dual complementary dynamic convolution operator to process the data. The researchers found that this method improved the detection accuracy for the signals of fine microbial activities, especially in complex and variable seawater conditions.

[0073] Step 103: Use the dynamic convolution kernel weight distribution to perform spatio-temporal degradation suppression on the spatial frequency domain distortion features, and combine the temporal correlation characteristics of the pulse network to generate an anti-interference enhanced feature sequence.

[0074] In this step, spatio-temporal degradation suppression is to suppress the degradation effects in the spatial domain and time domain on the spatially and spectrally distorted features through the dynamic convolution kernel weight distribution. The introduction of the spiking network is to complement the deficiency of the dynamic convolution kernel in suppressing temporal perturbations. The dynamic convolution kernel mainly deals with spatial and spectral distortion, while the spiking network focuses on suppressing temporal perturbations. The two work together to achieve comprehensive suppression of underwater degradation scenarios. The spiking network refers to a neural network or computational model based on spiking signal processing, which can capture the dynamic changes and correlation characteristics in temporal data. In the underwater target detection scenario, the spiking network is mainly used to process the temporal perturbation information in the light field distribution data (such as phase distortion caused by turbulence). The anti-interference enhanced feature sequence refers to the feature sequence after spatio-temporal degradation suppression and enhancement of temporal correlation characteristics, which has higher anti-interference ability.

[0075] In the embodiment of the present application, first, the spatially and spectrally distorted features are decomposed into multi-scale frequency band components, and then each frequency band component is subjected to layer-by-layer degradation suppression through the dynamic convolution kernel weight distribution. At the same time, combined with the temporal correlation characteristics of the spiking network, the temporal motion information of the target is extracted to generate an anti-interference enhanced feature sequence.

[0076] For example, with the in-depth research, the team further optimized the underwater target detection method, enabling it to accurately track the behavior patterns of target organisms even during strong currents or storms. This benefits from the application of spatio-temporal degradation suppression technology and the spiking network, which greatly improves the robustness and reliability of the system used in the underwater target detection method.

[0077] Step 104: Combine the anti-interference enhanced feature sequence to locate and classify the underwater target, and during the process of location and classification, eliminate the error of the underwater light transmission path to generate the location and classification result of the underwater target.

[0078] In this step, location and classification refer to determining the spatial position and category of the target based on the anti-interference enhanced feature sequence. The location and classification result includes location information and classification result. The location information includes the spatial coordinates, location accuracy, and motion trajectory of the target in three-dimensional space. The classification result includes the category label, category probability distribution, and classification confidence of the target.

[0079] In the embodiment of the present application, first, a spatial location model is constructed based on the spatial distribution characteristics of the anti-interference enhanced feature sequence to calculate the spatial coordinates of the target; then a temporal classification model is constructed based on the temporal motion characteristics to calculate the category probability distribution of the target; finally, the error of the underwater light transmission path is eliminated through closed-loop feedback correction to generate a high-precision location and classification result.

[0080] For example, in a shallow - sea environment, the team uses anti - interference enhanced feature sequences combined with target detection algorithms to identify and classify underwater targets. At the same time, in order to eliminate the error of the underwater light transmission path, the embodiments of the present application introduce a ray propagation correction technology based on physical modeling. By calculating the attenuation and scattering effects of light under different water quality conditions, the target position is corrected.

[0081] The present application obtains the light field distribution data through a photonic crystal sensing array, generates the dynamic convolution kernel weight distribution by using the dual - complementary dynamic convolution operator, combines the temporal correlation characteristics of the pulse network to generate anti - interference enhanced feature sequences, and finally realizes high - precision underwater target positioning and classification. This method effectively suppresses the degradation effects caused by path scattering and medium inhomogeneity, and improves the accuracy and robustness of underwater target detection.

[0082] To solve the problem of degradation effects caused by path scattering and medium inhomogeneity in underwater target detection and further improve the detection accuracy and robustness, in some embodiments, step 103: The space - frequency domain distortion features are suppressed for spatio - temporal degradation by using the dynamic convolution kernel weight distribution, and combined with the temporal correlation characteristics of the pulse network to generate anti - interference enhanced feature sequences, including:

[0083] Step 201: Adaptively correct the frequency - domain response characteristics of the dynamic convolution kernel weight distribution through the attenuation coefficient.

[0084] In step 201, the attenuation coefficient is a subset of the attenuation characteristics. The attenuation characteristics include the attenuation coefficient and other related parameters. The attenuation coefficient refers to the degree of attenuation of the optical signal in the underwater light transmission path, which is usually related to factors such as the concentration of suspended particles and the turbulence intensity in the water body. The larger the attenuation coefficient, the more serious the attenuation of the optical signal. The frequency - domain response characteristics refer to the response ability of the photonic crystal sensing in the frequency domain, which is used to extract the frequency - domain features in the light field distribution data.

[0085] In the embodiments of the present application, based on the attenuation coefficient of the underwater light transmission path, the amplitude - frequency response and phase - frequency response of the dynamic convolution kernel weight distribution are adjusted to adaptively match the attenuation characteristics of the current underwater environment, and a corrected dynamic convolution kernel weight distribution is generated.

[0086] Step 202: Based on the corrected dynamic convolution kernel weight distribution, suppress the spatio - temporal degradation of the space - frequency domain distortion features. The corrected dynamic convolution kernel weight distribution and the frequency - domain response characteristics of the photonic crystal sensing form a dual - complementary constraint.

[0087] In step 202, the dual - complementary constraint is used to enhance the degradation suppression effect.

[0088] In the embodiments of this application, the corrected dynamic convolution kernel weight distribution is applied to the spatial frequency domain distortion features. Through the synergistic effect of the frequency domain response characteristics and the dynamic convolution kernel, the high-frequency noise caused by the scattering of suspended particles and the low-frequency ghosting caused by the turbulent disturbance are suppressed, and the spatial frequency domain distortion features after degradation suppression are generated.

[0089] Step 203: Decompose the spatial frequency domain distortion features after spatio-temporal degradation suppression into multi-scale frequency band components.

[0090] In step 203, the multi-scale frequency band components refer to decomposing the spatial frequency domain distortion features into multiple frequency band components according to the frequency range, which is convenient for separately processing the degradation effects of different frequency ranges. For example, in the ocean environment, the light field distribution data is decomposed into high-frequency components (corresponding to the noise caused by the scattering of suspended particles, frequency range > 10 kHz), medium-frequency components (corresponding to the phase distortion caused by the turbulent disturbance, frequency range 1 kHz - 10 kHz), and low-frequency components (corresponding to the ghosting caused by the medium inhomogeneity, frequency range < 1 kHz), which is convenient for separately processing the degradation effects of different frequency ranges.

[0091] In the embodiments of this application, the spatial frequency domain distortion features are decomposed into high-frequency, medium-frequency, and low-frequency components through frequency domain filtering technology. Each component corresponds to different degradation effects, such as high-frequency noise, turbulent disturbance, etc., providing input for subsequent cross-domain interaction.

[0092] Step 204: Each of the frequency band components at each scale performs cross-domain interaction with the temporal correlation characteristics of the pulse network. During the cross-domain interaction, according to the attenuation coefficient, the frequency domain response characteristics of the photonic crystal sensing and the temporal correlation characteristics of the pulse network are dynamically allocated respectively to generate a frequency domain temporal weight ratio parameter.

[0093] In step 204, the frequency domain temporal weight ratio parameter is used to balance the contribution ratio of the frequency domain response characteristics and the temporal correlation characteristics in feature fusion.

[0094] In the embodiments of this application, the weight distribution of the frequency domain response characteristics and the temporal correlation characteristics is dynamically adjusted according to the attenuation coefficient to generate a frequency domain temporal weight ratio parameter. For example, when the turbulent disturbance is strong, the weight of the temporal correlation characteristics is increased; when the scattering of suspended particles is strong, the weight of the frequency domain response characteristics is increased. For example, when the turbulent disturbance is strong (such as the turbulence intensity > 0.5), the weight of the temporal correlation characteristics is increased (such as the weight is adjusted from 0.5 to 0.7) to better suppress the phase distortion caused by the turbulence; when the scattering of suspended particles is strong (such as the concentration of suspended particles > 1000 particles / mL), the weight of the frequency domain response characteristics is increased (such as the weight is adjusted from 0.5 to 0.7) to better suppress the high-frequency noise caused by the suspended particles.

[0095] Step 205: Weight and fuse the multi-scale frequency band components based on the frequency-domain time-series weight ratio parameter to generate an anti-interference enhanced feature sequence.

[0096] In the embodiment of the present application, each frequency band component is weighted and summed according to the weight ratio parameter to generate an anti-interference enhanced feature sequence, providing high-quality input for subsequent underwater target positioning and classification.

[0097] The following is a specific example:

[0098] In a shallow sea coral reef monitoring project, the research team deployed a set of photonic crystal sensing arrays to observe the biological activities in the coral reef area. First, the researchers determined the attenuation coefficient suitable for the local water quality conditions through on-site measurement and corrected the dynamic convolution kernel weight distribution accordingly. Then, the collected spatial frequency-domain distortion features were processed using the corrected weight distribution, reducing the noise interference caused by suspended particles and water flow fluctuations. Subsequently, these features were decomposed into frequency band components of multiple scales and interactively learned with the time-series data of the pulse network to obtain the optimal frequency-domain time-series weight ratio parameter for this environment. Finally, all frequency band components were fused according to their respective weights to form a high-quality anti-interference enhanced feature sequence. This process enables the researchers to accurately distinguish different species of fish and their swimming trajectories, maintaining high accuracy even in the shallow sea environment with frequent light changes.

[0099] An anti-interference enhanced feature sequence is generated by adaptively correcting the dynamic convolution kernel weight distribution through the attenuation coefficient, combining the frequency-domain response characteristics of the photonic crystal sensing and the time-series correlation characteristics of the pulse network. This method effectively suppresses the degradation effect caused by path scattering and medium inhomogeneity, improving the accuracy and robustness of underwater target detection, and is applicable to target recognition and positioning tasks in complex underwater environments.

[0100] To address the influence of the attenuation characteristics in the underwater light transmission path on the dynamic convolution kernel weight distribution and further improve the adaptive ability of the dynamic convolution kernel, in some embodiments, Step 201: The adaptive correction of the frequency-domain response characteristics of the dynamic convolution kernel weight distribution by the attenuation coefficient includes:

[0101] Step 301: Decompose the dynamic convolution kernel weight distribution into a high-frequency suppression kernel component and a low-frequency compensation kernel component according to the attenuation characteristics of the underwater light transmission path.

[0102] In step 301, the high-frequency suppression kernel component is for high-frequency signal components (such as spatial high-frequency noise caused by the scattering of suspended particles). The high-frequency noise is suppressed through the amplitude-frequency response characteristic of the dynamic convolution kernel, and the useful high-frequency information of the target is retained. The high-frequency suppression kernel component is for signal components with frequencies higher than a preset threshold (such as 10 Hz) and is used to suppress the spatial high-frequency noise caused by the scattering of suspended particles. The low-frequency compensation kernel component is for low-frequency signal components (such as low-frequency smear caused by turbulent perturbation). The low-frequency smear is compensated through the phase-frequency response characteristic of the dynamic convolution kernel, and the low-frequency structure information of the target is restored. The low-frequency compensation kernel component is for signal components with frequencies lower than a preset threshold (such as 10 Hz) and is used to compensate the low-frequency smear caused by turbulent perturbation.

[0103] In the embodiments of the present application, the weight distribution of the dynamic convolution kernel is decomposed in the frequency domain based on the attenuation characteristic. The high-frequency suppression kernel component is used to process the high-frequency noise caused by the scattering of suspended particles, and the low-frequency compensation kernel component is used to process the low-frequency smear caused by turbulent perturbation.

[0104] Step 302: Based on the attenuation characteristic, perform non-linear attenuation compensation on the amplitude-frequency response of the high-frequency suppression kernel component to generate a gain compensation coefficient.

[0105] In step 302, the amplitude-frequency response refers to the amplitude change characteristic of a system or filter at different frequencies, which describes the law of signal amplitude changing with frequency. For the high-frequency suppression kernel component, the amplitude-frequency response specifically refers to its suppression ability of the signal amplitude in the high-frequency band (such as higher than 10 Hz). Generation process: Based on the attenuation characteristic of the underwater light transmission path, a non-linear attenuation compensation function is designed to map the attenuation coefficient to the gain compensation coefficient in the high-frequency band, and the high-frequency signal components are suppressed through the frequency-domain response characteristic of the dynamic convolution kernel weight distribution to generate the amplitude-frequency response of the high-frequency suppression kernel component. The gain compensation coefficient is used to adjust the amplitude-frequency response of the high-frequency suppression kernel component to enhance the suppression ability of the high-frequency noise.

[0106] In the embodiments of the present application, first, the non-linear relationship between the light intensity and the distance under specific conditions is calculated, and then, based on this relationship, an adaptive gain control algorithm is used to compensate the amplitude-frequency response of the high-frequency suppression kernel component to determine the gain compensation coefficient.

[0107] Step 303: Based on the attenuation characteristic, perform time-varying correction on the phase-frequency response of the low-frequency compensation kernel component to obtain a phase compensation offset.

[0108] In step 303, the phase-frequency response refers to the phase change characteristics of a system or filter at different frequencies, which describes the law of signal phase change with frequency. Generation process: Based on the attenuation characteristics of the underwater optical transmission path, a time-varying attenuation correction function is designed, mapping the attenuation coefficient to the phase compensation offset in the low-frequency band, and compensating for the low-frequency signal components through the frequency-domain response characteristics of the dynamic convolution kernel weight distribution to generate the phase-frequency response of the low-frequency compensation kernel component. The phase compensation offset is used to adjust the phase-frequency response of the low-frequency compensation kernel component to eliminate the phase distortion caused by turbulence disturbances.

[0109] In the embodiment of the present application, by measuring the optical propagation time difference at different depths, a model is established, and this model is used to predict and correct the phase-frequency response of the low-frequency compensation kernel component, thereby obtaining the phase compensation offset.

[0110] Step 304: Fuse the gain compensation coefficient and the phase compensation offset to generate a corrected dynamic convolution kernel weight distribution.

[0111] In the embodiment of the present application, the gain compensation coefficient and the phase compensation offset are combined, and a weighted average method is adopted to generate the final corrected dynamic convolution kernel weight distribution.

[0112] The following is a specific example:

[0113] In a shallow sea coral reef monitoring project, researchers deployed a set of photonic crystal sensing array systems to track the activities of marine organisms near the coral reef in real time. First, the researchers decomposed the dynamic convolution kernel weight distribution into a high-frequency suppression kernel component and a low-frequency compensation kernel component by on-site measurement of water quality parameters such as suspended particle concentration and water transparency to separately handle noise interference and signal distortion problems. Then, using the non-linear attenuation compensation algorithm, the gain compensation coefficient was generated according to the change law of light intensity with distance and applied to the high-frequency suppression kernel component, reducing the influence of background noise. At the same time, by modeling the optical propagation time difference at different depths, the phase compensation offset was calculated and the phase-frequency response of the low-frequency compensation kernel component was corrected to ensure the time synchronization of the signal. Finally, the gain compensation coefficient and the phase compensation offset were fused to generate a corrected dynamic convolution kernel weight distribution, successfully improving the detection ability of the system for weak optical signals in the coral reef area. This technology enables the research team to more clearly identify the swimming trajectories of fish and the health status of corals, providing important data support for ecological protection.

[0114] By decomposing the dynamic convolution kernel weight distribution into a high-frequency suppression kernel component and a low-frequency compensation kernel component, and respectively performing amplitude-frequency response compensation and phase-frequency response correction based on the attenuation characteristics, a modified dynamic convolution kernel weight distribution is generated. This method improves the adaptive ability of the dynamic convolution kernel, effectively suppresses the degradation effects caused by suspended particle scattering and turbulent perturbations, and provides a high-quality dynamic convolution kernel weight distribution for underwater target detection.

[0115] In order to further improve the synergistic effect of the frequency-domain response characteristics and the temporal correlation characteristics, and optimize the accuracy and robustness of underwater target detection, in some embodiments, step 204: dynamically allocating the frequency-domain response characteristics of the photonic crystal sensing and the temporal correlation characteristics of the pulse network respectively according to the attenuation coefficient to generate a frequency-domain temporal weight ratio parameter, includes:

[0116] Step 401: Divide the multi-scale frequency band components into a high-frequency scattering-dominated frequency band and a low-frequency turbulence distortion-dominated frequency band according to the attenuation coefficient.

[0117] In step 401, the high-frequency scattering-dominated frequency band refers to the high-frequency noise caused by suspended particle scattering in the corresponding light field distribution data, usually manifested as rapidly changing amplitude fluctuations and spatial detail information. The frequency range of the high-frequency noise is usually related to the density and size distribution of the suspended particles. The low-frequency turbulence distortion-dominated frequency band refers to the low-frequency smear caused by turbulent perturbations in the corresponding light field distribution data, usually manifested as slowly changing phase distortions and amplitude fluctuations. The frequency range of the low-frequency smear is usually related to the intensity of the turbulent perturbations and the length of the light transmission path. Basis for high-low frequency division: The attenuation coefficient α is an important parameter of the underwater light transmission path, and its value range is [0, 1], where 0 represents no attenuation and 1 represents the maximum attenuation. According to the value of the attenuation coefficient, the multi-scale frequency band components can be divided into high-frequency and low-frequency: High-frequency scattering-dominated frequency band: The frequency band with a smaller attenuation coefficient α (such as α < 0.5), corresponding to the high-frequency noise caused by suspended particle scattering. Low-frequency turbulence distortion-dominated frequency band: The frequency band with a larger attenuation coefficient α (such as α ≥ 0.5), corresponding to the low-frequency smear caused by turbulent perturbations.

[0118] In the embodiments of the present application, the multi-scale frequency band components are divided into frequency bands based on the attenuation coefficient. The high-frequency scattering-dominated frequency band corresponds to the high-frequency noise caused by suspended particle scattering, and the low-frequency turbulence distortion-dominated frequency band corresponds to the low-frequency smear caused by turbulent perturbations.

[0119] Step 402: Extract the local detail features corresponding to the high-frequency scattering-dominated frequency band from the frequency-domain response characteristics of the photonic crystal sensing.

[0120] In step 402, the local detail features refer to the high-frequency detail information of the target in the high-frequency scattering-dominated frequency band, such as edges, textures, etc.

[0121] In the embodiments of the present application, local detail features in the high-frequency scattering dominant frequency band are extracted through frequency domain filtering technology, providing an input for subsequent non-linear gain compensation.

[0122] Step 403: Perform non-linear gain compensation on the local detail features based on the attenuation coefficient.

[0123] In step 403, non-linear gain compensation refers to adjusting the amplitude of the local detail features according to the attenuation coefficient to enhance the expression ability of high-frequency detail information.

[0124] In the embodiments of the present application, the attenuation coefficient is mapped to a gain compensation coefficient through a non-linear function to adjust the amplitude of the local detail features and enhance the expression ability of high-frequency detail information.

[0125] Step 404: Use a pre-trained multi-band attenuation compensation model to map the compensated local detail features to the weight assignment values of the frequency domain response characteristics.

[0126] In step 404, the multi-band attenuation compensation model refers to a frequency domain response characteristic optimization model based on the attenuation coefficient, which is used to generate the weight assignment values of the frequency domain response characteristics. The weight assignment values are used to adjust the contribution ratio of the frequency domain response characteristics in feature fusion.

[0127] In the embodiments of the present application, the compensated local detail features are input into the multi-band attenuation compensation model to generate the weight assignment values of the frequency domain response characteristics.

[0128] Step 405: Extract the motion trajectory features of the underwater target corresponding to the low-frequency turbulence distortion dominant frequency band from the temporal correlation characteristics of the pulse network.

[0129] In step 405, the motion trajectory features refer to the stability and consistency shown by the motion pattern of the target or scene in the time dimension in continuous multi-frame images or signals. Specifically, it describes the continuity and consistency of information such as the motion trajectory, speed, and direction of the target between different frames.

[0130] In the embodiments of the present application, the motion trajectory features in the low-frequency turbulence distortion dominant frequency band are extracted through temporal filtering technology, providing an input for subsequent phase distortion suppression and correction.

[0131] Step 406: Perform phase distortion suppression and correction on the motion trajectory features of the underwater target based on the attenuation coefficient.

[0132] In the embodiments of the present application, the attenuation coefficient is mapped to a phase compensation offset through a phase distortion suppression function to adjust the phase of the motion trajectory features and eliminate the phase distortion caused by turbulence disturbances.

[0133] Step 407: Based on the corrected motion trajectory features of the underwater target, calculate the weight assignment value of the temporal correlation characteristics through a preset time-varying attenuation correction function.

[0134] In step 407, the time-varying attenuation correction function refers to a temporal correlation characteristic optimization function based on the attenuation coefficient, which is used to generate the weight assignment value of the temporal correlation characteristics.

[0135] In the embodiment of the present application, the corrected motion trajectory features are input into the time-varying attenuation correction function to generate the weight assignment value of the temporal correlation characteristics.

[0136] Step 408: Integrate the weight assignment value of the frequency-domain response characteristics and the weight assignment value of the temporal correlation characteristics to generate a frequency-domain temporal weight ratio parameter.

[0137] In step 408, the frequency-domain temporal weight ratio parameter is used to control the contribution ratio of the photonic crystal sensing and the pulse network in the multi-scale frequency band components.

[0138] In the embodiment of the present application, the weight assignment value of the frequency-domain response characteristics and the weight assignment value of the temporal correlation characteristics are weighted and summed to generate a frequency-domain temporal weight ratio parameter, which is used to control the contribution ratio of the photonic crystal sensing and the pulse network in the multi-scale frequency band components.

[0139] The following is a specific example:

[0140] In a previous shallow coral reef monitoring project, the research team has successfully obtained the light field distribution data of underwater targets through a photonic crystal sensing array and generated a corrected dynamic convolution kernel weight distribution. To further improve the positioning and classification accuracy, the research team divided the multi-scale frequency band components into a high-frequency scattering-dominated frequency band and a low-frequency turbulence distortion-dominated frequency band according to the attenuation coefficient, corresponding to the effects of particle scattering and water flow movement on the optical signal in the coral reef area. Then, the wavelet transform was used to extract the local detail features of the high-frequency scattering-dominated frequency band from the wavelet decomposition results, and non-linear gain compensation was performed on them based on the measured attenuation coefficient on site. At the same time, they were mapped to the weight assignment value of the frequency-domain response characteristics through a pre-trained multi-band attenuation compensation model. Meanwhile, the motion trajectory features of underwater targets (such as fish) corresponding to the low-frequency turbulence distortion-dominated frequency band were extracted from the time series output of the pulse network, and phase distortion suppression correction was performed on them according to the attenuation coefficient. Subsequently, the weight assignment value of the temporal correlation characteristics was calculated through a preset time-varying attenuation correction function, and the two were integrated to generate a frequency-domain temporal weight ratio parameter. This process enables the system to control the contribution ratio of the photonic crystal sensing and the pulse network in the multi-scale frequency band components, ultimately improving the detection ability of weak optical signals and biological activity trajectories in the coral reef area and providing more comprehensive and reliable data support for ecological monitoring.

[0141] In the embodiments of the present application, by dividing the multi-scale frequency band components into a high-frequency scattering-dominated frequency band and a low-frequency turbulence distortion-dominated frequency band, and generating weight assignment values for the frequency-domain response characteristics and the time-series correlation characteristics respectively based on the attenuation coefficient, a frequency-domain time-series weight ratio parameter is finally generated through fusion. This method improves the synergistic effect between the frequency-domain response characteristics and the time-series correlation characteristics, optimizes the accuracy and robustness of underwater target detection, and is applicable to target recognition and positioning tasks in complex underwater environments.

[0142] In order to further improve the synergistic effect between the non-uniform scattering characteristics and the frequency-domain response characteristics of the photonic crystal sensing, in some embodiments, step 102: using the dual complementary dynamic convolution operator to perform dual complementary coupling on the non-uniform scattering characteristics and the frequency-domain response characteristics of the photonic crystal sensing to generate a dynamic convolution kernel weight distribution, including:

[0143] Step 501: Based on the spatial distribution characteristics of the non-uniform scattering characteristics and the frequency-domain response characteristics of the photonic crystal sensing, construct a dual complementary coupling model.

[0144] In step 501, the spatial distribution characteristics refer to the distribution law of the non-uniform scattering characteristics in space, manifested as high-frequency noise and low-frequency ghosting. The dual complementary coupling model is a synergistic model that combines the spatial distribution characteristics and the frequency-domain response characteristics and is used to generate a dynamic convolution kernel weight distribution.

[0145] In the embodiments of the present application, through the spatial frequency domain joint mapping function, a dual relationship is modeled between the spatial distribution characteristics and the frequency-domain response characteristics to generate a dual complementary coupling model.

[0146] Step 502: In the dual complementary coupling model, extract spatial high-frequency detail information and time-series perturbation information from the spatial distribution characteristics.

[0147] In step 502, the spatial high-frequency detail information refers to the rapidly changing detail features in the light field distribution data caused by the target reflection signal or environmental scattering, usually manifested as edges, textures, or fine structures in an image or signal. For example: if the sampling frequency of the light field distribution data is 100 Hz, the components in the spectrum with frequencies higher than 50 Hz can be defined as high-frequency detail information. The time-series perturbation information refers to the time-varying characteristics in the light field distribution data caused by target movement or environmental perturbations (such as turbulence, water flow, etc.), usually manifested as movement trajectories, phase changes, or amplitude fluctuations in a signal or image sequence. For example: assuming that the time sampling interval of the light field distribution data is 0.1 seconds, calculate the difference between adjacent frames, and the regions where the difference value is greater than a certain threshold (such as ΔI>10) can be defined as time-series perturbation information.

[0148] In the embodiments of the present application, spatial high-frequency detail information is extracted through frequency-domain filtering technology, and temporal perturbation information is extracted through temporal analysis technology, providing input for subsequent dual complementary fusion.

[0149] Step 503: Perform dual complementary fusion on the spatial high-frequency detail information and the temporal perturbation information through a dynamic weight allocation mechanism.

[0150] In step 503, the dynamic weight allocation mechanism refers to a method for dynamically adjusting the weights of spatial high-frequency detail information and temporal perturbation information according to environmental characteristics.

[0151] In the embodiments of the present application, the weight allocation ratio of spatial high-frequency detail information and temporal perturbation information is dynamically adjusted according to the attenuation coefficient of the underwater light transmission path to generate fused features.

[0152] Step 504: Input the dual complementary fusion result into a preset dynamic convolution kernel generation module to output the dynamic convolution kernel weight distribution. The preset dynamic convolution kernel generation module incorporates a dual complementary dynamic convolution operator.

[0153] In step 504, the dynamic convolution kernel generation module refers to a convolution kernel generation module based on a dual complementary dynamic convolution operator for generating the dynamic convolution kernel weight distribution.

[0154] In the embodiments of the present application, through the dual complementary dynamic convolution operator in the dynamic convolution kernel generation module, the fused features are mapped to the dynamic convolution kernel weight distribution for subsequent spatio-temporal degradation suppression.

[0155] The following is a specific example:

[0156] In a research project on a shallow-water coral reef ecosystem, the research team first established a dual complementary coupling model based on on-site measurement data to accurately describe the relationship between the spatial distribution characteristics of non-uniform scattering characteristics and the frequency-domain response characteristics of photonic crystal sensing. Then, the researchers used an edge detection algorithm to extract the spatial high-frequency detail information within the coral reef area from the wavelet transform results and used time series analysis technology to identify the temporal perturbation information caused by water flow movement. Then, through the designed dynamic weight allocation mechanism, the weight ratio of these two types of information was adaptively adjusted according to the real-time collected data, achieving optimal dual complementary fusion. Finally, the fusion result was input into a preset dynamic convolution kernel generation module to generate the dynamic convolution kernel weight distribution suitable for the local water quality conditions, improving the detection accuracy of the system for the coral reef and its surrounding biological activities and providing important technical support for ecological protection.

[0157] In the embodiments of the present application, a dual complementary coupling model is constructed to extract spatial high-frequency detail information and temporal perturbation information, and dual complementary fusion is performed through a dynamic weight allocation mechanism, and finally a dynamic convolution kernel weight distribution is generated. This method improves the synergistic effect between non-uniform scattering characteristics and the frequency-domain response characteristics of photonic crystal sensing. The generated adaptive dynamic convolution kernel weight distribution can effectively suppress the degradation effect in the underwater environment and provide high-quality feature input for underwater target detection.

[0158] In order to further improve the accuracy and robustness of underwater target positioning and classification, and eliminate the error of the underwater light transmission path during the positioning and classification process, in some embodiments, step 104: combining the anti-interference enhanced feature sequence to perform positioning and classification on the underwater target, and during the positioning and classification process, eliminating the error of the underwater light transmission path to generate a positioning and classification result of the underwater target, including:

[0159] Step 601: Calculate the spatial coordinates of the underwater target through the non-linear mapping relationship between the spatial distribution characteristics of the anti-interference enhanced feature sequence and the attenuation coefficient of the underwater light transmission path, and eliminate the spatial positioning deviation caused by the path scattering of the underwater light transmission path through iterative optimization during the calculation process.

[0160] In the embodiments of the present application, a spatial positioning model is constructed, the spatial distribution characteristics and the attenuation coefficient are input into the model, the spatial coordinates of the target are calculated through an iterative optimization algorithm, and the spatial positioning deviation caused by path scattering is eliminated during the optimization process. The specific optimization process: The non-linear least squares optimization algorithm is used to iteratively optimize the spatial positioning model, and the optimization goal is to minimize the residual between the artifact noise caused by path scattering and the target reflection signal. The iterative optimization conditions include: the initial parameter is set to the initial mapping value of the frequency-domain response characteristics of the photonic crystal sensing and the attenuation coefficient, and the iterative termination condition is that the residual is less than the preset threshold or the number of iterations reaches the maximum value. The algorithm parameters include a damping factor and a step size factor. The execution process is: first calculate the residual between the spatial distribution characteristics and the target reflection signal based on the initial parameters, then update the parameters through the non-linear least squares optimization algorithm, repeat calculating the residual and updating the parameters until the iterative termination condition is met, and finally output the spatial coordinates after eliminating path scattering.

[0161] Step 602: Calculate the class probability distribution of the underwater target through the non-linear mapping relationship between the temporal motion characteristics of the anti-interference enhanced feature sequence and the attenuation coefficient of the underwater light transmission path, and eliminate the classification error caused by the medium non-uniformity of the underwater light transmission path through iterative optimization during the calculation process.

[0162] In step 602, the temporal motion characteristic refers to the motion change law of the target reflection signal extracted from the anti-interference enhanced feature sequence in the time dimension, which is used to describe the dynamic behavior of the target in the underwater environment. The acquisition process includes: based on the anti-interference enhanced feature sequence, extracting the motion trajectory of the target through inter-frame difference analysis in the time dimension, and non-linearly correcting the motion trajectory in combination with the attenuation coefficient of the underwater light transmission path to eliminate the temporal distortion caused by turbulent perturbation and medium inhomogeneity, and finally generating the temporal motion characteristic that can accurately reflect the motion state of the target. The category probability distribution refers to the probability distribution of the target belonging to different categories and is used for classification tasks.

[0163] In the embodiment of the present application, a temporal classification model is constructed. The temporal motion characteristic and the attenuation coefficient are input into the model, and the category probability distribution of the target is calculated through an iterative optimization algorithm, and the classification error caused by medium inhomogeneity is eliminated during the optimization process. The specific optimization process: The gradient descent method is used as the iterative optimization algorithm, and the mean square error of the classification error is used as the optimization objective function. By calculating the gradient of the objective function with respect to the model parameters, the weight parameters in the temporal classification model are gradually adjusted until the error converges to a preset threshold or the maximum number of iterations is reached; the iterative optimization conditions include that the gradient norm of the objective function is less than the preset threshold (such as 1e-6) or the number of iterations reaches the maximum value (such as 1000 times), and the algorithm parameters include the learning rate (such as 0.01), the momentum factor (such as 0.9), and the weight decay coefficient (such as 1e-4); the execution process is: first, initialize the weight parameters of the temporal classification model, then calculate the mean square error between the category probability distribution output by the current model and the true label, then calculate the gradient of the error with respect to the model parameters and update the weight parameters, and finally repeat the above process until the iterative optimization conditions are met, and generate the category probability distribution after eliminating the classification error caused by medium inhomogeneity.

[0164] Step 603: Generate a positioning and classification result of the underwater target based on the spatial coordinates and the category probability distribution.

[0165] In the embodiment of the present application, the spatial coordinates and the category probability distribution are fused to generate the positioning information and classification result of the target, and the final positioning and classification result is output.

[0166] The following is a specific example:

[0167] In a research project in a shallow - water coral reef area, the research team first established a non - linear mapping model by using the anti - interference enhanced feature sequences obtained in the previous steps and combining with the attenuation coefficient of the underwater light transmission path. The researchers calculated the spatial coordinates of the target fish through an iterative optimization algorithm and successfully eliminated the positioning deviation caused by path scattering. Then, based on the temporal motion characteristics of these feature sequences, the researchers used a support vector machine classifier to determine the probability distribution of each potential target, and at the same time applied a medium non - uniformity correction model to reduce the classification error. Finally, by combining the optimized spatial coordinates with the category corresponding to the highest probability, an accurate positioning and classification result was obtained, which not only helped the researchers better understand the dynamic changes of the biological community around the coral reef, but also provided a scientific basis for protection and management.

[0168] Based on the spatial distribution characteristics and temporal motion characteristics of the anti - interference enhanced feature sequences, combined with the non - linear mapping relationship of the attenuation coefficient, calculate the spatial coordinates and category probability distribution of underwater targets, and eliminate the errors caused by path scattering and medium non - uniformity during the optimization process. This method improves the accuracy and robustness of underwater target positioning and classification and is applicable to target recognition and positioning tasks in complex underwater environments.

[0169] In order to further improve the adaptive ability and robustness of the underwater target detection system and form a closed - loop suppression mechanism, in some embodiments, after step 104: combining the anti - interference enhanced feature sequences to perform positioning and classification on the underwater target, and eliminating the error of the underwater light transmission path during the positioning and classification process to generate a positioning and classification result of the underwater target, the method further includes:

[0170] Step 701: Perform collaborative feedback on the frequency - domain response characteristics and the dynamic convolution kernel weight distribution according to the positioning and classification result and the attenuation coefficient.

[0171] In the embodiments of the present application, the target positioning and classification result and the attenuation coefficient are input into the feedback control module, and adjustment parameters of the frequency - domain response characteristics and the dynamic convolution kernel weight distribution are generated through a non - linear mapping function to optimize the system performance.

[0172] Step 702: Adjust the weight ratio parameter between the frequency - domain response characteristics and the temporal correlation characteristics according to the result of the collaborative feedback to generate optimized control parameters for suppressing complex degradation scenarios across domains. The optimized control parameters are used to correct the capture and processing process of the light - field distribution data to form a closed - loop suppression mechanism.

[0173] In step 702, cross-domain suppression of complex degradation scenarios refers to, during underwater target detection, realizing the collaborative suppression of various degradation factors in a complex underwater environment by simultaneously dealing with the degradation effects of the frequency-domain response characteristics in the spatial domain and the temporal correlation characteristics in the temporal domain. The optimization control parameters are used to correct the capture and processing process of the light field distribution data.

[0174] In the embodiments of this application, the collaborative feedback result is mapped into the weight ratio parameters of the frequency-domain response characteristics and the temporal correlation characteristics through a weight allocation function to generate optimization control parameters, which are used to correct the capture and processing process of the light field distribution data, forming a closed-loop suppression mechanism.

[0175] The following is a specific example:

[0176] In a shallow sea coral reef ecological monitoring project, after the research team completed underwater target positioning and classification, they conducted a collaborative feedback analysis on the frequency-domain response characteristics and the dynamic convolution kernel weight distribution according to the obtained positioning and classification results of the underwater target and the attenuation coefficient. The researchers found that under certain specific conditions, signal distortion caused by an increase in water flow velocity affected the detection accuracy. Therefore, the team used a reinforcement learning algorithm to re-evaluate the existing model and proposed targeted improvement measures. Then, the researchers adopted an adaptive adjustment algorithm to dynamically adjust the weight ratio parameters of the frequency-domain response characteristics and the temporal correlation characteristics according to the latest environmental conditions, generating optimization control parameters. This not only improved the anti-interference ability of the system but also enhanced the capture efficiency of weak signals. Finally, through this closed-loop suppression mechanism, the research team achieved high-precision and continuous monitoring of the coral reef and the biological activities around it, providing reliable data support for ecological protection.

[0177] By performing collaborative feedback on the frequency-domain response characteristics and the dynamic convolution kernel weight distribution according to the positioning and classification results of the underwater target and the attenuation coefficient, generating optimization control parameters, and correcting the capture and processing process of the light field distribution data, a closed-loop suppression mechanism is formed. This method improves the adaptive ability and robustness of the underwater target detection system and is applicable to target recognition and positioning tasks in complex underwater environments.

[0178] Figure 2 The following is a schematic structural diagram of an underwater target detection system based on dual complementary dynamic convolution provided by the embodiments of this application, as Figure 2 shown, the system includes:

[0179] An acquisition module 21, configured to acquire the light field distribution data of the underwater target through a photonic crystal sensing array, where the light field distribution data includes non-uniform scattering characteristics and spatial frequency domain distortion characteristics.

[0180] The coupling module 22 is used to perform dual-complementary coupling on the non-uniform scattering features and the frequency-domain response characteristics of the photonic crystal sensing by using the dual-complementary dynamic convolution operator, generate a dynamic convolution kernel weight distribution, and adaptively match the attenuation characteristics of the underwater optical transmission path during the generation process of the dynamic convolution kernel weight distribution.

[0181] The generation module 23 is used to perform spatio-temporal degradation suppression on the spatial frequency-domain distortion features by using the dynamic convolution kernel weight distribution, and generate an anti-interference enhanced feature sequence in combination with the timing correlation characteristics of the pulse network.

[0182] The positioning module 24 is used to perform positioning and classification on the underwater target in combination with the anti-interference enhanced feature sequence, and eliminate the error of the underwater optical transmission path during the positioning and classification process, and generate a positioning and classification result of the underwater target. Figure 2 The underwater target detection system based on dual-complementary dynamic convolution described above can execute Figure 1 The underwater target detection method based on dual-complementary dynamic convolution described in the above embodiments, the implementation principle and technical effects will not be elaborated. For the underwater target detection system based on dual-complementary dynamic convolution in the above embodiments, the specific ways for each module and unit to perform operations have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0183] In a possible design, Figure 2 The underwater target detection system based on dual-complementary dynamic convolution described in the above embodiments can be implemented as a computing device, such as Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32;

[0184] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 32.

[0185] The processing component 32 above Figure 1 The underwater target detection method based on dual-complementary dynamic convolution described in the above embodiments.

[0186] Among them, the processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above methods. Of course, the processing component may also be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components for executing the above methods.

[0187] The storage component 31 is configured to store various types of data to support the operation of the terminal. The storage component can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disc.

[0188] Of course, the computing device may also necessarily include other components, such as input / output interfaces, display components, communication components, etc.

[0189] The input / output interface provides an interface between the processing component and the peripheral interface module, and the above peripheral interface module may be an output device, an input device, etc.

[0190] The communication component is configured to facilitate wired or wireless communication between the computing device and other devices, etc.

[0191] Among them, the computing device may be a physical device or an elastic computing host provided by a cloud computing platform, etc. At this time, the computing device may refer to a cloud server, and the above processing component, storage component, etc. may be basic server resources leased or purchased from a cloud computing platform.

[0192] An embodiment of the present application also provides a computer storage medium storing a computer program, which when executed by a computer can implement the above Figure 1 underwater target detection method based on dual complementary dynamic convolution shown in the embodiment.

[0193] Those skilled in the art can clearly understand that for the convenience and conciseness of description, the specific working processes of the above-described systems, devices, and units can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0194] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative effort.

[0195] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0196] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not intended to limit them. Although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments or perform equivalent replacements for some of the technical features. These modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An underwater target detection method based on dual complementary dynamic convolution, characterized in that, Comprising: Obtaining the optical field distribution data of an underwater target through a photonic crystal sensing array, where the optical field distribution data includes non-uniform scattering characteristics and spatial frequency domain distortion characteristics; Using a dual complementary dynamic convolution operator to perform dual complementary coupling on the non-uniform scattering characteristics and the frequency domain response characteristics of photonic crystal sensing, generating a dynamic convolution kernel weight distribution. During the generation process of the dynamic convolution kernel weight distribution, adaptively matching the attenuation characteristics of the underwater optical transmission path; Using the dynamic convolution kernel weight distribution to perform spatio-temporal degradation suppression on the spatial frequency domain distortion characteristics, and combining the temporal correlation characteristics of the pulse network to generate an anti-interference enhanced feature sequence; Combining the anti-interference enhanced feature sequence to perform positioning and classification on the underwater target, and during the positioning and classification process, eliminating the error of the underwater optical transmission path to generate a positioning and classification result of the underwater target; The using the dynamic convolution kernel weight distribution to perform spatio-temporal degradation suppression on the spatial frequency domain distortion characteristics, and combining the temporal correlation characteristics of the pulse network to generate an anti-interference enhanced feature sequence includes: Adapting and correcting the frequency domain response characteristics of the dynamic convolution kernel weight distribution through an attenuation coefficient, where the attenuation coefficient is a subset of the attenuation characteristics; Based on the corrected dynamic convolution kernel weight distribution, performing spatio-temporal degradation suppression on the spatial frequency domain distortion characteristics, and the corrected dynamic convolution kernel weight distribution forms a dual complementary constraint with the frequency domain response characteristics of the photonic crystal sensing; Decomposing the spatially and temporally degraded suppressed spatial frequency domain distortion characteristics into multi-scale frequency band components; Each scale of the frequency band components respectively performs cross-domain interaction with the temporal correlation characteristics of the pulse network. During the cross-domain interaction process, according to the attenuation coefficient, dynamically allocating the frequency domain response characteristics of the photonic crystal sensing and the temporal correlation characteristics of the pulse network respectively to generate a frequency domain temporal weight ratio parameter; Based on the frequency domain temporal weight ratio parameter, performing weighted fusion on the multi-scale frequency band components to generate an anti-interference enhanced feature sequence; The using a dual complementary dynamic convolution operator to perform dual complementary coupling on the non-uniform scattering characteristics and the frequency domain response characteristics of photonic crystal sensing, generating a dynamic convolution kernel weight distribution includes: Constructing a dual complementary coupling model based on the spatial distribution characteristics of the non-uniform scattering characteristics and the frequency domain response characteristics of photonic crystal sensing; In the dual complementary coupling model, extracting spatial high-frequency detail information and temporal perturbation information from the spatial distribution characteristics; Performing dual complementary fusion on the spatial high-frequency detail information and the temporal perturbation information through a dynamic weight allocation mechanism; Inputting the dual complementary fusion result into a preset dynamic convolution kernel generation module, and outputting a dynamic convolution kernel weight distribution, where the preset dynamic convolution kernel generation module introduces a dual complementary dynamic convolution operator.

2. The method according to claim 1, wherein The adapting and correcting the frequency domain response characteristics of the dynamic convolution kernel weight distribution through an attenuation coefficient includes: Decomposing the dynamic convolution kernel weight distribution into a high-frequency suppression kernel component and a low-frequency compensation kernel component according to the attenuation characteristics of the underwater optical transmission path; Based on the attenuation characteristics, perform non-linear attenuation compensation on the amplitude-frequency response of the high-frequency suppression kernel component to generate a gain compensation coefficient; Based on the attenuation characteristics, perform time-varying correction on the phase-frequency response of the low-frequency compensation kernel component to obtain a phase compensation offset; Fuse the gain compensation coefficient and the phase compensation offset to generate a corrected dynamic convolution kernel weight distribution.

3. The method according to claim 1, characterized in that The dynamic allocation of the frequency-domain response characteristics of the photonic crystal sensing and the timing correlation characteristics of the pulse network according to the attenuation coefficient to generate a frequency-domain timing weight ratio parameter includes: Divide the multi-scale frequency band components into a high-frequency scattering dominant frequency band and a low-frequency turbulence distortion dominant frequency band according to the attenuation coefficient; Extract local detail features corresponding to the high-frequency scattering dominant frequency band from the frequency-domain response characteristics of the photonic crystal sensing; Perform non-linear gain compensation on the local detail features based on the attenuation coefficient; Use a pre-trained multi-band attenuation compensation model to map the compensated local detail features to the weight allocation values of the frequency-domain response characteristics; Extract the motion trajectory features of the underwater target corresponding to the low-frequency turbulence distortion dominant frequency band from the timing correlation characteristics of the pulse network; Perform phase distortion suppression and correction on the motion trajectory features of the underwater target based on the attenuation coefficient; Based on the corrected motion trajectory features of the underwater target, calculate the weight allocation value of the timing correlation characteristics through a preset time-varying attenuation correction function; Fuse the weight allocation value of the frequency-domain response characteristics and the weight allocation value of the timing correlation characteristics to generate a frequency-domain timing weight ratio parameter, and the frequency-domain timing weight ratio parameter is used to control the contribution ratio of the photonic crystal sensing and the pulse network in the multi-scale frequency band components.

4. The method according to claim 1, wherein The method of combining the anti-interference enhanced feature sequence to locate and classify the underwater target, and eliminating the error of the underwater light transmission path during the location and classification process to generate a location and classification result of the underwater target includes: Calculate the spatial coordinates of the underwater target through the non-linear mapping relationship between the spatial distribution characteristics of the anti-interference enhanced feature sequence and the attenuation coefficient of the underwater light transmission path, and eliminate the spatial positioning deviation caused by path scattering of the underwater light transmission path through iterative optimization during the calculation process; Calculate the category probability distribution of the underwater target through the non-linear mapping relationship between the timing motion characteristics of the anti-interference enhanced feature sequence and the attenuation coefficient of the underwater light transmission path, and eliminate the classification error caused by the medium inhomogeneity of the underwater light transmission path through iterative optimization during the calculation process; Generate a location and classification result of the underwater target based on the spatial coordinates and the category probability distribution.

5. The method according to claim 1, wherein After the method of combining the anti-interference enhanced feature sequence to locate and classify the underwater target, and eliminating the error of the underwater light transmission path during the location and classification process to generate a location and classification result of the underwater target, the method further includes: Perform collaborative feedback on the frequency-domain response characteristics and the dynamic convolution kernel weight distribution according to the location and classification result and the attenuation coefficient; Adjust the weight ratio parameter of the frequency-domain response characteristic and the timing correlation characteristic according to the result of collaborative feedback to generate an optimized control parameter for suppressing complex degradation scenarios across domains. The optimized control parameter is used to correct the acquisition and processing process of the optical field distribution data to form a closed-loop suppression mechanism.

6. An underwater target detection system based on dual complementary dynamic convolution, characterized in that, Including: An acquisition module, configured to acquire the optical field distribution data of an underwater target through a photonic crystal sensing array. The optical field distribution data includes non-uniform scattering characteristics and spatial frequency-domain distortion characteristics; A coupling module, configured to perform dual complementary coupling on the non-uniform scattering characteristic and the frequency-domain response characteristic of photonic crystal sensing by using a dual complementary dynamic convolution operator to generate a dynamic convolution kernel weight distribution. During the generation process of the dynamic convolution kernel weight distribution, adaptively match the attenuation characteristic of the underwater optical transmission path; A generation module, configured to perform spatio-temporal degradation suppression on the spatial frequency-domain distortion characteristic by using the dynamic convolution kernel weight distribution, and combine the timing correlation characteristic of a pulse network to generate an anti-interference enhanced feature sequence; A positioning module, configured to combine the anti-interference enhanced feature sequence to perform positioning and classification on the underwater target, and eliminate the error of the underwater optical transmission path during the positioning and classification process to generate a positioning and classification result of the underwater target; The performing spatio-temporal degradation suppression on the spatial frequency-domain distortion characteristic by using the dynamic convolution kernel weight distribution, and combining the timing correlation characteristic of a pulse network to generate an anti-interference enhanced feature sequence includes: Adaptive correction of the frequency-domain response characteristic of the dynamic convolution kernel weight distribution through an attenuation coefficient, where the attenuation coefficient is a subset of the attenuation characteristic; Based on the corrected dynamic convolution kernel weight distribution, perform spatio-temporal degradation suppression on the spatial frequency-domain distortion characteristic. The corrected dynamic convolution kernel weight distribution and the frequency-domain response characteristic of the photonic crystal sensing form a dual complementary constraint; Decompose the spatially and temporally degraded spatial frequency-domain distortion characteristic into multi-scale frequency band components; Each scale of the frequency band component respectively performs cross-domain interaction with the timing correlation characteristic of the pulse network. During the cross-domain interaction process, according to the attenuation coefficient, dynamically allocate the frequency-domain response characteristic of the photonic crystal sensing and the timing correlation characteristic of the pulse network respectively to generate a frequency-domain timing weight ratio parameter; Based on the frequency-domain timing weight ratio parameter, perform weighted fusion on the multi-scale frequency band components to generate an anti-interference enhanced feature sequence; The performing dual complementary coupling on the non-uniform scattering characteristic and the frequency-domain response characteristic of photonic crystal sensing by using a dual complementary dynamic convolution operator to generate a dynamic convolution kernel weight distribution includes: Construct a dual complementary coupling model based on the spatial distribution characteristic of the non-uniform scattering characteristic and the frequency-domain response characteristic of photonic crystal sensing; In the dual complementary coupling model, extract spatial high-frequency detail information and timing perturbation information from the spatial distribution characteristic; Perform dual complementary fusion on the spatial high-frequency detail information and the timing perturbation information through a dynamic weight allocation mechanism; Input the dual complementary fusion result into a preset dynamic convolution kernel generation module to output the dynamic convolution kernel weight distribution. The preset dynamic convolution kernel generation module incorporates a dual complementary dynamic convolution operator.

7. A computing device, characterized in that, It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement an underwater target detection method based on dual complementary dynamic convolution as described in any one of claims 1 to 5.

8. A computer storage medium, characterized in that, A computer program is stored, and when the computer program is executed by a computer, it implements an underwater target detection method based on dual complementary dynamic convolution as described in any one of claims 1 to 5.

Images