An underwater image depth estimation method based on adaptive physical information and saliency guidance for real-time underwater robots

Abstract

The application discloses a kind of underwater image depth estimation methods based on adaptive physical information and saliency guidance for real-time underwater robot, belong to computer vision technical field. Including: based on training set, underwater image depth estimation model based on adaptive physical information and saliency guidance is trained, model includes physical guided TMC representation module, coding-decoding module, attenuation perception target attention module and light weight visual transmission module;Acquire the underwater RGB image to be processed, the underwater RGB image to be processed is input into the underwater image depth estimation model based on adaptive physical information and saliency guidance after training, the depth estimation value of model output is acquired, to realize the depth estimation of underwater image.The application mainly utilizes underwater optical physical characteristics to combine salient object guidance, accurately carries out depth estimation in highly degraded and chaotic underwater scene, with high precision, high efficiency and high robustness.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to an adaptive physical information and saliency-guided underwater image depth estimation method for real-time underwater robots. Background Technology

[0002] In underwater environmental perception tasks, accurate depth estimation plays a fundamental role in achieving precise navigation, environmental monitoring, and complex operational tasks. While monocular depth estimation has achieved remarkable success on land, the underwater environment presents unique and severe challenges due to the complex interactions of light attenuation, scattering, and turbidity. These factors significantly degrade the quality of visual cues and obscure object boundaries necessary for reliable scene understanding.

[0003] Monocular depth estimation for underwater scenes provides rich, high-resolution scene information. Currently, common methods include traditional depth estimation relying on image forming models (IFM) or dedicated hardware. However, due to the variable underwater environment, the practical performance of existing methods is not ideal. In recent years, deep learning-based methods have attracted attention, but their generalization is limited by the difficulty in obtaining dense underwater RGB-D datasets. Furthermore, methods using synthetic datasets or air-to-water domain transfer techniques cannot fully capture the inherent physical characteristics and complex visual statistical properties of underwater scenes, further limiting their generalization performance. While unsupervised methods can alleviate the data scarcity problem, they still face challenges related to domain adaptation and ambiguity.

[0004] Recent underwater monocular depth estimation methods, such as UDepth and UWDepth, have incorporated underwater domain knowledge and depth priors to improve performance. However, these methods often struggle to accurately separate salient foreground objects from water columns, leading to incorrect background depth assignment and hindering their effectiveness in downstream robotic operations. Summary of the Invention

[0005] In view of the shortcomings of existing technologies, this invention provides an underwater image depth estimation method based on adaptive physical information and salient object guidance for real-time underwater robots. This invention mainly utilizes underwater optical-physical properties combined with salient object guidance to accurately estimate depth in highly degraded and cluttered underwater scenes, exhibiting high accuracy, high efficiency, and high robustness.

[0006] The technical means employed in this invention are as follows: An underwater image depth estimation method based on adaptive physical information and saliency guidance for real-time underwater robots includes: Obtain an underwater image dataset, and divide a portion of the data in the underwater image dataset into a training set; An underwater image depth estimation model based on adaptive physical information and saliency guidance is constructed. The model is trained on the training set. The model includes a physics-guided TMC representation module, an encoder-decoder module, an attenuation-aware target attention module, and a lightweight visual transmission module. The physics-guided TMC representation module generates a TMC input tensor from the input RGB image. The encoder-decoder module generates a hybrid feature set of fused data-driven depth cues and physical depth cues from the input RGB image. The attenuation-aware target attention module generates enhanced features from the input TMC input tensor and the hybrid feature set. The lightweight visual transmission module models the context of the enhanced features and obtains the final depth estimate and the Shannon entropy of the depth distribution based on the predicted depth distribution parameters. The Shannon entropy of the depth distribution represents the uncertainty of the prediction result. The underwater RGB image to be processed is acquired, and the underwater RGB image to be processed is input into the trained underwater image depth estimation model based on adaptive physical information and saliency guidance. The depth estimate value output by the model is obtained, thereby realizing the depth estimation of the underwater image.

[0007] Furthermore, the physically guided TMC representation module first processes the input RGB image through channel processing, and calculates and generates a transmission map, a model-based log-ratio decay feature, and a chromaticity feature respectively; then, it concatenates the transmission map, the model-based log-ratio decay feature, and the chromaticity feature to generate a three-channel TMC input tensor.

[0008] Furthermore, the encoder-decoder module is built on MobileNetV2 and includes an encoder and a decoder. The encoder first extracts features from the input three-channel RGB image through an initial convolutional layer with 32 convolutional kernels, and then extracts hierarchical features through 19 residual bottleneck layers. The decoder further generates decoded features with a resolution of 320×240 and 48 channels.

[0009] Furthermore, the attenuation-aware target attention module includes an object-aware branch, wherein the attenuation-aware branch includes an attenuation-aware modulation channel and a spatial attention channel; In the attenuation-sensing modulation channel, the three-channel TMC input tensor is first compressed to eight channels using 1×1 convolution, and then global context information is extracted through global average pooling to obtain an 8×1×1 feature vector. Subsequently, it is processed by two consecutive 1×1 convolutional layers to generate cue confidence weights. In the spatial attention channel, decoder features It is concatenated with the compressed TMC representation and processed through a 3×3 convolutional layer with a Sigmoid activation function to generate the initial spatial attention mask M; In the object perception branch, the initial spatial attention mask M is modulated based on the cue confidence weights to achieve spatial attention enhancement under physical constraints. The modulated attention mask is then convolved with a fixed Laplacian kernel to achieve edge enhancement, and refined output features are generated by combining the decoded features.

[0010] Furthermore, the lightweight visual transmission module will input feature maps. Divided into non-overlapping images Blocks to generate Each feature token is labeled and projected onto the embedding dimension. E Subsequently, these features are processed through a Transformer encoding layer that includes multi-head self-attention and a feedforward network, thereby enabling global contextual information exchange between different image patches. A set of attention maps generated by Transformer dot products and input feature maps The convolutional projection features are interacted with to generate a range attention map; at the same time, the output regression label of the Transformer is processed through two layers of MLP with a hidden dimension of 256 to predict the depth distribution parameters. By performing softmax normalization on the projection results of the Transformer's output regression labels, a pixel-wise ordered depth distribution is generated. :

[0011] Among them, MLP ( ) indicates a shared multi-layer perceptron. This represents the Sigmoid activation function. Indicates the projection of the regression label; Final depth estimate By adjusting the center of the dynamically adjusted depth interval The weighted expectation is calculated as follows:

[0012] in This represents the pixel depth projected onto the center of the i-th interval. The upper limit of the dimension representing the center of the interval; To quantify prediction confidence and identify uncertain regions, the Shannon entropy of the depth distribution is calculated. :

[0013] Furthermore, an underwater image depth estimation model based on adaptive physical information and saliency guidance is trained according to the following loss function:

[0014] in, For the total training loss, For scale-invariant logarithmic loss, For depth-weighted loss, For significance-weighted L1 loss, This is the entropy regularization loss.

[0015] Compared with the prior art, the present invention has the following advantages: The proposed method is applicable to complex underwater environments with limited visibility. It addresses the severe visual degradation caused by light attenuation and scattering, which disrupts depth cues in underwater scenes. It integrates a transmission model-based log-ratio attenuation contrast (Chromaticity TMC) input representation to explicitly encode underwater optical physical properties, and introduces an attenuation-aware object attention (AAOA) mechanism to dynamically select reliable visual cues and refine object boundaries. Furthermore, a lightweight mini-Vision Transformer (mViT) is used to optimize global spatial relationships, thereby ensuring depth consistency.

[0016] This invention reformulates traditional underwater transport estimation into a fast single-channel algorithm, achieving a 654x speedup and enabling real-time inference. Evaluation results on the USOD10K and FLSea benchmark datasets demonstrate that APUDepth achieves state-of-the-art performance, reducing absolute relative error (Abs Rel) by 37% and improving overall accuracy by 20% compared to CD-UDepth. It achieves real-time performance of 49 FPS on a server-side GTX3070 GPU and 35 FPS on an edge-side Jetson Orin GPU.

[0017] The method proposed in this invention has high precision, high efficiency and high robustness, and is suitable for real-time underwater robot navigation and operation tasks. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a diagram illustrating the underwater image depth estimation model architecture based on adaptive physical information and saliency guidance in an embodiment of the present invention.

[0020] Figure 2 This is a visualization comparison chart of transmission estimation in an embodiment of the present invention.

[0021] Figure 3 This invention presents a comparison of the visual effects of different methods in green-dominated and blue-dominated underwater scenes on the USOD10K dataset.

[0022] Figure 4 This invention presents a comparison of the visual effects of different methods in turbid and blurry underwater scenes on the USOD10K dataset, as described in this embodiment.

[0023] Figure 5 This is a visualization of the uncertainty map in an embodiment of the present invention.

[0024] Figure 6 This demonstrates the cross-domain generalization effect on the FLSea dataset in this embodiment of the invention.

[0025] Figure 7 This refers to the temporal consistency of video frames in this embodiment of the invention. Detailed Implementation

[0026] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0027] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0028] This invention provides an underwater image depth estimation method based on adaptive physical information and saliency guidance for real-time underwater robots, which mainly includes the following steps.

[0029] S1. Obtain the underwater image dataset and divide a portion of the data in the underwater image dataset into a training set.

[0030] This application uses two underwater datasets, USOD10K and FLSea, for training and evaluation. The USOD10K benchmark dataset contains 10,255 RGB images and their corresponding true depth maps, with a resolution of 640×480 pixels. This dataset is divided into a training set of 7,178 images, a validation set of 1,539 images, and a test set of 1,538 images. As a widely recognized and challenging underwater monocular depth estimation benchmark, USOD10K covers a variety of underwater environments with different lighting conditions, turbidity levels, depth ranges, and rich distributions of object sizes and categories. This diversity provides a sufficient data foundation for the main training and evaluation of the model. The FLSea benchmark dataset consists of two main subsets, reflecting different underwater environments: Canyons and the Red Sea. The Canyons subset includes four scenes (U Canyon, Horse Canyon, Tiny Canyon, and Flatiron), totaling 8,592 frames. The Red Sea subset comprises eight scenes (Big Dice Loop, CoralTable Loop, Cross Pyramid Loop, Dice Path, Northeast Path, Landward Path, PierPath, and Sub Pier), totaling 22,451 frames. In the cross-domain evaluation, 50 images were randomly selected from each scene in the FLSea benchmark dataset, resulting in 600 test images. This evaluation scheme was used to verify the model's generalization ability in unseen underwater environments.

[0031] S2. Construct an underwater image depth estimation model based on adaptive physical information and saliency guidance, and train the underwater image depth estimation model based on the training set.

[0032] Specifically, such as Figure 1 As shown, the underwater image depth estimation model based on adaptive physical information and saliency guidance includes a physically guided TMC representation module, an encoder-decoder module, an attenuated sensing target attention module, and a lightweight visual transmission module.

[0033] A physically guided TMC (Transmission, Model-based log-ratio attenuation contrast, Chromaticity) representation module is used to generate TMC input tensors from the input RGB image. This application proposes a physically-based input transformation method—TMC representation—which is a combination of Transmission, Model-based log-ratio attenuation contrast, and Chromaticity. This representation method is based on the Jaffe-McGlamery model, which describes the attenuation-chromaticity at wavelengths... The observed pixel intensity is used as a function of scene radiance, depth, light attenuation, and background light:

[0034] in, Represents the radiance of a real-world scene. Indicates scene depth. Indicates the attenuation coefficient. Indicates the light attenuation coefficient. This represents the background light intensity. The model reveals key challenges in underwater imaging, including strong red light attenuation (…). And the problem of backscattering being dominant.

[0035] The core idea of ​​this application is that no single clue can guarantee high reliability in all situations. Therefore, TMC representation maintains a monotonic relationship with depth by fusing multiple complementary clues, thus ensuring robust operation under various water quality conditions. TMC inputs are dynamically generated on-the-fly for each sample, allowing for efficient integration into real-time deep learning transport pipeline results. Its refined transport graph... This method is used to capture depth-related light attenuation and scattering information. The computation employs a local block size of 15, a weighting factor of 0.95, and refinement using a guided filter with a filter radius of 60 and a regularization parameter of 0.001. This reconstruction method allows for single-pass, parallelizable computation, is suitable for GPU acceleration, and can generate high-quality transport maps that preserve depth structure.

[0036] The original transmission diagram calculation formula is expressed as follows:

[0037] in, Indicates background light. Indicates the strength of the green channel. Indicates the intensity of the blue channel. Indicates the weight value. This is used to ensure the stability of numerical calculations. After guided filtering, the final refined transport map can be obtained. .

[0038] To address the rapid attenuation of the red channel in blue-green water, this invention, based on the Beer-Lambert law, derives a model-based logarithmic attenuation comparison term from the intensity ratio of the green to blue channels. Neglecting backscattering, the observed light intensity exhibits an exponential decay with depth:

[0039] Taking the logarithm of the intensity ratio of the green and blue channels yields a depth-dependent feature:

[0040] in, Indicates the green channel in pixels The intensity of the position, Indicates the blue channel in pixels The intensity of the location.

[0041] Subsequently, instance normalization is performed on the feature to stabilize its dynamic range and improve the model's learning efficiency.

[0042] Chromaticity achieves light invariance by emphasizing relative color balance, and it is calculated as follows:

[0043] Indicated in pixels The chromaticity value of the location, Indicates the red channel in pixels The intensity of the location.

[0044] The resulting three single-channel feature maps are concatenated together to form a three-channel TMC input tensor.

[0045] This physically-guided parameterized input ensures at least one reliable depth-related cue exists under any type of water body condition, while chromaticity provides robustness to illumination variations. This design supports efficient, real-time, per-sample generation, providing a stable foundation for subsequent depth estimation and learning-based fusion.

[0046] The encoder-decoder module generates a hybrid feature set that fuses data-driven depth cues and physical depth cues from the input RGB image. This application employs an encoder-decoder backbone based on MobileNetV2. The encoder first extracts features from the three-channel RGB input through an initial convolutional layer with 32 kernels, followed by 19 residual bottleneck layers to extract hierarchical features. The decoder further generates a feature map with a resolution of 320×240 and 48 channels. The resulting feature map is concatenated with the physically guided TMC representation to form a hybrid feature set that fuses data-driven depth cues and physical depth cues, which is then robustly processed by the subsequent attenuated perceived target attention (AAOA) module. The entire architecture is optimized for computational efficiency, enabling real-time inference on resource-constrained underwater robotic platforms.

[0047] The Attenuation-Aware Object Attention (AAOA) module generates enhanced features based on the input TMC input tensor and the mixed feature set. An AAOA module is introduced to address the unreliable foreground-background separation problem in underwater scenes by dynamically recalibrating the importance of features using physical attenuation cues in the TMC representation. This module feeds the input features from the decoder and the TMC features into the attenuation-aware branch and the spatial attention branch respectively for parallel processing, and finally generates refined output features through a physics-conditioned gating mechanism. .

[0048] In the decay-aware branch, the three-channel TMC input is first compressed to eight channels using a 1×1 convolution (ReLU activation), and then global contextual information is extracted through global average pooling to obtain an 8×1×1 feature vector. This vector is then processed through two consecutive 1×1 convolutional layers to generate cue confidence weights.

[0049] in, These represent the confidence estimates for the transmission cue and the log-ratio decay cue, respectively, with their values ​​determined by the Sigmoid activation function. It is constrained to the range [0,1].

[0050] Meanwhile, the spatial attention branch incorporates decoder features. Compared with the compressed TMC representation The data is concatenated and processed through a 3×3 convolutional layer with a sigmoid activation function to generate an initial spatial attention mask. The mask is then modulated by confidence weights based on a physics model, thereby achieving spatial attention enhancement under physical constraints:

[0051] This modulation mechanism can dynamically adjust the focus of attention under different water quality conditions: enhancing transmission cues in clear waters, while highlighting log-ratio contrast cues in blue-green or turbid waters.

[0052] The modulated attention mask is then convolved with a fixed Laplacian kernel (L) to achieve edge enhancement, further improving the detail representation of the target boundary.

[0053] Final refined output features By using the original decoder features Combined with attenuation mask and edge enhancement results The combined mask is obtained by element-wise multiplication.

[0054] This method can adaptively enhance stable depth cues and sharpen target boundaries, thereby effectively overcoming the common problem of foreground-background separation in heterogeneous underwater environments.

[0055] The lightweight vision transport (mini Vision Transformer, mViT) module is used to model the contextual relationships of the input augmented features and obtain the final depth estimate and the Shannon entropy of the depth distribution based on the predicted depth distribution parameters. The Shannon entropy of the depth distribution is used to represent the uncertainty of the prediction result. This application introduces a MiniVision Transformer (mViT) module, which refines the features output by the AAOA module through contextual relationship modeling, maintaining computational efficiency suitable for real-time operation while ensuring local accuracy and global consistency.

[0056] The mViT module will input feature maps Divided into non-overlapping images Blocks are used to generate feature tags. tokens and project them onto the embedding dimension. E Subsequently, these features are processed by a Transformer encoding layer that includes multi-head self-attention and a feedforward block, thereby enabling global contextual information exchange between different image patches to solve the depth blur problem that occurs in murky or occluded areas.

[0057] Furthermore, a set of learnable query vectors interacts with the convolutional projection features of the original input to generate range attention maps, highlighting key depth-related information. Simultaneously, a dedicated regression token is processed through two layers of an MLP with a hidden dimension of 256 to predict depth distribution parameters.

[0058] mViT generates a pixel-wise ordered depth distribution by performing softmax normalization on the projection results of the regression tokens. :

[0059] Among them, MLP ( ) indicates a shared multi-layer perceptron. This represents the Sigmoid activation function. This represents the regression label projection.

[0060] Final depth estimate By adjusting the center of the dynamically adjusted depth interval The weighted expectation is calculated as follows:

[0061] in This represents the pixel depth projected onto the center of the i-th interval. This represents the upper dimension of the center of the interval.

[0062] To quantify prediction confidence and identify uncertain regions, this module calculates the Shannon entropy of the depth distribution. :

[0063] Higher entropy values ​​indicate regions where predictions are uncertain; these regions typically correspond to homogeneous water backgrounds or areas affected by strong scattering. This indicator provides a basis for assessing the reliability of prediction results in the subsequent robot decision-making process.

[0064] Furthermore, the underwater image depth estimation model constructed in this application, based on adaptive physical information and saliency guidance, employs a multi-component loss function during training. By weightedly combining multiple complementary objective terms, it achieves comprehensive constraints on physical consistency, saliency perception, and uncertainty handling. Its overall training objective function is defined as follows:

[0065] Among them, the weighting coefficients of each loss term Cross-validation is used to determine the balance of contributions of each component to the overall loss.

[0066] Regarding scale-invariant logarithmic loss The calculation formula is as follows:

[0067] in, Indicates the number of valid pixels. and To predict the actual depth value, This refers to the scale-invariant control weights. Scale-invariant logarithmic loss. This enhances the model's robustness to global scale variations in depth prediction.

[0068] Regarding depth-weighted loss The calculation formula is as follows:

[0069] Where Ω represents the set of valid pixels. Depth-weighted loss. By employing a weighting design that is inversely proportional to the true depth, the prediction accuracy of the foreground target region is improved, using a significance-weighted L1 loss. By using the saliency map generated by TC-USOD, the network is guided to focus on visually salient regions to further optimize the depth estimation performance.

[0070]

[0071] Where Ω represents the pixel domain, This represents the significance value after bilinear interpolation. This represents the predicted depth of pixel x. This represents the true depth value of pixel x. Represents pixels Significance weight at each position Ω represents the number of pixels with effective depth in the pixel field.

[0072] A key innovation of this invention lies in the introduction of an entropy regularization term. This item utilizes the prediction entropy map generated by the mViT module. Minimize prediction uncertainty in well-constrained regions, thereby improving the stability and reliability of the model in depth estimation:

[0073] in, Indicates pixel position The prediction entropy at a given location is used to quantify the uncertainty of deep prediction.

[0074] This design stabilizes the training process in ambiguous underwater regions by suppressing high-entropy predictions in areas with sufficient visual information, while preserving necessary uncertainty in truly challenging regions. This integrated loss structure enables the model to simultaneously optimize scale invariance, foreground accuracy, salient region attention, and prediction confidence under various underwater environmental conditions.

[0075] S3. Obtain the underwater RGB image to be processed, input the underwater RGB image to be processed into the trained underwater image depth estimation model based on adaptive physical information and saliency guidance, obtain the depth estimate value output by the model, thereby realizing the depth estimation of the underwater image.

[0076] The following specific application examples will further illustrate the solution and effects of the present invention.

[0077] This embodiment presents a physically guided underwater depth estimation method, the implementation steps of which are as follows.

[0078] Step 1: Dataset Configuration and Partitioning This invention uses two comprehensive underwater benchmark datasets: USOD10K and FLSea.

[0079] Step 2 Parameter Configuration and Model Building The training process is implemented using PyTorch 2.0.1, employing the AdamW optimizer, with an initial learning rate of 1×10⁻⁶. -4 The depth estimation method decreases exponentially by 0.9 every 10 epochs. Training is performed on an NVIDIA RTX3070 GPU with a batch size of 16. During inference, the depth estimation method disclosed in this application achieves a real-time performance of 35 FPS on the Jetson AGX Orin platform without model compression or quantization, validating its practicality and real-time performance in underwater robotic tasks. To enhance model robustness, a saliency mask generated by the TC-USOD saliency detector is optionally introduced as an auxiliary supervision signal during training; this is omitted during inference to avoid any latency. The complete underwater image depth estimation model framework based on adaptive physical information and saliency guidance contains approximately 15.67 million trainable parameters, distributed as follows: MobileNetV2 encoder: 6.8M parameters.

[0080] Decoder: 5.4M parameters.

[0081] Mini Vision Transformer (mViT) and regression head: 3.4M parameters.

[0082] Each forward inference requires 8.11 GFLOPs, achieving an optimal balance between model complexity and computational efficiency.

[0083] Step 3 Experimental Evaluation and Performance Analysis This method was trained and evaluated on the USOD10K and FLSea datasets, and compared with methods such as IBLA, UWNet, LapDepth, Lite-Mono, BTS, WaterMono, CD-UDepth, and UDepth to verify its advantages in accuracy, generalization ability, and computational efficiency.

[0084] The method of this invention exhibits excellent performance in various underwater environments, including blue-green waters, green waters, and turbid environments. For example... Figure 3 and Figure 4As shown, compared with existing methods, APUDepth generates results with clear target boundaries, consistent depth levels, and stable foreground-background separation. Furthermore, the saliency-guided training mechanism, which introduces the TC-USOD saliency mask, effectively improves the model's depth estimation performance, enabling it to accurately focus on salient underwater structures without increasing inference latency.

[0085] Experimental results (see Tables 1 and 2) show that the method of the present invention is significantly better than the comparative model in all indicators.

[0086] Table 1 Quantitative performance metrics on the USOD10K dataset

[0087] Table 2 Quantitative performance metrics on the FLSea dataset

[0088] Specifically: On the USOD10K dataset, the method of this application achieves the following results: Absolute relative error (Abs Rel): 0.43 (37% improvement over UDepth and 12% improvement over CD-UDepth); Squared relative error (Sq Rel): 0.06; Root mean square error (RMSE): 0.11; Log10 error: 0.15.

[0089] In cross-domain evaluation on the FLSea dataset (without retraining), APUDepth achieved AbsRel=0.96, surpassing all baseline methods and demonstrating excellent cross-environment generalization ability. This result is particularly important because changes in underwater visibility and turbidity significantly affect depth sensing performance.

[0090] In contrast, other methods often produce incorrect depth predictions in areas lacking effective visual cues, while this invention achieves reliable cue selection and robust prediction under various water quality conditions through a physically driven TMC representation and an adaptive AAOA module.

[0091] In addition to accuracy and generalization ability, this invention also has significant advantages in computational efficiency, as shown in Table 3.

[0092]

[0093] Table 3 Comparison of Estimated Transmission Speeds (UDCP vs. APUDepth Fast Version) The performance of the method in this application is as follows: 49.13 FPS (High-end GPU, NVIDIA RTX 3070) 35 FPS (Embedded platform NVIDIA Jetson AGX Orin) Computational requirements: 19.30 GFLOPs In comparison, the LapDepth model (73.57M parameters) is more than four times larger and slower, while CD-UDepth requires 46.18 GFLOPs. APUDepth, through its lightweight design, achieves the best balance between performance and efficiency, making it an ideal choice for real-time depth estimation in resource-constrained underwater environments.

[0094] Furthermore, an ablation experiment was conducted on the method of this application in this embodiment. Systematic ablation experiments (see Table 4) further verified the effectiveness of each module: Removed the AAOA module: Abs Rel increased from 0.43 to 0.55. Removing the TMC input resulted in the most significant performance degradation, with Abs Rel reaching 0.58. The complete TMC triplet (T+M+C) achieves optimal performance, with all loss terms contributing to improved training stability and accuracy.

[0095] Table 4 shows the ablation results of each component of the TMC input space on the USOD10K test set.

[0096] In summary, the method proposed in this invention effectively solves key problems such as underwater image degradation, foreground-background confusion, and excessive model complexity. Experimental results show that the method of this application surpasses existing methods in both accuracy and efficiency, achieving high-speed and accurate depth estimation, and is particularly suitable for practical underwater robot navigation and operation tasks.

[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for underwater image depth estimation based on adaptive physical information and saliency guidance for real-time underwater robots, characterized in that, include: Obtain an underwater image dataset, and divide a portion of the data in the underwater image dataset into a training set; An underwater image depth estimation model based on adaptive physical information and saliency guidance is constructed. The underwater image depth estimation model based on adaptive physical information and saliency guidance is trained on the training set. The underwater image depth estimation model based on adaptive physical information and saliency guidance includes a physical guidance TMC representation module, an encoder-decoder module, an attenuated perceptual target attention module, and a lightweight visual transmission module. The physical guidance TMC representation module is used to generate a TMC input tensor based on the input RGB image. The encoder-decoder module is used to generate a hybrid feature set of fused data-driven depth cues and physical depth cues based on the input RGB image. The attenuation-aware target attention module is used to generate enhanced features based on the input TMC input tensor and the mixed feature set; The lightweight visual transmission module is used to model the contextual relationships of the input enhanced features and obtain the final depth estimate and the Shannon entropy of the depth distribution based on the predicted depth distribution parameters. The Shannon entropy of the depth distribution is used to represent the uncertainty of the prediction result. The underwater RGB image to be processed is acquired, and the underwater RGB image to be processed is input into the trained underwater image depth estimation model based on adaptive physical information and saliency guidance. The depth estimate value output by the model is obtained, thereby realizing the depth estimation of the underwater image. The attenuation-aware target attention module includes an attenuation-aware branch and an object-aware branch, wherein the attenuation-aware branch includes an attenuation-aware modulation channel and a spatial attention channel; In the attenuation-sensing modulation channel, the three-channel TMC input tensor is first compressed to eight channels using 1×1 convolution, and then global context information is extracted through global average pooling to obtain an 8×1×1 feature vector. Subsequently, it is processed by two consecutive 1×1 convolutional layers to generate cue confidence weights. In the spatial attention channel, decoder features It is concatenated with the compressed TMC representation and processed through a 3×3 convolutional layer with a Sigmoid activation function to generate the initial spatial attention mask M; In the object perception branch, the initial spatial attention mask M is modulated based on the cue confidence weights to achieve spatial attention enhancement under physical constraints. The modulated attention mask is then convolved with a fixed Laplacian kernel to achieve edge enhancement, and refined output features are generated by combining the decoded features. The lightweight visual transmission module will input feature maps. Divided into non-overlapping images Blocks to generate Each feature token is labeled and projected onto the embedding dimension. E Subsequently, these features are processed through a Transformer encoding layer that includes multi-head self-attention and a feedforward network, thereby enabling global contextual information exchange between different image patches. A set of attention maps generated by Transformer dot products and input feature maps The convolutional projection features are interacted with to generate a range attention map; at the same time, the output regression label of the Transformer is processed through two layers of MLP with a hidden dimension of 256 to predict the depth distribution parameters. By performing softmax normalization on the projection results of the Transformer's output regression labels, a pixel-wise ordered depth distribution is generated. : Among them, MLP ( ) indicates a shared multilayer perceptron. This represents the softmax activation function. Indicates the projection of the regression label; Final depth estimate By adjusting the center of the dynamically adjusted depth interval The weighted expectation is calculated as follows: in This represents the pixel depth projected onto the center of the i-th interval. The upper dimension of the interval center; To quantify prediction confidence and identify uncertain regions, the Shannon entropy of the depth distribution is calculated. : ； The underwater image depth estimation model is trained using the following loss function: in, For the total training loss, For scale-invariant logarithmic loss, For depth-weighted loss, For significance-weighted L1 loss, For entropy regularization loss, The weighting coefficients represent the scale-invariant logarithmic loss. The weighting coefficients represent the depth-weighted loss. The weighting coefficients represent the significance-weighted L1 loss. The weighting coefficients represent the entropy regularization loss.

2. The underwater image depth estimation method based on adaptive physical information and saliency guidance for real-time underwater robots according to claim 1, characterized in that, The physically guided TMC representation module first performs channel processing on the input RGB image, and calculates and generates the transmission map, the model-based log ratio decay feature, and the chromaticity feature respectively. The transmission graph, the model-based logarithmic ratio attenuation feature, and the chromaticity feature are then spliced ​​together to generate a three-channel TMC input tensor.

3. The underwater image depth estimation method based on adaptive physical information and saliency guidance for real-time underwater robots according to claim 1, characterized in that, The encoder-decoder module is built on MobileNetV2 and includes an encoder and a decoder. The encoder first extracts features from the input three-channel RGB image through an initial convolutional layer with 32 convolutional kernels, and then passes it through 19 residual bottleneck layers to extract hierarchical features. The decoder further generates decoding features with a resolution of 320×240 and 48 channels.

Images