An underwater visible light signal recovery method and communication system based on transfer learning

Abstract

The application relates to an underwater visible light signal recovery method based on transfer learning, aiming at the problems of signal noise, distortion and communication performance decline of an underwater visible light communication (UVLC) system caused by turbidity, flow rate, environmental light intensity and other disturbances of a water body. In combination of transfer learning and a generative adversarial network (GAN), data is collected through a self-developed hardware platform, background data is generated through MATLAB simulation, conditional generative adversarial network (cGAN) is used to fuse simulation and real image features to generate pseudo-real data, a U-NetGAN model is used to train signal recovery capability, and finally, the model is deployed on a receiving end. Compared with related technologies, the scheme overcomes the poor adaptability of traditional modulation under the disturbance of water turbidity, flow rate and the like, and the difficulty in dealing with complex underwater environments; and solves the problems of deep learning methods, such as dependence on large-scale data and weak cross-scene generalization. Through the combination of transfer learning and GAN, the communication performance of the UVLC system is significantly optimized, the delay is low, the complexity is low, and the scheme is more suitable for actual deployment requirements.

Description

Technical Field

[0001] This application relates to the field of underwater visible light communication, and in particular to an underwater visible light signal recovery method and communication system based on transfer learning, as well as a computer device and a computer-readable storage medium. Background Technology

[0002] Underwater wireless communication is widely used due to its advantages such as low cost and high flexibility. Its main technological carriers are sound waves, radio frequency (RF) waves, and light waves. While underwater acoustic communication can achieve long-distance transmission of up to 15km, it suffers from high latency and low data rate, limiting it to simple responses. RF communication offers high data rates, but its electromagnetic wave attenuation in water is severe and increases exponentially with distance, restricting it to short-range communication. Visible light frequencies attenuate relatively less underwater, making underwater optical wireless communication (UOWC) a popular choice due to its high bandwidth, low power consumption, and resistance to electromagnetic interference. Among these, underwater visible light communication (UVLC) is the most widely used because of its minimal attenuation and mature device technology.

[0003] In underwater visible light communication (UVLC) systems, environmental disturbances such as turbidity, current velocity, and light intensity variations can trigger absorption, scattering, and turbulence effects, leading to signal attenuation and structural distortion. This, in turn, causes recognition errors, increases the bit error rate, exacerbates system energy consumption, and reduces efficiency. Existing receivers mostly use photodiodes (PDs), which, while supporting high speeds, only output one-dimensional electrical signals, resulting in limited signal recovery capabilities at low signal-to-noise ratios. Although CMOS image sensors have limited speeds, they can acquire two-dimensional image data containing spatial and textural features, thus providing conditions for interference suppression and signal recovery. However, the original image is also affected by noise and distortion, and without effective processing, the bit error rate remains difficult to reduce.

[0004] While traditional UVLC technology can counteract interference by optimizing modulation methods (such as PPM and PAM) or improving hardware, it requires complex equipment and precise algorithms, resulting in high costs, difficult maintenance, and poor environmental adaptability. In recent years, deep learning methods have been introduced, which can reduce hardware dependence and enhance system robustness, but they rely on large-scale training data, and the complex underwater environment makes data coverage difficult and the model's generalization ability insufficient.

[0005] Therefore, a new method is urgently needed to solve the problem of poor generalization ability of underwater visible light communication methods in related technologies. Summary of the Invention

[0006] This application provides a method, system, computer device, and computer-readable storage medium for recovering underwater visible light signals based on transfer learning, in order to at least solve the problem of poor generalization ability of underwater visible light communication methods in related technologies.

[0007] In a first aspect, embodiments of this application provide a method for recovering underwater visible light signals based on transfer learning, the method comprising:

[0008] Based on real images from underwater visible light communication scenarios, and combined with preset simulation parameters, data simulation is performed to generate a simulated image similar to the real image.

[0009] Through the data augmentation module and style transfer mechanism, the deep structural features of the real image are mapped to the shallow background features of the simulated image to generate realistic images covering a variety of complex environments, thus obtaining an extended dataset.

[0010] Furthermore, the performance recovery network is trained using the extended dataset as the disturbed input and the undisturbed standard reference image as the supervision target to obtain an image performance recovery model for denoising and restoring the structure of the disturbed input.

[0011] A trained image performance recovery model is deployed at the receiving end to optimize the interfered communication images captured in underwater visible light scenes and output the recovered images.

[0012] In some embodiments, data simulation is performed in conjunction with preset simulation parameters, including:

[0013] By using a two-dimensional Gaussian function, the intensity distribution of the main aperture and the secondary aperture of the real image is generated to simulate the real image. An exponential decay model is used to generate a halo structure that simulates the real image caused by the scattering effect of water outside the aperture, thus obtaining the aperture layer of the simulated image.

[0014] By modeling with a periodic function, the real image is simulated, and modulation stripes formed by the camera sampling the OOK signal light intensity are generated to obtain the simulated image stripe layer. The contrast of the modulation stripes is determined according to the total attenuation coefficient corresponding to the turbidity of the water and the propagation distance.

[0015] Based on the signal strength of the real image and the signal-to-noise ratio calculated with the ambient background noise, a simulated image occlusion layer is generated to simulate the occlusion mask of the real image. The range and transparency gradient of the occlusion mask are derived in reverse based on the signal-to-noise ratio to simulate the degree of damage to the imaging quality caused by the ambient noise in the real image.

[0016] The simulated image is obtained by superimposing the aperture layer, the stripe layer, and the occlusion layer of the simulated image onto a solid color background layer.

[0017] In some embodiments, a generator maps the deep structural features of the real image to the shallow background features of the simulated image to obtain the realistic image, specifically including:

[0018] The shallow background features are obtained by extracting style features at each level of the simulated image through the first encoder.

[0019] The deep structural features are obtained by extracting the structural features of each level of the real image through the second encoder.

[0020] Obtain the parameter labels of the underwater environment corresponding to the real image, and input them into a multilayer perceptron for multilayer fully connected mapping to obtain a high-dimensional label vector, wherein the parameter labels include turbidity, flow velocity and illuminance;

[0021] The deep structural features, the shallow background features, and the high-dimensional label vector are concatenated to obtain a fusion tensor.

[0022] Using a decoder, based on the fusion tensor and guided by the high-dimensional label vector, the deep structural features of the real image are mapped to the shallow background features of the simulated image to obtain the simulated image.

[0023] In some embodiments, mapping the deep structural features of the real image to the shallow background features of the simulated image based on the fusion tensor includes:

[0024] The encoder employs deconvolution layers to perform stepwise upsampling based on the fused tensor.

[0025] During the progressive upsampling process, based on the adaptive instance normalization skip connection mechanism, the deep structural features and shallow style features are connected by AdaIN skip connection at each level, and the shallow style features are applied to the deep structural features to output the realistic image.

[0026] In some embodiments, after generating the realistic image, the method further includes:

[0027] The high-dimensional label vector is linearly mapped to the spatial feature map of the simulated image to obtain a conditional map with the same spatial size as the simulated image. The simulated image and the conditional map are then concatenated to obtain a concatenated tensor.

[0028] The discriminator performs multi-layer convolutional downsampling based on the spliced ​​tensor, outputting a real-fake discrimination score representing the difference between the simulated image and the real image, as well as an output feature matching loss.

[0029] Based on the discrimination score and multiple loss functions, the generator and discriminator of the data augmentation module are jointly optimized for parameters, wherein...

[0030] The overall loss of the discriminator is the binary cross-entropy loss, which is used to constrain the discriminator to maximize the probability of identifying a real image as real and minimize the probability of identifying a simulated image as real.

[0031] The loss function of the generator includes at least:

[0032] Adversarial loss used to constrain the realistic image to improve the discrimination score in the discriminator;

[0033] Structural loss used to ensure that the simulated image maintains consistency with the simulated image at the pixel level;

[0034] The perceptual loss used to improve the semantic level matching of the simulated image with the high-level features of the real image;

[0035] Style loss used to ensure that the simulated image remains consistent with the real image in terms of texture and frequency features;

[0036] Feature matching loss is used to enhance the intermediate semantic consistency between the generator and the discriminator.

[0037] In some embodiments, the performance recovery network is built based on a U-Net neural network. Training the performance recovery network using the extended dataset as disturbed input and an undisturbed standard reference image as the supervised target includes:

[0038] The simulated image is processed by a performance recovery encoder to extract multi-level features of the simulated image;

[0039] The performance recovery decoder upsamples the multi-level features step by step and fuses the features of the corresponding level in the performance recovery encoder to reconstruct the performance recovery image after denoising and structural restoration.

[0040] The performance recovery discriminator divides the input performance recovery image into multiple local regions, and after performing local discrimination in conjunction with the interference-free standard reference image, outputs a discrimination score that indicates that each local region conforms to the real image.

[0041] Furthermore, based on the discrimination score and the composite loss function consisting of adversarial loss, pixel loss, structural similarity loss, perceptual loss and edge loss, the parameters of the performance recovery decoder and the performance recovery encoder are jointly optimized until the network model converges.

[0042] In some embodiments, processing the simulated image to extract multi-level features of the simulated image includes:

[0043] Multi-scale convolutional structures are used to capture local and global features at different scales;

[0044] The local features and the global features are added together to obtain the fused features;

[0045] The fused features are spatially modeled using a multi-scale filter, and the channel weights of the spatial modeling result are dynamically adjusted using a channel attention mechanism to obtain a weighted feature map.

[0046] The weighted feature map is downsampled by max pooling to output the multi-level features, and the multi-level features are then connected to the decoder via skip connections.

[0047] In some embodiments, after the trained image performance restoration model is deployed at the receiving end, the method further includes:

[0048] Preprocessing of the interfered communication images captured in real time generates input tensors that are normalized in size and color;

[0049] The input tensor is sent to the trained image performance restoration model, and the encoder and decoder of the image performance restoration model are used to restore the input tensor to obtain the compensated image;

[0050] The compensated image is superimposed pixel-by-pixel with the input tensor to obtain the restored image;

[0051] The striped regions in the restored image are grayscaled and projected, and threshold judgment and bit extraction are performed according to preset encoding rules to demodulate the original bit sequence.

[0052] Secondly, embodiments of this application provide an underwater visible light communication system based on transfer learning, the system comprising a transmitter and a receiver, wherein:

[0053] The transmitter is used to emit visible light signals by modulating a light source, and after encoding data in the visible light signals, transmit them to an underwater channel environment;

[0054] The receiving end is used to receive the visible light signal through a CMOS camera and process the visible light signal using the underwater visible light signal recovery method of the first aspect to obtain the recovered image.

[0055] In some embodiments, the transmitter includes a transmitter module, an encoding module, and a driver module, wherein:

[0056] The transmitting module is used to output the original bit sequence to be modulated, wherein the original bit sequence contains the data information to be transmitted;

[0057] The encoding module is used to expand each bit in the original bit sequence into two state bit changes using Manchester encoding to enhance bit boundary clarity and eliminate DC offset, and to convert the Manchester-encoded sequence into a drive signal using OOK modulation.

[0058] The driving module is used to generate a light intensity signal according to the driving signal and send the light intensity signal to the underwater visible light communication environment.

[0059] Thirdly, embodiments of this application provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described in the first aspect above.

[0060] Fourthly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect above.

[0061] Compared to related technologies, this application provides an underwater visible light signal recovery method based on transfer learning, addressing the problems of signal noise, distortion, and communication performance degradation in underwater visible light communication (UVLC) systems caused by disturbances such as water turbidity, current velocity, and ambient light intensity. Background data is generated using MATLAB simulation, and simulated data is generated by fusing simulation and real image features through a conditional generative adversarial network (cGAN). The signal recovery capability is then trained using a U-NetGAN model and finally deployed at the receiver. This method overcomes the shortcomings of traditional modulation methods (poor adaptability) and existing deep learning models (reliant on large amounts of data and weak generalization ability). The trained model can accurately eliminate interference such as scattering and noise in the underwater environment, effectively restoring key structural features of the communication signal (such as modulation fringes and light source morphology), providing high-quality image input for subsequent bit sequence demodulation, and significantly improving the signal recovery capability and communication reliability of underwater visible light communication systems in complex environments. Attached Figure Description

[0062] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0063] Figure 1 This is a flowchart of an underwater visible light signal recovery method based on transfer learning according to an embodiment of this application;

[0064] Figure 2 It is MATALB simulation image data under interference-free conditions according to the embodiments of this application;

[0065] Figure 3 This is a structural diagram of the cGAN model according to an embodiment of this application;

[0066] Figure 4 This is a schematic diagram illustrating the principle of training a U-Net-based neural network according to an embodiment of this application.

[0067] Figure 5 This is a schematic diagram of the communication physical link of the recovery method according to an embodiment of this application;

[0068] Figure 6a This is a schematic diagram of the auxiliary and control circuit area of ​​the transmitter circuit in an embodiment of this application;

[0069] Figure 6b This is a schematic diagram of the interface and power management area of ​​the transmitter circuit according to an embodiment of this application;

[0070] Figure 6c This is a schematic diagram of the core functional circuit area of ​​the transmitter circuit in an embodiment of this application;

[0071] Figure 7 This is a schematic diagram of a laboratory platform according to an embodiment of this application;

[0072] Figure 8 This is a schematic diagram comparing the signal-to-noise ratio according to an embodiment of this application;

[0073] Figure 9 This is a schematic diagram comparing bit error rates according to an embodiment of this application;

[0074] Figure 10 This is a schematic diagram of the overall performance curve according to an embodiment of this application;

[0075] Figure 11 This is a schematic diagram illustrating the impact of data volume on overall performance according to embodiments of this application;

[0076] Figure 12 This is a schematic diagram illustrating the impact of the data expansion module on performance improvement according to an embodiment of this application;

[0077] Figure 13 This is a schematic diagram of the internal structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0078] To make the objectives, technical solutions, and advantages of this application clearer, the application is described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application.

[0079] Obviously, the accompanying drawings described below are merely some examples or embodiments of this application. Those skilled in the art can apply this application to other similar scenarios based on these drawings without any inventive effort. Furthermore, it is understood that although the efforts made in this development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, any changes to design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as insufficient disclosure of the content of this application.

[0080] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0081] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms “a,” “an,” “an,” “the,” and similar words used in this application do not indicate quantity limitation and may indicate singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. The terms “connected,” “linked,” “coupled,” and similar words used in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following objects are in an "or" relationship. The terms "first," "second," and "third" used in this application are merely to distinguish similar objects and do not represent a specific ordering of the objects.

[0082] Compared to wired communication, underwater wireless communication has been widely used and developed due to its advantages such as low cost, high flexibility, long distance, and convenient deployment. Since the visible light band in the electromagnetic wave band has relatively low attenuation underwater, underwater visible light communication (UVLC) is the most widely used due to its minimal attenuation and mature devices.

[0083] In practical applications, UVLC systems face environmental disturbances, especially in complex and variable underwater environments, where communication performance is severely affected. In harsh underwater environments, the communication effect of UVLC systems becomes very poor.

[0084] Specifically, changes in water turbidity, flow velocity, and ambient light intensity can all affect the optical link through absorption, scattering, and turbulence effects, leading to increased signal noise, structural distortion, and consequently, increased communication identification errors, higher bit error rates, increased system energy consumption, and decreased operating efficiency.

[0085] To address this issue, related technologies utilize UVLC, which improves hardware and modulation methods. However, these technologies still have shortcomings. On the one hand, they require complex hardware and precise modulation and demodulation algorithms, increasing system cost and maintenance difficulty. On the other hand, these methods are poorly adaptable to environmental changes and cannot effectively cope with the ever-changing underwater environment.

[0086] In recent years, with the rise of deep learning algorithms, various communication optimization schemes based on deep learning algorithms have been proposed to address the environmental interference problem in underwater optical communication (UVLC). However, these schemes generally suffer from limitations such as limited generalization ability or reliance on large-scale training data. Specifically, these include:

[0087] Defoaming and Correction Scheme in Strong Foam Environments: This scheme combines a defoaming algorithm with binary stripe correction technology to achieve robust transmission in strong foam interference scenarios, with a transmission rate of up to 7.2 Kbit / s, which is 8.6 times higher than traditional schemes. However, the application scenarios of this scheme are highly limited, and it is only effective in foam environments. It performs poorly in other types of interference (such as turbulence and turbidity changes), and its generalization ability is significantly lacking.

[0088] A stripe brightness sequence processing scheme based on LSTM (Long Short-Term Memory) networks: This scheme utilizes the "temporal memory" characteristic of LSTM networks to analyze the stripe brightness sequence in image frames, thereby reconstructing the 0 / 1 bit stream. At a transmission rate of 2.7 kbps, this scheme can control the bit error rate (BER) to ≤3.8 × 10⁻⁶. -3 Within the range, this solution only works for specific pre-trained environments, and its performance drops significantly when faced with untrained, interfering scenarios.

[0089] A unified model optimization scheme considering turbulence and alignment errors is proposed: To improve the model's generalization ability, a unified Gamma-Gamma model considering turbulence and alignment errors is proposed, and a multi-antenna (SIMO) architecture and spatial diversity techniques are introduced at the receiver. This method significantly reduces the system's bit error rate (BER) and outage probability. Meanwhile, Monte Carlo simulations verify the robustness of this scheme to signals with varying environmental flow velocity and turbidity. However, the scheme's performance is highly dependent on large-scale data training, resulting in high data acquisition costs.

[0090] A Deep Unfolding Network Scheme Based on Compressed Sensing (CS): This paper proposes a framework based on compressed sensing (CS) that combines a deep unfolding network with the Approximate Message Passing (AMP) algorithm to achieve sparse estimation of UVLC channels. This scheme maintains high estimation accuracy even in scenarios with significant multipath interference. However, its core drawback, similar to the aforementioned Gamma-Gamma model scheme, is that it requires a large amount of training data, limiting its applicability in data-scarce scenarios.

[0091] In summary, these deep learning-based algorithms rely on massive amounts of data for training. However, the complexity of the underwater environment makes it extremely difficult to obtain data covering all real-world environments, resulting in poor generalization ability and insufficient robustness of the models. Even with ample data support, existing deep learning models still exhibit poor generalization ability when faced with new underwater environments. A key technical challenge in this field is that while deep learning algorithms have addressed the hardware requirements and other shortcomings of traditional methods, they still cannot overcome the inherent weaknesses of poor generalization and adaptability. These issues limit the widespread application of UVLC systems, necessitating a new approach to address these challenges.

[0092] In view of this, this application provides an underwater visible light signal recovery method based on transfer learning. It utilizes a style transfer mechanism to fuse the feature differences between simulated and measured images, and combines a conditional generative adversarial network to model disturbance environmental parameters (turbidity, current velocity, ambient light intensity). This generates high-fidelity simulated communication images under various complex environmental combinations with only a small number of measured images, thus effectively expanding the training data. This reduces the reliance on large-scale real data. The generative model constructed in this application generates a large amount of simulated data with the assistance of a small number of labeled samples, replacing the high-cost and difficult field collection process. This effectively reduces the dependence of deep learning communication models on data volume, and high-performance training can still be achieved under conditions of sparse data or missing samples. Furthermore, this application significantly expands the model's training domain by simulating disturbance combination samples not appearing in the training set, enhancing its adaptability to unknown underwater environments and improving the robustness and generalization ability of the communication image recovery model in actual deployment.

[0093] Figure 1 This is a flowchart of an underwater visible light signal recovery method based on transfer learning according to an embodiment of this application, such as... Figure 1 As shown, the process includes the following steps:

[0094] S101 generates a simulated image similar to the real image by combining real images in underwater visible light communication scenarios with preset simulation parameters.

[0095] This step aims to address the limitations of limited real-world data acquisition and insufficient coverage of diverse environments in underwater visible light communication scenarios. By referencing the visual characteristics of real images (such as light source shape, stripe distribution, and noise interference patterns) and combining them with preset multi-dimensional environmental parameters (such as water turbidity, propagation distance, signal strength, and background noise level), physical simulation and feature synthesis are performed. This ultimately generates simulated images that highly match real images in visual style and signal characteristics. These simulated images can be combined with a small number of real images to provide a rich and generalizable data source for model training in the subsequent performance recovery module, improving the model's adaptability to complex underwater environments.

[0096] In this embodiment, underwater visible light communication data is simulated using MATLAB. The simulation parameters are expanded to cover most of the changing scenarios in real water environments, specifically including:

[0097] S1, through a two-dimensional Gaussian function, generates the intensity distribution of the main aperture and the secondary aperture of the simulated real image, and uses an exponential decay model to generate the halo structure of the simulated real image produced by the scattering effect of water outside the aperture, thus obtaining the aperture layer of the simulated image;

[0098] Among them, the two-dimensional Gaussian function is used to simulate the core light spot characteristics formed by direct underwater light source (such as LED): the main aperture corresponds to the high intensity area of ​​direct light source, and its intensity distribution conforms to the Gaussian curve (the center is the brightest and it decreases smoothly towards the edge); the secondary aperture simulates the secondary high intensity area formed by the light source after a small amount of scattering by water particles, and forms a nested structure with the main aperture to match the multi-layer brightness distribution of the light source in the real image.

[0099] Furthermore, when light propagates underwater, some photons are scattered by water molecules and suspended particles, resulting in a blurred halo around the light source (an effect of superimposed non-direct light). By using an exponential function attenuation model (the attenuation coefficient is positively correlated with the turbidity of the water), the characteristic of the halo intensity rapidly decreasing with increasing distance from the center of the light source is simulated, making the blurred area around the aperture layer consistent with the real scattering effect and enhancing the physical realism of the simulation.

[0100] S2, by modeling a periodic function, simulates a real image and generates modulation stripes formed by the camera sampling the intensity of the OOK (On-Off Keying) signal light, thus obtaining a simulated image stripe layer. The contrast of the modulation stripes is determined based on the total attenuation coefficient corresponding to the turbidity of the water and the propagation distance.

[0101] Periodic functions (such as square waves and sine waves) are used to reconstruct the time-domain characteristics of the OOK modulated signal. In this system, the transmitter performs OOK modulation at a frequency of 5.682 kHz, with a bit period Tb ≈ 176 μs. Bit information is transmitted by switching the light intensity between "on" and "off". The CMOS camera uses a line-by-line scanning method. When sampling this periodic light intensity change at a certain line readout frequency, the time information is mapped onto the image line direction, forming alternating bright and dark stripes with a line spacing that matches the bit period Tb. By modeling with periodic functions, the density and period of the generated stripes can be made consistent with the actual transmission modulation frequency, thereby reconstructing the modulation characteristics of the signal.

[0102] Furthermore, the total attenuation coefficient is jointly determined by the water turbidity (affecting the degree of scattering / absorption) and the propagation distance (affecting light intensity loss). The larger the coefficient, the more significant the light intensity attenuation, and the weaker the difference between the "brightness" and "dullness" of the signal. Therefore, the contrast of the stripe layer (the brightness difference between bright and dark stripes) decreases as the total attenuation coefficient increases, accurately simulating the impact of signal intensity attenuation on stripe clarity in the real environment, making the simulated stripes closer to actual communication scenarios.

[0103] S3 generates a simulated image occlusion layer based on the signal-to-noise ratio calculated from the signal strength and ambient background noise. The range and transparency gradient of the occlusion mask are derived in reverse based on the signal-to-noise ratio to simulate the damage of ambient noise to image quality.

[0104] The signal-to-noise ratio (SNR) is the ratio of signal strength to ambient background noise, which directly reflects the degree of interference to the communication signal: when the SNR is high, the signal is less affected by noise; when the SNR is low, noise (such as suspended particles blocking the image or background stray light) has a more significant impact on the image.

[0105] In this embodiment, the generation logic of the occlusion mask is linked to the signal-to-noise ratio (SNR). The occlusion mask is a local interference area (such as randomly distributed blurred patches or dark spots) covered on the stripe layer, used to simulate the occlusion or blurring of signal stripes by noise. The following is derived in reverse using SNR: the lower the SNR, the higher the noise ratio, the larger the occlusion mask area (covering more stripe areas), and the steeper the transparency gradient (the transition between the mask area and the normal area is more abrupt), simulating the effect of the signal being "severed" and "blurred" under strong noise; the higher the SNR, the smaller the mask area and the gentler the transparency gradient, corresponding to weak noise scenarios, making the interference level of the occlusion layer consistent with the real noise level.

[0106] S4, the simulated image aperture layer, simulated image stripe layer and simulated image occlusion layer are superimposed on the solid color background layer to obtain the simulated image.

[0107] A solid color background layer serves as the base background color, providing a unified background benchmark for each feature layer and avoiding visual conflicts when features are superimposed. This solid color background can be, but is not limited to, black.

[0108] Specifically, the overlay process employs pixel-level fusion: the aperture layer (light source spot) is located at the bottom layer, providing a spatial reference for the light source's position in the image; the stripe layer (modulation signal) is superimposed on the aperture layer, reflecting the signal's modulation information; and the occlusion layer (noise interference) covers the top layer, simulating the damage noise causes to the signal. The overlay weights of each layer are adjusted according to the visual proportion of each feature in the real image. The final generated image integrates the core features of the light source, signal, and noise, and is highly similar to real underwater communication images in terms of visual appearance and feature distribution.

[0109] Preferably, the present application uses the MATLAB simulation tool to generate highly similar simulation data. Alternatively, the simulation data can also be generated using the following alternative methods:

[0110] Python simulation tools: Simulation tools in Python (such as NumPy, SciPy, etc.) can be used to replace MATLAB for data generation. Python has wide applications in data processing and machine learning, and can provide a more flexible simulation environment.

[0111] Commercial simulation software: Commercial simulation software (such as ANSYS, COMSOL, etc.) can be used to replace MATLAB for data generation. These software programs have higher accuracy in simulating complex physical phenomena and can generate data that is closer to the real environment.

[0112] In an exemplary embodiment, taking MATLAB simulation as an example, data simulation specifically includes:

[0113] Setting simulation parameters: The parameters of the control experimental platform in this embodiment are shown in Table 1 below, which is the simulation parameter data table:

[0114] Table 1 Simulation Parameter Data Table

[0115]

[0116]

[0117] Furthermore, through MATLAB simulation, signal images that are highly similar to real images are generated. Although these generated images are structurally identical to real images, they do not contain any relevant knowledge at the signal level. In other words, these simulation data provide a rich "background" for transfer learning.

[0118] In this embodiment, underwater visible light communication images are synthesized layer by layer using a four-layer structure. Figure 2These are MATALB simulation image data under interference-free conditions according to embodiments of this application; such as Figure 2 As shown, the layers are: Layer 1 (pure black background layer), Layer 2 (aperture layer), Layer 3 (modulation stripe layer), and Layer 4 (blue masking layer).

[0119] Figure 2 The simulation image data is shown under conditions without interference from turbidity, flow velocity, or ambient light intensity. The last three layers are all generated independently by physical modeling and dynamically adjusted according to the simulation environment parameters to ensure that the image has continuous, controllable, and physically consistent interference characteristics.

[0120] Layer 2 (aperture layer) in this embodiment simulates the intensity distribution of the main aperture and its four symmetrical sub-apertures using a two-dimensional Gaussian function. The brightness of the main aperture is determined by the emission power. LED light effect Glass transmittance Divergence angle loss and system receiving efficiency The decision was made jointly, and the formula for calculating the center brightness is as follows:

[0121] (1)

[0122] The intensity distribution of the main aperture is as follows:

[0123] (2)

[0124] in, Control the radius of the main aperture. The secondary aperture is set to 0.4 times the brightness of the main aperture and positioned near it to simulate the refraction effect caused by water waves and glass refraction. The coordinate offset distance is... Its distribution is as follows:

[0125] (3)

[0126] A halo structure is superimposed around the main and secondary apertures to simulate the edge-emitting diffusion of LEDs in highly scattering water, using exponential decay modeling.

[0127] (4)

[0128] in, The initial brightness of the halo. The extent of halo expansion is controlled by the water scattering coefficient. Influence:

[0129] (5)

[0130] As water turbidity (Turb), depth h, and LED power increase, the halo range expands and the edges become more blurred, resulting in a diffuse structure that is strong at the center and weak at the edges.

[0131] Layer 3 (modulation stripes) is used to simulate the light intensity sampling texture of the OOK signal modulated by a CMOS camera. This layer has a periodic alternating bright and dark modulation structure, and the modeling function is:

[0132] (6)

[0133] in, The modulation period length, For pixel scanning speed, For LED modulation frequency, The modulation contrast factor is affected by the total attenuation coefficient. and transmission distance Influence:

[0134] (7)

[0135] (8)

[0136] (9)

[0137] Therefore, as turbidity increases and distance increases, the contrast of the stripes decreases significantly, ultimately resulting in a gradual transition of the stripes from sharp and clear to blurred and faded in layer 3.

[0138] Layer 4 (occlusion layer) is generated by deducing the SNR from physical calculations, rather than directly setting an image threshold or blur mask. Its core is derived from noise-driven signal degradation; the background noise is modeled as follows:

[0139] (10)

[0140] The total noise is:

[0141] (11)

[0142] The signal current is:

[0143] (12)

[0144] (13)

[0145] The SNR expression is:

[0146] (14)

[0147] Based on this SNR value, the occlusion range and transparency gradient are derived in reverse, specifically using the following mapping relationship:

[0148] (15)

[0149] (16)

[0150] When SNR < 2dB, the occlusion area will cover the entire main aperture and even the secondary aperture area, making the stripes completely invisible, with only weak texture remaining at the edges. This layer does not preset any occlusion pattern, but rather uses a mask growth process driven by a physical model, realistically reflecting the degree of environmental noise's impact on image quality.

[0151] S102, through the data augmentation module, based on the style transfer mechanism, maps the deep structural features of real images to the shallow background features of simulated images, generating realistic images covering a variety of complex environments, and obtaining an extended dataset;

[0152] In this embodiment, this step aims to address the limitations of the amount of real-world data collected and the insufficient coverage of environmental diversity in underwater visible light communication scenarios. By referencing the visual characteristics of real images (such as light source shape, stripe distribution, and noise interference patterns) and combining them with preset multi-dimensional environmental parameters, physical modeling and feature synthesis are performed. This ultimately generates simulated images that highly match the real images in terms of visual style and signal characteristics. These numerous simulated images can be fused with a small number of real images, providing a rich and generalizable data source for model training in the subsequent performance recovery module, thereby improving the model's adaptability to complex underwater environments.

[0153] It should be noted that the data augmentation process in this application can use conditional generative adversarial networks (CGAN) for data augmentation. Alternatively, variational autoencoders (VAEs) and autoencoders can also be used to achieve data augmentation.

[0154] Specifically:

[0155] Variational Autoencoders (VAEs): VAEs can be used as an alternative to CGANs for data augmentation. By learning the latent representation of data, VAEs can generate realistic data that resembles real images. While VAEs may not be as diverse as GANs in terms of the variety of data they generate, they can effectively address the problem of insufficient data volume in certain situations.

[0156] Autoencoders: Data augmentation can be performed using autoencoders. An autoencoder compresses input data into a low-dimensional representation using an encoder, and then reconstructs the data using a decoder. While the data generated by an autoencoder may not be as realistic as that from a GAN or VAE, it can still provide effective data augmentation in certain application scenarios.

[0157] In this embodiment, the data augmentation module maps the deep structural features of the real image to the shallow background features of the simulated image, specifically including the following steps:

[0158] S1, the style features of each level of the simulation image are extracted through the first encoder to obtain the shallow background features;

[0159] S2, extract the structural features of each level of the real image through the second encoder to obtain the deep structural features;

[0160] S3, obtain the parameter labels of the underwater environment corresponding to the real image, and input them into the multilayer perceptron for multilayer fully connected mapping to obtain a high-dimensional label vector;

[0161] First, for each real image, the corresponding underwater environmental parameters at the time of capture were collected, namely turbidity, current velocity, and illumination. These parameters directly affect the image quality and are a quantitative representation of the underwater environmental characteristics.

[0162] Furthermore, these parameter labels are fed as input into a multilayer perceptron (MLP). An MLP consists of multiple fully connected layers, where neurons in each fully connected layer are connected to all neurons in the previous layer. A linear transformation is performed through a weight matrix, and non-linearity is introduced through an activation function. During this process, the MLP extracts and combines features from the input environmental parameters layer by layer, mapping the original low-dimensional environmental parameter vector into a high-dimensional label vector. This high-dimensional label vector contains complex relationships and potential features between different environmental parameters. For example, a certain combination of turbidity and illumination may map to a specific location in high-dimensional space, reflecting a specific stylistic tendency that the image may have under that environment, providing environment-related semantic information for subsequent image feature fusion.

[0163] S4 concatenates the deep structural features, shallow background features, and high-dimensional label feature vectors to obtain a fusion tensor.

[0164] This step concatenates the shallow background feature vector of the simulated image obtained in S1, the deep structural feature vector of the real image obtained in S2, and the high-dimensional label feature vector generated in S3, according to their dimensions. This concatenation method integrates information from different sources into a unified data structure, enabling the subsequent decoder to simultaneously utilize the structural information of the simulated image, the stylistic information of the real image, and the corresponding environmental semantic information.

[0165] For example, in the fusion tensor, the structural features retain the geometric features of the modulation stripes, the style features carry color shift information under specific turbidity, and the environmental label feature vector suggests the flow rate of the environment in which the image is located, providing a comprehensive data foundation for generating highly realistic images.

[0166] S5, through the decoder, based on the fusion tensor and guided by the high-dimensional label vector, maps the deep structural features of the real image to the shallow background features of the simulated image, thus obtaining a realistic image.

[0167] This step specifically includes the following sub-steps:

[0168] S5.1, based on fusion tensors, maps the deep structural features of the real image to the shallow background features of the simulated image to generate a realistic image, including:

[0169] S5.2, through the encoder, uses deconvolutional layers to perform step-by-step upsampling based on fusion tensors; and, during the step-by-step upsampling process, utilizes the adaptive instance normalization skip connection mechanism, by performing AdaIN (Adaptive Instance Normalization) skip connections between deep structural features and shallow style features at each level, applying shallow style features to deep structural features, and thus outputting a realistic image.

[0170] It is understood that the decoder in this embodiment adopts a network structure composed of deconvolution, upsampling, and fusion modules. Taking the fusion tensor obtained in step S4 as input, the fusion tensor is first deconvolved to gradually restore the spatial resolution of the image, converting high-dimensional vector information into feature maps with spatial structure. In this process, the structural feature information in the fusion tensor is used to guide the reconstruction of the image structure, ensuring the accuracy of key signal structures such as modulation stripes; furthermore, the style of the image is adjusted using style features and environmental label feature vectors, such as adjusting the image's sharpness, brightness, and color style based on the features corresponding to turbidity and illumination.

[0171] For example, if the environmental label features in the fused tensor imply high turbidity, the decoder will add blurring and color attenuation when generating the image. Simultaneously, the fusion module optimizes and combines features from different levels to ultimately generate a realistic image with stylistic features of a real image while retaining structural information of the simulated image, completing the conversion from fused features to a realistic image and expanding the training dataset.

[0172] In addition, after obtaining the simulated image, step S6 is included to optimize the model parameters to improve the realism and environmental adaptability of the simulated image. Specifically, the "realism" and "environmental adaptability" of the simulated image are evaluated by a discriminator, and the model parameters are adjusted in reverse by combining a multi-dimensional loss function. This ultimately improves the consistency between the simulated image and the real image (including pixel structure, semantic features, style texture, etc.), while ensuring that the simulated image can accurately reflect the characteristics of the corresponding underwater environmental parameters (turbidity, flow velocity, illuminance), making the extended dataset closer to the real communication scenario.

[0173] Specifically, step S6 includes the following sub-steps:

[0174] S6.1, linearly map the high-dimensional label vector to the spatial feature map of the simulated image to obtain a conditional map with the same spatial size as the simulated image. Then, concatenate the simulated image and the conditional map to obtain the concatenated tensor.

[0175] Specifically, the high-dimensional label vector containing environmental parameters such as turbidity and flow velocity is transformed into a conditional map of the same spatial size as the simulated image (each position corresponds to the intensity of environmental influence) through linear mapping, and then stitched with the simulated image along the channels. This step allows the discriminator to simultaneously consider the visual features of the image and the rationality of environmental parameters, avoiding the generation of simulated images that contradict the environment.

[0176] S6.2, through the discriminator, multi-layer convolutional downsampling is performed based on the concatenated tensor to output the real and fake discrimination score to characterize the real and fake images, as well as the feature matching loss;

[0177] In this embodiment, the discriminator concatenates tensors through multi-layer convolutional downsampling: shallow layers extract details such as texture, deep layers aggregate semantic features, and finally output the "real probability score" of the simulated image (the closer to 1, the more like the real image). At the same time, it calculates the difference between the simulated image and the real image in the intermediate layer features (feature matching loss), and evaluates the consistency between the two from the overall to the intermediate layer semantics.

[0178] S6.3, based on the discriminant score and multiple loss functions, performs joint parameter optimization on the generator and discriminator of the data augmentation module, whereby...

[0179] The overall loss of the discriminator is the binary cross-entropy loss, which is used to constrain the data augmentation module to maximize the probability of the real image being identified as real, and to minimize the probability of the simulated image being identified as real.

[0180] The generator's loss function includes at least: adversarial loss to guide the simulated image to obtain a high discrimination score in the discriminator; structural loss to ensure that the simulated image maintains consistency with the simulated image in terms of pixel-level structure; perceptual loss to improve the simulated image's ability to match high-level features of the real image at the semantic level; style loss to ensure that the simulated image maintains consistency with the real image in terms of texture and frequency features; and feature matching loss to enhance the intermediate semantic consistency between the generator and the discriminator.

[0181] Through the above joint optimization, the generator produces realistic images that not only retain the signal structure of the simulated images but also approximate real images in terms of vision, semantics, and environmental adaptability. The discriminator continuously improves its discrimination ability, forming a closed loop of "generation-discrimination-optimization," and finally outputs a high-quality extended dataset, providing a reliable training foundation for the performance recovery module.

[0182] Specifically, in an exemplary embodiment, style transfer is achieved through a Conditional Generative Adversarial Network (cGAN). The aim is to map the deep signal features of a small number of real images (real water data changes rapidly, and the data cannot fully cover all areas) to the shallow background features of simulated "background" data, generating a realistic image. This process, through transfer learning, addresses the challenge of data sparsity in real water environments. The deep signal structural features are high-level features of the signal ontology, including communication-related information such as modulation stripes, texture structure, and frequency domain contours. These features are derived from the higher semantic levels of the global structure captured in later layers of the model. The shallow background features refer to the features extracted from earlier layers of the model, representing stylistic information such as brightness, color, and background texture, expressing the background appearance and image style.

[0183] Figure 3 This is a structural diagram of the cGAN model according to an embodiment of this application, such as... Figure 3 As shown, the GAN model in this application utilizes a ResNet50 encoder and a multi-loss collaborative training mechanism to enable the generator to learn to transfer the structural content of simulated images to the distribution of the target's real style. The entire algorithm flow of the multi-input conditional generative adversarial network can be divided into the following main steps:

[0184] First, in the data preparation stage, let the simulated image be X. sim The real-style image is X style The environmental parameter labels are l=[t,v,e] (representing turbidity, flow rate, and illuminance respectively) as a three-dimensional vector. The environmental parameter labels are used to supplement the image style transfer process. The overall training samples are organized into triples. .

[0185] At the input end, the simulation and style images are respectively processed by a ResNet50 encoder to obtain multi-layer feature representations. , Each layer has the following features:

[0186] (17)

[0187] These characteristics belong to:

[0188] , , , (18)

[0189] These correspond to abstract representations at different resolutions. The shallow features of the signal are represented as follows:

[0190] (19)

[0191] The deep features are:

[0192]

[0193] Environmental parameter labels The input is a multi-layer fully connected mapping (two layers of ReLU + reshape), first through...

[0194] (20)

[0195] Then from:

[0196] ,(twenty one)

[0197] The output feature vector is transformed layer by layer into a high-dimensional vector; where... , For learnable parameters, This is the ReLU activation function. The final output is... It is a latent vector that is reshaped and tiled to encode a feature map size-matching tensor. This allows it to participate in the integration in the spatial dimension.

[0198] Simulated image structural features and real image style features The splicing forms a fused tensor:

[0199] ,(twenty two)

[0200] The fused tensor enters the decoder, undergoes progressive upsampling through deconvolutional layers (transposed convolutions), and at each level, an AdaIN skip connection is made with the simulated features and style features from the encoder to apply the real style to the simulated structure. Specifically, each skip connection is as follows:

[0201] ,(twenty three)

[0202] The AdaIN operator is:

[0203] (twenty four)

[0204] in, These are the mean and standard deviation in the feature map space, respectively, and S is the feature map of the simulated image. , representing "shallow background features", and c is the feature map of the real image. This represents "deep structural content," which integrates image generation driven by realistic style, communication structure, and physical conditions, and finally decodes to obtain the output generated image.

[0205] (25)

[0206] The image output by the generator After activation by Sigmoid, the range of pixel values ​​is constrained to [0,1] to ensure that the image has the numerical distribution of a natural image and is aligned with the dynamic range of a real image.

[0207] The image is then input into the decision processor, which uses a conditional PatchGAN (Patch Generative Adversarial Network) architecture for spatial local perception and achieves finer-grained texture constraints through local discrimination.

[0208] First, regarding the parameter labels Extend the linear mapping to a spatial feature map, and perform channel dimensionality upscaling and spatial broadcasting:

[0209] (26)

[0210] A conditional map with the same spatial size as the input image is obtained, which is used to introduce guidance information on physical parameters in the spatial dimension. The conditional map is then concatenated with the input image along the channel dimension to form the input tensor.

[0211] (27)

[0212] The discriminator accepts the stitched image as input, passes it through multiple convolutional downsampling modules (N convolutional modules in PatchGAN), and outputs a spatial mapping map. The true / false discrimination score represents the spatial distribution of the image. At the same time, the intermediate features of each layer are preserved. Used for feature matching loss, and for subsequent feature matching losses. use.

[0213] The decision maker also accepts realistic style images. With tags The input, the overall decisionr loss is in the form of binary cross-entropy (BCE):

[0214] (28)

[0215] The goal is to maximize the probability of identifying a real image as real and minimize the probability of identifying a generated image as real, thereby increasing the difficulty of identifying the generated image.

[0216] The generator loss is a combination of multiple factors. The first is the adversarial loss, which guides the generated image to obtain a true score in the discriminator, thus approximating the image distribution.

[0217] (29)

[0218] in, To generate an image, The input environment parameter labels are used, and BCE is the binary cross-entropy. The more difficult it is for the discriminator to distinguish between different images, the better the image generation quality. To ensure the structural accuracy of the generated image, pixel-level L1 and L2 losses are introduced.

[0219] L1 structural loss:

[0220] (30)

[0221] L2 loss:

[0222] (31)

[0223] and The input simulated image is compared with another enhanced simulated image to provide supervision of image structure. L1 emphasizes edge consistency, while L2 emphasizes overall smoothness.

[0224] Furthermore, considering that signal features are concentrated in the central region, a central mask loss is introduced to pass through the central region. Multiplication and weighting emphasize accurate reproduction of the main ring stripe area:

[0225] (32)

[0226] The generated graph is semantically matched with features of real images, and a perceptual loss for VGG19 feature extraction is introduced.

[0227] (33)

[0228] in, Indicates VGG19 j The layer feature map output is used to measure the consistency of the high-level perception structure.

[0229] To ensure consistency in image style (texture distribution, frequency features, etc.), a style loss is introduced, described using a Gram matrix:

[0230] (34)

[0231] (35)

[0232] Meanwhile, to enhance the intermediate semantic consistency with the discriminator, a feature matching loss is introduced:

[0233] (36)

[0234] in, This represents the intermediate feature output of the i-th layer of the discriminator. This term improves semantic consistency stability and reduces training oscillations.

[0235] Because the generated images are generally dark, the brightness can easily become unbalanced during training, causing the model to crash. Therefore, a brightness constraint is introduced:

[0236] (37)

[0237] β is a preset light intensity threshold to suppress deviations such as overexposure or underexposure in the image. Furthermore, to suppress a large number of overexposed (close to 1) or underexposed (close to 0) pixels in the generated image, an extreme pixel penalty loss term is designed:

[0238] (38)

[0239] Used to filter out extreme cases and prevent gradient explosion or image saturation during training.

[0240] Finally, the weights of all losses The generator is dynamically adjusted based on the moving average, and its optimization objective is... :

[0241] (39)

[0242] The discriminator is optimized to Similarly, in each iteration, training first uses simulation / style / label forward generation. The total loss is calculated sequentially, and the parameters of the generator and decision maker are optimized through joint backpropagation. Statistical metrics include SSIM (structural similarity) and PSNR (peak signal-to-noise ratio). Samples and weights are saved periodically. Finally, during the generation phase... , Same group:

[0243] (40)

[0244] This embodiment applies conditions to arbitrary simulation images and style images, batch-generates style transfer images, and saves them, completing an end-to-end parametric image generation process from multi-dimensional conditional simulation to the real style domain. It provides a large number of realistic datasets for model training, solving the problem that MATLAB simulation images have "images" but lack communication features.

[0245] In addition, it should be noted that in this embodiment, transfer learning is used to transfer an existing model trained in a similar environment to a new underwater environment, reducing the need for a large amount of new environmental data; those skilled in the art will understand that the following optional schemes can also be used to generate an extended dataset.

[0246] Domain Adaptation: Domain adaptation can be used as an alternative to transfer learning. It reduces the difference between the source and target domains by adjusting the model to fit the new environment. This approach may be more effective than traditional transfer learning in some cases, especially when there is a significant difference between the source and target domains.

[0247] Meta-learning: Meta-learning methods can be used as an alternative to transfer learning. By learning how to learn, meta-learning enables models to quickly adapt to new tasks and environments. This method is particularly effective when data is limited, allowing for rapid model adjustments to suit new underwater environments.

[0248] S103, using the extended dataset as the disturbed input and the undisturbed standard reference image as the supervision target, the performance recovery network is trained to obtain an image performance recovery model for denoising and restoring the structure of the disturbed input.

[0249] The core of this step is to train the performance recovery network to have the ability to recover a clear and complete communication signal from the interfered signal through supervised learning of simulated images (simulated interference scenarios) and interference-free images (ideal signal images).

[0250] Specifically, through the co-constraint of the encoder-decoder structure and discriminator of U-Net, the network can simultaneously learn denoising (eliminating interference such as water scattering and noise) and structure restoration (repairing the effects of water absorption and scattering on the periodicity of stripes and the structure of the light source after the signal passes through the underwater channel), ultimately adapting to the complex interference scenarios in real underwater visible light communication.

[0251] This step specifically includes the following steps:

[0252] S1 processes the simulated image through a performance recovery encoder to extract multi-level features of the simulated image;

[0253] S1.1 employs a multi-scale convolutional structure to capture local and global features at different scales;

[0254] The performance recovery encoder employs the same convolutional downsampling structure as the U-Net encoder, consisting of multiple convolutional layers (such as 3×3 convolutions), activation functions (ReLU), and max pooling layers (or stride convolutions). Its core function is to perform "shallow-to-deep" feature extraction on the simulated image (a signal image containing interference).

[0255] Shallow background features (first 2-3 layers): capture low-level details of the image, such as the edges of modulation stripes, the outline of the light source aperture, and the distribution of local noise (corresponding to the basic visual features of signals in underwater communication).

[0256] Mid-layer features (layers 4-6): Aggregate shallow background features and extract structural features such as the periodicity of stripes and the spatial correlation between aperture and stripes (e.g., the correspondence between the position of the light source and the distribution of stripes);

[0257] Deep features (7 layers): abstracted into high-order semantic features, such as "overall blurred pattern of stripes affected by high turbidity interference" and "local signal distortion features caused by turbulence", reflecting the global relationship between interference and signal structure.

[0258] This approach solves the problem that single-scale convolution cannot simultaneously capture "details" and "global awareness," ensuring that the encoder can comprehensively extract signals and interference features at different scales in the simulated image.

[0259] S1.2, add the local features and global features to obtain the fused features;

[0260] It is understandable that by adding the local features extracted by multi-scale convolution to the global features element-wise along the same dimension, the local features are "embedded" into the spatial framework of the global features. The fused features contain both the fine structure of the signal and reflect the overall distribution pattern, providing a more comprehensive feature foundation for subsequent optimization.

[0261] S1.3, spatial modeling of the fused features is performed using a multi-scale filter, and the channel weights of the spatial modeling results are dynamically adjusted using a channel attention mechanism to obtain a weighted feature map;

[0262] Optionally, filters with different receptive fields (such as pooling or convolution based on different sizes) can be used to model the spatial relationships of the fused features, capturing the dependencies of features in different spatial ranges (such as the positional association between a stripe and adjacent stripes, and the spatial attenuation relationship between the light source region and the surrounding halo). This step enhances the spatial contextual information of the features and avoids viewing local features in isolation (such as mistakenly identifying stripe breaks as noise).

[0263] Furthermore, SEBlock (channel attention module) is introduced to perform channel-level optimization on the features after spatial modeling; first, global average pooling is performed on the feature map to compress spatial information and obtain the global descriptor for each channel;

[0264] Furthermore, the importance weights of the channels (weight values ​​0-1) are learned through the fully connected layer. Channels that are critical to signal recovery (such as channels carrying stripe periodic information) are given high weights, while redundant channels that are sensitive to interference (such as noise-dominated channels) are given low weights. Finally, the weights are multiplied by the original feature map channel by channel to obtain a weighted feature map.

[0265] Understandably, this step allows the encoder to "focus" on the feature channels that are more important to the recovery task, suppress interfering channels, and improve the discriminative power of the features.

[0266] S1.4, the optimized feature map is downsampled by max pooling to obtain multi-level features, and the multi-level features are then connected to the decoder via skip connections.

[0267] In this embodiment, max pooling (e.g., 2×2 pooling with a stride of 2) is performed on the weighted feature map through max pooling downsampling. This reduces the spatial resolution (e.g., 224×224 → 112×112) while preserving the strongest response features of each local region (e.g., the maximum gradient value of the stripe edges), generating the feature map for that level. The above process (multi-scale convolution → fusion → spatial modeling → attention weighting → downsampling) is repeated to obtain multi-level features from shallow to deep (resolution is halved layer by layer, and the number of channels is increased layer by layer).

[0268] Furthermore, it should be noted that in this embodiment, the feature maps of each layer are directly passed to the corresponding layer of the decoder through skip connections (e.g., the features of the second layer of the encoder are passed to the second layer of the decoder). Shallow background features (high resolution) provide the decoder with a basis for detail restoration (e.g., stripe edge correction), while deep features (low resolution, high semantics) provide global structural constraints (e.g., stripe period consistency). The combination of the two ensures the integrity and accuracy of the image spacing reconstructed by the decoder.

[0269] Through step S1, the encoder completes the transformation from the original simulated image to structured multi-level features, providing reliable feature support for the signal recovery of the decoder.

[0270] S2, through the performance recovery decoder, upsamples multi-level features step by step, and fuses the features of the corresponding level in the encoder to reconstruct the performance recovery image after denoising and structural restoration;

[0271] In this embodiment, the decoder uses deconvolution to achieve step-by-step upsampling, and simultaneously fuses features from corresponding levels of the encoder through skip connections to finally reconstruct the restored image. The specific logic is as follows:

[0272] Stepwise upsampling: Starting from the deepest features (low resolution, high semantics) output by the encoder, the spatial resolution is gradually increased through deconvolution (e.g., from 14×14→28×28→56×56), and the high-order semantic features are "mapped" back to the pixel space.

[0273] After each upsampling, the features of the corresponding layer of the encoder (such as the 3rd layer of the decoder and the 3rd layer of the encoder) are concatenated along the channel; the shallow background features (from the encoder) provide "fine-grained information" such as stripe edges and local details, while the decoder upsampled features provide global structure and semantic information after denoising. The fusion of the two can avoid the "loss of details" in deep networks (such as stripe edges being over-smoothed).

[0274] The final output "performance restored image" needs to simultaneously achieve denoising (eliminating scattering, blurring / spots caused by noise) and structure restoration (restoring the periodicity of the stripes and the clear outline of the aperture), approaching the signal characteristics of an interference-free image.

[0275] S3 inputs the performance restoration image and the interference-free image into the performance restoration discriminator. The discriminator divides the input performance restoration image into multiple local regions and performs local discrimination. The output is a discrimination score representing whether each local region conforms to the real image.

[0276] In this embodiment, the performance recovery discriminator adopts the PatchGAN structure, the core of which is to perform "fine-grained authenticity judgment" on the recovered image.

[0277] First, the input restored image and the interference-free image (real reference) are divided into multiple 16×16 (or 32×32) local regions (patch), each corresponding to local details of the signal (such as a stripe or a part of the aperture). Then, the features of each local region are extracted through multi-layer convolution to determine whether it conforms to the local features of the interference-free image (such as the local contrast of the stripe, edge sharpness, and absence of noise). Finally, a score of 0-1 is output for each local region (the closer to 1, the closer the region is to the local features of the real interference-free image), rather than a single judgment on the entire image.

[0278] This approach can specifically constrain the quality of local details in the recovered image. For example, if a stripe becomes partially broken after being recovered due to interference, the discriminator will output a low score for that area, prompting the network to optimize that detail.

[0279] S4 optimizes the performance recovery decoder and performance recovery encoder based on the discriminant score and a composite loss function consisting of adversarial loss, pixel loss, structural similarity loss, perceptual loss and edge loss until the network model converges.

[0280] It is understandable that by co-constraining the discriminant score with the multi-dimensional loss function, the parameters of the encoder and decoder are adjusted in reverse, ultimately ensuring that the recovered image is close to an interference-free image at the pixel, structural, and semantic levels.

[0281] Specifically, the loss function of the data recovery module in this embodiment is as follows:

[0282] Adversarial loss: Based on the local score calculation of the discriminator, the goal is to make each local region of the restored image judged as "real" (score close to 1), thus driving the network to generate a detailed restored image;

[0283] Pixel loss: Using L1 / L2 loss, the difference in pixel values ​​between the restored image and the undisturbed image is calculated to ensure the overall consistency of the two at the pixel level (such as matching the brightness of stripes).

[0284] Structural similarity loss (SSIM loss): measures the similarity between two signals in terms of brightness, contrast, and structure, avoiding focusing only on pixel differences while ignoring the overall signal structure (such as whether the stripe period is consistent).

[0285] Perceptual loss: High-level features are extracted using pre-trained networks (such as VGG) to constrain the semantic feature matching between the recovered image and the interference-free image (such as "integrity of the stripe region" and "spatial logic of aperture and stripes").

[0286] Edge loss: Extract edge features (such as stripe edges and aperture contours) from both images using edge detection operators (such as Sobel), calculate the differences, and ensure that the key edges of the recovered image are clear (blurred edges will affect subsequent bit demodulation).

[0287] It is understandable that the optimization process of the model uses the backpropagation algorithm to propagate the composite loss (weighted sum of each loss) back to the encoder and decoder, adjust parameters such as convolution kernel weights and biases, and iterate repeatedly until the loss value stabilizes (network convergence). At this point, the performance restoration network can stably output high-quality restored images.

[0288] In step S103, an extended dataset (simulated images) is used as the input to be disturbed, and an interference-free standard reference image is used as the supervised target. A performance recovery network is trained based on the U-Net architecture, ultimately resulting in an image performance recovery model capable of denoising and restoring the structure of the disturbed input. The trained model can accurately eliminate interference such as scattering and noise in the underwater environment, effectively restore the key structural features of the communication signal (such as modulation fringes and light source morphology), provide high-quality image input for subsequent bit sequence demodulation, and significantly improve the signal recovery capability and communication reliability of underwater visible light communication systems in complex environments.

[0289] Furthermore, it should be noted that the above embodiments disclose the use of generative adversarial networks (GANs) for communication performance recovery, restoring interfered data to clean data; it should be understood that other alternative solutions can also be used to achieve the above effects, including but not limited to:

[0290] CycleGAN: CycleGAN can be used as an alternative to GAN for communication performance recovery. CycleGAN learns the mapping between two domains, enabling unsupervised data transformation. This approach may be more effective than traditional GANs in some cases, especially when paired data is unavailable.

[0291] Pix2Pix: Pix2Pix can be used as an alternative to GANs for communication performance recovery. Pix2Pix is ​​a conditional generative adversarial network that learns the mapping between input and target images to generate high-quality output images. This method is particularly effective when paired data is available, producing high-quality reconstructed images.

[0292] Specifically, in an exemplary embodiment, a generative adversarial network is used to improve communication performance by employing a neural network with a U-Net structure to recover the underlying original modulation information structure from an image affected by environmental interference, while preserving the consistency of the background optical texture.

[0293] Figure 4 This is a schematic diagram illustrating the principle of training a U-Net-based neural network according to an embodiment of this application, such as... Figure 4 As shown, a clean reference image is used as the supervision target to guide the model to learn denoising and structure restoration capabilities from a large number of disturbed signal images.

[0294] The network architecture employs an enhanced U-Net structure, embedding multi-scale convolutional structures, a channel attention mechanism (SEBlock), and residual connection modules to ensure the capture of high-frequency information and the reconstruction of multi-level textures. The input batch consists of noisy image signals.

[0295] (41)

[0296] With a unique standard reference image:

[0297] (42)

[0298] N This represents the batch size, with the three channels corresponding to the RGB color dimensions. Both are normalized and scaled before being fed into the network. The generator structure is based on U-Net, with each level containing multi-scale convolutions and channel attention mechanisms.

[0299] The encoder section employs three parallel convolutional paths of different scales (3x3, 5x5, 7x7) to capture local and global features in the spatial perception of the image, specifically:

[0300] (43)

[0301] (44)

[0302] (45)

[0303] The fused feature is obtained by adding the three features together:

[0304] (46)

[0305] By spatially modeling the same feature using multi-scale filters, the model's adaptability to interference at different scales in striped images is enhanced. Subsequently, to improve the model's ability to perceive key signal feature channels, a Squeeze-and-Excitation (SE) module is employed to process the fused features. The process for channel weighting is as follows:

[0306] First, global average pooling (GAP) is performed to obtain z. Then, two fully connected layers are used to construct a non-linear mapping to generate attention weights.

[0307] (47)

[0308] (48)

[0309] Attention weights s and Perform channel multiplication output:

[0310] (49)

[0311] This mechanism dynamically adjusts channel weights, making the model focus more on structural channels that play a decisive role in communication recovery and suppressing irrelevant background.

[0312] Each encoder stage is followed by a 3×3 convolution, ReLU, and then a SEBlock layer, followed by max pooling downsampling. The output feature channels are 64, 128, 256, and 512 respectively, with the spatial resolution halved with each layer. The outputs e1, e2, e3, and e4 of each encoder stage are used for skip connections.

[0313] In the decoder section, the deepest structural feature input is upsampled:

[0314] (50)

[0315] splicing with e3:

[0316] (51)

[0317] After multi-scale convolution, SEBlock, convolution, and ReLU, d3 is obtained, and similarly, upsampling is performed sequentially:

[0318] (52)

[0319] splicing with e2 Output d2, then from:

[0320] (53)

[0321] The image is concatenated with e1 and cat1. During the decoding stage, reconstruction and skip connections are combined to maintain shallow spatial resolution while gradually fusing deep semantic features. Finally, after multi-scale convolution, SEBlock, convolution, and sigmoid activation, a three-channel image 'out' is output. This residual structure is then added back to the input image. Perform pixel-by-pixel addition to form a residual image, and output:

[0322] (54)

[0323] The shape remains N×3×224×224;

[0324] The discriminator employs a PatchGAN structure, which divides the input image into multiple local regions and achieves finer-grained texture constraints through local discrimination. The input is a three-channel image, regardless of the residual image. The image y is processed sequentially through four convolutional blocks:

[0325] (55)

[0326] Where the convolution stride is s = 2, 2, 2, 1, and the dimension of the output feature map is... Each element in the output represents the true probability of the corresponding receptive field patch (approximately 16×16 region). The discriminator is designed to improve the model's sensitivity to local textures and edge structures, enhancing its ability to distinguish details with high fidelity. It guides the generator to learn realistic, structurally accurate, and physically plausible image reconstruction maps. Define the composite loss function:

[0327] (56)

[0328] Detailed explanation: Countering losses :

[0329] (57)

[0330] The ability of a generated image to be identified as real by a discriminator is measured by the binary cross-entropy (BCE) metric, driving... Fit the true image distribution. Pixel L1 loss:

[0331] (58)

[0332] To ensure pixel-level reconstruction accuracy, the focus is on constraining the position and brightness restoration of fringes. Structural similarity loss:

[0333] (59)

[0334] SSIM is used to measure the structure, texture, and brightness contrast of an image, improving the accuracy of structural region reconstruction. Perceptual loss:

[0335] (60)

[0336] in, Indicates the VGG network's... j Layer feature output. This measure measures high-level semantic consistency, constraining the generated image to visually approximate a real image. Edge loss:

[0337] (61)

[0338] in, This indicates the Sobel operation. This term emphasizes the restoration of edge structure and texture contours, which helps improve the clarity of the demodulated image. In each training iteration, all x and y values ​​in a batch are processed by G to output hat{y}, which is then fed into the discriminator D. The discriminator uses standard adversarial loss.

[0339] (62)

[0340] All losses are calculated sequentially, and the generator and discriminator parameters are optimized independently. This structure ensures that the network can be constrained from multiple perspectives, including multi-scale spatial feature extraction, channel adaptive enhancement, deep residual detail compensation, and local spatial discrimination by the discriminator. Under the collaborative approach of multi-task loss, it can directly map noisy signal images into high-fidelity restored signal images.

[0341] S104 deploys a trained image performance recovery model at the receiving end to optimize the interfered communication images captured in underwater visible light scenes and outputs the recovered image.

[0342] This step is the core of applying the trained performance recovery model to a real-world underwater visible light communication scenario. By processing the interference-affected images captured in real time by the receiver, signal recovery and demodulation of the original data are achieved, ultimately leading to reliable communication. Its core function is to translate the model's generalization ability into practical communication performance, solving the signal distortion problem caused by underwater environmental interference and ensuring the accuracy of data transmission. Specifically, it includes the following steps:

[0343] S1 preprocesses the interfered communication images captured in real time by the image sensor to generate input tensors that are normalized in size and color.

[0344] In this process, the image sensor (such as a CMOS camera) at the receiving end captures underwater communication scenes in real time. The resulting images may have issues such as inconsistent size and color deviation due to hardware differences (such as resolution and exposure time) or environmental fluctuations (such as sudden changes in lighting). In this embodiment, the preprocessing stage standardizes the input through the following operations:

[0345] 1) Scale the image to the same size as when the model was trained (e.g., 224×224) to ensure that the spatial dimension of the input tensor matches the input requirements of the model and avoid feature extraction deviation due to size incompatibility;

[0346] 2) Perform mean subtraction or normalization on the RGB channels of the image (e.g., map pixel values ​​to the 0-1 range) to eliminate global brightness fluctuations caused by changes in illumination, so that the input feature distribution is consistent with the training data and ensures the stability of the model processing;

[0347] 3) Convert the preprocessed image into a tensor format that meets the requirements of the deep learning framework (e.g., add batch dimensions) and use it as input to the performance recovery model.

[0348] It is understandable that the preprocessing process includes, but is not limited to, the above steps. The core of preprocessing is to eliminate the influence of non-signal interference factors (such as hardware differences and lighting fluctuations) on the model, provide standardized and consistent input for subsequent processing, and ensure that the model can stably perform.

[0349] S2, input the input tensor into the trained image performance recovery model, and obtain the compensated image through the model's encoder and decoder;

[0350] The trained performance recovery model processes the standardized input tensor in real time, specifically:

[0351] By employing multi-scale convolution, feature fusion, and channel attention mechanisms, we can quickly extract interference features (such as scattered halos and noise spots) and signal features (such as stripe contours) from the input tensor to generate multi-level features.

[0352] Based on the features of the encoder output, shallow details and deep semantics are fused through upsampling and skip connections to generate a targeted compensation map. For example, the pixel values ​​of this image correspond to the correction amount of the disturbed region in the input tensor (such as adding a contrast-enhancing compensation value to the blurred stripe region and a suppression compensation value to the noise spot region).

[0353] In this embodiment, the essence of the compensation image is the model's reverse correction of the "interference part". Its generation process relies entirely on the interference-recovery rules learned during the training phase, requiring no manual intervention and meeting the low latency requirements of the communication system.

[0354] S3, superimpose the compensated image and the input tensor pixel by pixel to form the final restored image; and, perform grayscale and projection processing on the striped areas in the restored image, and perform threshold judgment and bit extraction according to the preset encoding rules to demodulate the original bit sequence.

[0355] This step is the final stage of signal recovery and data extraction, and it can be divided into two parts:

[0356] Image restoration generation: The compensated image is superimposed pixel by pixel on the input tensor (the original disturbed image). The correction in the compensated image cancels out the interference components in the input tensor (such as noise suppression and enhancement of blurred stripes), and finally a clear and complete restored image is obtained. This image retains the key structure of the original signal (such as the periodicity of OOK modulation stripes and the clear outline of the light source) and is highly similar to the interference-free reference image.

[0357] Bit sequence demodulation: First, the color image is converted into a single-channel grayscale image by grayscale conversion to highlight the contrast between the light and dark stripes, and then converted into a NumPy array with a shape of (480, 640), and all pixel values ​​are uniformly in the range of [0, 255], in preparation for subsequent brightness transitions;

[0358] Furthermore, in Manchester modulation, each bit corresponds to a brightness transition; the number of brightness transitions in the vertical direction of the entire image equals the number of bits (or the number of levels); by vertically segmenting, each bit region of the image is processed independently. Given that a frame of the image corresponds to 22 transitions, each transition occupies an average of 480 / 22 = 22 rows. The key features of Manchester coding are: each bit contains one brightness transition, i.e., the brightness is reversed above and below the center line; the direction of the transition is determined to be bit 0 or bit 1; 22 rows correspond to 11 bits per frame.

[0359] Finally, the bit information is extracted, and the communication performance is calculated.

[0360] This process transforms "interferenced images" into "raw data," and through real-time processing of the performance recovery model, it counteracts the complex interference of the underwater environment, enabling the demodulation stage to accurately identify signal characteristics, significantly reduce the bit error rate, and ensure the reliability and effectiveness of underwater visible light communication.

[0361] In step S104 above, the performance recovery module is deployed at the receiving end. Through a coherent process of standardized preprocessing, real-time model correction, image overlay, and data demodulation, the trained performance recovery model is transformed into a core component of the actual communication system, achieving efficient recovery and data extraction of real-time interfered signals. This overcomes the limitations of underwater environments on visible light communication, improves the system's robustness in complex scenarios, and provides key technical support for high-speed data transmission between underwater devices.

[0362] In an exemplary embodiment, the model deployment module can run on the receiving PC, and an example of the module's operation is as follows:

[0363] Figure 5 This is a schematic diagram of the communication physical link of the underwater visible light signal recovery method according to an embodiment of this application, such as... Figure 5 As shown, after the CMOS camera at the receiving end captures the light signal, it first enters the image preprocessing module. This module performs image center cropping, size normalization, and color normalization, ultimately forming an RGB tensor of size 224x224, denoted as:

[0364] (63)

[0365] The input is a pre-trained U-Net model. Within the model, the encoder first performs multi-level feature extraction, specifically including:

[0366] The first layer, enc1, includes multi-scale convolutions (3×3, 5×5, 7×7) and a SE attention module to extract shallow texture features from the original image. The output is a tensor of size 224×224 with 64 channels. Then, the image spatial size is reduced to 112×112 through the first pooling (pool1) and fed into the enc2 layer to continue extracting mid-level perturbation features, increasing the number of channels to 128. After another pooling (pool2), the size becomes 56×56 and is fed into the enc3 layer to further learn perturbation structures such as blurred regions and local occlusions, expanding the number of channels to 256. Finally, it passes through pool3 into the deepest layer, enc4, to extract high semantic-level information at a spatial resolution of 28×28, ultimately resulting in a compressed feature tensor with 512 channels.

[0367] The decoder path reconstructs and enhances the aforementioned deep structural features step by step, specifically including:

[0368] The deep structure feature map is upsampled back to 56×56 using a deconvolutional layer up3, and then concatenated with the output of enc3 in the encoder to form a 512-channel fused feature. After fusion, it enters dec3 for feature restoration and compression to 128 channels. Similarly, up2 upsamples to 112×112, concatenates the feature map of enc2, and then reconstructs the texture structure through dec2. Up1 continues to upsample to the original size of 224×224, concatenates it with the output of enc1, and then the dec1 layer restores the low-level structure of the image, such as edges and brightness. The network tail uses the Sigmoid activation function to normalize the pixels of the output image to the [0,1] interval, and finally outputs a restored image with the same size as the input and 3 channels.

[0369] (64)

[0370] To maintain the integrity of information in undisturbed areas, a residual mechanism is used at the output. The network only needs to learn the compensated image of the disturbed area, and finally superimposes it pixel by pixel with the input image to form the final reconstructed image.

[0371] (65)

[0372] Among them, G( ) represents the main mapping of U-Net. The image is the restored image. This design allows the model to focus on learning the restoration of perturbed regions while preserving the undisturbed information in the original image.

[0373] After the image is restored, the central stripe area is automatically extracted. The signal intensity map is then converted to grayscale. To extract the time-series features of the modulated signal, the system obtains a one-dimensional brightness vector by longitudinally averaging the image:

[0374] , (66)

[0375] The resulting binary vector Indicates the on / off state of the stripes; identifies the positions of adjacent transition points based on Manchester coding rules. Complete bit extraction:

[0376] (67)

[0377] Final output intra-frame bit sequence Multiple image frames are decoded to form a complete bitstream.

[0378] Through steps S101 to S104 above, the problems of signal noise, distortion, and communication performance degradation caused by disturbances such as water turbidity, current velocity, and ambient light intensity in underwater visible light communication (UVLC) systems are addressed. Background data is generated using MATLAB simulation, and simulated data is generated by fusing simulation and real image features through a conditional generative adversarial network (cGAN). The U-NetGAN model is then used to train signal recovery capabilities, and finally deployed at the receiver. This overcomes the shortcomings of traditional modulation methods (poor adaptability) and existing deep learning models (reliant on large amounts of data and weak generalization ability). The trained model can accurately eliminate interference such as scattering and noise in the underwater environment, effectively restore the key structural features of the communication signal, provide high-quality image input for subsequent bit sequence demodulation, and significantly improve the signal recovery capability and communication reliability of underwater visible light communication systems in complex environments.

[0379] Secondly, embodiments of this application provide an underwater visible light communication system based on transfer learning, the system including a transmitter and a receiver. Figure 5 This is a framework diagram of an optical underwater communication system according to an embodiment of this application, such as... Figure 5 As shown,

[0380] The transmitter 50 is used to emit visible light signals by modulating a light source, and to encode data in the visible light signals and send them to the underwater channel environment.

[0381] The transmitter 50 is the core of the underwater visible light communication system. Its main function is to load the data to be transmitted (such as detection commands, sensor data, etc.) into a visible light signal and transmit it to the underwater channel via a light source. The specific process is as follows:

[0382] The transmitter 50's communication principle involves modulating an LED light source to emit visible light signals, encoding data within the light waves, which are then received, decoded, and reconstructed by a CMOS camera in the water. The transmitter consists of three parts: a transmitter PC, a UNO microcontroller, and an LED driver circuit.

[0383] Figure 6a This is a schematic diagram of the auxiliary and control circuit area of ​​the transmitter circuit in an embodiment of this application; Figure 6b This is a schematic diagram of the interface and power management area of ​​the transmitter circuit according to an embodiment of this application; Figure 6c This is a schematic diagram of the core functional circuit area of ​​the transmitter circuit in an embodiment of this application; such as... Figure 6a , Figure 6b and Figure 6c As shown, the transmission process is as follows: the host computer sends the bit sequence to be modulated to the UNO via the Bluetooth module. UNO employs OOK modulation and introduces Manchester coding to expand each bit into two state bit changes, enhancing bit boundary clarity and eliminating DC offset. Its modulation frequency is 5.682 kHz, corresponding to one bit period of... In this way, the light intensity signal output by the LED light source can be represented as a bit-driven rectangular pulse sequence, i.e.:

[0384] (68)

[0385] in, For standard rectangular pulse functions, This represents the on / off state after Manchester encoding. During propagation in the water channel, the modulated optical signal will be affected by various interferences such as absorption, scattering, and turbulence, resulting in composite loss in the underwater channel. The channel transfer function can be derived from Beer... Lambert's law can be simplified to:

[0386] (69)

[0387] in, For transmission distance, The wavelength of LED light emission and These are the absorption coefficient and scattering coefficient at this wavelength, respectively. The final received signal intensity is:

[0388] (70)

[0389] in, This represents the equivalent environmental background noise, which follows a Gaussian distribution.

[0390] The receiver 51 is used to receive visible light signals through a CMOS camera and process the visible light signals through the underwater visible light signal recovery method of the first aspect to obtain the recovered image.

[0391] The receiver 51 consists of a CMOS image sensor and an image processing PC. The CMOS camera operates at a frame rate... The data was collected, and the exposure time for each frame was... Because the exposure time is much longer than a single bit period, brightness changes within multiple bit periods can be captured simultaneously in a single frame image. Therefore, the signal will appear as periodic bright and dark stripes or regional brightness differences on the image. (Any pixel on the image...) The grayscale value can be represented as the convolution of the signal and the optical path spatial spread function over a period of time, that is:

[0392] (71)

[0393] in, It is the spatial perturbation response of light propagating in turbulence. This refers to pixel noise from the camera device. Under interference-free conditions, the system measured a frame image signal-to-noise ratio of approximately 7.87 dB. The signal-to-noise ratio in an image is defined as:

[0394] (72)

[0395] in, The difference between the mean gray values ​​of bright and dark areas in the image. The background noise variance. Each image frame can record an average of approximately 11 bits. The image signal is demodulated using a line-by-line scanning method to achieve the following:

[0396] (73)

[0397] This is for bit recovery at the theoretical rate. Actual statistics show the system's actual communication rate is 630 bps, and the bit error rate is... Demodulation accuracy is measured using the bit error rate (BER), which is defined as follows:

[0398] (74)

[0399] in, For index functions, The bits obtained from image decoding and demodulation.

[0400] also, Figure 7 This is a schematic diagram of a laboratory platform according to an embodiment of this application; as shown Figure 7 As shown, a laboratory platform simulating an underwater environment was constructed in a box-shaped experimental pool with dimensions of 2m x 1.2m x 2.4m. The glass thickness was 2mm, the water depth was 1.2m, the horizontal distance between the transmitter and receiver was 2.4m, and the vertical distance from the bottom of the pool was 1m. Influencing factors were set as turbidity, flow velocity, and ambient light intensity. Turbidity was altered by soil (10g of soil was left to stand in 10L of clean water for 5 minutes, resulting in a turbidity of approximately 50 NTU, with a ratio of approximately 1g:5 NTU). Flow velocity was achieved by replacing the turbulence generator. Ambient light intensity was controlled by switching on and off an unmodulated light source 1.6 meters vertically above the water surface. Currently, 70 interference environments were set up. Turbidity ranged from 0-300 NTU in 50 NTU intervals, flow velocities were 0m / s, 0.1m / s, 0.4m / s, 0.7m / s, and 1.2m / s, and ambient light intensities were 0 Lux and 100 Lux. The values ​​for turbidity, flow velocity, and ambient light intensity are approximate, with the mean differing from the actual value by less than 5%. Data was collected using a laboratory platform, with 100 images taken for each environment, and communication performance was calculated for each environment.

[0401] Table 2 below shows communication performance data under some environmental disturbance scenarios:

[0402]

[0403] Table 2 shows communication performance data under some environmental disturbance scenarios.

[0404] Table 2 shows the communication performance under interference conditions in some laboratory platforms. 70 sets of data were divided into 10 style guidance groups and 60 validation groups (not used in training). The 10 sets simulated scenarios where rapid underwater changes prevented large-scale data collection. MATLAB was used to expand the data volume using the reference parameters, generating 70 sets of data. These simulation data, along with the 10 style guidance group sets, were fed into the data expansion model to generate 70 sets of realistic data. This data was then used to train the performance recovery module. The best-performing set of data was used as the indicator to generate the model. The remaining 60 sets of data were then fed into the model to calculate the communication performance of the data after performance recovery.

[0405] Using the system provided in this embodiment, the experimental results include:

[0406] A. Communication performance recovery test

[0407] Figure 8 This is a schematic diagram comparing the signal-to-noise ratio according to an embodiment of this application. Figure 9 This is a schematic diagram comparing bit error rates according to an embodiment of this application. Figure 10 This is a schematic diagram of the overall performance curves according to an embodiment of this application. For example... Figure 8 , Figure 9 and Figure 10 As shown, comparing the communication performance using the proposed solution with that without, the overall performance improvement is represented by the average increase in communication rate, bit error rate, and signal-to-noise ratio. The communication rate is calculated based on the base communication rate and bit error rate, and can be obtained using the formula... Calculate, where, It is the actual effective communication rate (unit: bit / s, bits per second), which refers to the effective data transmission rate that the receiving end can successfully demodulate without errors, and is used to reflect the actual communication capability of the system; Basic communication rate (unit: bit / s) refers to the theoretical transmission rate of the transmitting end under ideal error-free conditions (such as the rate agreed upon by the modem protocol, such as 150Mbps for 4G LTE, 10Gbps for fiber optic communication, etc.); BER is the bit error rate, which is the ratio of the number of incorrectly received bits to the total number of transmitted bits in a communication system (e.g., BER=10). -6 This indicates that approximately one erroneous bit is sent for every million bits transmitted, and its value ranges from 0 to 1.

[0408] like Figure 10As shown, the proposed solution significantly improves underwater visible light communication performance in response to environmental changes, reaching a maximum improvement of 90.5%. Optionally, this system uses the Mechrevo Aurora X5060 laptop as the model deployment platform. When inputting a 224×224 pixel CMOS signal image, the single-frame inference latency can be stably achieved within 8ms on the CPU and within 1ms on the GPU, which is far less than the CMOS acquisition frame interval of 16.7ms (60fps).

[0409] B. Data Volume Test

[0410] Figure 11 This is a schematic diagram illustrating the impact of data volume on overall performance according to embodiments of this application, such as... Figure 11 As shown, a comparison was made between training with different amounts of data. 1, 5, 10, 20, 30, and 50 groups of data were selected as the style guidance group; 20 groups served as the validation group. This means that 88.4% of the effect of training with the full amount of data can be achieved with only 1 / 7 of the total data.

[0411] C. Data Expansion Module Performance Test

[0412] Figure 12 This is a schematic diagram illustrating the performance improvement effect of the data expansion module according to an embodiment of this application, such as... Figure 12 As shown, to verify whether the model needs data augmentation, it was divided into two groups. Ten sets of original data were extracted. One group served as a style control group and underwent data augmentation before entering the performance recovery module. The other group directly used the original data for training in the performance recovery module. The comparison showed that the group that underwent data augmentation demonstrated significantly better communication performance recovery than the group trained directly. Therefore, the proposed solution can significantly reduce the model's dependence on the amount of data.

[0413] D. Background Technology Comparison

[0414] Table 3 below is a comparison table of the average overall performance improvement based on the solution in this application.

[0415] Table 3 Average comprehensive performance parameters of various schemes

[0416]

[0417] This paper compares the communication performance of the proposed solution with traditional modulation methods and traditional deep learning algorithms. The experiment uses the background technology of SIM-PWC (a traditional modulation method) and compares it with the LSTM deep learning algorithm. Since SIM-PWC modulation is currently not feasible, simulation is used as the performance indicator. The LSTM algorithm was trained with only 10 style control groups (simulated data is scarce), and the communication performance was tested on 60 groups. Generalization scenario coverage refers to the communication performance of untrained data reaching 95% of the ideal communication rate. The ideal (interference-free) communication rate of SIM-PWC is 1440 bps, while the ideal communication rate of the proposed solution and LSTM is 630 bps. Table 3 shows that the bit error rate of the proposed solution is lower than that of the traditional SIM-PWC modulation method and the LSTM deep learning algorithm, and the generalization scenario coverage is higher than these two methods, reaching 90%. Therefore, the proposed solution improves the communication performance and generalization ability of existing technologies to a certain extent.

[0418] In addition, Table 4 below is the transmitter parameter configuration data table:

[0419] Table 4 Main parameters of the transmitter

[0420]

[0421] Table 5 below shows the environmental parameter data for the water tank:

[0422] Table 5 Main parameters of the pool environment

[0423]

[0424] Table 6 below is the channel parameter configuration data table:

[0425] Table 6 Main parameters of the channel

[0426]

[0427] Table 7 below shows the receiver parameter configuration data:

[0428] Table 7 Main parameters of the receiver

[0429]

[0430] It must be noted that the solution used in this application is a self-developed system, including the transmitter and receiver. Under this architecture, data acquisition is not limited to using exactly the same equipment described above; data acquisition can also be achieved through the following methods:

[0431] Multi-transmitter and multi-receiver systems: These systems replace single-transmitter and single-receiver systems with multi-transmitter and multi-receiver systems. Such systems provide more data samples, improving the model's generalization ability.

[0432] Underwater robot platforms: Underwater robot platforms can be used to replace UVLC systems built in the laboratory. Underwater robots can collect data in actual underwater environments, providing data samples that are closer to real-world scenarios.

[0433] Compared with existing technologies, this system achieves the following beneficial effects: Through the data augmentation module, only 1 / 7 of the total environmental data volume is needed to achieve 88.4% of the total data volume, while maintaining high communication performance, significantly reducing dependence on data volume; simultaneously, by expanding the parameter range of this module, it can broadly cover environmental changes, enabling the model to maintain good communication even in untrained environments, significantly improving generalization ability; the performance recovery module effectively enhances communication performance under interference environments, showing outstanding performance in dealing with environmental disturbances; furthermore, compared with the high complexity caused by traditional technologies relying on changes in transmitter modulation methods or large-scale training, this application's solution can stably achieve processing latency of less than 8ms on CPUs and less than 1ms on GPUs, with a maximum tolerance latency of less than 16.7ms (60fps), achieving near-zero latency communication and significantly simplifying system complexity.

[0434] In one embodiment, Figure 13 This is a schematic diagram of the internal structure of an electronic device according to an embodiment of this application, such as... Figure 13 As shown, an electronic device is provided, which can be a server, and its internal structure diagram can be as follows. Figure 11 As shown, the electronic device includes a processor, a network interface, internal memory, and non-volatile memory connected via an internal bus. The non-volatile memory stores the operating system, computer programs, and a database. The processor provides computing and control capabilities, the network interface communicates with external terminals via a network, the internal memory provides an environment for the operating system and computer programs to run, the computer programs are executed by the processor to implement an underwater visible light communication method based on transfer learning, and the database stores data.

[0435] Those skilled in the art will understand that Figure 13 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the electronic device to which the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0436] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0437] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0438] The above embodiments merely illustrate several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for recovering underwater visible light signals based on transfer learning, characterized in that, The method includes: Based on real images in underwater visible light communication scenarios, data simulation is performed using preset simulation parameters to generate a simulated image similar to the real image. Specifically, this includes: generating the intensity distribution of the main aperture and sub-aperture of the real image using a two-dimensional Gaussian function, and generating the halo structure of the real image generated by water scattering effect around the aperture using an exponential decay model, thus obtaining the aperture layer of the simulated image. By modeling with a periodic function, the real image is simulated, and modulation stripes formed by the camera sampling the OOK signal light intensity are generated to obtain the stripe layer of the simulated image. The contrast of the modulation stripes is determined according to the total attenuation coefficient corresponding to the turbidity of the water and the propagation distance. Based on the signal strength of the real image and the signal-to-noise ratio calculated with the ambient background noise, a simulated image occlusion layer is generated to simulate the occlusion mask of the real image. The range and transparency gradient of the occlusion mask are derived in reverse based on the signal-to-noise ratio to simulate the degree of damage to the imaging quality caused by the ambient noise in the real image. The simulated image is obtained by superimposing the simulated image aperture layer, the simulated image stripe layer, and the simulated image occlusion layer onto a solid color background layer. Through the data augmentation module and style transfer mechanism, the deep structural features of the real image are mapped to the shallow background features of the simulated image to generate realistic images covering a variety of complex environments, thus obtaining an extended dataset. Furthermore, the performance recovery network is trained using the extended dataset as the disturbed input and the undisturbed standard reference image as the supervision target to obtain an image performance recovery model for denoising and restoring the structure of the disturbed input. A trained image performance recovery model is deployed at the receiving end to optimize the interfered communication images captured in underwater visible light scenes and output the recovered images.

2. The method according to claim 1, characterized in that, The generator maps the deep structural features of the real image to the shallow background features of the simulated image to obtain the realistic image, specifically including: The shallow background features are obtained by extracting style features at each level of the simulated image through the first encoder. The deep structural features are obtained by extracting the structural features of each level of the real image through the second encoder. Obtain the parameter labels of the underwater environment corresponding to the real image, and input them into a multilayer perceptron for multilayer fully connected mapping to obtain a high-dimensional label vector, wherein the parameter labels include turbidity, flow velocity and illuminance; The deep structural features, the shallow background features, and the high-dimensional label vector are concatenated to obtain a fusion tensor. Using a decoder, based on the fusion tensor and guided by the high-dimensional label vector, the deep structural features of the real image are mapped to the shallow background features of the simulated image to obtain the simulated image.

3. The method according to claim 2, characterized in that, Based on the fusion tensor, mapping the deep structural features of the real image to the shallow background features of the simulated image includes: The encoder employs deconvolution layers to perform stepwise upsampling based on the fused tensor. During the progressive upsampling process, based on the adaptive instance normalization skip connection mechanism, the deep structural features and shallow style features are connected by AdaIN skip connection at each level, and the shallow style features are applied to the deep structural features to output the realistic image.

4. The method according to any one of claims 2 and 3, characterized in that, After generating the simulated image, the method further includes: The high-dimensional label vector is linearly mapped to the spatial feature map of the simulated image to obtain a conditional map with the same spatial size as the simulated image. The simulated image and the conditional map are then concatenated to obtain a concatenated tensor. The discriminator performs multi-layer convolutional downsampling based on the spliced ​​tensor, outputting a real-fake discrimination score representing the difference between the simulated image and the real image, as well as an output feature matching loss. Based on the discrimination score and multiple loss functions, the generator and discriminator of the data augmentation module are jointly optimized for parameters, wherein... The overall loss of the discriminator is the binary cross-entropy loss, which is used to constrain the discriminator to maximize the probability of identifying a real image as real and minimize the probability of identifying a simulated image as real. The loss function of the generator includes at least: Adversarial loss used to constrain the realistic image to improve the discrimination score in the discriminator; Structural loss used to ensure that the simulated image maintains consistency with the simulated image at the pixel level; The perceptual loss used to improve the semantic level matching of the simulated image with the high-level features of the real image; Style loss used to ensure that the simulated image remains consistent with the real image in terms of texture and frequency features; Feature matching loss is used to enhance the intermediate semantic consistency between the generator and the discriminator.

5. The method according to claim 1, characterized in that, The performance recovery network is built based on the U-Net neural network. Training the performance recovery network involves using the extended dataset as the disturbed input and an undisturbed standard reference image as the supervised target, and includes: The simulated image is processed by a performance recovery encoder to extract multi-level features of the simulated image; The multi-level features are upsampled level by level by the performance recovery decoder and the features of the corresponding level in the performance recovery encoder are fused to reconstruct the performance recovery image after denoising and structural restoration. The performance recovery discriminator divides the input performance recovery image into multiple local regions, and after performing local discrimination in conjunction with the interference-free standard reference image, it outputs a discrimination score that indicates that each local region conforms to the real image. Furthermore, based on the discrimination score and the composite loss function consisting of adversarial loss, pixel loss, structural similarity loss, perceptual loss and edge loss, the parameters of the performance recovery decoder and the performance recovery encoder are jointly optimized until the network model converges.

6. The method according to claim 5, characterized in that, in, Processing the simulated image to extract its multi-level features includes: Multi-scale convolutional structures are used to capture local and global features at different scales; The local features and the global features are added together to obtain the fused features; The fused features are spatially modeled using a multi-scale filter, and the channel weights of the spatial modeling result are dynamically adjusted using a channel attention mechanism to obtain a weighted feature map. The weighted feature map is downsampled by max pooling to output the multi-level features, and the multi-level features are then connected to the decoder via skip connections.

7. The method according to claim 1, characterized in that, After deploying the trained image performance restoration model at the receiving end, the method further includes: Preprocessing of the interfered communication images captured in real time generates input tensors that are normalized in size and color; The input tensor is sent to the trained image performance restoration model, and the encoder and decoder of the image performance restoration model are used to restore the input tensor to obtain the compensated image; The compensated image is superimposed pixel-by-pixel with the input tensor to obtain the restored image; The striped regions in the restored image are grayscaled and projected, and threshold judgment and bit extraction are performed according to preset encoding rules to demodulate the original bit sequence.

8. An underwater visible light communication system based on transfer learning, characterized in that, The system includes a transmitter and a receiver, wherein: The transmitter is used to emit visible light signals by modulating a light source, and after encoding data in the visible light signals, transmit them to an underwater channel environment; The receiving end is used to receive the visible light signal through a CMOS camera and process the visible light signal using the underwater visible light signal recovery method according to any one of claims 1-7 to obtain the recovered image.

9. The system according to claim 8, characterized in that, The transmitter includes a transmitter module, an encoding module, and a driver module, wherein: The transmitting module is used to output the original bit sequence to be modulated, wherein the original bit sequence contains the data information to be transmitted; The encoding module is used to expand each bit in the original bit sequence into two state bit changes using Manchester encoding to enhance bit boundary clarity and eliminate DC offset, and to convert the Manchester encoded sequence into a drive signal using OOK modulation. The driving module is used to generate a light intensity signal according to the driving signal and send the light intensity signal to the underwater visible light communication environment.

Images