An underwater image despeckling enhancement method based on active polarization sampling and modal complementary reasoning

Abstract

The application discloses an underwater image de-scattering enhancement method based on active polarization sampling and modal complementary reasoning, and belongs to the technical field of intelligent visual processing. The method comprises the following steps: an image of a target underwater scene is collected by a polarization camera to obtain a polarization original image; a light residual convolutional neural network is constructed to perform de-mosaic processing on the polarization original image to obtain a polarization image after de-mosaic processing; polarization prior information is constructed based on the polarization image after de-mosaic processing; a light external enhancement model trained by knowledge distillation is constructed by taking a total intensity image as input to obtain external enhancement prior information; a modal complementary reasoning network is constructed, and joint reasoning is performed on the polarization prior information and the external enhancement prior information to obtain an initial de-scattering image; Gamma correction is performed on the initial de-scattering image to obtain a final enhancement image; underwater image quality multi-dimensional evaluation is performed on the final enhancement image; and the enhancement image after evaluation is provided to a downstream task module for calling.

Description

Technical Field

[0001] This invention belongs to the field of intelligent vision processing technology, and relates to an underwater image descattering enhancement method based on active polarization sampling and modal complementarity inference. Background Technology

[0002] With the continuous development of underwater robots, autonomous underwater vehicles, marine observation platforms, and underwater operational equipment, underwater imagery has become a crucial source of information for environmental detection, target identification, attitude estimation, path planning, and operational decision-making. However, suspended particles, dissolved substances, and non-uniform lighting conditions in water bodies collectively alter the light propagation process, leading to common problems in acquired images such as enhanced scattering fog, brightness attenuation, decreased contrast, texture blurring, and color distortion, significantly impacting subsequent visual processing.

[0003] Existing underwater image enhancement or restoration methods mainly include physical model-based methods, empirical enhancement methods, and deep learning-based methods. Physical model-based methods typically rely on background light, scene depth, or medium parameter estimation. While offering some interpretability, they are prone to parameter mismatch in complex and turbid environments. Empirical enhancement methods can improve visual appearance, but they do not adequately consider the scattering formation mechanism, easily leading to over-enhancement, pseudo-textures, and color instability. Deep learning-based methods, while possessing strong representational capabilities, often simply regress the enhancement result directly from the degraded image, lacking a unified modeling of the active polarization imaging mechanism and engineering processing flow.

[0004] Furthermore, existing solutions often focus more on the restoration result itself, while neglecting post-processing correction, quality assessment, adaptation to downstream tasks, and continuous optimization based on feedback. For example, the descattering output often suffers from localized darkening, channel brightness imbalance, or inconsistent visual style, and lacks a multi-dimensional quality evaluation mechanism for underwater scenes, making it difficult to directly support downstream tasks such as detection, segmentation, recognition, and navigation. Therefore, there is an urgent need to propose an underwater image descattering enhancement method with a complete engineering chain and the ability to form a closed-loop optimization mechanism.

[0005] In practical deployments, underwater vision systems typically need to operate long-term under varying turbidity, water depth, lighting conditions, and target categories. If the enhancement method is only effective under offline experimental conditions and lacks coordinated design with the sampling device, image preprocessing module, task invocation module, and feedback optimization module, it will be difficult to stably implement in engineering scenarios. Especially in autonomous robot operations and long-endurance observation tasks, the enhancement results should not only have good subjective visual effects but also maintain the stability of target edges, local textures, and structural relationships as much as possible to provide consistent input for subsequent algorithm modules.

[0006] Therefore, there is a need for an underwater image descattering enhancement method that can cover the entire process of sampling, polarization reconstruction, prior construction, joint inference, post-processing, quality assessment, task invocation, and feedback update, so that underwater image descattering enhancement is no longer an isolated single-point processing step, but an engineering link that can be continuously invoked, evaluated, and optimized. Summary of the Invention

[0007] This invention aims to address the problems in existing technologies, such as insufficient utilization of polarization information, incomplete polarization original image processing flow, insufficient detail recovery capability, poor brightness stability of output results, and insufficient connection between result evaluation and downstream tasks. It provides an underwater image descattering enhancement method based on active polarization sampling and modal complementarity inference.

[0008] This invention provides an underwater image descattering enhancement method based on active polarization sampling and modal complementarity inference, comprising:

[0009] Step 1: Use active polarized light illumination and a polarization camera to acquire images of the target underwater scene, obtaining original polarized images in multiple polarization directions;

[0010] Step 2: Construct a lightweight residual convolutional neural network to perform demosaic processing on the acquired original polarization image to obtain demosaic polarization images in each polarization direction;

[0011] Step 3: Construct polarization prior information based on the de-mosaiced polarization image;

[0012] Step 4: Construct a lightweight external augmentation model trained by knowledge distillation using the total intensity image as input to obtain prior information for external augmentation;

[0013] Step 5: Construct a modal complementary inference network and use polarization prior information and external enhancement prior information for joint inference to obtain the initial descattering image;

[0014] Step 6: Perform Gamma correction on the initial descattered image to obtain the final enhanced image;

[0015] Step 7: Perform a multi-dimensional evaluation of underwater image quality on the final enhanced image;

[0016] Step 8: Provide the enhanced image after evaluation to the downstream task module for invocation.

[0017] The underwater image descattering enhancement method based on active polarization sampling and modal complementarity inference of the present invention has the following beneficial effects:

[0018] 1) Integrate active polarization sampling, polarization demosaicing, polarization prior construction, external prior generation, modal complementarity reasoning, Gamma correction, and quality assessment into a complete engineering closed loop to improve overall feasibility.

[0019] 2) By introducing external enhancement priors and polarization prior information for modal complementarity, the loss of texture and structure caused by polarization feature degradation in highly turbid environments can be effectively compensated.

[0020] 3) By performing Gamma correction on the initial descattered image, problems such as local darkening, channel imbalance and visual inconsistency can be improved, and the brightness balance and color naturalness of the image can be enhanced.

[0021] 4) By introducing quality assessment and downstream task invocation mechanisms, the stability and applicability of the enhanced results in underwater detection, identification, segmentation and navigation perception applications can be improved. Attached Figure Description

[0022] Figure 1 This is a flowchart of an underwater image descattering enhancement method based on active polarization sampling and modal complementarity inference according to the present invention;

[0023] Figure 2 This is a schematic diagram of the modal complementary inference network of the present invention;

[0024] Figure 3 This is a schematic diagram of the structure of the first inference branch of the modal complementary inference network of the present invention;

[0025] Figure 4 This is a schematic diagram of the adaptive fusion module of the modal complementary inference network of the present invention. Detailed Implementation

[0026] like Figure 1 As shown, an underwater image descattering enhancement method based on active polarization sampling and modal complementarity inference according to the present invention includes:

[0027] Step 1: Illuminate the target underwater scene using active polarized light and combine it with a polarization camera to acquire images of the target underwater scene, obtaining raw polarized images in multiple polarization directions, specifically:

[0028] Step 1.1: Before acquiring the image, adjust the exposure parameters, gain parameters, and active light source output parameters of the polarization camera in sequence.

[0029] Step 1.2: Under active polarized light illumination, acquire original polarized images with polarization directions of 0°, 45°, 90°, and 135° respectively. , , , .

[0030] Step 2: Construct a lightweight residual convolutional neural network to perform demosaic processing on the acquired original polarization image, obtaining demosaiced polarization images in each polarization direction, specifically:

[0031] Construct a lightweight residual convolutional neural network, with the original polarized image as input. , , , The output is a de-mosaiced polarization image. , , , .

[0032] The lightweight residual convolutional neural network comprises three parts: a feature extraction layer, a feature transformation layer, and a residual output layer. The feature extraction layer consists of a single-layer convolution and a ReLU activation function. The feature transformation layer consists of several stacked convolutional layers, each containing convolution operations, batch normalization, and ReLU activation. The residual output layer consists of a single-layer convolution with 12 output channels, corresponding to residual maps in four polarization directions. All convolutional kernels in the network are set to a size of 3×3, and all pooling operations are removed. The total number of layers in the network is set to 8, and the number of feature channels in each convolutional layer is set to 32.

[0033] Step 3: Construct polarization prior information based on the de-mosaiced polarization image, specifically as follows:

[0034] Step 3.1: Calculate the Stokes vector based on the de-mosaiced polarization image:

[0035]

[0036]

[0037]

[0038] in, The first Stokes vector, The second Stokes vector, This is the third Stokes vector.

[0039] Step 3.2: Extract four types of polarization prior information based on Stokes vectors: total intensity image, parallel polarization component image, perpendicular polarization component image, and linear polarization degree image;

[0040]

[0041]

[0042]

[0043]

[0044] in, The image shows the total intensity and the image shows the parallel polarization components. Image with vertical polarization component These are defined as components that are parallel to and perpendicular to the polarization direction of the active light source, respectively. This is a linear polarization degree image, describing the proportion of the linear polarization component to the total intensity at each pixel in the scene.

[0045] Step 4: Construct a lightweight external augmentation model trained by knowledge distillation using the total intensity image as input to obtain prior information for external augmentation, specifically:

[0046] A lightweight external augmentation model trained by knowledge distillation is constructed, which consists of four parts: an encoder, a decoder, an input residual branch, and an output layer.

[0047] The encoder consists of five cascaded adaptive downsampling blocks that extract multi-scale features from the input image step by step. Each adaptive downsampling block consists of two consecutive 3×3 convolutions and a ReLU activation function. Except for the first adaptive downsampling block, each of the other adaptive downsampling blocks applies a max pooling operation with a stride of 2 before the convolution to achieve spatial downsampling. The number of output channels of the five encoders are 32, 64, 128, 256, and 512, respectively, corresponding to spatial sizes reduced to 1 / 2, 1 / 4, 1 / 8, and 1 / 16 of the input, thus constructing a complete multi-scale feature pyramid.

[0048] The decoder consists of four upsampled convolutional blocks cascaded sequentially, progressively restoring the spatial resolution of the feature maps. Each upsampled convolutional block first upsamples the low-resolution feature map to the same spatial size as the corresponding encoder feature map using bilinear interpolation. Then, it compresses the number of channels to the target number of channels through two layers of 3×3 convolutions. Next, it is spliced ​​and fused with the skip connection features of the corresponding layer of the encoder along the channel dimension. Finally, it completes feature refinement through two consecutive layers of 3×3 convolutions and ReLU activation. The output channel numbers of the four-level decoder are 256, 128, 64, and 32, respectively, with the spatial size progressively restored to the input resolution.

[0049] The input residual branch is set up in parallel with the backbone network consisting of the encoder and decoder. This branch consists of two 3×3 convolutional layers and directly acts on the network input. The output of the input residual branch is a shallow feature with the same size as the final layer feature map of the decoder. It is then added element-wise with the output of the decoder to obtain fused features, which are then fed into the output layer.

[0050] The output layer consists of a 1×1 convolution and a sigmoid activation function, which maps the fused features to the range of 0 to 1 to obtain three channels of external enhancement prior information.

[0051] When training the lightweight external augmentation model, a multimodal large language model is used as the teacher model. The lightweight external augmentation model is obtained through knowledge distillation training of the multimodal large language model. During distillation training, the multimodal large language model is used to train the total intensity image. The generated scene semantic estimation results are used as teacher soft labels. The output of the lightweight external augmentation model, i.e., the external augmentation prior information. Supervised optimization by applying a joint loss function:

[0052]

[0053]

[0054]

[0055] in, For pixel-level reconstruction loss, constrain the low-frequency consistency between external augmentation prior information and teacher soft labels in pixel space; H is the image height of external augmentation prior information; W is the image width of external augmentation prior information. To perceive the loss, the distance between the intermediate layer features of the pre-trained feature extraction network VGG16 is used to constrain the high-frequency consistency between the external augmentation prior information and the teacher soft labels in terms of semantic content, texture details and structural contours. , and These represent the number of channels, height, and width of the feature map at layer k, respectively. Here, K represents the weight coefficients of the k-th layer; K is the total number of selected feature layers. Indicates As input, the feature map output by the kth layer of the pre-trained feature extraction network; Indicates As input, the feature map output by the kth layer of the pre-trained feature extraction network; , The weighting coefficients are used to balance the two losses.

[0056] Step 5: Construct a modal complementary inference network and use polarization prior information and external enhancement prior information for joint inference to obtain the initial descattering image.

[0057] like Figure 2 As shown, the modal complementarity inference network includes a first inference branch, an adaptive fusion module, and a second inference branch. The results of the first and second inference branches are used to obtain the initial descattered image through physical calculations.

[0058] Step 5.1: Design the first reasoning branch. For example... Figure 3 As shown, the first inference branch consists of four functional modules connected in sequence: a shallow feature extractor, a deep feature aggregation chain, a global cross-block fusion processor, and a physical constraint output layer. After the polarization prior information is input into the first inference branch, it is first mapped to a 64-channel shallow semantic feature space by a shallow feature extractor constructed from two consecutive convolutional blocks, thus obtaining shallow features. The output of the first convolutional layer These features, carrying low-frequency spatial information, are retained for use in the global residual path. Shallow features Then it enters a deep feature aggregation chain consisting of N residual dense blocks connected in series, each residual dense block Internally, a layer-by-layer dense connection mechanism is employed. The output of each layer is concatenated with all preceding features within the residual dense block along the channel dimension and then propagated backward, forming a continuously growing dense feature flow within the block. At the end of the residual dense block, a local feature fusion convolution is configured to compress and integrate the wide-channel dense features. The result is then combined with the input of the residual dense block. Construct local residual connections and output the features of the current residual dense block. .

[0059] Block-level characteristics of the output of each residual dense block After being independently cached, it is sent to the global cross-block merging unit. Global fusion convolution and Refined convolutions perform unified cross-layer weighted integration of all block-level features to obtain deep semantic features. Then, with shallow reference features Element-wise addition is performed, and shallow spatial details are organically integrated with deep high-level semantics through global residual connections, forming a comprehensive feature representation that combines local details and global semantics. The comprehensive feature representation is input into the system... The physically constrained output layer, consisting of output convolution and a sigmoid activation function, is mapped to a three-channel backscattering ratio estimation map. As output, the Sigmoid activation function will... Strictly constrained Within the physically effective range, the estimation results of the network at any spatial location are ensured to satisfy the physical boundedness of the backscattering ratio, thereby guaranteeing the effectiveness and numerical stability of the subsequent physical calculation process.

[0060] Step 5.2: Design the adaptive fusion module. For example... Figure 4 As shown, the adaptive fusion module takes externally enhanced prior information and polarization prior information as dual inputs, and is composed of five functional units: a high-level polarization feature extractor, a high-level semantic feature extractor, a deformable convolutional semantic correction module, a cross-attention fusion module, and a gated feedforward module.

[0061] Externally enhanced prior information is fed into a high-level semantic feature extractor constructed from a pre-trained ResNet50. The two intermediate layers, layer 2 and layer 3, extract semantic features at two scales, with output channels of 512 and 1024, and spatial sizes of 1 / 8 and 1 / 16 of the input, respectively. Then, the number of channels is compressed to 256 and 512, respectively, through two independent 1×1 projection convolutions, resulting in high-level semantic features at two scales.

[0062] Polarization prior information is input into the high-level polarization feature extractor, which extracts high-level polarization features at two scales through five cascaded downsampling convolutional blocks. Each convolutional block consists of a max-pooling layer, two consecutive 3×3 convolutions, and ReLU activation. The outputs of the fourth and fifth downsampling convolutional blocks are 256 channels (1 / 8 of the spatial size) and 512 channels (1 / 16 of the spatial size), respectively, which correspond to the aforementioned high-level semantic features in terms of channel dimension and semantic level.

[0063] The deformable convolutional semantic correction module takes the concatenation of high-level semantic features and high-level polarization features of the same scale as input, and simultaneously predicts the spatial offset and modulation mask through a 3×3 convolution. The offset field describes the spatial displacement of each sampling point, and the modulation mask is constrained between 0 and 1 after Sigmoid activation, serving as the importance weight of each sampling point. The high-level semantic features are then adaptively resampled through deformable convolution with mask modulation, and the output is a corrected semantic feature that is aligned with the high-level polarization features in the spatial semantic distribution.

[0064] The corrected semantic features and the corresponding high-level polarization features are then fed into a cross-attention fusion unit. The cross-attention fusion unit takes the two features as input after being normalized by layers, generates keys and values ​​with polarization features, generates queries with corrected semantic features, calculates cross-modal attention weights through a multi-head transpose attention mechanism, and superimposes the attention output onto the high-level polarization features in a residual connection manner to complete the semantically guided cross-modal feature selection and fusion, and obtains fused features.

[0065] Meanwhile, the high-level semantic features output by the high-level semantic feature extractor are input into the gated feedforward module. After being mapped to twice the hidden layer dimension by a 1×1 convolution, the high-level semantic features are equally split into two features along the channel dimension. One feature is generated by the Sigmoid activation function to generate spatial adaptive gate weights, and the other feature is transformed by the GELU activation function to perform nonlinear feature transformation. The two features are multiplied element by element to obtain the gated feedforward features.

[0066] The final high-level fused feature is obtained by adding the gated feedforward feature and the fused feature together.

[0067] Step 5.3: Design the second inference branch, which consists of four levels of upsampling convolutional blocks connected in series. Taking high-level fused features as input, it progressively passes them down to shallower layers to restore spatial details. Each upsampling convolutional block first performs bilinear interpolation upsampling on the high-level fused features, then compresses the number of channels using a 3×3 convolution, concatenates it along the channel dimension with features of the same scale output from the downsampling convolutional block, and then performs two consecutive 3×3 convolutions to complete cross-scale feature fusion before inputting it to the next level of upsampling convolutional block. The output of the last level of upsampling convolutional block is mapped to a three-channel polarization bounded variable through a 1×1 convolution and Sigmoid activation. Its output value is strictly constrained within the range of (0,1), which satisfies the requirement of the polarization physics model for the boundedness of the direct transmission component.

[0068] Step 5.4: Backscattering ratio estimation plot of the first inference branch output The polarization bounded variable output by the second inference branch The image is input into the physics calculation module to obtain the initial descattering image.

[0069] The physical calculation module relies on an underwater active polarization imaging mechanism to obtain the total intensity image of the scene. Modeled as a polarization-bounded variable Plot with backscattering ratio estimation Linear superposition:

[0070]

[0071]

[0072] The initial descattered image is directly obtained by the following equation:

[0073]

[0074] in, The recovered real-world image, i.e., the initial descattered image; This is a transmittance diagram. The background scattered light intensity.

[0075] Step 6: Perform Gamma correction on the initial descattered image to obtain the final enhanced image.

[0076] In practice, the Gamma parameters of the red, green, and blue channels can be predicted separately using an independent Gamma correction module or a parameter prediction head attached to the main network. Brightness mapping is then performed for each channel to improve issues such as localized darkening, channel imbalance, and inconsistent visual style. The correction process can be represented by the following formula:

[0077]

[0078] in, This indicates the initial descattered image in the color channels. pixel values ​​on This represents the Gamma parameter for the corresponding color channel. This represents the output after Gamma correction. This step enhances brightness uniformity and color naturalness while preserving structural details.

[0079] Step 7: Perform a multi-dimensional evaluation of the underwater image quality of the final enhanced image, specifically:

[0080] Step 7.1: Use the underwater color image quality assessment index UCIQE and the underwater image quality metric UIQM for multi-dimensional evaluation.

[0081] UCIQE is based on the standard deviation of colorimetry. Brightness and contrast and mean saturation A linear combination of these parameters quantitatively characterizes three types of degradation: color cast, blur, and low contrast.

[0082]

[0083] The weighting coefficients are: .

[0084] UIQM is a weighted average of the underwater image color metric UICM, the sharpness metric UISM, and the contrast metric UIConM.

[0085]

[0086] The weighting coefficients are: .

[0087] Step 7.2: Map the multi-dimensional evaluation results to a unified score. After normalizing the above-mentioned indicators, perform weighted fusion to form a comprehensive quality score. :

[0088]

[0089]

[0090] in, This represents the normalized value of the underwater color image quality assessment index. This represents the normalized value of an underwater image quality metric. and The corresponding weights are the weights of the respective evaluation indicators.

[0091] Step 8: Provide the enhanced image after evaluation to the downstream task module for invocation.

[0092] The downstream task module may include: an underwater target detection module, a target recognition module, a semantic segmentation module, an instance segmentation module, a key target measurement module, a navigation perception module, and a scene understanding module. By introducing the enhanced image of this invention before the downstream task, the discernibility of target edges, the separability of foreground and background, and the stability of task execution can be improved. This step reflects that this invention not only focuses on the enhancement result itself, but also on the way the enhancement result is invoked and its application value in a practical underwater vision system. In one embodiment, the enhancement module can be deployed before the task module as a unified visual preprocessing unit to reduce the burden on different task models to adapt to complex underwater degradation conditions. For scenarios requiring real-time processing, enhancement configurations of different complexities can be selected according to task priority to balance effectiveness and latency.

[0093] The underwater image descattering enhancement method based on active polarization sampling and modal complementarity inference of this invention is particularly suitable for scenarios such as target observation in turbid waters, marine life monitoring, underwater structure detection, underwater robot navigation, complex environment perception, and visual preprocessing. Compared with methods that rely solely on single image enhancement or single polarization restoration, this invention can better balance the integrity of the engineering process, the stability of the restoration effect, and the ability to support practical tasks.

[0094] The above description is only a preferred embodiment of the present invention and is not intended to limit the ideas of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An underwater image descattering enhancement method based on active polarization sampling and modal complementarity inference, characterized in that, include: Step 1: Use active polarized light illumination and a polarization camera to acquire images of the target underwater scene, obtaining original polarized images in multiple polarization directions; Step 2: Construct a lightweight residual convolutional neural network to perform demosaic processing on the acquired original polarization image to obtain demosaic polarization images in each polarization direction; Step 3: Construct polarization prior information based on the de-mosaiced polarization image; Step 4: Construct a lightweight external augmentation model trained by knowledge distillation using the total intensity image as input to obtain prior information for external augmentation; Step 5: Construct a modal complementary inference network and use polarization prior information and external enhancement prior information for joint inference to obtain the initial descattering image; Step 6: Perform Gamma correction on the initial descattered image to obtain the final enhanced image; Step 7: Perform a multi-dimensional evaluation of underwater image quality on the final enhanced image; Step 8: Provide the enhanced image after evaluation to the downstream task module for invocation.

2. The underwater image descattering enhancement method based on active polarization sampling and modal complementarity inference according to claim 1, characterized in that, Step 1 specifically involves: Step 1.1: Before acquiring the image, adjust the exposure parameters, gain parameters, and active light source output parameters of the polarization camera in sequence; Step 1.2: Under active polarized light illumination, acquire original polarized images with polarization directions of 0°, 45°, 90°, and 135° respectively. , , , .

3. The underwater image descattering enhancement method based on active polarization sampling and modal complementarity inference according to claim 1, characterized in that, Step 2 specifically involves: Construct a lightweight residual convolutional neural network, with the original polarized image as input. , , , The output is a de-mosaiced polarization image. , , , ; The lightweight residual convolutional neural network comprises three parts: a feature extraction layer, a feature transformation layer, and a residual output layer. The feature extraction layer consists of a single-layer convolution and a ReLU activation function. The feature transformation layer consists of several stacked convolutional layers, each containing convolution operations, batch normalization, and ReLU activation. The residual output layer consists of a single-layer convolution with 12 output channels, corresponding to residual maps in four polarization directions. All convolutional kernels in the network are set to a size of 3×3, and all pooling operations are removed. The total number of layers in the network is set to 8, and the number of feature channels in each convolutional layer is set to 32.

4. The underwater image descattering enhancement method based on active polarization sampling and modal complementarity inference according to claim 1, characterized in that, Step 3 specifically involves: Step 3.1: Calculate the Stokes vector based on the de-mosaiced polarization image: in, The first Stokes vector, The second Stokes vector, The third Stokes vector; Step 3.2: Extract four types of polarization prior information based on Stokes vectors: total intensity image, parallel polarization component image, perpendicular polarization component image, and linear polarization degree image; in, The image shows the total intensity and the image shows the parallel polarization components. Image with vertical polarization component These are defined as components that are parallel to and perpendicular to the polarization direction of the active light source, respectively. This is a linear polarization degree image, describing the proportion of the linear polarization component to the total intensity at each pixel in the scene.

5. The underwater image descattering enhancement method based on active polarization sampling and modal complementarity inference according to claim 4, characterized in that, Step 4 specifically involves: A lightweight external augmentation model trained by knowledge distillation is constructed, which consists of four parts: an encoder, a decoder, an input residual branch, and an output layer. The encoder consists of five cascaded adaptive downsampling blocks that extract multi-scale features from the input image step by step. Each adaptive downsampling block consists of two consecutive 3×3 convolutions and a ReLU activation function. Except for the first adaptive downsampling block, each of the other adaptive downsampling blocks applies a max pooling operation with a stride of 2 before the convolution to achieve spatial downsampling. The number of output channels of the five-level encoder is 32, 64, 128, 256, and 512, respectively, corresponding to spatial sizes reduced to 1 / 2, 1 / 4, 1 / 8, and 1 / 16 of the input, thus constructing a complete multi-scale feature pyramid. The decoder consists of four upsampled convolutional blocks cascaded sequentially, progressively restoring the spatial resolution of the feature maps. Each upsampled convolutional block first upsamples the low-resolution feature map using bilinear interpolation to the same spatial size as the corresponding encoder feature map. Then, it compresses the number of channels to the target number of channels through two 3×3 convolutions. Next, it is concatenated and fused with the skip connection features of the corresponding layer of the encoder along the channel dimension. Finally, it completes feature refinement through two consecutive 3×3 convolutions and ReLU activation. The output channel numbers of the four-stage decoder are 256, 128, 64, and 32 respectively, with the spatial size progressively restored to the input resolution. The input residual branch is set up in parallel with the backbone network consisting of the encoder and decoder. This branch is composed of two layers of 3×3 convolutions and directly acts on the network input. The output of the input residual branch is a shallow feature with the same size as the feature map of the final layer of the decoder. It is then added element by element to the output of the decoder to obtain fused features, which are then fed into the output layer. The output layer consists of a 1×1 convolution and a sigmoid activation function, which maps the fused features to the range of 0 to 1 to obtain three channels of external enhancement prior information.

6. The underwater image descattering enhancement method based on active polarization sampling and modal complementarity inference according to claim 5, characterized in that, When training the lightweight external augmentation model, the multimodal large language model is used as the teacher model. The lightweight external augmentation model is obtained by knowledge distillation training on the multimodal large language model. During the distillation training process, a multimodal large language model is used to target the total intensity image. The generated scene semantic estimation results are used as teacher soft labels. The output of the lightweight external augmentation model, i.e., the external augmentation prior information. Supervised optimization by applying a joint loss function: in, For pixel-level reconstruction loss, constrain the low-frequency consistency between external augmentation prior information and teacher soft labels in pixel space; H is the image height of external augmentation prior information; W is the image width of external augmentation prior information. To perceive the loss, the distance between the intermediate layer features of the pre-trained feature extraction network VGG16 is used to constrain the high-frequency consistency between the external augmentation prior information and the teacher soft labels in terms of semantic content, texture details and structural contours. , and These represent the number of channels, height, and width of the feature map at layer k, respectively. Here, K represents the weight coefficients of the k-th layer; K is the total number of selected feature layers. Indicates As input, the feature map output by the kth layer of the pre-trained feature extraction network; Indicates As input, the feature map output by the kth layer of the pre-trained feature extraction network; , The weighting coefficients are used to balance the two losses.

7. The underwater image descattering enhancement method based on active polarization sampling and modal complementarity inference according to claim 1, characterized in that, The modal complementary inference network in step 5 includes a first inference branch, an adaptive fusion module, and a second inference branch. Step 5.1: Design the first inference branch, which consists of four functional modules connected in sequence: a shallow feature extractor, a deep feature aggregation chain, a global cross-block fusion unit, and a physical constraint output layer. After the polarization prior information is input into the first inference branch, it is first mapped to a 64-channel shallow semantic feature space by a shallow feature extractor constructed from two consecutive convolutional blocks, thus obtaining shallow features. The output of the first convolutional layer It is retained as a benchmark feature carrying low-frequency spatial information; Shallow features The system then enters a deep feature aggregation chain consisting of N residual dense blocks connected in series. Each residual dense block employs a layer-by-layer dense connection mechanism, where the output of each layer is concatenated with all preceding features within the residual dense block along the channel dimension and then passed forward. At the end of each residual dense block, a local feature fusion convolution is configured to compress and integrate the dense features. The result is then combined with the input of that residual dense block. Construct local residual connections and output the features of the current residual dense block. ; Block-level characteristics of the output of each residual dense block After being independently cached, it is sent to the global cross-block merging unit. Global fusion convolution and Refined convolutions perform unified cross-layer weighted integration of all block-level features to obtain deep semantic features. Then, with shallow reference features Perform element-wise addition to obtain the comprehensive feature representation. ; Comprehensive feature representation input to by The physically constrained output layer, consisting of output convolution and a sigmoid activation function, is mapped to a three-channel backscattering ratio estimation map. As output; Step 5.2: Design an adaptive fusion module. The adaptive fusion module takes external enhancement prior information and polarization prior information as dual inputs and consists of five functional units: a high-level polarization feature extractor, a high-level semantic feature extractor, a deformable convolutional semantic correction module, a cross-attention fusion module, and a gated feedforward module. Externally enhanced prior information is fed into a high-level semantic feature extractor constructed from pre-trained ResNet50. Semantic features at two scales are extracted through its two intermediate layers, layer 2 and layer 3, respectively. Then, high-level semantic features at two scales are obtained through two independent 1×1 projection convolutions. The polarization prior information is input into the high-level polarization feature extractor, which extracts high-level polarization features at two scales that are the same as the high-level semantic feature scale through five-level cascaded downsampling convolutional blocks. The deformable convolutional semantic correction module takes the concatenation of high-level semantic features and high-level polarization features of the same scale as input, and simultaneously predicts the spatial offset and modulation mask through a 3×3 convolution. The offset field describes the spatial displacement of each sampling point, and the modulation mask is constrained between 0 and 1 after being activated by Sigmoid, serving as the importance weight of each sampling point. The high-level semantic features are then adaptively resampled via deformable convolution with mask modulation, outputting corrected semantic features that are aligned with the high-level polarization features in the spatial semantic distribution. The corrected semantic features and the corresponding high-level polarization features are then fed into a cross-attention fusion unit. The cross-attention fusion unit takes the two features as input after being normalized by layers, generates keys and values ​​with polarization features, generates queries with corrected semantic features, calculates cross-modal attention weights through a multi-head transpose attention mechanism, and superimposes the attention output onto the high-level polarization features in a residual connection manner to obtain fused features. Meanwhile, the high-level semantic features output by the high-level semantic feature extractor are input into the gated feedforward module. After the high-level semantic features are mapped to twice the hidden layer dimension by a 1×1 convolution, they are equally split into two features along the channel dimension. One feature is generated by the Sigmoid activation function to generate spatial adaptive gate weights, and the other feature is transformed by the GELU activation function. The two features are multiplied element by element to obtain the gated feedforward features. The obtained gated feedforward features and fused features are added together to obtain the final high-level fused features; Step 5.3: Design the second inference branch, which consists of four levels of upsampling convolutional blocks connected in series. Taking the high-level fused features as input, each upsampling convolutional block first performs bilinear interpolation upsampling on the high-level fused features. Then, after 3×3 convolution to compress the number of channels, it is concatenated along the channel dimension with the features of the same scale output by the downsampling convolutional block. After two consecutive 3×3 convolutions to complete cross-scale feature fusion, it is input to the next level of upsampling convolutional block. The output of the last level of upsampling convolutional block is mapped to a three-channel polarization bounded variable through a 1×1 convolution and Sigmoid activation. Its output value is strictly constrained within the range of (0,1), which satisfies the requirement of the polarization physics model for the boundedness of the direct transmission component; Step 5.4: Backscattering ratio estimation plot of the first inference branch output The polarization bounded variable output by the second inference branch Input into the physics calculation module to obtain the initial descattering image: The physical calculation module relies on an underwater active polarization imaging mechanism to obtain the total intensity image of the scene. Modeled as a polarization-bounded variable Plot with backscattering ratio estimation Linear superposition: The initial descattered image is directly obtained by the following equation: in, The recovered real-world image, i.e., the initial descattered image; This is a transmittance diagram. The background scattered light intensity.

8. The underwater image descattering enhancement method based on active polarization sampling and modal complementarity inference according to claim 1, characterized in that, Step 7 specifically involves: Step 7.1: Use the underwater color image quality assessment index UCIQE and the underwater image quality metric UIQM for multi-dimensional evaluation; UCIQE is based on the standard deviation of colorimetry. Brightness and contrast and mean saturation A linear combination of these parameters quantitatively characterizes three types of degradation: color cast, blur, and low contrast. The weighting coefficients are: ; UIQM is a weighted average of the underwater image color metric UICM, the sharpness metric UISM, and the contrast metric UIConM. The weighting coefficients are: ; Step 7.2: Map the multi-dimensional evaluation results to a unified score. After normalizing the above-mentioned indicators, perform weighted fusion to form a comprehensive quality score. : in, This represents the normalized value of the underwater color image quality assessment index. This represents the normalized value of an underwater image quality metric. and The corresponding weights are the weights of the respective evaluation indicators.

Images