Filtering method and device based on forward-looking sonar and storage medium and underwater carrier

Abstract

This invention discloses a filtering method, device, storage medium, and underwater vehicle based on forward-looking sonar. The filtering method employs the following steps: sonar physical detection area modeling and mask mapping; first-level global adaptive benchmark extraction and coarse filtering; acoustic attenuation sensing and radial block partitioning; second-level non-uniform sensitivity local fine filtering; spatial smoothing fusion and image reconstruction, outputting a denoised forward-looking sonar image. By cleverly combining an acoustic radial distance model with the local statistical features of a single-frame image, and through pure mathematical cascade matrix operations and adaptive soft thresholding, a perfect balance between reverberation suppression and weak target protection is achieved with extremely low computational overhead. It preserves the features of weak targets in the far field, solves the denoising distortion problem caused by drastic changes in the dynamic range of sonar images, reduces the temporal and spatial complexity of the algorithm, ensures real-time deployment at the edge of the underwater unmanned platform, and ensures smooth edges and no geometric distortion in the denoised image, providing high-fidelity input for subsequent target recognition and path planning.

Description

Technical Field

[0001] This invention relates to the fields of image processing and underwater detection technology, specifically to a lightweight cascaded block adaptive filtering method, apparatus, computer-readable storage medium, and underwater vehicle based on forward-looking sonar. Background Technology

[0002] With the deepening development of marine resources, unmanned underwater vehicles (UUVs) and remotely operated vehicles (ROVs) play a central role in subsea pipeline inspection, search and rescue, and topographic mapping. In these missions, forward-looking sonar (FLS), as an active detection device based on acoustic imaging, can generate continuous video images in murky waters with zero visibility, serving as the "eyes" of underwater sensing systems. However, limited by the physical characteristics of the underwater acoustic channel, the imaging quality of FLS images is far lower than that of optical images, posing a significant challenge to subsequent visual processing tasks.

[0003] Specifically, the formation of forward-looking sonar images depends on the emission and reception of sound waves. Due to the complexity of the underwater acoustic environment, strong reverberation noise, large-scale multiplicative speckle noise, and geometric diffusion and absorption attenuation during sound wave propagation are inevitably generated during the imaging process. Existing technologies face three main technical bottlenecks in denoising FLS images: First, the complex acoustic noise interference and the failure of traditional spatial consistency filtering algorithms. Traditional digital image processing methods, such as median filtering, Gaussian filtering, and even more complex nonlocal mean filtering (NLM) and BM3D algorithms, often fail to address sonar noise effectively. These algorithms are typically based on the spatial consistency assumption, which uses a globally uniform filter strength or a fixed hard threshold to process the entire image. However, forward-looking sonar imaging exhibits strong distance dependence: in the near field, sound energy is strong, resulting in extremely high reverberation and background noise; while in the far field, sound energy decays exponentially, and the target echo signal is extremely weak. If a global high threshold filter is used, although it can filter out near-field reverberation, it will directly erase the already weak target edges in the far field; if a global low threshold is used, it cannot suppress the strong noise in the near field. In addition, traditional hard threshold truncation is prone to introducing unnatural step artifacts into the image, destroying the texture continuity of the target and seriously affecting the feature extraction of subsequent target recognition algorithms. Furthermore, sonar images are fan-shaped regions, and traditional algorithms cannot effectively process fan-shaped images.

[0004] Second, deep learning denoising methods face the computational nightmare and data hunger. In recent years, deep learning denoising methods based on convolutional neural networks (CNNs) or Transformer architectures have made significant progress. However, these supervised learning models are heavily reliant on training with pairs of noisy and clean image data (Ground Truth). In real underwater dynamic environments, obtaining noise-free sonar (often used as sonar) "ground truth" images is virtually impossible. Even with unsupervised or weakly supervised deep learning networks, their massive number of parameters and huge floating-point operation overhead (FLOPs) make them extremely dependent on high-performance GPU computing power. For edge devices powered by embedded chips and subject to stringent power consumption and heat dissipation constraints, deploying such heavy networks and achieving high frame rate real-time video processing is highly impractical in engineering.

[0005] Third, there is a lack of mechanisms to protect target features. In the field of lightweight denoising, many existing methods often adopt a one-size-fits-all smoothing strategy to reduce computational complexity. Underwater targets in sonar images typically appear as local brightness abrupt changes or textured regions with specific variance. Existing lightweight algorithms lack the ability to perceive the local activity of the image and cannot distinguish between clean background noise and effective edges containing the target, resulting in the denoising process often being accompanied by excessive smoothing caused by severe target blurring.

[0006] In summary, the current field of underwater visual perception urgently needs a sonar image denoising method with extremely low complexity that does not require prior collection of ground truth datasets for training, does not rely on heavy hardware computing power, and can perfectly fit the underwater acoustic physics attenuation model. Summary of the Invention

[0007] One of the technical problems to be solved by this application is to overcome the defects of the above-mentioned related technologies and provide a filtering method, device and computer-readable storage medium and underwater vehicle based on forward-looking sonar. The filtering method cleverly combines the acoustic radial distance model and the local statistical features of a single frame image. Through cascaded matrix operations in pure mathematical form and adaptive soft threshold mapping, it achieves a perfect balance between reverberation suppression and weak target protection with extremely low computing power.

[0008] The technical solution adopted by this filtering method to solve the technical problem is as follows: a filtering method based on forward-looking sonar, which specifically includes the following steps: Step S1, Modeling and Mask Mapping of the Physical Detection Area of ​​Sonar: Acquire the original video frame image of the forward-looking sonar, construct a mask in the Cartesian coordinate system according to the physical parameters of the sonar sensor, and simultaneously calculate the radial distance mapping map of the real physical corresponding effective acoustic pixels within the mask; wherein, the physical parameters include, but are not limited to, beam opening angle, detection radius and installation attitude. Step S2, First-level global adaptive benchmark extraction and coarse filtering: Extract the sonar echo intensity within the mask, take the global maximum echo intensity in the effective detection area as the benchmark, and calculate the global dynamic threshold in combination with the preset global environment configuration ratio coefficient; use the logistic regression smoothing function to construct the first-level weight map to initially suppress global speckle noise and bottom noise caused by suspended matter in water. Step S3, Acoustic Attenuation Sensing and Radial Block Division: Based on the physical characteristics of underwater acoustic signal attenuation with distance, and using the radial distance mapping obtained in Step S1, the effective acoustic region of the sector is divided radially from near to far. n A separate sector-shaped block region; Step S4, Second-level Non-uniform Sensitivity Local Fine Filtering: For each segmented interval divided in Step S3, calculate the local maximum echo intensity and introduce a set of lengths... n The non-uniform sensitivity ratio array is used to calculate the local absolute threshold of the corresponding block interval; the local weight in each block interval is calculated using the logistic regression smoothing function, and the two blocks are spliced ​​together to obtain the second-level segmented weight map. Step S5, Spatial Smoothing Fusion and Image Reconstruction: Gaussian smoothing filtering is applied to the stitched second-level segmented weight map to eliminate weight step abrupt changes at the block boundaries; then, the smoothed final weight map is multiplied with the original sonar image to reconstruct and output the denoised forward-looking sonar image.

[0009] The technical solution adopted by this filtering device to solve the technical problem is as follows: a filtering device based on forward-looking sonar, used to implement the above-mentioned filtering method, including: Physical detection area modeling module: used to analyze the physical parameters of sonar equipment's attitude deflection angle and opening angle, and generate a geometric sector mask and a real physical radial distance mapping map; Global adaptive coarse filtering module: used to perform basis statistics based on the global extrema of the effective detection area, calculate the first global dynamic threshold and generate a coarse filtering weight distribution map with smooth transition; Radial array block fine filtering module: It is used to divide the entire image into multiple block intervals radially according to the attenuation characteristics of physical sound waves with distance, and to calculate the local adaptive absolute threshold and corresponding local weight for each block interval using a non-uniform sensitivity ratio array. Boundary smoothing and reconstruction fusion module: used to smoothly stitch together the weight maps of each segment and perform dot product fusion with the original image to output a clear sonar image with high fidelity target.

[0010] Compared with related technologies, the advantages of this filtering method and device are: First, a radial block adaptive gain mechanism based on the physical characteristics of sonar was constructed: In view of the physical characteristics of forward-looking sonar signals attenuating with propagation distance, with strong energy in the near field and weak energy in the far field, the mechanism breaks through the limitation of traditional algorithms that use a uniform threshold to process the entire frame image; and a radial block strategy is provided, which can manually or dynamically configure the threshold ratio of different distance intervals to achieve strong suppression of near-field interference noise while maximizing the preservation of the characteristics of weak targets in the far field, thus solving the problem of noise distortion caused by drastic changes in the dynamic range of sonar images.

[0011] Second, a lightweight processing architecture was established to implement cascaded coarse and fine filtering: a two-stage linked filtering process was designed to obtain high-quality images with extremely low computational power consumption. The first stage performs fast coarse filtering through global adaptive statistics to extract background benchmarks and initially suppress global speckle noise; the second stage introduces geometrically aware radial block segmentation for fine filtering, and achieves smooth weight transition through a logistic regression smoothing function. This architecture eliminates the need for complex iterative calculations or model inference, greatly reducing the time and space complexity of the algorithm and ensuring real-time deployment at the edge of underwater unmanned platforms.

[0012] Third, it solves the problem of geometric consistency denoising in irregular fan-shaped regions: For the fan-shaped detection field of view unique to forward-looking sonar, a set of precise geometric mask and radial distance mapping algorithms is defined. By fusing image processing in the Cartesian coordinate system with acoustic beam geometry in polar coordinates, the filtering operation can be accurately applied to the effective acoustic region and supports various complex conditions such as mirror flipping and background preservation. This ensures that the edges of the denoised image are smooth and free of geometric distortion, providing high-fidelity input for subsequent target recognition and path planning.

[0013] Preferably, the formula for calculating the radial distance mapping in step S1 is: For any pixel in the image matrix ( x, y ), its radial distance R from the sonar beam transmission center x, y ) and its deflection relative to the beam center axis θ ( x, y )satisfy: in,( c x ,c y () represents the coordinates of the forward-looking sonar probe relative to the origin within the current image frame; y' Compensation coordinates are introduced to account for the potential physical upside-down installation characteristics of forward-looking sonar image sensors. y' = H - y+1, H The physical total height of the image; the fan-shaped mask is defined by satisfying R. min ≤R( x, y )≤R max and θ min ≤ θ ( x, y )≤ θ max It consists of a set of pixels.

[0014] Preferably, the first-level coarse filtering process in step S2 uses the following logistic regression smoothing function to calculate the continuous global smoothing weights. W g ( x, y ): in, I ( x, y () indicates the original sonar echo intensity within the mask; TH g The dynamic threshold for global acoustic noise floor is determined by the maximum intensity value (max) of all pixels within the mask. I Multiply by the global environment configuration scaling factor P g get; k g The preset first-level transition smoothing coefficient reflects the sharpness of the signal at the threshold boundary.

[0015] Preferably, the non-uniform sensitivity ratio array introduced in step S3 is represented as follows: P local = P 1, P 2,..., P n Its length is strictly equal to the number of partitioned intervals along the radial distance. n, P i The first division along the radial distance i The sensitivity ratio values ​​for each segmented interval; based on the physical law that forward-looking sonar exhibits strong reverberation in near-field water and rapid attenuation of sound wave energy in the far-field, the ratio values ​​in the array gradually decrease along the sonar detection radial direction, satisfying: .

[0016] Preferably, the local fine filtration process in step S4 includes the following physical steps: For those classified as the first i A sub-mask is used to extract the local maximum intensity value of pixels within a radial fan-ring interval. max ( Ii ); Combining the array described in step S4, calculate the first... i Adaptive absolute threshold for each block interval TH i = max ( I i )× P i The continuous smoothing weights for this block interval are calculated using the following logistic regression smoothing function. W l ( x , y ): in, k l The preset second-level transition smoothing coefficient, and k l < k g .

[0017] Preferably, the spatial smoothing fusion in step S5 uses Gaussian filtering to process the stitched second-level segmented weight map: in, G ( x , y , σ ) is a Gaussian convolution kernel, σ To control the physical standard deviation of smoothness, This represents a two-dimensional convolution operation.

[0018] Preferably, the image reconstruction in step S5 further includes external background mask logic processing: For the non-acoustic effective areas not covered by the mask generated in step S1, forced weight values ​​are assigned to these pixels. If configured for pure black mode, their reconstruction weight is always 0; if configured to preserve the background, their reconstruction weight is always 1. The final output formula for the reconstructed image is: .

[0019] The technical solution adopted by this storage medium to solve the technical problem is as follows: a storage medium storing a computer-executable program, which, when executed by an underwater device processor, implements the above-mentioned filtering method.

[0020] The technical solution adopted by this underwater vehicle to solve the technical problem is as follows: an underwater vehicle, including a forward-looking sonar hardware system and an airborne edge computing device, wherein the airborne edge computing device is equipped with the aforementioned filtering device; the filtering device directly accesses the forward-looking sonar continuous frames from underwater acoustic communication or a cable channel, and realizes real-time noise reduction and target enhancement processing with low computing power and zero latency at the device end, providing high-fidelity environmental data for subsequent automatic obstacle avoidance or path planning modules. Attached Figure Description

[0021] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0022] First, those skilled in the art should understand that these embodiments are merely used to explain the technical principles of the embodiments of this application and are not intended to limit the scope of protection of the embodiments of this application. Those skilled in the art can make adjustments as needed to adapt to specific application scenarios.

[0023] This invention aims to solve the technical problems existing in forward-looking sonar image denoising techniques, such as the loss of far-field targets due to a one-size-fits-all approach to global thresholding, the inability of traditional spatial filtering to adapt to the fan-shaped geometry of sonar, and the difficulty of deploying highly complex algorithms in real-time on underwater edge devices. Specifically, this invention strives to achieve the following objectives: First, a radial segmentation adaptive gain mechanism based on the physical characteristics of sonar is constructed: Addressing the physical characteristic of forward-looking sonar signals attenuating with propagation distance, with strong near-field energy and weak far-field energy, this invention overcomes the limitation of traditional algorithms that use a uniform threshold to process the entire frame. A radial segmentation strategy is provided, which, by manually or dynamically configuring the threshold ratio (Percentage Array) for different distance intervals, achieves strong suppression of near-field interference noise while maximizing the preservation of features of weak far-field targets, thus solving the noise distortion problem caused by drastic changes in the dynamic range of sonar images. Second, a lightweight processing architecture of cascaded coarse and fine filtering is implemented: To obtain high-quality images with extremely low computational power consumption, this invention designs a two-stage linked filtering process. The first stage performs rapid coarse filtering through global adaptive statistics to extract background benchmarks and initially suppress global speckle noise; the second stage introduces geometrically aware radial segmentation for fine filtering, achieving a smooth transition of weights through a logistic regression function (Sigmoid-like Softness). This architecture eliminates the need for complex iterative calculations or model inference, significantly reducing the algorithm's temporal and spatial complexity and ensuring real-time deployment at the edge of underwater unmanned platforms (AUVs / ROVs). Third, it addresses the problem of geometrically consistent denoising within irregular fan-shaped regions: A precise geometric mask and radial distance mapping (R-Map) algorithm is defined for the fan-shaped detection field of view unique to forward-looking sonar. By fusing image processing in Cartesian coordinates with acoustic beam geometry in polar coordinates, filtering operations can be precisely applied to the effective acoustic region. It also supports various complex conditions such as upside-down and background-keeping, ensuring smooth edges and no geometric distortion in the denoised image, providing high-fidelity input for subsequent target recognition and path planning.

[0024] Specifically, the present invention provides a lightweight cascaded block adaptive filtering method based on forward-looking sonar, comprising the following steps: Step S1: Modeling and Masking of the Physical Detection Area of ​​Sonar. Acquire the original video frame images of the forward-looking sonar. Based on the physical parameters such as the beam opening angle, detection radius and installation attitude of the sonar sensor, construct a mask (Sector Mask, two-dimensional geometric mask) in the Cartesian coordinate system, and simultaneously calculate the radial distance mapping map (R-map) of the real physical corresponding effective acoustic pixels within the mask.

[0025] The formula for calculating the true physical radial distance map (R-map) is: for any pixel in the image matrix ( x, y ), its radial distance R from the sonar beam transmission center x, y ) and its deflection relative to the beam center axisθ ( x, y )satisfy: (1) (2) In equations (1) and (2), ( c x ,c y The coordinates of the forward-looking sonar probe relative to the origin within the current image frame are the acoustic emission point. These origin coordinates are calculated from the offset of the probe's physical installation. y' Compensation coordinates are introduced to account for the potential upside-down physical mounting characteristics of forward-looking sonar image sensors, i.e. y' = H - y +1, in the formula H Let R be the physical total height of the image. The sector mask is defined by satisfying R... min ≤R( x, y )≤R max and θ min ≤ θ ( x, y )≤ θ max It consists of a set of pixels, where R min R is the minimum radial distance from a pixel in the image matrix to the center of the sonar beam. max This represents the maximum radial distance from each pixel in the image matrix to the center of the sonar beam transmission. θ min This represents the minimum offset angle of a pixel in the image matrix relative to the beam center axis. θ max This represents the maximum deflection angle of a pixel in the image matrix relative to the beam center axis.

[0026] Step S2: First-level global adaptive benchmark extraction and coarse filtering. The sonar echo intensity within the mask is extracted. Using the global maximum echo intensity (Intensity) within the effective detection area as the benchmark, and combined with a preset global environment configuration scaling factor, a global dynamic threshold is calculated. A first-level weight map is constructed using a logistic regression smoothing function to initially suppress global speckle noise and background noise caused by suspended matter in the water.

[0027] The first-level global coarse filtering process uses the following logistic regression smoothing function to calculate the continuous global smoothing weights. W g ( x, y ): (3) (4) In equations (3) and (4), I ( x, y () indicates the original sonar echo intensity within the mask; TH g The dynamic threshold for global acoustic noise floor is determined by the maximum intensity value of all pixels within the mask. [[ID=�2]]max ( I Multiply by the global environment configuration scaling factor P g get; k g The preset first-level transition smoothing coefficient reflects the sharpness of the signal at the threshold boundary. This logistic regression smoothing function can retain high-intensity main target echoes while continuously and gently filtering out low-intensity speckle and suspended particle noise.

[0028] Step S3: Acoustic Attenuation Sensing and Radial Block Division. Based on the physical characteristics of underwater acoustic signal attenuation with distance, the effective acoustic region in the sector is divided radially from near to far using the radial distance mapping (R-map) obtained in Step S1. n Each independent sector-shaped block interval (also called a local interval).

[0029] The introduced non-uniform sensitivity ratio array is represented as follows: P local = P 1, P 2,..., P n Its length is strictly equal to the number of partitioned intervals along the radial distance. n, P i The first division along the radial distance i The sensitivity ratio values ​​for each segmented interval; based on the physical law that forward-looking sonar exhibits strong reverberation in near-field water and rapid attenuation of sound wave energy in the far-field, the ratio values ​​in the array gradually decrease along the sonar detection radial direction, satisfying: (5) This allows for higher threshold suppression of near-field reverberation interference and more lenient signal-to-noise ratio extraction conditions for weak targets in the far field.

[0030] Step S4: Second-level non-uniform sensitivity local fine filtering. For each segmented interval divided in step S3, the local maximum echo intensity is calculated, and a set of lengths is introduced. n The non-uniform sensitivity ratio array is used to calculate the local absolute threshold of the corresponding block interval; the local weight within each block interval is calculated using the logistic regression smoothing function, and the results are concatenated to obtain the second-level segmented weight map. W local_merged ( x, y).

[0031] The local fine filtration process includes the following physical steps: for the part divided into the first... i A sub-mask is used to extract the local maximum intensity value of pixels within a radial fan-ring interval. max ( I i ); Combining the array described in step S4, calculate the first... i Adaptive absolute threshold for each block interval TH i = max ( I i )× P i The continuous smoothing weights for this block interval are calculated using the following logistic regression smoothing function. W l ( x , y ): (6) In equation (6), k l The preset second-level transition smoothing coefficient, and k l < k g This allows for the preservation of richer target morphological edges during local fine filtering compared to global coarse filtering.

[0032] Step S5: Spatial Smoothing Fusion and Image Reconstruction. Gaussian smoothing filtering is applied to the stitched second-level segmented weight map to eliminate abrupt weight jumps at block boundaries; subsequently, the smoothed final weight map (… W final ( x, y ), W final , W f ) with the original sonar image I ( x, y Perform dot product mapping to reconstruct and output the denoised forward-looking sonar image. I out ( x, y ).

[0033] In this process, spatial smoothing fusion employs Gaussian filtering on the stitched second-level segmented weight map. W local_merged ( x, y The formula for processing is: (7) In equation (7), G (x , y , σ ) is a Gaussian convolution kernel, σ It should be noted that there may be some inaccuracies in the original text, especially in the line where "max " seems to have a misspelling as "�2". This has been maintained as is in the translation for the purpose of following the rules. To control the physical standard deviation of smoothness, This represents a two-dimensional convolution operation. This step eliminates the discontinuity in the weight space caused by radial block division, preventing concentric circular structural artifacts from appearing in the reconstructed image after denoising.

[0034] In addition, the image reconstruction in step S5 also includes external background mask logic processing: for non-acoustic effective areas not covered by the mask generated in step S1, forced weight values ​​are assigned to these pixels. If configured for pure black mode, their reconstruction weight is always set to 0 to eliminate the fixed border interference introduced by the original device; if configured to retain the background, their reconstruction weight is always set to 1. The final output formula for the reconstructed image is: .

[0035] A lightweight cascaded block adaptive filtering device based on forward-looking sonar for implementing the above filtering method includes: Physical detection area modeling module: used to analyze the physical parameters of sonar equipment's attitude deflection angle and opening angle, and generate a geometric sector mask and a real physical radial distance mapping map; Global adaptive coarse filtering module: used to perform basis statistics based on the global extrema of the effective detection area, calculate the first global dynamic threshold and generate a coarse filtering weight distribution map with smooth transition; Radial array block fine filtering module: It is used to divide the entire image into multiple block intervals radially according to the attenuation characteristics of physical sound waves with distance, and to calculate the local adaptive absolute threshold and corresponding local weight for each block interval using a non-uniform sensitivity ratio array. Boundary smoothing and reconstruction fusion module: used to smoothly stitch together the weight maps of each segment and perform dot product fusion with the original image to output a clear sonar image with high fidelity target.

[0036] A computer-readable storage medium having a computer program stored thereon, which, when executed by an underwater device processor, implements the filtering method described above.

[0037] An underwater vehicle includes a forward-looking sonar hardware system and an airborne edge computing device. The airborne edge computing device is equipped with the aforementioned filtering device. The filtering device directly accesses continuous frames from the forward-looking sonar via underwater acoustic communication or a wired channel, and performs low-computing-power, zero-latency real-time noise reduction and target enhancement processing at the device end, providing high-fidelity environmental data for subsequent automatic obstacle avoidance or path planning modules.

[0038] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A filtering method based on forward-looking sonar, characterized in that, Specifically, the steps include the following: Step S1, Modeling and Mask Mapping of the Physical Detection Area of ​​Sonar: Acquire the original video frame image of the forward-looking sonar, construct a mask in the Cartesian coordinate system according to the physical parameters of the sonar sensor, and simultaneously calculate the radial distance mapping map of the real physical corresponding effective acoustic pixels within the mask; wherein, the physical parameters include, but are not limited to, beam opening angle, detection radius and installation attitude. Step S2, First-level global adaptive benchmark extraction and coarse filtering: Extract the sonar echo intensity within the mask, take the global maximum echo intensity in the effective detection area as the benchmark, and calculate the global dynamic threshold in combination with the preset global environment configuration ratio coefficient; use the logistic regression smoothing function to construct the first-level weight map to initially suppress global speckle noise and bottom noise caused by suspended matter in water. Step S3, Acoustic Attenuation Sensing and Radial Block Division: Based on the physical characteristics of underwater acoustic signal attenuation with distance, and using the radial distance mapping obtained in Step S1, the effective acoustic region of the sector is divided radially from near to far. n Each independent sector-shaped block interval; Step S4, Second-level Non-uniform Sensitivity Local Fine Filtering: For each segmented interval divided in Step S3, calculate the local maximum echo intensity and introduce a set of lengths... n The non-uniform sensitivity ratio array is used to calculate the local absolute threshold of the corresponding block interval; the local weight in each block interval is calculated using the logistic regression smoothing function, and the two blocks are spliced ​​together to obtain the second-level segmented weight map. Step S5, Spatial Smoothing Fusion and Image Reconstruction: Gaussian smoothing filtering is applied to the stitched second-level segmented weight map to eliminate weight step abrupt changes at the block boundaries; then, the smoothed final weight map is multiplied with the original sonar image to reconstruct and output the denoised forward-looking sonar image.

2. The filtering method based on forward-looking sonar according to claim 1, characterized in that, The formula for calculating the radial distance mapping in step S1 is as follows: For any pixel in the image matrix ( x,y ), its radial distance R from the sonar beam transmission center x,y ) and its deflection relative to the beam center axis θ ( x,y )satisfy: in,( c x ,c y () represents the coordinates of the forward-looking sonar probe relative to the origin within the current image frame; y' Compensation coordinates are introduced to account for the potential physical upside-down installation characteristics of forward-looking sonar image sensors. y' = H - y +1, H The physical total height of the image; the fan-shaped mask is defined by satisfying R. min ≤R( x,y )≤R max and θ min ≤ θ ( x,y )≤ θ max It consists of a set of pixels.

3. The filtering method based on forward-looking sonar according to claim 1, characterized in that, The first-level coarse filtering process in step S2 uses the following logistic regression smoothing function to calculate the continuous global smoothing weights. W g ( x,y ): in, I ( x,y () indicates the original sonar echo intensity within the mask; TH g The dynamic threshold for global acoustic noise floor is determined by the maximum intensity value (max) of all pixels within the mask. I Multiply by the global environment configuration scaling factor P g get; k g The preset first-level transition smoothing coefficient reflects the sharpness of the signal at the threshold boundary.

4. The filtering method based on forward-looking sonar according to claim 1, characterized in that, The non-uniform sensitivity ratio array introduced in step S3 is represented as follows: P local = P 1, P 2,..., P n Its length is strictly equal to the number of partitioned intervals along the radial distance. n, P i The first division along the radial distance i The sensitivity ratio values ​​for each segmented interval; based on the physical law that forward-looking sonar exhibits strong reverberation in near-field water and rapid attenuation of sound wave energy in the far-field, the ratio values ​​in the array gradually decrease along the sonar detection radial direction, satisfying: .

5. The filtering method based on forward-looking sonar according to claim 4, characterized in that, The local fine filtration process in step S4 includes the following physical steps: For those classified as the first i A sub-mask is used to extract the local maximum intensity value of pixels within a radial fan-ring interval. max ( I i ); Combining the array described in step S4, calculate the first... i Adaptive absolute threshold for each block interval TH i = max ( I i )× P i The continuous smoothing weights for this block interval are calculated using the following logistic regression smoothing function. W l ( x , y ): in, k l The preset second-level transition smoothing coefficient, and k l < k g .

6. The filtering method based on forward-looking sonar according to claim 1, characterized in that, In step S5, spatial smoothing fusion uses Gaussian filtering to process the stitched second-level segmented weight map. in, G ( x , y , σ ) is a Gaussian convolution kernel, σ To control the physical standard deviation of smoothness, This represents a two-dimensional convolution operation.

7. The filtering method based on forward-looking sonar according to claim 1, characterized in that, The image reconstruction in step S5 also includes external background mask logic processing: For the non-acoustic effective areas not covered by the mask generated in step S1, forced weight values ​​are assigned to these pixels. If configured for pure black mode, their reconstruction weight is always 0; if configured to preserve the background, their reconstruction weight is always 1. The final output formula for the reconstructed image is: 。 8. A forward-looking sonar-based filtering device for implementing the filtering method according to any one of claims 1 to 7, characterized in that, include: Physical detection area modeling module: used to analyze the physical parameters of sonar equipment's attitude deflection angle and opening angle, and generate a geometric sector mask and a real physical radial distance mapping map; Global adaptive coarse filtering module: used to perform basis statistics based on the global extrema of the effective detection area, calculate the first global dynamic threshold and generate a coarse filtering weight distribution map with smooth transition; Radial array block fine filtering module: It is used to divide the entire image into multiple block intervals radially according to the attenuation characteristics of physical sound waves with distance, and to calculate the local adaptive absolute threshold and corresponding local weight for each block interval using a non-uniform sensitivity ratio array. Boundary smoothing and reconstruction fusion module: used to smoothly stitch together the weight maps of each segment and perform dot product fusion with the original image to output a clear sonar image with high fidelity target.

9. A storage medium having stored thereon a computer executable program, characterized in that: When the program is executed by the underwater device processor, it implements the filtering method as described in any one of claims 1 to 7.

10. An underwater vehicle characterized by: It includes a forward-looking sonar hardware system and an airborne edge computing device, wherein the airborne edge computing device is equipped with a filtering device as described in claim 8; the filtering device directly accesses the forward-looking sonar continuous frames from underwater acoustic communication or a cable channel, and performs low-computing-power, zero-latency real-time noise reduction and target enhancement processing at the device end, providing high-fidelity environmental data for subsequent automatic obstacle avoidance or path planning modules.

Images