Image enhancement type marine floating object ai visual recognition method and system based on multispectral fusion and light suppression

Abstract

The present application relates to the technical field of visual identification of marine floating objects, in particular to an image-enhanced marine floating object AI visual identification method and system based on multispectral fusion and reflection suppression, comprising the following steps: multispectral data acquisition and registration; multispectral feature decoupling and reflection suppression; multispectral feature fusion and target detection; trajectory verification and result output. The present application fuses the visible light and near-infrared dual-band spectral absorption differences and the sea surface spatial texture information to construct a roughness perception factor, which is used to guide the multispectral image decoupling into three feature flows of floating object diffuse reflection body, specular reflection and white wave scattering, and the roughness is used to realize the dynamic weight reorganization of each feature flow, so as to suppress the sea surface reflection and wave disturbance from the imaging source, while the details of the floating object are completely retained, the defects that a single spectrum cannot simultaneously eliminate interference and retain targets are broken through, and the image clarity under complex sea conditions such as strong light and broken waves is improved.

Description

Technical Field

[0001] This invention relates to the field of visual recognition technology for marine floating objects, and more specifically, to an image-enhanced AI visual recognition method and system for marine floating objects based on multispectral fusion and reflection suppression. Background Technology

[0002] With the increasing demand for automated monitoring in maritime regulation, waterway safety, and marine ecological governance, floating object detection systems based on optical imaging and artificial intelligence are gradually becoming the mainstream configuration. These systems analyze sea surface images in real time using high-resolution cameras and embedded processing platforms, separating floating objects from the background by utilizing the differences in optical properties between targets and water bodies in the visible or infrared bands. They have become a standard technical means for maritime safety and environmental protection operations.

[0003] Among the solutions that have been deployed on a large scale, two types of technologies are particularly typical. One type is a deep neural network detection system based on visible light images. This system continuously acquires sea surface images through shore-based or shipborne cameras, and the trained detection network directly outputs the target bounding box and confidence score. This solution has controllable hardware costs, strong daytime detail resolution capabilities, and is widely used in port and waterway monitoring. The other type is a floating object sensing system based on infrared thermal imaging. This system utilizes the difference in thermal radiation between the target and the water body in the long-wave infrared band for imaging. It has good adaptability to backlight and shadow scenes and stable detection capabilities at night and in low-light conditions.

[0004] Although the aforementioned solutions have been deployed, common technical bottlenecks remain when facing complex sea surface optical environments. Sea surface imaging is inevitably affected by specular reflection caused by solar flares and white wave scattering from broken waves, which manifests as high-intensity, cluttered bright spots in the image, severely obscuring or distorting the target's features. While infrared thermal imaging can avoid some visible light reflection problems, its spatial resolution is low, texture details are insufficient, and its effectiveness in detecting small floating objects and targets with slight temperature differences from the background is limited. Furthermore, it cannot utilize the rich color and texture information of the visible light band. This leads to two prominent technical problems: first, under conditions of strong reflection and wave scattering, the false alarm rate and false negative rate of the detection system remain high, making it difficult to guarantee alarm reliability; second, a single spectral mode cannot simultaneously address reflection suppression and target detail preservation, resulting in insufficient adaptability of the system to varying ocean lighting conditions. These limitations directly compress the system's effective operating window in high sea states, oblique lighting, and open sea areas, increasing deployment and maintenance costs. In view of this, we propose an image-enhanced AI visual recognition method and system for marine floating objects based on multispectral fusion and reflection suppression. Summary of the Invention

[0005] The purpose of this invention is to provide an image-enhanced AI visual recognition method and system for marine floating objects based on multispectral fusion and reflection suppression, so as to solve the problems mentioned in the background art, such as insufficient detection reliability under strong reflection and wave scattering conditions, and the difficulty of simultaneously achieving reflection suppression and target detail preservation with a single spectral mode.

[0006] To address the aforementioned technical problems, one objective of this invention is to provide an image-enhanced AI visual recognition method for marine floating objects based on multispectral fusion and reflection suppression, comprising: S1. Multispectral data acquisition and registration: Acquire visible light and near-infrared images, and based on multi-scale phase consistency, perform adaptive spatial registration of the near-infrared images with the visible light images as the reference to obtain a multispectral registered image set. S2. Multispectral Feature Decoupling and Reflection Suppression: Based on the local absorption differences and texture gradients of different spectral bands in the multispectral registered image set, a roughness perception factor is constructed. Using the roughness perception factor as a guiding condition, the multispectral registered image set is decoupled into diffuse reflection body feature flow, specular reflection environment feature flow, and white wave scattering interference feature flow. Based on the roughness perception factor, dynamic weight allocation and feature recombination are performed on the diffuse reflection body feature flow, specular reflection environment feature flow, and white wave scattering interference feature flow to obtain the reflection suppression enhanced image and the roughness distribution map corresponding to the roughness perception factor. S3. Multi-scale feature fusion and target detection: The reflection suppression enhancement image and roughness distribution map are input into the multi-scale feature fusion neural network model. The roughness distribution map is used to guide the multi-scale feature fusion neural network model to reduce the attention weight of the high roughness area of ​​the sea surface and output the target detection result of floating objects on the sea surface. S4. Trajectory Verification and Result Output: Obtain target detection results for multiple consecutive frames. Based on the temporal continuity of physical kinematic constraints and spectral features, perform trajectory association and consistency verification on the target detection results for multiple consecutive frames to obtain the final sea surface floating object detection results.

[0007] As a further improvement to this technical solution, the multispectral data acquisition and registration in S1 includes the following steps: S1.1. Through the synchronous triggering mechanism of the multispectral imaging system, visible light band images and near-infrared band images are acquired simultaneously to ensure that the visible light images and near-infrared images have the same timestamp and overlapping spatial coverage area. S1.2 Perform nonlocal mean filtering on the visible light image and the near-infrared image respectively, and then correct them by adaptive histogram equalization to output the preprocessed visible light image and near-infrared image. S1.3. Multi-scale convolution operations are performed on the preprocessed visible light image and near-infrared image using a multi-scale Log-Gabor filter bank to calculate the phase consistency response at each scale, thereby obtaining the multi-scale phase consistency feature map of the visible light image and the multi-scale phase consistency feature map of the near-infrared image. S1.4 Based on the multi-scale phase consistency feature map of visible light image and the multi-scale phase consistency feature map of near-infrared image, the feature point similarity is calculated using the mutual information maximization criterion, and a dynamic matching threshold is set to filter reliable feature point pairs to form a feature point matching set; S1.5. Use the RANSAC algorithm to remove mismatched points from the feature point matching set and estimate the rotation angle, scaling ratio and offset of the affine transformation matrix. S1.6 Apply the affine transformation matrix to the near-infrared image for bilinear interpolation resampling to obtain the registered near-infrared image; treat the registered near-infrared image and the visible light image as two independent channels to form a multispectral registered image set.

[0008] As a further improvement to this technical solution, the multispectral feature decoupling and reflection suppression in S2 includes the following steps: S2.1 Extract the spectral absorption features and spatial gradient features of the multispectral registered image set, and generate a roughness perception factor through multi-channel feature fusion and edge-preserving smoothness constraints; S2.2. The roughness perception factor is injected as a conditional modulation parameter into three parallel convolutional neural network branches. Each branch first performs an invertible wavelet transform on the multispectral registered image set to decompose the image into a frequency domain multi-scale space. Then, guided by the roughness perception factor, through the feature space orthogonalization constraint, it decouples and outputs the diffuse reflection body feature flow, the specular reflection environment feature flow, and the white wave scattering interference feature flow in the frequency domain. S2.3. The numerical distribution of roughness perception factor is transformed into three sets of basic weights using a piecewise mapping function. After adaptive correction of the three sets of basic weights by combining the channel attention mechanism, the three feature streams are weighted and fused to obtain the recombined feature representation. S2.4 Perform frequency domain inverse transform on the reconstructed feature representation to reconstruct the reflection suppression and enhancement image, and output the roughness distribution map after performing gray quantization and spatial alignment verification on the roughness perception factor.

[0009] As a further improvement to this technical solution, the generation of the roughness perception factor in S2.1 includes the following steps: S2.11 Calculate the spectral absorption difference pixel by pixel for the visible light band image and the near-infrared band image in the multispectral registered image set to obtain the local absorption difference feature map; S2.12. Perform gradient operations on the multispectral registered image set along the horizontal and vertical directions respectively to synthesize the texture gradient magnitude map and the texture gradient direction map; S2.13. Normalize the local absorption difference feature map and the texture gradient magnitude map, and then weight and fuse them according to a preset ratio to generate an initial roughness response map. S2.14. Guided by the texture gradient direction map, perform anisotropic diffusion smoothing on the initial roughness response map to output a spatially continuous roughness perception factor.

[0010] As a further improvement to this technical solution, the output diffuse reflection body characteristic flow, specular reflection environment characteristic flow, and white wave scattering interference characteristic flow in S2.2 include the following steps: S2.21. The roughness perception factor is injected as a conditional input into the diffuse reflection ontology feature extraction branch, and the diffuse reflection ontology feature flow is separated from the multispectral registered image set through spectral-spatial joint convolution. S2.22. The roughness perception factor is injected as a conditional input into the specular reflection environment feature extraction branch. The specular reflection environment feature flow is separated from the multispectral registered image set by using the intensity correlation model based on the difference in spectral intensity between visible light and near-infrared bands. S2.23. The roughness perception factor is injected as a conditional input into the white wave scattering interference feature extraction branch, and the white wave scattering interference feature flow is separated from the multispectral registered image set based on the high frequency spatial fluctuation characteristics. S2.24 Calculate the inner product of the feature space between each pair of diffuse reflection body feature flow, specular reflection environment feature flow and white wave scattering interference feature flow. When the inner product exceeds the orthogonality threshold, perform Gram-Schmidt orthogonalization correction.

[0011] As a further improvement to this technical solution, obtaining the recombination feature representation in S2.3 includes the following steps: S2.31. Divide the roughness perception factor into three levels: low roughness, medium roughness, and high roughness, and obtain the first basic weight of the diffuse reflection body feature flow in each level by looking up the table. S2.32. Based on the same numerical range, look up the table to obtain the second basic weight of the specular reflection environment characteristic flow at each level; S2.33. Based on the same numerical range, look up the table to obtain the third basic weight of the white wave scattering interference characteristic flow at each level; S2.34 Multiply the first basic weight, the second basic weight, and the third basic weight by the channel attention coefficient of the corresponding feature stream to obtain the final dynamic weight, and perform weighted summation on the diffuse reflection body feature stream, the specular reflection environment feature stream, and the white wave scattering interference feature stream to output the reconstructed feature representation.

[0012] As a further improvement to this technical solution, the output of the roughness distribution map in step S2.4 includes the following steps: S2.41. Perform an inverse wavelet transform on the recombined feature representation to restore it to the spatial domain and generate a preliminary reflection suppression and enhancement image; S2.42. Linearly quantize and encode the roughness perception factor according to the preset gray level to generate an initial roughness distribution map with the same resolution as the initial reflection suppression and enhancement image. S2.43. Compare the coordinate indices of the preliminary reflection suppression enhancement image and the initial roughness distribution map pixel by pixel, and perform sub-pixel level resampling correction on the offset pixels to ensure that the two are strictly aligned in space. S2.44. Pack the aligned reflection suppression and enhancement image and roughness distribution map and output them to S3.

[0013] As a further improvement to this technical solution, the multi-scale feature fusion and target detection in S3 includes the following steps: S3.1 Input the reflection suppression and enhancement image into the multi-scale feature fusion neural network model, and extract multi-scale spatial feature maps layer by layer through the multi-scale feature fusion neural network model; S3.2 Calculate the attention attenuation coefficient at each spatial location based on the roughness distribution map and generate a roughness suppression weight map, wherein the attention attenuation coefficient corresponding to the high roughness region of the sea surface is smaller than the attention attenuation coefficient corresponding to the low roughness region of the sea surface. S3.3. Use the roughness suppression weight map to perform layer-by-layer weighting processing on the multi-scale spatial feature map, reduce the attention weight of the multi-scale feature fusion neural network model to the high roughness region of the sea surface, and obtain the roughness-suppressed multi-scale spatial feature map. S3.4 Perform cross-level multi-scale feature fusion on the roughness-suppressed multi-scale spatial feature map to generate a fused feature map; S3.5. Based on the fused feature map, perform target regression and classification calculation, and output the target detection results of floating objects on the sea surface. The target detection results include the bounding box coordinates, category label and confidence score of the floating objects.

[0014] As a further improvement to this technical solution, the trajectory verification and result output in S4 includes the following steps: S4.1 Obtain the target detection results of multiple consecutive frames, extract the bounding box position, motion state vector and spectral features of each detected target in each frame, and construct a candidate target set; S4.2. Based on the motion state vector, establish a physical kinematic state transition model to predict the bounding box position of each detected target in the candidate target set in the next frame, and obtain the predicted position of each detected target. S4.3 Calculate the kinematic association cost between the predicted position and the actual detection position of the detected target in adjacent frames. When the kinematic association cost is lower than the preset association threshold, associate the corresponding detection results with the same trajectory sequence. S4.4 Calculate the temporal similarity between the spectral features of the detected targets in adjacent frames of the same trajectory sequence. When the temporal similarity is lower than the preset continuity threshold, mark the corresponding trajectory sequence as a consistency anomaly. S4.5 Remove trajectory sequences marked as inconsistent, and integrate the remaining trajectory sequences that pass the consistency check to output the final sea surface floating object detection result. The second objective of this invention is to provide an image-enhanced AI visual recognition system for marine floating objects based on multispectral fusion and reflection suppression, for performing the steps of the aforementioned image-enhanced AI visual recognition method for marine floating objects based on multispectral fusion and reflection suppression, including: A multispectral data acquisition and registration unit is used to acquire visible light images and near-infrared images, perform adaptive spatial registration on the near-infrared images, and output a multispectral registered image set; the multispectral data acquisition and registration unit includes: The synchronous acquisition module is used to synchronously trigger the multispectral imaging system to acquire visible light and near-infrared images that are aligned in time and space. The preprocessing module is used to denoise and enhance the contrast of the image; The phase consistency feature extraction module is used to extract multi-scale phase consistency features of images; The feature matching and registration module is used to match based on the features and complete spatial registration; A multispectral feature decoupling and reflection suppression unit, connected to the multispectral data acquisition and registration unit, is used to receive the multispectral registration image set, construct a roughness perception factor, decouple image features and suppress reflection, and output a reflection suppression enhanced image and a roughness distribution map; the multispectral feature decoupling and reflection suppression unit includes: The roughness perception factor generation module is used to fuse spectral absorption differences and spatial gradient information to generate roughness perception factors. The multispectral feature decoupling module is used to decouple the image into diffuse reflection body feature stream, specular reflection environment feature stream, and white wave scattering interference feature stream, guided by the roughness perception factor; The dynamic weight allocation and recombination module is used to dynamically adjust the weights of the three feature streams and recombine them according to the roughness perception factor. The image reconstruction and distribution map output module is used to reconstruct the reflection suppression and enhancement image and output the corresponding roughness distribution map; A multi-scale feature fusion and target detection unit, connected to the multispectral feature decoupling and reflection suppression unit, is used to perform target detection using the reflection suppression enhanced image and roughness distribution map, and outputs the target detection result of floating objects on the sea surface; the multi-scale feature fusion and target detection unit includes: The multi-scale feature extraction module is used to extract multi-scale spatial features of images; The attention weight modulation module is used to reduce the attention weight of high-roughness regions based on the roughness distribution map. The feature fusion and detection output module is used to fuse multi-scale features and output target detection results; A trajectory verification and result output unit, connected to the multi-scale feature fusion and target detection unit, is used to perform trajectory association and consistency verification on target detection results of multiple consecutive frames, and output the final sea surface floating object detection result; the trajectory verification and result output unit includes: The candidate target construction and prediction module is used to extract the target state and predict its location; The trajectory association and verification module is used to perform trajectory association and consistency determination based on kinematic constraints and spectral continuity. The results integration and output module is used to remove abnormal trajectories and output the final floating object detection results.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a roughness perception factor to decouple the multispectral image of the sea surface into three feature flows: diffuse reflection, specular reflection, and white wave scattering. The latter two are dynamically suppressed and recombined, so that reflection and wave interference are effectively separated from the imaging source, improving the image recognizability under complex lighting conditions and providing clear input with controlled interference for subsequent detection. 2. This invention integrates the spectral absorption differences and spatial texture information of the visible and near-infrared bands, which preserves the detailed features of the floating object while suppressing specular reflection and white wave scattering. It overcomes the limitation of a single spectral mode in balancing reflection suppression and target retention, and enhances the system's adaptability and detection robustness under varying marine lighting conditions. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall process flow of the present invention; Figure 2 This is a schematic diagram of the overall system framework of the present invention; The meanings of the labels in the diagram are as follows: 1. Multispectral data acquisition and registration unit; 2. Multispectral feature decoupling and reflection suppression unit; 3. Multi-scale feature fusion and target detection unit; 4. Trajectory verification and result output unit. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0018] like Figure 1 As shown, this embodiment provides an image-enhanced AI visual recognition method for marine floating objects based on multispectral fusion and reflection suppression. This method can be applied to automated visual recognition of targets such as floating debris and algae in marine environmental monitoring. The hardware environment supporting this method includes a multispectral synchronous imaging system composed of visible light and near-infrared cameras, and computing equipment running the recognition algorithm and detection network.

[0019] The core idea of ​​this method is to construct a roughness perception factor by utilizing the absorption differences and texture characteristics of visible light and near-infrared images at different wavelengths. This factor guides the decoupling of multispectral images into diffuse reflection body feature streams, specular reflection environment feature streams, and white wave scattering interference feature streams. Dynamic weight recombination generates a reflection suppression enhancement image and a roughness distribution map. These two images are then jointly input into a multi-scale feature fusion detection network. The roughness distribution map guides the network to reduce attention to high-roughness regions, and physical kinematic constraints and spectral temporal consistency checks are applied to the detection results across multiple consecutive frames to obtain the final floating object detection result. The roughness perception factor and roughness distribution map serve as information links between each step, forming a collaborative processing chain from multispectral acquisition, reflection suppression, target detection, and trajectory verification. This method includes the following steps: S1. Multispectral Data Acquisition and Registration: Acquire visible light and near-infrared images. Based on multi-scale phase consistency, adaptive spatial registration is performed on the near-infrared images using the visible light images as a reference to obtain a multispectral registered image set. This includes the following steps: S1.1. Through the synchronous triggering mechanism of the multispectral imaging system, visible light band images and near-infrared band images are acquired simultaneously to ensure that the visible light images and near-infrared images have the same timestamp and overlapping spatial coverage area. In this step, this embodiment uses a visible light camera and a near-infrared camera to form a multispectral imaging system. The two cameras are mounted on the same rigid bracket, with their optical axes parallel, and arranged side-by-side, either vertically or horizontally adjacent, with their lenses both pointing towards the sea surface area to be monitored. Depending on the monitoring platform, this bracket can be deployed on a shore-based observation tower, the deck of an unmanned surface vessel, or a UAV gimbal, etc. During installation, the focal length and pitch angle of the two cameras are adjusted so that their field of view coverage areas on the sea surface are substantially overlapped; the target overlap rate is generally set above 90%.

[0020] For imaging control, the two cameras synchronize their exposures using a unified hardware trigger signal. This hardware trigger can be provided by an external synchronization signal generator, or one camera can act as the master device, outputting a synchronization pulse to the other. This is to ensure that each pair of visible light and near-infrared images has the exact same timestamp, preventing shifts in wave texture and the position of floating objects between the two images due to asynchronous shooting times. The acquired visible light images will be denoted as follows: Near-infrared images are denoted as .

[0021] S1.2 Perform nonlocal mean filtering on the visible light image and the near-infrared image respectively, and then correct them by adaptive histogram equalization to output the preprocessed visible light image and near-infrared image. In this step, nonlocal mean filtering is first used for denoising: a local image patch is taken centered on each pixel, and the patch is slid across a set search window. The weighted Euclidean distance between each patch and the central patch is calculated to measure the structural similarity. The similarity is used as the weight to perform a weighted average of the pixels within the search window to obtain the denoised image. This step can effectively suppress random noise while preserving object edges and sea surface textures. Subsequently, adaptive histogram equalization is performed on the denoised image. The image is divided into several sub-blocks, and contrast-limited histogram equalization is performed independently within each sub-block to improve local brightness differences caused by uneven lighting or atmospheric scattering. The preprocessed visible light and near-infrared images are denoted as follows: and .

[0022] S1.3. Multi-scale convolution operations are performed on the preprocessed visible light image and near-infrared image using a multi-scale Log-Gabor filter bank to calculate the phase consistency response at each scale, thereby obtaining the multi-scale phase consistency feature map of the visible light image and the multi-scale phase consistency feature map of the near-infrared image. In this step, to extract stable structural features insensitive to illumination changes and contrast reversal, a multi-scale Log-Gabor filter bank is used to calculate the phase consistency feature map. The filter bank covers... Spatial scale and The radial component of the component exhibits a Gaussian distribution on the logarithmic frequency coordinate, while the angular component exhibits a Gaussian distribution on the directional coordinate. In one specific implementation, the number of spatial scales You can choose 3 to 5, the number of directions Four to eight filters can be selected; the center frequencies of the filters are set at equal intervals according to the scale, and the radial and angular bandwidths are taken from conventional empirical values ​​to balance the multi-scale coverage capability and computational efficiency of feature extraction. and After transforming to the frequency domain, multiplying with filters at each scale and direction, and then inversely transforming back to the spatial domain, we obtain the results for each scale. Each direction The response components are used to calculate the position of each pixel. Local energy at the point and local amplitude And synthesize the phase consistency feature map according to the following formula: ; in, pixel position Phase consistency response value at the location; For this location at scale ,direction The difference between the local phase and the scale-averaged phase. As a direction-weighted factor, This is the noise compensation threshold. This means setting negative values ​​to zero. To prevent small constants with denominators of zero, a typical value is 10. -6 Up to 10 -8 The visible light phase consistency feature map was obtained through pixel-by-pixel calculation. Near-infrared phase consistency feature map In the feature map, the response values ​​are higher at the edges and corners of objects, and lower in flat water surfaces.

[0023] S1.4 Based on the multi-scale phase consistency feature map of visible light image and the multi-scale phase consistency feature map of near-infrared image, the feature point similarity is calculated using the mutual information maximization criterion, and a dynamic matching threshold is set to filter reliable feature point pairs to form a feature point matching set; In this step, and Local maxima are detected as candidate feature points. A fixed-size neighborhood window is taken centered on each feature point; in one specific implementation, the neighborhood window can be 7×7 pixels. The phase consistency response values ​​of each pixel within the window are arranged into a one-dimensional vector in a row-major predetermined order, serving as the description vector for that point. For any feature point in visible light... and any feature point in the near-infrared. The mutual information value between the description vectors of two vectors is calculated to measure the degree of similarity in their local structures. Mutual information can be obtained by statistically analyzing the joint gray-level histogram and edge histogram of the two description vectors, and then calculating it using the conventional mutual information formula.

[0024] Considering the significant differences in image texture richness under different sea conditions, this embodiment employs dynamic thresholding to screen reliable matching pairs. The distribution of mutual information values ​​for all candidate feature point pairs in the current frame is statistically analyzed, and the median is calculated. and median absolute deviation Determine the matching threshold as follows: : ; In the formula: The median of the mutual information values ​​for all candidate feature point pairs; The median absolute deviation of mutual information values ​​is the median of the absolute values ​​of the differences between all mutual information values ​​and their median, and is a robust statistic of dispersion. This is the threshold adjustment coefficient, typically ranging from 1.5 to 3.0. Mutual information values ​​exceeding... Feature point pairs included in the matching set .

[0025] S1.5. Use the RANSAC algorithm to remove mismatched points from the feature point matching set and estimate the rotation angle, scaling ratio and offset of the affine transformation matrix. In this step, the matching set There may still be mismatches, requiring the use of the RANSAC algorithm to simultaneously eliminate mismatches and estimate affine transformation parameters. In each iteration, from... Three pairs of non-collinear feature points are randomly selected from the data, and the candidate affine transformation matrix is ​​calculated: ; In the formula: These are the matrix elements representing the linear part of the affine transformation, describing rotation and scaling; , These represent the translation amounts in the horizontal and vertical directions, respectively. Substituting the remaining feature point pairs into this transformation, the Euclidean distance between the transformed near-infrared feature point and its corresponding visible light feature point is calculated. The distance is less than a preset interior point threshold. The points are denoted as interior points. After repeating the above process for a preset number of iterations, the model with the most interior points is selected. In one specific implementation, the preset number of iterations can be between 100 and 500, and the interior point distance threshold is used. The number of interior points can be 3 to 5 pixels. If the number of interior points in the optimal model is lower than the preset minimum number of interior points (e.g., no less than 10 pairs), the registration reliability is deemed insufficient, and the transformation matrix of the previous frame can be used or a registration anomaly flag can be output. The final affine transformation matrix is ​​obtained by resolving the matrix using the least squares method with all its interior points. .

[0026] S1.6 Apply the affine transformation matrix to the near-infrared image for bilinear interpolation resampling to obtain the registered near-infrared image; treat the registered near-infrared image and the visible light image as two independent channels to form a multispectral registered image set.

[0027] Specifically, using the obtained The original near-infrared image Resample to the visible light image coordinate space. Create an image with... Blank images of the same size, for each pixel coordinate ,pass The inverse transform is used to calculate its value in... The corresponding sub-pixel coordinates If the coordinates fall within the valid boundaries of the image, they are obtained by linear interpolation twice, once in the horizontal direction and once in the vertical direction, using the gray values ​​of the surrounding four integer pixels. The pixel value at each location is used; if it exceeds the boundary, a preset padding value is assigned. After pixel-by-pixel processing, a registered near-infrared image that is perfectly aligned with the visible light image space is obtained. .

[0028] Will and Merging along the channel dimension to form a dual-channel multispectral registered image set Then it is sent to step S2 for subsequent multispectral feature decoupling and reflection suppression processing.

[0029] S2. Multispectral Feature Decoupling and Reflection Suppression: Based on the local absorption differences and texture gradients of different spectral bands in the multispectral registered image set, a roughness perception factor is constructed. Guided by this roughness perception factor, the multispectral registered image set is decoupled into diffuse reflection volumetric feature streams, specular reflection environmental feature streams, and white wave scattering interference feature streams. Based on the roughness perception factor, dynamic weight allocation and feature recombination are performed on these streams to obtain a reflection suppression enhanced image and a roughness distribution map corresponding to the roughness perception factor. Specifically, this step follows the multispectral registered image set output from step S1 and is responsible for the physically interpretable separation and suppression of specular reflection and white wave interference in the image. In marine monitoring scenarios, the light signal reaching the camera sensor can be considered as a superposition of three parts: diffuse reflection components from floating objects and the water itself, specular reflection flare components generated by the smooth water surface, and white wave scattering components generated by wave-breaking areas. Among these, specular reflection and white wave scattering are the main interference sources affecting the accuracy of floating object detection. The basic idea of ​​this step is to construct a perception factor reflecting the roughness of the sea surface, and use this factor to guide the image to be decomposed in the frequency domain into three feature streams corresponding to the three physical components mentioned above. Then, the recombination weights of each feature stream are dynamically adjusted according to the roughness, thereby effectively suppressing specular reflection and white wave scattering interference while preserving the diffuse reflection information of floating objects. This includes the following steps: S2.1 Extract spectral absorption features and spatial gradient features from the multispectral registered image set, and generate a roughness perception factor through multi-channel feature fusion and edge-preserving smoothness constraints; including the following steps: S2.11 Calculate the spectral absorption difference pixel by pixel for the visible light band image and the near-infrared band image in the multispectral registered image set to obtain the local absorption difference feature map; In this step, the multispectral registered image set is... Visible light images Near-infrared images Pixel-by-pixel calculation of spectral absorption differences. Water bodies have different absorption coefficients in the visible and near-infrared bands, and the spectral responses of floating objects, specular reflection areas, and white wave areas also differ. By calculating the difference in pixel values ​​at the same spatial location in the two bands, these areas with significant differences in spectral response can be initially highlighted. In one specific embodiment, a local absorption difference feature map is generated. Calculate using the following formula: ; In the formula, and At pixel locations in visible light images and registered near-infrared images, respectively. Pixel value at; The absolute value of the spectral absorption difference at this location is used to preserve the degree of deviation between the responses of the two bands. The spatial dimensions are consistent with the input image.

[0030] S2.12. Perform gradient operations on the multispectral registered image set along the horizontal and vertical directions respectively to synthesize the texture gradient magnitude map and the texture gradient direction map; In this step, to further capture the texture roughness information brought about by the undulations of sea waves, gradient calculations are performed on the multispectral registered image set along both the horizontal and vertical directions. Taking visible light images as an example, the horizontal gradient is calculated using the Sobel operator or central difference. and vertical gradient Then, synthesize the texture gradient magnitude map according to the following formula. and texture gradient direction map : ; ; In the formula, and Pixel positions Gray-level gradient values ​​along the horizontal and vertical directions, in units of gray-level values ​​per pixel; The gradient magnitude reflects the degree of texture change at that location; The gradient direction angle, in radians, indicates the main direction of the texture.

[0031] Repeat the same operation on the near-infrared image to obtain the gradient amplitude map in the near-infrared band. and gradient direction pattern In one specific implementation, the gradient magnitudes of the two bands are averaged pixel-by-pixel and fused into a final texture gradient magnitude map: ; Texture gradient maps typically show little difference between visible and near-infrared light, so the calculation results from visible light images can be used. The final texture gradient direction map can also be the pixel-wise average direction of the two. This averaging fusion method can take into account the texture information captured by both bands and reduce the impact of local texture loss that may occur in a single band on roughness perception.

[0032] S2.13. Normalize the local absorption difference feature map and the texture gradient amplitude map, and then weight and fuse them according to a preset ratio to generate an initial roughness response map. In this step, the local absorption difference feature map is generated. Texture gradient magnitude map Normalization is performed separately to unify the numerical range of both values ​​to the [0,1] interval. In one specific implementation, a minimum-maximum normalization method is used: ; ; In the formula, and These are the minimum and maximum values ​​of the corresponding feature map at all pixel locations, respectively. and These are the normalized local absorption difference feature map and texture gradient magnitude map, respectively; To prevent extremely small constants with a denominator of zero, a typical value is 10. -6 To avoid calculation errors when the feature values ​​of the entire image are consistent, the two are weighted and fused according to a preset ratio to generate an initial roughness response map. : ; In the formula, The fusion weighting coefficient, with a value ranging from 0 to 1, is used to adjust the relative contribution of spectral difference information and texture gradient information. The value of can be pre-calibrated based on the spectral characteristics and wave features of the sea surface in the actual application scenario. In one specific implementation, is taken as . This allows the two types of information to participate in the roughness response with equal weight.

[0033] S2.14. Guided by the texture gradient direction map, perform anisotropic diffusion smoothing on the initial roughness response map to output a spatially continuous roughness perception factor.

[0034] In this step, the initial roughness response map The texture may contain local noise and discontinuities, and direct use of it can affect the stability of subsequent feature decoupling. Therefore, texture gradient direction maps are used instead. Guided by, Perform anisotropic diffusion smoothing. The basic principle of anisotropic diffusion is to suppress diffusion in directions with larger image gradients (i.e., directions across edges) to prevent edges from being blurred; and to allow diffusion in directions with smaller gradients (i.e., along edges) to smooth local noise. The diffusion process is iteratively updated using the following formula: ; In the formula, For the first Roughness response map after the next iteration. ; The diffusion step size is set to a positive value less than 0.25 to ensure numerical stability. for Spatial gradient; This represents the gradient magnitude. As a function of the diffusion coefficient, we can take... In the form of, This is the edge sensitivity parameter, which controls the degree of sensitivity to edges. The typical value range is 0.05 to 0.15 (corresponding to the normalized roughness response value range). This represents the divergence operation. The iteration termination condition for diffusion is reaching a preset maximum number of iterations (e.g., 20 to 50 times), or the change in structural similarity between two adjacent iterations being lower than a preset threshold. In embodiments using structural similarity as the termination condition, the conventional structural similarity (SSIM) calculation formula in the art can be used, with a threshold value between 0.995 and 0.999. Convergence is determined when the SSIM of two adjacent iterations is higher than this threshold, and iteration stops. After smoothing, a spatially continuous roughness perception factor is output. .

[0035] S2.2. The roughness-aware factor is injected as a conditional modulation parameter into three parallel convolutional neural network branches. Each branch first performs an invertible wavelet transform on the multispectral registered image set to decompose the image into a multi-scale frequency domain space. Then, guided by the roughness-aware factor and through feature space orthogonalization constraints, it decouples and outputs the diffuse reflection body feature flow, the specular reflection environment feature flow, and the white wave scattering interference feature flow in the frequency domain. Specifically, the roughness-aware factor is obtained. Subsequently, this sub-step uses it as a conditional modulation parameter to decompose the multispectral registered image set into three feature flows in the frequency domain: diffuse reflection body, specular reflection environment, and white wave scattering interference. Decoupling is performed in the multi-scale space of the frequency domain, with the three branches sharing the same invertible wavelet transform front end.

[0036] First, the multispectral registered image set Visible light images and near-infrared images Each image is decomposed from the spatial domain to a multi-scale frequency domain using reversible wavelet transforms, yielding its respective wavelet sub-band coefficients. In one specific implementation, a discrete wavelet transform based on a lifting scheme is employed, with the Cohen-Daubechies-Feauveau (CDF) 9 / 7 wavelet chosen as the wavelet basis function, exhibiting linear phase characteristics. The number of wavelet decomposition layers is determined based on the image resolution and the scale of the target floating object, typically ranging from 2 to 4 layers, generating a total of [number missing] waves. Individual band ( The decomposition layer contains one low-frequency sub-band and three high-frequency sub-bands in each direction. Reversible wavelet transform does not lose information, ensuring accurate image reconstruction in subsequent inverse transforms. The corresponding sub-band coefficients obtained after wavelet decomposition of two images are stacked along the channel dimension. The sub-bands are organized in the order of low-frequency sub-band first, followed by high-frequency sub-bands in each layer arranged horizontally, vertically, and diagonally, forming a frequency-domain multispectral representation shared by the subsequent three branches, denoted as . It includes the following steps: S2.21. The roughness perception factor is injected as a conditional input into the diffuse reflection ontology feature extraction branch, and the diffuse reflection ontology feature flow is separated from the multispectral registered image set through spectral-spatial joint convolution. In this step, the diffuse ontological feature extraction branch will incorporate the roughness perception factor. As a conditional modulation parameter, it is represented from the frequency domain multispectral representation. Separate the diffuse volumetric feature flow in the middle. The input to this branch is the complete... That is, all subband coefficients obtained by wavelet decomposition.

[0037] In one specific implementation, the spectral-spatial joint convolution is specifically implemented as follows: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] For each subband in the process, the subband coefficients of the visible light channel and the subband coefficients of the near-infrared channel are first stacked in the channel dimension to form a dual-channel input. Then, a 3×3 spatial convolution kernel is used to perform convolution operation on the stacked dual-channel features. While extracting spatial texture features, the information of the two spectral channels is fused to achieve joint processing of the spectral and spatial dimensions.

[0038] This branch is It consists of alternating stacks of convolutional layers and a corresponding number of conditional normalization layers. Two to four layers can be used, with each layer having a 3×3 kernel size. The scaling parameters for the conditional normalization layers are also specified. Translation parameters Roughness perception factor It is generated through a separate, small mapping network. This mapping network first... Global average pooling is performed to obtain a global roughness scalar. This scalar is then passed through two fully connected layers. The first fully connected layer maps the scalar to the latent variable space. The number of hidden layer channels can be scaled from 1 / 2 to 1 / 4 of the number of channels in the convolutional layer. The activation function is ReLU. The second fully connected layer outputs scaling and translation vectors corresponding to the number of channels in the convolutional layer. The scaling vectors are then mapped to the interval [−1, 1] by the Tanh function and used as... The translation vector, after identity mapping, becomes The conditional normalization layer performs an affine transformation on each channel of the convolutional output feature map: multiplying each channel of the feature map by the corresponding scaling parameter and adding a translation parameter, thereby injecting roughness information into the feature extraction process. Smooth water surface areas with lower values Take a larger value to enhance the diffuse response related to the floating object material; in High-value roughness region A smaller value is chosen to suppress potential specular reflection and scattering components. The output of this branch is denoted as the diffuse volumetric characteristic flow. .

[0039] S2.22. The roughness perception factor is injected as a conditional input into the specular reflection environment feature extraction branch. The specular reflection environment feature flow is separated from the multispectral registered image set by using the intensity correlation model based on the difference in spectral intensity between visible light and near-infrared bands. In this step, the specular reflection environment feature extraction branch also uses... As a conditional input, its input is complete. .

[0040] Specular reflection primarily occurs on relatively smooth water surfaces, and its intensity relationship between the visible and near-infrared bands follows a certain pattern: specular reflection produces a strong response in both bands, but the response in the visible band is typically significantly higher than that in the near-infrared band. In one specific implementation, this characteristic is used to construct an intensity correlation model: for For each sub-band, the ratio of the visible light channel coefficient to the near-infrared channel coefficient is calculated based on its position. When this ratio exceeds a preset ratio threshold... (For example At step 1.5), the sub-band location is marked as a candidate region for specular reflection. The marking result is binarized into a mask image with the same spatial size as the sub-band coefficients. This mask image serves as the prior attention map, and... The corresponding sub-band coefficients are multiplied element by element, so that the branch network focuses on the features of the specular reflection candidate region in subsequent processing.

[0041] This branch is It consists of alternating convolutional layers and conditional normalization layers. Two to four layers can be used, with a structural layout similar to the diffuse branch. Each layer has a 3×3 kernel size, but the parameters are learned independently and not shared with other branches. The scaling parameters of the conditional normalization layer... Translation parameters Similarly, it is composed of roughness perception factors It is generated via a separate, small mapping network, the structure of which is the same as described in S2.21. Smooth regions with lower values A larger value is chosen to enhance the extraction of specular reflection features; in High-value roughness region A smaller value is selected to suppress extraction. Combining the spatial prior information provided by the intensity correlation model with the roughness modulation information injected by the conditional normalization layer, this branch separates the specular reflection environment feature flow from the frequency domain wavelet coefficients. .

[0042] S2.23. The roughness perception factor is injected as a conditional input into the white wave scattering interference feature extraction branch, and the white wave scattering interference feature flow is separated from the multispectral registered image set based on the high frequency spatial fluctuation characteristics. In this step, the white wave scattering interference feature extraction branch also uses... As a conditional input, white waves are generated by wave breaking, which manifests in the image as local high-frequency, multi-directional disordered textures, mainly concentrated in the higher-frequency subbands in the wavelet domain.

[0043] In one specific implementation, the input to this branch is All subbands are used, but a subband selection gate is added at the input to highlight high-frequency information: the top three directional subbands (highest frequency) are assigned a channel weight of 1.0, the next top three directional subbands are assigned a channel weight of 0.5, and the lower-level subbands and low-frequency subbands are assigned a channel weight of 0.1. This gate weight is applied in a subband-by-subband multiplication manner. The coefficients, after weighting, remain in the frequency domain representation and all participate in subsequent convolution operations. High-frequency subbands enter the feature extraction process with a larger amplitude, while low-frequency subbands participate with a weakened amplitude, highlighting high-frequency fluctuation information while preserving necessary contextual references. This gating method selectively enhances high-frequency information rather than truncating or discarding it, matching the physical characteristics of white wave scattering.

[0044] This branch is It consists of alternating convolutional layers and conditional normalization layers. Two to four layers can be used, with each layer having a 3×3 kernel size, and the parameters are learned independently. The scaling parameters for the conditional normalization layers... Translation parameters Similarly, it is composed of roughness perception factors It is generated via a separate, small mapping network, the structure of which is the same as described in S2.21. High-value wave breaking areas A larger value is chosen to enhance the extraction of white wave scattering features; in Smooth regions with lower values A smaller value is chosen to suppress extraction. The output of this branch is denoted as the white wave scattering interference feature flow. .

[0045] It should be noted that the small mapping networks in the three branches mentioned above are independent and do not share parameters. This design is based on the consideration of the roughness perception factor. Different modulation strategies should be applied to feature extraction of different physical components. Diffuse reflection needs to be enhanced in low roughness regions, specular reflection needs to be suppressed after identification in low roughness regions, and white wave scattering needs to be extracted in high roughness regions. The three respond in different directions to the same roughness value. Therefore, using independent mapping networks can learn their respective optimal modulation curves, making the parameter generation of the conditional normalization layer more in line with the task requirements of each branch.

[0046] S2.24 Calculate the inner product of the feature space between each pair of diffuse reflection body feature flow, specular reflection environment feature flow and white wave scattering interference feature flow. When the inner product exceeds the orthogonality threshold, perform Gram-Schmidt orthogonalization correction.

[0047] In this step, to ensure the independence of the three feature streams and to avoid the situation where one feature stream contains components of another, the following steps are taken: , and Orthogonality checks and corrections are performed. In one specific implementation, the three feature flows are expanded pixel-by-pixel into feature vectors according to their spatial locations. After calculating the pixel-by-pixel inner product of each pair of feature flows along the channel dimension, the average of the inner products across all spatial locations in the entire image is taken as the correlation measure between the two feature flows. When the correlation measure of any pair of feature flows exceeds a preset orthogonality threshold... (For example When the value is between 0.1 and 0.3, Gram-Schmidt orthogonalization correction is performed on both. and For example, the correction: ; In the formula, and These represent the spatial positions of the specular reflection feature flow and the diffuse reflection bulk feature flow, respectively. ,aisle Eigenvalues ​​at; This represents the inner product operation along the channel dimension. (Corrected) and Orthogonal in the channel dimension. This correction can be performed iteratively or sequentially on the three pairs of feature streams in any order, until all pairwise correlations are lower than [value missing]. The correlation between the three feature flows is effectively suppressed after the correction, and the purity of feature decoupling is guaranteed.

[0048] S2.3. The numerical distribution of the roughness perception factor is transformed into three sets of basic weights using a piecewise mapping function. After adaptively correcting the three sets of basic weights using a channel attention mechanism, a weighted fusion is performed on the three feature streams to obtain the reconstructed feature representation; this includes the following steps: S2.31. Divide the roughness perception factor into three levels: low roughness, medium roughness, and high roughness, and obtain the first basic weight of the diffuse reflection body feature flow in each level by looking up the table. In this step, based on the roughness perception factor The numerical values ​​at various locations in space are divided into three levels: low roughness, medium roughness, and high roughness, according to preset thresholds.

[0049] In one specific implementation, low roughness corresponds to Medium roughness corresponds to High roughness corresponds to .in and The preset threshold values ​​can be calibrated based on the statistical characteristics of roughness distribution in actual sea conditions, for example, by taking... The 33rd and 67th percentiles of the cumulative distribution function over the entire dataset, or simply taking empirical values. For diffuse reflection volumetric feature flow Preset the first basic weight for each gear level. A value table. In one specific embodiment, the value table is as follows: Low roughness setting: ; Medium roughness setting: ; High roughness setting: .

[0050] S2.32. Based on the same numerical range, look up the table to obtain the second basic weight of the specular reflection environment characteristic flow at each level; In this step, the same roughness grading method is used to define the characteristic flow of the specular reflection environment. Look up the table to obtain the second basic weight for each gear level. .

[0051] In one specific implementation: Low roughness setting: (Smooth water surfaces have strong specular reflections, so they are assigned low weights to suppress them). Medium roughness setting: ; High roughness setting: (Specular reflections are negligible in such areas).

[0052] S2.33. Based on the same numerical range, look up the table to obtain the third basic weight of the white wave scattering interference characteristic flow at each level; In this step, the same roughness grading method is used to classify the white wave scattering interference characteristic flow. Look up the table to obtain the third basic weight for each level. .

[0053] In one specific implementation: Low roughness setting: (Few white waves); Medium roughness setting: ; High roughness setting: .

[0054] The three sets of basic weights are all non-negative values ​​at their respective levels, and are distributed in space in a pixel-by-pixel manner.

[0055] S2.34 Multiply the first basic weight, the second basic weight, and the third basic weight by the channel attention coefficient of the corresponding feature stream to obtain the final dynamic weight, and perform weighted summation on the diffuse reflection body feature stream, the specular reflection environment feature stream, and the white wave scattering interference feature stream to output the reconstructed feature representation.

[0056] In this step, to compensate for the shortcomings of fixed-level lookup tables in spatial adaptability, a channel attention mechanism is introduced to correct the three sets of basic weights pixel by pixel. In one specific implementation, the channel attention module adopts a compression-excitation structure.

[0057] diffuse reflection volumetric feature flow For example: First, regarding Global average pooling is performed to compress the feature map of each channel into a scalar, resulting in a channel description vector. This vector is then passed through two fully connected layers. The first fully connected layer reduces the number of channels back to their original value. ( To reduce the proportion, for example After ReLU activation, the second fully connected layer restores the number of channels to the original number, and the attention coefficients of each channel are obtained after mapping by the Sigmoid function. .

[0058] Furthermore, regarding and Perform the same operation as above to obtain and The final dynamic weights are obtained by multiplying the base weights of each feature stream by their corresponding attention coefficients. ; ; ; In the formula, , , The final dynamic weights for the diffuse reflection ontology, specular reflection environment, and white wave scattering interference feature flows are all spatial-channel three-dimensional weight tensors consistent with the dimensions of the corresponding feature flows. Since the base weights and attention coefficients are non-negative, the final dynamic weights also automatically remain non-negative, ensuring the physical interpretability of the three feature flows during the recombination process. A pixel-wise weighted summation is performed on the three feature flows to obtain the recombined feature representation. : .

[0059] S2.4. Perform inverse frequency domain transform on the reconstructed feature representation to obtain a reflection suppression and enhancement image, and output a roughness distribution map after performing gray quantization and spatial alignment verification on the roughness perception factor. This includes the following steps: S2.41. Perform an inverse wavelet transform on the recombined feature representation to restore it to the spatial domain and generate a preliminary reflection suppression and enhancement image; In this step, the recombination features are represented. Perform the inverse reversible wavelet transform. Using the reconstruction filter bank corresponding to the inverse wavelet transform in step S2.2, [the process is repeated]. From frequency domain multi-scale spatial reconstruction to spatial domain, a preliminary reflection suppression and enhancement image is generated. This image has the same spatial resolution as the input multispectral registered image set.

[0060] S2.42. Linearly quantize and encode the roughness perception factor according to the preset gray level to generate an initial roughness distribution map with the same resolution as the initial reflection suppression and enhancement image. In this step, the roughness perception factor is... Linear quantization encoding is performed according to a preset grayscale level. In one specific implementation, [the following is omitted as it is not relevant to the main text]. The numerical range [0,1] is linearly mapped to the integer range from 0 to 255: ; In the formula, This indicates a rounding operation. The generated initial roughness distribution map. and They have the same spatial resolution.

[0061] S2.43. Compare the coordinate indices of the preliminary reflection suppression enhancement image and the initial roughness distribution map pixel by pixel, and perform sub-pixel level resampling correction on the offset pixels to ensure that the two are strictly aligned in space. In this step, pixel-by-pixel comparison is performed. and The coordinate indices are used to confirm a strict spatial correspondence between the two. Since the reversible wavelet transform and inverse transform do not change the spatial coordinates while ensuring reversibility, there is no spatial offset between them under normal circumstances. If a sub-pixel level offset is detected due to numerical precision factors, then... Based on the coordinate grid, for Resampling correction is performed using bilinear interpolation to ensure strict spatial alignment between the two.

[0062] S2.44. Pack the aligned reflection suppression and enhancement image and roughness distribution map and output them to S3.

[0063] In this step, the aligned reflection suppression enhancement image is... Roughness distribution map As a data package output, it is fed into the multi-scale feature fusion and object detection network in step S3. The enhanced image, which has effectively suppressed specular reflection and white wave scattering interference, can more clearly present the physical characteristics of the floating object; This identifies the sea surface roughness at each spatial location, which is used by step S3 to generate a roughness suppression weight map, guiding the detection network to reduce the attention weight for high roughness regions.

[0064] Through the processing in step S2 above, specular reflection and wave scattering components unrelated to floating object detection are effectively separated and suppressed in the multispectral registered image set, while the diffuse reflection information of the floating object itself is preserved and enhanced. The roughness perception factor is used throughout the entire process of feature decoupling, weight allocation, and roughness distribution map generation, enabling the reflection suppression strategy to adaptively adjust according to the sea surface state, providing higher-quality input data for the multi-scale feature fusion detection in step S3 and the trajectory verification in step S4.

[0065] S3. Multi-scale Feature Fusion and Target Detection: The reflection suppression and enhancement image and roughness distribution map are input into a multi-scale feature fusion neural network model. The roughness distribution map guides the multi-scale feature fusion neural network model to reduce the attention weight of high-roughness areas on the sea surface, and outputs the target detection results of floating objects on the sea surface; including the following steps: S3.1 Input the reflection suppression and enhancement image into the multi-scale feature fusion neural network model, and extract multi-scale spatial feature maps layer by layer through the multi-scale feature fusion neural network model; In this step, the reflection suppression and enhancement image output from step S2 is... As input, the data is fed into a multi-scale feature fusion neural network model. This neural network model can employ a convolutional neural network architecture based on a feature pyramid structure, such as using ResNet or CSPDarknet as the backbone network, performing convolution and downsampling operations layer by layer from bottom to top to extract feature maps with different semantic levels and spatial resolutions. As the number of network layers increases, shallow feature maps have higher spatial resolution and retain more locational details; deep feature maps have a larger receptive field and contain richer semantic information. Assume the backbone network generates... The feature maps at different scales are denoted as follows, from highest to lowest spatial resolution: ,in This is the shallowest feature map, with the highest spatial resolution; These are the deepest feature maps, with the lowest spatial resolution. These feature maps constitute a multi-scale spatial feature map set, which is used for subsequent attention suppression and cross-layer fusion.

[0066] S3.2 Calculate the attention attenuation coefficient for each spatial location based on the roughness distribution map, and generate a roughness suppression weight map. The attention attenuation coefficient for high-roughness areas of the sea surface is less than that for low-roughness areas. The roughness suppression weight map assigns a weight coefficient between 0 and 1 to each spatial location based on the sea surface roughness. High-roughness areas (wave-broken zones, areas with strong reflective residue) have low weights, while low-roughness areas (smooth water surfaces, areas with floating objects) have high weights, thus guiding the detection network to focus its attention on areas with higher confidence. Generating the roughness suppression weight map includes the following steps: S3.21 Perform a global pooling operation on the roughness distribution map to obtain a global roughness scalar that characterizes the roughness of the sea surface in the entire image; In this step, the roughness distribution map output in step S2 is... Perform a global average pooling operation. The space dimensions are Each pixel value represents the sea surface roughness at its corresponding spatial location; a larger value indicates a rougher sea surface at that location. The global roughness scalar is calculated using the following formula. : ; In the formula, The roughness distribution map is represented on the coordinate system. Pixel value at; This is a dimensionless scalar reflecting the overall roughness level of the sea surface in the current frame. When the sea surface is generally calm, The value is relatively small; when there is a large area of ​​broken waves or strong reflective areas on the sea surface. The value increases accordingly.

[0067] S3.22. Concatenate the roughness distribution map with the global roughness scalar, extract the spatial distribution features of roughness through convolution operation, and generate an initial roughness attention map. In this step, the global roughness scalar is... Expand to Single-channel matrices with the same spatial dimensions ,in All elements take the value .Will and By splicing along the channel dimension, a structure with dimensions of The input tensor is a two-channel input. This two-channel input tensor is fed into a small convolutional subnetwork consisting of several convolutional layers. For example, this subnetwork contains two 3×3 convolutional layers, each followed by a ReLU activation function. The convolutional subnetwork extracts the spatial distribution features of roughness by learning the relative relationship between local roughness values ​​and global roughness levels. The output size is... Initial roughness attention map .exist In the model, the response values ​​at each spatial location reflect the significance of the roughness relative to the global roughness level and the local neighborhood roughness distribution: regions with significantly higher local roughness than the global level (such as isolated white wave areas) produce higher responses, while regions with local roughness close to or lower than the global level produce lower responses.

[0068] S3.23. Perform a nonlinear activation function mapping on the initial roughness attention map to normalize the values ​​to a preset interval, and obtain a roughness suppression weight map, so that high roughness regions are mapped to lower weight values ​​and low roughness regions are mapped to higher weight values.

[0069] In this step, the initial roughness attention map is... Performing a sigmoid function mapping compresses the response values ​​at each spatial location into the (0,1) interval, resulting in an intermediate attention map. : ; In the formula, and The initial roughness attention map and the intermediate attention map are respectively located in the coordinate system. The value at the specified location is then determined. Subsequently, based on the requirements for the degree of suppression of high roughness regions in practical applications, the intermediate attention map is inverted to obtain the final roughness suppression weight map. : ; In the formula, For roughness suppression weight map in coordinate The weight value at that location. In this weight map, the high roughness region is... The response is relatively high, after Sigmoid mapping Approaching 1, then inverted Approaching 0, i.e., assigning lower attention weight; low roughness regions in The response is low in the middle, after Sigmoid mapping Approaching 0, then inverted A value close to 1 indicates a higher attention weight. The space dimensions are ,and The spatial dimensions are consistent.

[0070] S3.3. Use the roughness suppression weight map to perform layer-by-layer weighting processing on the multi-scale spatial feature map, reduce the attention weight of the multi-scale feature fusion neural network model to the high roughness region of the sea surface, and obtain the roughness-suppressed multi-scale spatial feature map. In this step, a roughness suppression weight map is used. The multi-scale spatial feature map extracted in step S3.1 Perform layer-by-layer weighting. Since the spatial resolution of the feature maps in each layer is different, it is necessary to first... Downsampled to the same spatial size as the feature maps of each layer. For the th layer... Layer feature map Its spatial dimensions are ,Will Downsampling to bilinear interpolation The suppression weight map of the corresponding layer is obtained. The downsampling process maintains a consistent coordinate system, ensuring a strict correspondence between the spatial pixel positions of the weight map and the corresponding feature map. Then, for... Multiply each channel pixel by The feature map after roughness suppression is obtained. : ; In the formula, For the first Layer feature map in spatial location ,aisle Eigenvalues ​​at; This represents the value of the downsampled suppression weight map at the corresponding position. These are the suppressed eigenvalues. After layer-by-layer weighted processing, the feature responses of high-roughness regions are attenuated, while the feature responses of low-roughness regions are preserved, resulting in a multi-scale spatial feature map after roughness suppression. .

[0071] S3.4 Perform cross-level multi-scale feature fusion on the roughness-suppressed multi-scale spatial feature map to generate a fused feature map; In this step, the multi-scale spatial feature map after roughness suppression is processed. Perform multi-scale feature fusion across layers. In one specific implementation, a top-down path enhancement is performed using a Feature Pyramid Network (FPN) structure: starting from the deepest feature map... Initially, upsampling is used to enlarge deep, low-resolution feature maps to the same spatial size as adjacent shallow feature maps. Then, strong semantic information from the deeper layers is transferred to the shallower layers through element-wise addition or channel concatenation. Simultaneously, bottom-up paths can be added to transfer fine-grained location information from the shallower layers to the deeper layers, forming a bidirectional feature pyramid. After cross-layer fusion, each layer yields a fused feature map that contains both the spatial resolution of that layer and semantic information from other layers, forming a set of fused feature maps. .

[0072] S3.5. Based on the fused feature map, perform target regression and classification calculation, and output the target detection results of floating objects on the sea surface. The target detection results include the bounding box coordinates, category label and confidence score of the floating objects.

[0073] In this step, the feature map set will be fused. The feature maps from each layer are input into the detection head network. The detection head network contains two parallel sub-branches: a classification branch and a regression branch. The classification branch outputs the probability score of each spatial location belonging to each category after passing through several convolutional layers. After normalization by the Softmax function, the class label and the corresponding classification confidence score are obtained. The regression branch outputs the bounding box offset corresponding to each spatial location after passing through several convolutional layers. Combined with the preset anchor box size and position, the center coordinates of the bounding box are decoded to obtain the bounding box center coordinates. ,width and height Non-maximum suppression is performed on all candidate bounding boxes output by the detection head to remove highly overlapping duplicate detection boxes. Bounding boxes with a confidence level higher than a preset detection threshold are retained as the final detection results for floating objects on the sea surface. The detection results are output in list form, and each detected floating object includes the following information: bounding box coordinates. Category labels and confidence scores .

[0074] Through the processing in step S3 above, the roughness distribution map is passed from step S2 to this point and injected into the multi-scale feature fusion neural network model in the form of a suppressed weight map. This enables the detection network to automatically reduce the feature response in the high roughness region of the sea surface and maintain a normal response in the low roughness region of the sea surface, thereby effectively reducing false detections in the wave break zone and residual reflective area, and improving the accuracy and robustness of floating object target detection.

[0075] S4. Trajectory Verification and Result Output: Obtain target detection results for multiple consecutive frames. Based on the temporal continuity of physical kinematic constraints and spectral features, perform trajectory association and consistency verification on the target detection results for multiple consecutive frames to obtain the final sea surface floating object detection result. This includes the following steps: S4.1 Obtain the target detection results of multiple consecutive frames, extract the bounding box position, motion state vector and spectral features of each detected target in each frame, and construct a candidate target set; In this step, the detection results of floating objects on the sea surface obtained from processing multiple consecutive frames of images in step S3 are acquired. Let the current processing window contain... Frame-by-frame images The typical value ranges from 5 to 15 frames, which are sequentially labeled as frame 1, frame 2, ..., frame 3 in chronological order. Frame; During processing, the sliding window is updated frame by frame: each newly acquired frame is added to the end of the window, while the oldest historical frame at the front of the window is removed, maintaining a constant window length and achieving frame-by-frame processing of the continuous video stream. For the first... frame Its test results include Each detection target It carries the following information: bounding box center coordinates , bounding box width and height Category tags and confidence score .

[0076] To construct a candidate target set, a motion state vector and spectral features are extracted for each detected target. The motion state vector describes the target's position and motion trend in the current frame, denoted as: ; In the formula, For the first Frame number The motion state vector of each detected target; and These are the horizontal and vertical coordinates of the target in the image coordinate system, respectively, in pixels; and These represent the target's horizontal and vertical movement velocities, respectively, in pixels per frame. For the first frame ( The velocity component is initialized to zero or obtained from differential estimation of subsequent frames.

[0077] Spectral features are used to characterize the spectral response of a target in the visible and near-infrared bands. In one specific embodiment, for each detected target, a multispectral registration image set is generated using the target's bounding box output in step S1. Visible light image of the corresponding region and near-infrared images Based on this, the mean pixel values ​​of the two bands in this region are calculated respectively, forming the spectral feature vector of the target: ; In the formula, For the first Frame number The spectral feature vector of each detected target; and These are the average pixel values ​​in the visible and near-infrared bands within the bounding box region of the target, respectively.

[0078] The bounding box positions, motion state vectors, and spectral features of all detected targets in each frame are summarized to form a candidate target set. If no target is detected in a frame, the candidate target subset corresponding to that frame is empty. During trajectory association, it is automatically determined that all existing trajectories have no matching results in that frame, and the trajectory termination rules are applied. ; Each candidate target Contains attribute tuples .

[0079] S4.2. Based on the motion state vector, establish a physical kinematic state transition model to predict the bounding box position of each detected target in the candidate target set in the next frame, and obtain the predicted position of each detected target. In this step, a physical kinematic state transition model is established based on the motion state vector. In one specific implementation, a uniform motion model is used as the state transition model, assuming that the floating object on the sea surface maintains approximately uniform motion between adjacent frames. For the first... The first frame A detection target, its motion state vector via state transition matrix Predict its in the Frame motion state: ; Where the state transition matrix The form is: ; In the formula, The time interval between adjacent frames, under the condition of equal sampling intervals in consecutive frames. (In units of frames). Therefore, the predicted state vector is... The first two components are the target at the th... Predicted position of the frame .

[0080] S4.3 Calculate the kinematic association cost between the predicted position and the actual detection position of the detected target in adjacent frames. When the kinematic association cost is lower than the preset association threshold, associate the corresponding detection results with the same trajectory sequence. In this step, trajectory association is performed between adjacent frames. For the first... Detection target of frame and the Detection target of frame The kinematic correlation cost between the two is calculated. In one specific implementation, the correlation cost is the Mahalanobis distance between the predicted location and the actual detection location: ; In the formula, The kinematic correlation cost between the two detection targets is dimensionless. For the first Frame target The predicted state vector; For the first Frame target The actual state vector; The covariance matrix of the state vector can be preset according to the positioning accuracy of the detector, for example, by taking the diagonal matrix. ,in and The standard deviation of the location prediction is typically 5 to 15 pixels. and The standard deviation of the velocity estimate is typically 2 to 5 pixels per frame.

[0081] Furthermore, the kinematic cost Associated with preset threshold Comparison: when At that time, it was assumed that the two detection targets were kinematically correlated, and Associated with The trajectory sequence it belongs to; when At that time, it was considered that the two did not match; If the first A certain detected target in the frame and the first If the association cost of any existing trajectory in a frame is not lower than the threshold, then a new trajectory sequence may be created for a newly appearing target. If an existing trajectory fails to match a new detected target in several consecutive frames, then that trajectory is terminated. Typical values ​​can be selected from the chi-square distribution critical value table. For example, the chi-square critical value corresponding to 95% confidence level and 4 degrees of freedom is about 9.49.

[0082] After traversing all adjacent frames and completing the trajectory association, a trajectory sequence is obtained. Each trajectory sequence It contains a list of associated detection targets arranged in chronological order. This represents the total number of trajectory sequences.

[0083] S4.4 Calculate the temporal similarity between the spectral features of the detected targets in adjacent frames of the same trajectory sequence. When the temporal similarity is lower than the preset continuity threshold, mark the corresponding trajectory sequence as a consistency anomaly. In this step, the consistency of spectral features is checked for each trajectory sequence obtained in step S4.3. For the trajectory sequences... The Middle Frame target and the Frame target Calculate the spectral eigenvectors of both. and The temporal similarity between them. In one specific implementation, cosine similarity is used as the metric: ; In the formula, The spectral temporal similarity between two detected targets is defined, with a value ranging from [−1, 1]. It is the inner product of two spectral eigenvectors; and Let be the Euclidean norms of the two vectors.

[0084] Spectral temporal similarity With preset continuity threshold The spectral similarity of any two adjacent frames in the trajectory sequence is compared. When the trajectory shows discontinuity in its spectral characteristics, it may be caused by reflective flicker, white wave transients, or incorrect association of different objects, and the trajectory sequence is marked as having an inconsistency. The typical value is 0.85 to 0.95, which means that targets on the same trajectory in adjacent frames are required to maintain a high degree of similarity in multispectral space.

[0085] S4.5 Remove trajectory sequences marked as inconsistent, and integrate the remaining trajectory sequences that pass the consistency check to output the final sea surface floating object detection result.

[0086] In this step, from the set of trajectory sequences All trajectory sequences marked as inconsistent are removed, retaining only those that simultaneously satisfy both physical kinematic constraints and spectral feature continuity constraints. The detected targets in each retained trajectory sequence are then integrated frame-by-frame to form the final detection result output. Each valid detection result includes: a trajectory identifier, bounding box coordinates for each frame, a class label, and an average confidence score.

[0087] Through the processing in step S4 above, the physical kinematic constraints and temporal consistency of spectral features between consecutive frames are used for dual verification, effectively filtering out false targets caused by specular reflection flicker, white wave transients, and isolated false detections. The output is a stable and consistent detection result of floating objects on the sea surface in the time dimension, which further improves the reliability of the entire recognition method under actual sea conditions.

[0088] like Figure 2 As shown, this embodiment also provides an image-enhanced AI visual recognition system for marine floating objects based on multispectral fusion and reflection suppression. Based on the aforementioned image-enhanced AI visual recognition method for marine floating objects based on multispectral fusion and reflection suppression, it includes: Multispectral data acquisition and registration unit 1 is used to acquire visible light images and near-infrared images, perform adaptive spatial registration on the near-infrared images, and output a multispectral registered image set; multispectral data acquisition and registration unit 1 includes: The synchronous acquisition module is used to synchronously trigger the multispectral imaging system to acquire visible light and near-infrared images that are aligned in time and space. The preprocessing module is used to denoise and enhance the contrast of the image; The phase consistency feature extraction module is used to extract multi-scale phase consistency features of images; The feature matching and registration module is used to perform feature-based matching and spatial registration. Multispectral feature decoupling and reflection suppression unit 2, connected to multispectral data acquisition and registration unit 1, is used to receive multispectral registered image sets, construct roughness perception factors, decouple image features and suppress reflection, and output a reflection suppression enhanced image and a roughness distribution map; multispectral feature decoupling and reflection suppression unit 2 includes: The roughness perception factor generation module is used to fuse spectral absorption differences and spatial gradient information to generate roughness perception factors. The multispectral feature decoupling module is used to decouple an image into diffuse reflection volumetric feature stream, specular reflection environment feature stream, and white wave scattering interference feature stream, guided by the roughness perception factor. The dynamic weight allocation and recombination module is used to dynamically adjust the weights of the three feature flows and recombine them according to the roughness perception factor. The image reconstruction and distribution map output module is used to reconstruct the reflection suppression and enhancement image and output the corresponding roughness distribution map; Multi-scale feature fusion and target detection unit 3, connected to multispectral feature decoupling and reflection suppression unit 2, is used to perform target detection using reflection suppression enhancement image and roughness distribution map, and outputs the target detection result of floating objects on the sea surface; multi-scale feature fusion and target detection unit 3 includes: The multi-scale feature extraction module is used to extract multi-scale spatial features of images; The attention weight modulation module is used to reduce the attention weight of high-roughness regions based on the roughness distribution map. The feature fusion and detection output module is used to fuse multi-scale features and output target detection results; The trajectory verification and result output unit 4, connected to the multi-scale feature fusion and target detection unit 3, is used to perform trajectory association and consistency verification on the target detection results of multiple consecutive frames, and output the final sea surface floating object detection result; the trajectory verification and result output unit 4 includes: The candidate target construction and prediction module is used to extract the target state and predict its location; The trajectory association and verification module is used to perform trajectory association and consistency determination based on kinematic constraints and spectral continuity. The results integration and output module is used to remove abnormal trajectories and output the final floating object detection results.

[0089] Those skilled in the art will understand that the process of implementing all or part of the steps of the above embodiments can be carried out by hardware or by a program instructing the relevant hardware.

[0090] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. An image-enhanced AI visual recognition method for marine floating objects based on multispectral fusion and reflection suppression, characterized in that, include: S1. Multispectral data acquisition and registration: Acquire visible light and near-infrared images, and based on multi-scale phase consistency, perform adaptive spatial registration of the near-infrared images with the visible light images as the reference to obtain a multispectral registered image set. S2. Multispectral Feature Decoupling and Reflection Suppression: Based on the local absorption differences and texture gradients of different spectral bands in the multispectral registered image set, a roughness perception factor is constructed. Using the roughness perception factor as a guiding condition, the multispectral registered image set is decoupled into diffuse reflection body feature flow, specular reflection environment feature flow, and white wave scattering interference feature flow. Based on the roughness perception factor, dynamic weight allocation and feature recombination are performed on the diffuse reflection body feature flow, specular reflection environment feature flow, and white wave scattering interference feature flow to obtain the reflection suppression enhanced image and the roughness distribution map corresponding to the roughness perception factor. S3. Multi-scale feature fusion and target detection: The reflection suppression enhancement image and roughness distribution map are input into the multi-scale feature fusion neural network model. The roughness distribution map is used to guide the multi-scale feature fusion neural network model to reduce the attention weight of the high roughness area of ​​the sea surface and output the target detection result of floating objects on the sea surface. S4. Trajectory Verification and Result Output: Obtain target detection results for multiple consecutive frames. Based on the temporal continuity of physical kinematic constraints and spectral features, perform trajectory association and consistency verification on the target detection results for multiple consecutive frames to obtain the final sea surface floating object detection results.

2. The image-enhanced AI visual recognition method for marine floating objects based on multispectral fusion and reflection suppression according to claim 1, characterized in that, The multispectral data acquisition and registration in S1 includes the following steps: S1.

1. Through the synchronous triggering mechanism of the multispectral imaging system, visible light band images and near-infrared band images are acquired simultaneously to ensure that the visible light images and near-infrared images have the same timestamp and overlapping spatial coverage area. S1.2 Perform nonlocal mean filtering on the visible light image and the near-infrared image respectively, and then correct them by adaptive histogram equalization to output the preprocessed visible light image and near-infrared image. S1.

3. Multi-scale convolution operations are performed on the preprocessed visible light image and near-infrared image using a multi-scale Log-Gabor filter bank to calculate the phase consistency response at each scale, thereby obtaining the multi-scale phase consistency feature map of the visible light image and the multi-scale phase consistency feature map of the near-infrared image. S1.4 Based on the multi-scale phase consistency feature map of visible light image and the multi-scale phase consistency feature map of near-infrared image, the feature point similarity is calculated using the mutual information maximization criterion, and a dynamic matching threshold is set to filter reliable feature point pairs to form a feature point matching set; S1.

5. Use the RANSAC algorithm to remove mismatched points from the feature point matching set and estimate the rotation angle, scaling ratio and offset of the affine transformation matrix. S1.6 Apply the affine transformation matrix to the near-infrared image for bilinear interpolation resampling to obtain the registered near-infrared image; treat the registered near-infrared image and the visible light image as two independent channels to form a multispectral registered image set.

3. The image-enhanced AI visual recognition method for marine floating objects based on multispectral fusion and reflection suppression according to claim 1, characterized in that, The multispectral feature decoupling and reflection suppression in S2 includes the following steps: S2.1 Extract the spectral absorption features and spatial gradient features of the multispectral registered image set, and generate a roughness perception factor through multi-channel feature fusion and edge-preserving smoothness constraints; S2.

2. The roughness perception factor is injected as a conditional modulation parameter into three parallel convolutional neural network branches. Each branch first performs an invertible wavelet transform on the multispectral registered image set to decompose the image into a frequency domain multi-scale space. Then, guided by the roughness perception factor, through the feature space orthogonalization constraint, it decouples and outputs the diffuse reflection body feature flow, the specular reflection environment feature flow, and the white wave scattering interference feature flow in the frequency domain. S2.

3. The numerical distribution of roughness perception factor is transformed into three sets of basic weights using a piecewise mapping function. After adaptive correction of the three sets of basic weights using a channel attention mechanism, the three feature streams are weighted and fused to obtain the recombined feature representation. S2.4 Perform frequency domain inverse transform on the reconstructed feature representation to reconstruct the reflection suppression and enhancement image, and output the roughness distribution map after performing gray quantization and spatial alignment verification on the roughness perception factor.

4. The image-enhanced AI visual recognition method for marine floating objects based on multispectral fusion and reflection suppression according to claim 3, characterized in that, The generation of the roughness perception factor in S2.1 includes the following steps: S2.11 Calculate the spectral absorption difference pixel by pixel for the visible light band image and the near-infrared band image in the multispectral registered image set to obtain the local absorption difference feature map; S2.

12. Perform gradient operations on the multispectral registered image set along the horizontal and vertical directions respectively to synthesize the texture gradient magnitude map and the texture gradient direction map; S2.

13. Normalize the local absorption difference feature map and the texture gradient magnitude map, and then weight and fuse them according to a preset ratio to generate an initial roughness response map. S2.

14. Guided by the texture gradient direction map, perform anisotropic diffusion smoothing on the initial roughness response map to output a spatially continuous roughness perception factor.

5. The image-enhanced AI visual recognition method for marine floating objects based on multispectral fusion and reflection suppression according to claim 3, characterized in that, The output diffuse reflection body feature stream, specular reflection environment feature stream, and white wave scattering interference feature stream in S2.2 include the following steps: S2.

21. The roughness perception factor is injected as a conditional input into the diffuse reflection ontology feature extraction branch, and the diffuse reflection ontology feature flow is separated from the multispectral registered image set through spectral-spatial joint convolution. S2.

22. The roughness perception factor is injected as a conditional input into the specular reflection environment feature extraction branch. The specular reflection environment feature flow is separated from the multispectral registered image set by using an intensity correlation model based on the difference in spectral intensity between visible light and near-infrared bands. S2.

23. The roughness perception factor is injected as a conditional input into the white wave scattering interference feature extraction branch, and the white wave scattering interference feature flow is separated from the multispectral registered image set based on the high frequency spatial fluctuation characteristics. S2.24 Calculate the inner product of the feature space between each pair of diffuse reflection body feature flow, specular reflection environment feature flow and white wave scattering interference feature flow. When the inner product exceeds the orthogonality threshold, perform Gram-Schmidt orthogonalization correction.

6. The image-enhanced AI visual recognition method for marine floating objects based on multispectral fusion and reflection suppression according to claim 3, characterized in that, The process of obtaining the recombination feature representation in S2.3 includes the following steps: S2.

31. Divide the roughness perception factor into three levels: low roughness, medium roughness, and high roughness, and obtain the first basic weight of the diffuse reflection body feature flow in each level by looking up the table. S2.

32. Based on the same numerical range, look up the table to obtain the second basic weight of the specular reflection environment characteristic flow at each level; S2.

33. Based on the same numerical range, look up the table to obtain the third basic weight of the white wave scattering interference characteristic flow at each level; S2.34 Multiply the first basic weight, the second basic weight, and the third basic weight by the channel attention coefficient of the corresponding feature stream to obtain the final dynamic weight, and perform weighted summation on the diffuse reflection body feature stream, the specular reflection environment feature stream, and the white wave scattering interference feature stream to output the reconstructed feature representation.

7. The image-enhanced AI visual recognition method for marine floating objects based on multispectral fusion and reflection suppression according to claim 3, characterized in that, The output of the roughness distribution map in S2.4 includes the following steps: S2.

41. Perform an inverse wavelet transform on the recombined feature representation to restore it to the spatial domain and generate a preliminary reflection suppression and enhancement image; S2.

42. Linearly quantize and encode the roughness perception factor according to the preset gray level to generate an initial roughness distribution map with the same resolution as the initial reflection suppression and enhancement image. S2.

43. Compare the coordinate indices of the preliminary reflection suppression enhancement image and the initial roughness distribution map pixel by pixel, and perform sub-pixel level resampling correction on the offset pixels; S2.

44. Pack the aligned reflection suppression and enhancement image and roughness distribution map and output them to S3.

8. The image-enhanced AI visual recognition method for marine floating objects based on multispectral fusion and reflection suppression according to claim 1, characterized in that, The multi-scale feature fusion and target detection in S3 includes the following steps: S3.1 Input the reflection suppression and enhancement image into the multi-scale feature fusion neural network model, and extract multi-scale spatial feature maps layer by layer through the multi-scale feature fusion neural network model; S3.2 Calculate the attention attenuation coefficient at each spatial location based on the roughness distribution map and generate a roughness suppression weight map, wherein the attention attenuation coefficient corresponding to the high roughness region of the sea surface is smaller than the attention attenuation coefficient corresponding to the low roughness region of the sea surface. S3.

3. Use the roughness suppression weight map to perform layer-by-layer weighting processing on the multi-scale spatial feature map, reduce the attention weight of the multi-scale feature fusion neural network model to the high roughness region of the sea surface, and obtain the roughness-suppressed multi-scale spatial feature map. S3.4 Perform cross-level multi-scale feature fusion on the roughness-suppressed multi-scale spatial feature map to generate a fused feature map; S3.

5. Based on the fused feature map, perform target regression and classification calculation, and output the target detection results of floating objects on the sea surface. The target detection results include the bounding box coordinates, category label and confidence score of the floating objects.

9. The image-enhanced AI visual recognition method for marine floating objects based on multispectral fusion and reflection suppression according to claim 1, characterized in that, The trajectory verification and result output in S4 include the following steps: S4.1 Obtain the target detection results of multiple consecutive frames, extract the bounding box position, motion state vector and spectral features of each detected target in each frame, and construct a candidate target set; S4.

2. Based on the motion state vector, establish a physical kinematic state transition model to predict the bounding box position of each detected target in the candidate target set in the next frame, and obtain the predicted position of each detected target. S4.3 Calculate the kinematic association cost between the predicted position and the actual detection position of the detected target in adjacent frames. When the kinematic association cost is lower than the preset association threshold, associate the corresponding detection results with the same trajectory sequence. S4.4 Calculate the temporal similarity between the spectral features of the detected targets in adjacent frames of the same trajectory sequence. When the temporal similarity is lower than the preset continuity threshold, mark the corresponding trajectory sequence as a consistency anomaly. S4.5 Remove trajectory sequences marked as inconsistent, and integrate the remaining trajectory sequences that pass the consistency check to output the final sea surface floating object detection result.

10. An image-enhanced AI visual recognition system for marine floating objects based on multispectral fusion and reflection suppression, characterized in that, When the computer program of the image-enhanced AI visual recognition system for marine floating objects based on multispectral fusion and reflection suppression is executed, it performs the steps of the image-enhanced AI visual recognition method for marine floating objects based on multispectral fusion and reflection suppression as described in any one of claims 1-9, including: A multispectral data acquisition and registration unit (1) is used to acquire visible light images and near-infrared images, and to perform adaptive spatial registration on the near-infrared images to output a multispectral registered image set; the multispectral data acquisition and registration unit (1) includes: The synchronous acquisition module is used to synchronously trigger the multispectral imaging system to acquire visible light and near-infrared images that are aligned in time and space. The preprocessing module is used to denoise and enhance the contrast of the image; The phase consistency feature extraction module is used to extract multi-scale phase consistency features of images; The feature matching and registration module is used to match based on the features and complete spatial registration; A multispectral feature decoupling and reflection suppression unit (2) is connected to the multispectral data acquisition and registration unit (1) and is used to receive the multispectral registered image set, construct a roughness perception factor, decouple image features and suppress reflection, and output a reflection suppression enhanced image and a roughness distribution map; the multispectral feature decoupling and reflection suppression unit (2) includes: The roughness perception factor generation module is used to fuse spectral absorption differences and spatial gradient information to generate roughness perception factors. The multispectral feature decoupling module is used to decouple the image into diffuse reflection body feature stream, specular reflection environment feature stream, and white wave scattering interference feature stream, guided by the roughness perception factor; The dynamic weight allocation and recombination module is used to dynamically adjust the weights of the three feature streams and recombine them according to the roughness perception factor. The image reconstruction and distribution map output module is used to reconstruct the reflection suppression and enhancement image and output the corresponding roughness distribution map; A multi-scale feature fusion and target detection unit (3) is connected to the multi-spectral feature decoupling and reflection suppression unit (2) and is used to perform target detection using the reflection suppression enhancement image and roughness distribution map, and output the target detection result of floating objects on the sea surface; the multi-scale feature fusion and target detection unit (3) includes: The multi-scale feature extraction module is used to extract multi-scale spatial features of images; The attention weight modulation module is used to reduce the attention weight of high-roughness regions based on the roughness distribution map. The feature fusion and detection output module is used to fuse multi-scale features and output target detection results; The trajectory verification and result output unit (4) is connected to the multi-scale feature fusion and target detection unit (3) and is used to perform trajectory association and consistency verification on the target detection results of multiple consecutive frames, and output the final sea surface floating object detection result; the trajectory verification and result output unit (4) includes: The candidate target construction and prediction module is used to extract the target state and predict its location; The trajectory association and verification module is used to perform trajectory association and consistency determination based on kinematic constraints and spectral continuity. The results integration and output module is used to remove abnormal trajectories and output the final floating object detection results.

Images