An underwater target search method based on task decoupling and dynamic normalization

Abstract

The application discloses a kind of underwater target search method based on task decoupling and dynamic normalization, belong to underwater target search technical field.The method includes: constructing lightweight fidelity evaluation branch, using physical imaging model to assist training, real-time extraction low-level visual features and quantification input sample fidelity coefficient;In backbone network, set by general feature module, detection enhancement module and identification robust module the mixed expert architecture of composition, based on task instruction and fidelity signal generation routing weight and extraction path;By mapping function, generate and water quality state and task type match affine parameter, carry out adaptive correction to feature distribution;Using fidelity adjustment based on joint loss function to carry out end-to-end optimization.The application can relieve the representation conflict between detection and identification tasks, improve the search accuracy, robustness and real-time processing capability of the model under turbidity and illumination change conditions, suitable for underwater robot online perception and other application scenarios.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and underwater robot perception technology, specifically to an underwater target search method based on task decoupling and dynamic normalization. Background Technology

[0002] Underwater target search is a marine perception task that combines underwater target detection and underwater target re-identification, used to detect targets and retrieve specified objects in complex underwater scenes. With the increasing application of autonomous underwater vehicles (AUVs) and remotely operated vehicles (ROVs) in marine resource exploration, underwater engineering inspection, shipwreck search and rescue, and environmental monitoring, underwater target search places higher demands on the stability and real-time performance of algorithms. This task typically includes two sub-tasks: target detection and target re-identification, requiring the system to perform target localization and cross-viewpoint identity matching in different environments such as deep sea, nearshore, and ports. Underwater imaging is affected by medium absorption, scattering, and suspended particles, often resulting in degradation phenomena such as color cast, reduced contrast, blurred edges, and increased noise. These degradation factors alter the distribution of the input image and weaken the ability of depth models to extract stable features.

[0003] Currently, conventional underwater target search algorithms mainly face the following technical bottlenecks: Conflicts exist in multi-task feature representation: In the joint target search architecture, detection and recognition tasks have different emphases on features. Detection tasks focus on the global structure, geometric contours, and spatial location of the target, and are more sensitive to edges and local structures; recognition tasks focus more on discriminative, fine-grained semantic information, such as texture, color, and identity identifiers. Existing multi-task models often use shared backbone networks, and the training gradients of different tasks can easily influence each other within the same feature space. This problem is more pronounced under low light or turbid water conditions, making it difficult for the model to simultaneously achieve both localization accuracy and identity recognition accuracy.

[0004] Static models are ill-suited to adapting to dynamic environments: underwater environments exhibit significant spatiotemporal variations, with substantial differences in water quality across different sea areas, depths, and time periods. Traditional deep learning models typically maintain fixed network parameters and normalization layer statistics after training. While these models perform relatively stably on samples near the training distribution, they struggle to adjust feature extraction processes promptly based on degradation levels when AUVs enter regions with abrupt changes in water quality. Existing image enhancement methods are often used as standalone preprocessing modules, potentially increasing computational overhead and lacking sufficient coupling with backend detection or recognition tasks. Furthermore, artifacts in the enhancement results can interfere with subsequent feature extraction.

[0005] Normalization mechanisms suffer from domain bias: the batch normalization layers commonly used in standard deep networks rely on global statistics formed during the training phase. In variable underwater scenarios, the distribution of input data shifts with changes in water quality fidelity. Existing normalization methods struggle to simultaneously incorporate differentiated adjustments based on task type and degradation level, leading to decreased robustness of the model under conditions such as turbidity and low illumination.

[0006] Therefore, how to enable the model to achieve task decoupling, path scheduling, and dynamic parameter adjustment at the feature level while perceiving environmental quality is a problem that needs to be solved to improve target search performance in complex underwater environments.

[0007] Furthermore, underwater target search systems are typically deployed on underwater platforms with limited computing power and energy consumption. The models not only need to possess good detection and recognition accuracy, but also require control over computational load and inference latency. Therefore, simply increasing network depth or stacking image enhancement modules is insufficient to meet the engineering requirements of online search scenarios; a collaborative design approach is needed, encompassing feature splitting, parameter modulation, and loss constraints. Summary of the Invention

[0008] In view of the shortcomings of existing technologies, this invention provides an underwater target search method based on task decoupling and dynamic normalization. This method, through fidelity evaluation, conditional routing, and dynamic normalization, enables the feature extraction process to be adjusted according to the degree of water quality degradation and the type of task.

[0009] The technical solution adopted in this invention is as follows: An underwater target search method based on task decoupling and dynamic normalization is characterized by the following steps: Collect and construct a dataset of underwater images under simulated conditions with different water qualities, and generate contrast sample pairs with different fidelity using physical imaging models or enhancement techniques; By reading the degradation features of underwater images, a lightweight fidelity evaluation module is designed. The fidelity evaluation module calculates the fidelity coefficient of the input image in real time, reflecting the degree of influence of scattering and noise on the image. Design a feature decoupling backbone network, which includes a general feature module, a detection enhancement module, and a recognition robust module; Based on the fidelity coefficient and the currently executed task, detection or target recognition is performed, and the parameters of the normalization layer of the feature decoupling backbone network are dynamically adjusted. In the detection task, structural denoising is enhanced, and in the recognition task, color information is compensated. Based on the task instructions and fidelity signal input conditions, the routing weights corresponding to the general feature module, detection enhancement module and recognition robust module are calculated, and the output features of each module are weighted and fused according to the routing weights to determine the feature extraction path of the current sample. A joint loss function is constructed by weighting the target detection loss, target recognition comparison loss and fidelity regression loss. The fidelity coefficient is used to weight the recognition loss to reduce the negative impact of extremely poor water quality samples on the model's discrimination power.

[0010] Furthermore, when the fidelity evaluation module calculates the fidelity coefficient of the input image in real time, the output fidelity coefficient is used to quantify the scattering intensity, color cast, and noise level of the underwater image. The fidelity coefficient is calculated as follows: in For fidelity coefficient, This represents the low-level visual features extracted from the original image II. The weight matrix is ​​a learnable matrix. This is a bias term.

[0011] Furthermore, when dynamically adjusting the parameters of the normalization layer of the feature decoupling backbone network, the following method is used: The adaptive adjustment of the feature distribution is achieved using the following formula:

[0012] in, As input features, and The mean and variance are given. and Based on fidelity coefficient and task vector A dynamic mapping function for the independent variable, used to perform structural denoising or color compensation according to task requirements.

[0013] Furthermore, the routing weights corresponding to the general feature module, detection enhancement module, and recognition robustness module are calculated in the following manner:

[0014]

[0015] Where G is the routing mapping function, w is the weight vector, and wgen, wdet, and wrec correspond to the activation weights of the general feature module, the detection enhancement module, and the recognition robust module, respectively.

[0016] Furthermore, the joint loss function Represented as:

[0017] Where Ldet is the target detection loss, Lrec is the target recognition comparison loss, and Lfid is the fidelity evaluation regression loss; α(n) is an adjustment factor based on the fidelity coefficient, used to reduce the negative impact of extremely poor water quality samples on the recognition and discrimination ability.

[0018] In the method of this invention, the fidelity coefficient is used as a unified control quantity throughout data construction, feature normalization, routing distribution, and loss weighting. This control quantity does not directly replace the detection or recognition results, but is used to describe the degradation state of the input samples, enabling subsequent network branches to select appropriate feature processing methods based on sample quality and task type.

[0019] Compared with the prior art, the present invention has the following advantages: 1. Alleviate feature expression conflicts in multi-tasks: By using a hybrid expert architecture (MoE), the structural features required for the detection task and the identity features required for the recognition task are handled by different branches, reducing mutual interference in multi-task training.

[0020] 2. Achieve dynamic environmental compensation at the feature level: The dynamic normalization layer (DN) adjusts the mean and variance of the feature map according to the fidelity coefficient, enabling the model to complete denoising and color compensation in the feature space.

[0021] 3. Improve adaptability to complex water quality changes: Introduce fidelity assessment as a control signal to increase the participation of related denoising and structure enhancement branches when water quality degradation is severe, thereby improving the stability of underwater target search. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of an underwater target search method based on task decoupling and dynamic normalization in an embodiment of the present invention.

[0024] Figure 2 This is a diagram illustrating the overall architecture of the underwater target search model with decoupling and dynamic normalization in this embodiment of the invention. Detailed Implementation

[0025] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be described below with reference to the accompanying drawings of the embodiments. The described embodiments are only a part of the embodiments of the present invention, and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort should all fall within the scope of protection of the present invention.

[0026] It should be noted that the terms "first," "second," etc., in the specification, claims, and drawings of this invention are used only to distinguish similar objects and are not intended to limit a specific order or sequence. Where appropriate, the above terms can be used interchangeably. Furthermore, the terms "comprising," "having," and their variations are intended to indicate non-exclusive inclusion, that is, a process, method, system, product, or device that comprises a series of steps or units, and may include other related steps or units in addition to those explicitly listed.

[0027] like Figure 1 As shown, this invention provides an underwater target search method based on task decoupling and dynamic normalization. This method adjusts the feature flow direction and normalization distribution within the model using environmental perception signals to improve the feature conflict and insufficient generalization ability of AUVs in complex underwater environments caused by water quality degradation. Specifically, it includes the following steps: S1. Construct simulated and real underwater datasets containing different water qualities, and generate comparative sample pairs with different "fidelity", including the following steps.

[0028] S101. Multi-source data acquisition: Utilizing the optical camera mounted on an autonomous underwater vehicle (AUV), images containing specific underwater targets, such as fish, pipelines, and waste, are acquired at different times and in different sea areas, including clear nearshore waters, turbid ports, and deep-sea low-light areas. Manual bounding box annotation and identification ID labeling are then performed. For each original image frame, a two-stage manual annotation process is implemented. First, target location is marked using the minimum bounding rectangle to determine the target's spatial position. Second, identification ID association is performed, assigning a unique ID index to a sequence of images of the same object from different viewpoints and under different turbidity levels, providing a supervisory signal for subsequent task decoupling learning.

[0029] S102. Physical Degradation Simulation: A controlled degradation process is performed on sharp samples using an underwater optical imaging model. Here, J(x) represents the sharp image to be generated, and d(x) represents scene depth information. The absorption coefficient 'a' is adjusted to simulate the color shift of light in water as depth increases by changing the absorption weights of different wavelengths (R, G, B). The scattering coefficient 'b' is adjusted to simulate the deflection of light by suspended particles in the water. Increasing the value of 'b' reduces the global contrast of the image, creating a fog-like degradation effect. By applying different combinations of 'a' and 'b' parameters to the same sharp sample, contrasting sample pairs with continuous degradation gradients are generated, thereby expanding the dataset and establishing a controlled degradation path from sharp to murky.

[0030] S103. Fidelity Labeling: Construct a comparison pair between the degraded image and the original image. Based on the degree of image degradation, such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM), assign a continuous fidelity truth value η to each degraded image. ∈[0,1], where 1.0 represents a clear image without distortion and 0 represents an extremely murky image.

[0031] S2. Design a lightweight fidelity evaluation module to calculate the fidelity coefficient η in real time, including the following steps.

[0032] S201, Lightweight Network Construction: This module consists of several layers of depthwise separable convolutions (DSVs) and global average pooling layers. To meet the real-time processing requirements of the AUV, the input image is first downsampled to 128×128.

[0033] S202, Low-level visual feature extraction: This module extracts statistical features of the image, including color histogram distribution, local contrast gradient, and edge energy.

[0034] S203, Fidelity Prediction: The extracted feature vector is input into the fully connected layer and mapped to the final evaluation value through the Sigmoid activation function. This coefficient will serve as a control variable for subsequent steps S4 and S5, guiding the model's feature correction and route allocation.

[0035] S3. Construct a feature-decoupled backbone network based on a hybrid expert architecture (MoE), including the following steps.

[0036] S301, Submodule Pooling Decomposition: In the intermediate layers of the backbone network (such as a ResNet branch), the computational unit is decomposed into three parallel expert submodules: 1. General Feature Module (Mgen): Employing standard residual convolution, this module extracts fundamental semantic features shared across tasks, such as the overall outline of an object. As a base mapping, it maintains the stability of the target's abstract category features under different water quality conditions.

[0037] 2. Detection Enhancement Module (Mdet): Addressing the requirements of detection tasks regarding edges and geometry, this module introduces Deformable Convolution (Mdet) to adaptively capture the boundaries of blurred underwater targets by learning offsets. This module handles edge blurring caused by underwater light refraction, water flow, or turbidity, improving the accuracy of the detection box localization.

[0038] 3. Robust Recognition Module (Mrec): To address the requirement for detail consistency in recognition tasks, a channel attention mechanism (SE-Net) is introduced to extract identity features that are relatively stable to changes in illumination. This module can suppress channels heavily affected by noise and enhance discriminative local texture features, thereby providing stable feature descriptions for identity recognition tasks.

[0039] S302. Task-guided dynamic weighting and aggregation of features: This embodiment sets up a lightweight conditional router, which takes the fidelity signal η output from S2 and the current task instruction T as input to generate routing weights for the general feature module, detection enhancement module, and recognition robust module. When η is low and a detection task is performed, the participation of the detection enhancement module is increased to capture weak edges using deformable convolution; when an identity comparison task is performed, the participation of the recognition robust module is increased to extract cross-frame stable texture and semantic features.

[0040] In one alternative implementation, the input of the conditional router may further include intermediate layer global pooling features to supplement the semantic state of the current sample. The weights output by the router are normalized to satisfy non-negativity constraints and a summation of one constraint, thereby preventing a particular branch from being suppressed for a long period in the early stages of training and ensuring that all three expert modules can obtain effective gradient updates.

[0041] S4. Implement adaptive parameter modulation for dynamic domain-guided normalization, including the following steps.

[0042] S401, Task instruction vectorization: Define a task vector T=[t1,t2], where T=[1,0] indicates that the current system is in a wide-area search or detection state, and T=[0,1] indicates that the current system is in a target confirmation or identification state.

[0043] S402, Dynamic Mapping Function Learning: Using the concatenated vector of fidelity coefficients η and task vector T as input, normalized affine parameters are dynamically generated through two small mapping networks: S403, Feature Distribution Reconstruction: Real-time adjustment of intermediate layer feature x:

[0044] In low-fidelity scenarios, the γ parameter is used to suppress high-frequency noise components, and the β parameter is used to compensate for color cast caused by underwater absorption, so that the feature distribution moves closer to the stable feature domain required by the current task.

[0045] The dynamic domain-guided normalization can be set in one or more intermediate stages of the backbone network. Affine parameters can be generated per channel or per group of channels to achieve a balance between accuracy and computational cost. For detection tasks, normalization modulation prioritizes preserving edge and contour responses; for recognition tasks, normalization modulation prioritizes preserving texture and color consistency.

[0046] S5. The dynamic feature scheduling process is executed using a conditional router, which mainly includes the following steps.

[0047] S501, Route Weight Generation: The router receives the fidelity signal η and the task signal T, and obtains the activation weights of each expert module through linear mapping and Softmax normalization. Where G is a lightweight routing gating function, and the output weight vector represents the relative contributions of Mgen, Mdet, and Mrec to the feature fusion of the current sample.

[0048] To improve the stability of routing results, smoothing constraints or temperature coefficients can be added to the routing weights during the training phase to ensure continuous weight changes within adjacent fidelity intervals. This setting can reduce frequent path switching caused by slight fluctuations in water quality, enabling the model to achieve more stable outputs across consecutive video frames or image sequences.

[0049] S502, Dynamic scheduling logic: If the detected fidelity η < 0.4 and the task is detection, increase the weight of wdet so that the detection enhancement branch participates in feature extraction; if η > 0.8 and the task is recognition, increase the participation of the corresponding branch of Mrec to obtain more stable identity discrimination features.

[0050] S503, Adaptive Feature Fusion: The final output feature is a weighted sum of the features from each expert module. Through this mechanism, the model can call the corresponding branches according to the water quality status and task requirements, thereby achieving task decoupling and adaptive feature fusion.

[0051] S6. Construct a joint dynamic loss function to achieve end-to-end optimization.

[0052] S601, Multi-target loss definition: Define detection loss Ldet, which includes CIOU bounding box regression loss and classification loss, recognition contrast loss Lrec, and uses Triplet Loss to optimize identity distance and fidelity regression loss.

[0053] S602, Fidelity-Based Loss Weighting: To reduce the impact of noise labels in extremely poor water quality samples on the recognition network, a dynamic adjustment factor α(n) is introduced: S603, Minimizing Total Loss: The final joint loss function is:

[0054] Wherein, λ is a preset hyperparameter. During training, the system reduces the gradient contribution of poor-quality samples to the recognition head in order to maintain the stability and robustness of the recognition feature space.

[0055] During training, detection samples, recognition samples, and fidelity samples can come from the same batch or be input into the network through alternating sampling. Model parameters are updated uniformly through backpropagation. The fidelity evaluation module, dynamic normalization mapping network, and conditional router are jointly optimized with the backbone network, thus forming an end-to-end learning framework for underwater target search tasks.

[0056] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described 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 therein. 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. An underwater target search method based on task decoupling and dynamic normalization, characterized in that, include: Collect and construct a dataset of underwater images under simulated conditions with different water qualities, and generate contrast sample pairs with different fidelity using physical imaging models or enhancement techniques; By reading the degradation features of underwater images, a lightweight fidelity evaluation module is designed. The fidelity evaluation module calculates the fidelity coefficient of the input image in real time, reflecting the degree of influence of scattering and noise on the image. Design a feature decoupling backbone network, which includes a general feature module, a detection enhancement module, and a recognition robust module; Based on the fidelity coefficient and the currently executed task, detection or target recognition is performed, and the parameters of the normalization layer of the feature decoupling backbone network are dynamically adjusted. In the detection task, structural denoising is enhanced, and in the recognition task, color information is compensated. Based on the task instructions and fidelity signal input conditions, the routing weights corresponding to the general feature module, detection enhancement module and recognition robust module are calculated, and the output features of each module are weighted and fused according to the routing weights to determine the feature extraction path of the current sample. A joint loss function is constructed by weighting the target detection loss, target recognition comparison loss and fidelity regression loss. The fidelity coefficient is used to weight the recognition loss to reduce the negative impact of extremely poor water quality samples on the model's discrimination power.

2. The underwater target search method based on task decoupling and dynamic normalization according to claim 1, characterized in that: When the fidelity evaluation module calculates the fidelity coefficient of the input image in real time, its output fidelity coefficient is used to quantify the scattering intensity, color cast, and noise level of the underwater image. The fidelity coefficient is calculated as follows: in For fidelity coefficient, This represents the low-level visual features extracted from the original image II. The weight matrix is ​​a learnable matrix. This is a bias term.

3. The underwater target search method based on task decoupling and dynamic normalization according to claim 1, characterized in that: When dynamically adjusting the parameters of the normalization layer of the feature decoupling backbone network, the following method is used: The adaptive adjustment of the feature distribution is achieved using the following formula: in, As input features, and The mean and variance are given. and Based on fidelity coefficient and task vector A dynamic mapping function for the independent variable, used to perform structural denoising or color compensation according to task requirements.

4. The underwater target search method based on task decoupling and dynamic normalization according to claim 1, characterized in that: The routing weights corresponding to the general feature module, detection enhancement module, and recognition robustness module are calculated in the following way: Where G is the routing mapping function, w is the weight vector, and wgen, wdet, and wrec correspond to the activation weights of the general feature module, the detection enhancement module, and the recognition robust module, respectively.

5. The underwater target search method based on task decoupling and dynamic normalization according to claim 1, characterized in that: The joint loss function Represented as: Where Ldet is the target detection loss, Lrec is the target recognition comparison loss, and Lfid is the fidelity evaluation regression loss; α(n) is an adjustment factor based on the fidelity coefficient, used to reduce the negative impact of extremely poor water quality samples on the recognition and discrimination ability.

Images