Underwater rotating target detection method and system, storage medium and underwater intelligent equipment

Abstract

The application belongs to the technical field of image processing and underwater detection, and discloses a kind of underwater rotating target detection method, system and storage medium and underwater intelligent equipment, this underwater rotating target detection method includes sonar image acquisition and high-dimensional physical feature coding, scene target distribution density perception and reasoning parameter adaptive scheduling, five degrees of freedom rotating target state space initialization, decoupling scaling and multi-step progressive reverse denoising fine filtering and rotating frame fusion and real physical target output.The underwater rotating target detection method is realized based on adaptive rotating frame diffusion model, the accurate generation of rotating bounding box is realized by reconstructing diffusion state space and decoupling signal scaling mechanism, and the adaptive reasoning engine based on single-frame spatial distribution characteristics is ingeniously combined, which greatly improves the accuracy of dense and slender target counting, breaks the algorithmic bottleneck caused by fixed iteration steps, and realizes the best dynamic balance between speed and accuracy of underwater intelligent equipment in edge deployment.

Description

Technical Field

[0001] This invention relates to the fields of image processing and underwater detection technology, specifically to an underwater rotating target detection method, system, storage medium, and underwater intelligent equipment based on an adaptive rotating frame diffusion model. Background Technology

[0002] With the continuous development of underwater visual perception technology, forward-looking sonar (FLS) has become a core sensor for detection equipment such as unmanned underwater vehicles due to its ability to perform active acoustic imaging in zero-visibility and turbid waters. In practical applications such as marine life detection and intensive aquaculture, such as acoustic sampling and counting of pond eels, target detection algorithms need to extract and classify targets with high accuracy from continuous dynamic sonar images. However, underwater acoustic images are highly susceptible to reverberation and speckle noise interference, and target edges are already blurred. In addition, the special characteristics of specific biological morphologies and the drastic fluctuations in scene density pose significant challenges to existing target detection technologies.

[0003] Specifically, existing technologies face three main technical bottlenecks when dealing with the detection and counting of such complex underwater targets: First, the unique morphology of slender, rotating targets makes matching with traditional horizontal bounding boxes ineffective. In actual underwater sampling, targets such as eels often exhibit extreme aspect ratios (reaching over 10:1) and rotate arbitrarily in three-dimensional water. Existing mainstream target detection technologies, such as Faster R-CNN and the YOLO series, generally use traditional horizontal bounding boxes to represent the spatial location of targets. When predicting the horizontal bounding boxes of such slender targets, the bounding boxes often contain more than 80% useless background regions. This extremely loose physical representation not only severely interferes with the extraction of deep features, but more critically, in the non-maximum suppression stage of post-detection processing, the densely intersecting slender targets generate a severely inflated cross-union ratio (CUI). This directly leads to the algorithm mistakenly treating the detection boxes of valid targets as redundant boxes and eliminating them, resulting in a large number of missed detections and counting errors. Although some anchor-based rotating box detection methods can avoid missed detections, they require exhaustively enumerating and pre-setting a large number of anchor boxes with multiple angles and proportions, greatly increasing computational complexity.

[0004] Second, object detection frameworks based on diffusion models cannot directly adapt to rotated bounding box representations. With breakthroughs in generative models in the field of vision, a novel object detection framework called DiffusionDet has been proposed, which establishes object detection as a denoising diffusion process from noisy boxes to target boxes. During the training phase, target boxes diffuse from ground truth boxes to a random distribution, and the model learns to reverse this denoising process; during the inference phase, the model progressively refines a set of randomly generated boxes into the output results. However, the existing DiffusionDet architecture is only designed for horizontal bounding boxes, and its data samples are only composed of center coordinates (…). c x , c y ) and the width and height of the frame ( w , h It consists of four parameters. Due to the rotation angle parameter... i The physical distribution and periodicity of underwater targets are drastically different from their coordinate scale. If angle parameters are simply spliced ​​together directly into the existing Markov forward noise addition and reverse denoising processes, the original signal scaling factor will become invalid, and the signal-to-noise ratio imbalance will make it difficult for the denoising network to converge. Therefore, existing generative detection methods cannot be directly transferred to underwater slender target detection tasks that heavily rely on rotated bounding boxes.

[0005] Third, there is a contradiction between the static evaluation sampling mechanism and the computing power bottleneck of edge devices. Underwater target detection scenarios often face extreme dynamic changes: in the same sonar video stream, some frames may contain only 1-2 swimming targets, while others may contain dozens of densely overlapping targets. While DiffusionDet is attractive for its flexibility, enabling dynamic box counts and iterative evaluation, and significant performance gains can be achieved by increasing the number of evaluation boxes and iteration steps in dense and crowded scenarios, current practical applications still use a one-size-fits-all static configuration during training and testing, such as uniformly using 300 boxes and a fixed number of sampling steps for evaluation. This static inference mechanism leads to serious resource mismatch: in sparse target scenarios, high-frequency iterations are a huge waste of computing power; while in dense target scenarios, a fixed number of steps cannot guarantee sufficient recall. Especially when deploying actual products on edge computing platforms such as FPGAs or Ascend, which have strict limitations on power consumption and heat generation, such heavy networks that lack scene awareness are simply unable to achieve high frame rate real-time detection and cannot meet the standards for industrial-grade delivery.

[0006] In summary, the current field of underwater visual perception urgently needs a lightweight, high-precision detection method that can accurately fit the shape of slender rotating targets, avoid background IoU contamination, and achieve on-demand allocation of computing power on resource-constrained edge hardware. Summary of the Invention

[0007] One technical problem this application aims to solve is to overcome the shortcomings of the above-mentioned related technologies and provide an underwater rotating target detection method, system, storage medium, and underwater intelligent equipment. The underwater rotating target detection method is based on an adaptive rotating box diffusion model. It not only achieves accurate generation of rotating bounding boxes by reconstructing the diffusion state space and decoupling the signal scaling mechanism, but also cleverly combines an adaptive inference engine based on single-frame spatial distribution features. This greatly improves the counting accuracy of dense and slender targets while breaking the computing power bottleneck caused by the fixed number of iteration steps, and achieves the best dynamic balance between speed and accuracy when deploying underwater intelligent equipment at the edge.

[0008] The technical solution adopted by this underwater rotating target detection method to solve the technical problem is as follows: An underwater rotating target detection method, comprising the following steps: S1. Sonar Image Acquisition and High-Dimensional Physical Feature Encoding: Acquire the continuous dynamic video stream acquired by the forward-looking sonar, preprocess the current sonar image frame and input it into a deep convolutional neural network or Transformer backbone network to construct a physical feature map containing a multi-scale physical feature pyramid. S2. Scene Target Distribution Density Perception and Inference Parameter Adaptive Scheduling: A lightweight density estimation strategy is used to assess the density of targets within the current sonar image frame. A pre-defined resource scheduling table is consulted based on the scene density level to determine the number of candidate rotated bounding boxes to be generated for the current sonar image frame. N eval and the number of iterations for reverse denoising. S ; S3. Initialization of the five-degree-of-freedom rotating target state space: Generated in a purely random manner within the physical detection space of the current sonar image frame. N eval A set of candidate rotated bounding boxes that follow a standard Gaussian distribution is used to form an initial set of rotated bounding boxes with five degrees of freedom, including the physical center coordinates, width, height, and rotation angle. S4. Decoupled Scaling and Multi-Step Progressive Inverse Denoising Refinement: The initial rotated bounding box set and physical feature map are input into the cascaded detection decoder, and the filtering is performed according to the number of iterations. S Perform Markov inverse denoising; S5. Rotated Bounding Box Fusion and Real Physical Target Output: After iterative inverse denoising, the final set of rotated bounding boxes is processed by applying the maximum value suppression of the rotation intersection-union ratio to remove redundant rotated bounding boxes, and the physical coordinates and classification labels of the real rotated bounding boxes of the targets are output.

[0009] Preferably, the lightweight density estimation strategy in step S2 is calculated using an acoustic feature proportion model, specifically as follows: For the effective foreground mask region of the input current sonar image frame, set an absolute threshold for acoustic noise floor determination. t bg Calculate the density parameter of the acoustic activity of the entire image. r Based on density parameters r The scene density level of the current sonar image frame is divided and dynamically mapped to a predefined configuration pair for that scene density level. N eval , S ); in, , I ( x , y ) represents coordinates ( x , y Pixel acoustic echo intensity at () This is an indicator function that takes the value 1 when the condition is met, and 0 otherwise. Area Mask The total pixel area of ​​the effective acoustic region in the sector. Mask For effective masking.

[0010] Preferably, the data structure of the five-degree-of-freedom rotating target state space in step S3 is defined as follows: Regarding the first i The state vector of each target b i Represented as ; in,( c i x , c i y ) as the target i The physical center coordinates in the two-dimensional physical coordinate system of the image, ( w i , h i ) as the target i The physical width and height, For the goal i The physical deflection angle of the central vertical axis relative to the horizontal axis.

[0011] Preferably, in step S4, the cascaded detection decoder employs a decoupled signal scaling mechanism during the forward noise addition process in model training, as follows: During the forward physical diffusion process, the true target reference box is... z 0 Add Gaussian noise to generate the first t Moment-by-moment noise-adding boxes z t , ; in, For the Markov forward process in the following t The cumulative noise variance scheduling coefficient, which is monotonically decreasing, is ò, representing standard Gaussian noise. Scaling the diagonal matrix to decouple the signal , s pos , s size , s angle These are independent signal scaling factors for the physical center coordinates, aspect ratio, and rotation angle, respectively. s angle > s pos .

[0012] Preferably, in each denoising loop of the Markov inverse denoising step S4, directional features are precisely extracted by rotating the region of interest to regress the noiseless rotated bounding box of the current step, and dynamic bounding box update is performed according to the preset confidence threshold to replace the falsely detected background box with random Gaussian noise.

[0013] Preferably, the dynamic bounding box update includes the following physical logic judgment mechanism: For the s After the denoising iteration output, evaluate the generated N eval The physical classification confidence score C of each temporary bounding box is used by the filter to remove boxes that are below the local adaptive threshold. t conf bad bounding boxes, retain M Each physical feature points to a well-defined, high-quality target bounding box. M ≤ N eval ; Introduction N eval - M A newly sampled, randomly rotated bounding box from a standard Gaussian distribution is concatenated with the retained high-quality bounding boxes along the tensor dimension and then input into the first... s + 1 step of noise reduction iteration in progress.

[0014] The technical solution adopted by this underwater target detection system to solve the technical problem is as follows: an underwater target detection system for implementing the above-mentioned underwater rotating target detection method, comprising: Multi-dimensional physical feature encoding module: used to receive continuous dynamic video streams from underwater forward-looking sonar sensors in real time, and extract multi-scale physical feature maps through built-in convolution calculation units; Adaptive inference control unit: used to perform pixel-level activity calculations, dynamically issue control commands based on scene density level division results, and adjust the number of parallel computing channels and the cycle clock period of subsequent processing modules; Random Rotation Bounding Box Generator: Used in response to control commands to generate a corresponding number of high-dimensional floating-point random matrices that follow a Gaussian distribution within a specific physical memory region; Cascaded denoising decoder: Deployed with rotation region of interest alignment operator and candidate box update logic gate, it uses the weights of the decoding network to perform inverse Markov denoising process, gradually restores the physical morphological parameters of the slender underwater target and completes the localization.

[0015] Compared with related technologies, the advantages of this underwater rotating target detection method and system are: First, a multi-dimensional state space diffusion and decoupling scaling mechanism for slender rotating targets is constructed: a rotating box diffusion strategy is provided to smoothly expand the state space of the diffusion network from the traditional 4 degrees of freedom to 5 degrees of freedom including angle parameters; at the same time, a decoupling signal scaling mechanism is set for coordinate, width, height and angle parameters. While accurately enclosing the target and solving the problem of non-compact morphological representation, it overcomes the signal-to-noise ratio imbalance caused by severe angle perturbation, ensuring stable convergence in complex underwater acoustic environments. For slender underwater targets, especially eels and other objects with extremely high aspect ratios and arbitrary orientations in three-dimensional water, this approach overcomes the limitations of traditional horizontal bounding boxes, which contain a large amount of background noise, leading to serious inaccuracies in IoU matching and accidental deletion due to non-maximum suppression.

[0016] Second, an adaptive dynamic inference and computing power scheduling architecture based on scene density awareness is implemented: To achieve high frame rate processing of continuous sonar video streams on an underwater edge computing platform with extremely limited computing power and power consumption, a scene-adaptive dynamic inference engine is designed. A lightweight spatial distribution density estimation module is introduced to perceive the target density of the current sonar image frame in advance during the inference stage, and dynamically look up a table according to the scene density level to adaptively adjust the number of candidate rotated bounding boxes generated in real time. N eval The number of iterations for reverse denoising S This architecture enables on-demand allocation of edge computing power: reducing the number of steps in sparse scenarios to significantly reduce power consumption and increase speed, while increasing the number of steps and frames in dense scenarios to ensure high recall, perfectly achieving a dynamic balance between speed and accuracy. Addressing the characteristic of target density fluctuating drastically over time in real underwater scenarios, it breaks away from the traditional DiffusionDet's static configuration of a fixed number of evaluation frames and iteration steps.

[0017] Third, it addresses the challenges of high-fidelity feature extraction and weak target protection under complex underwater acoustic characteristics: Leveraging the progressive denoising properties of a generative diffusion model, it defines a depth-adaptive rotating region-of-interest alignment and dynamic bounding box update mechanism that adapts to the acoustic characteristics of sonar. By accurately extracting directional sonar image features during the iterative denoising inference process and dynamically filtering and resetting low-confidence redundant prediction boxes, the model can effectively distinguish between clean background noise and weak target edges in extremely high background noise. This mechanism addresses the unavoidable strong reverberation, large-size multiplicative speckle noise, and extremely blurred far-field target edges in forward-looking sonar images, avoiding the severe missed detections of traditional lightweight target detection algorithms on sonar images. It provides high-fidelity sensing input with an extremely low false negative rate for underwater intelligent equipment.

[0018] The technical solution adopted by this storage medium to solve the technical problem is as follows: a storage medium storing a computer-readable underwater target detection program, which, when executed by the edge-side processor of an underwater intelligent equipment, implements the above-mentioned underwater rotating target detection method.

[0019] The technical solution adopted by this underwater intelligent equipment to solve the technical problem is as follows: an underwater intelligent equipment, including a forward-looking sonar hardware system and an airborne edge computing device, wherein the airborne edge computing device deploys the aforementioned underwater target detection system; the underwater intelligent equipment directly reads data streams from the main control bus, and implements an ultra-low power sleep mode in sparse target waters or a full-load computing power operation mode in dense target waters through a dynamic peak-shaving configuration module, and outputs rotating frame detection results for underwater fine grasping or fish biomass statistics.

[0020] Preferably, the underwater intelligent equipment is an underwater intelligent detection wearable device or an unmanned underwater vehicle. Attached Figure Description

[0021] Figure 1 This is a flowchart of the underwater rotating target detection method of the present invention.

[0022] Figure 2 This is a schematic diagram of the hardware structure of the underwater intelligent equipment of the present invention. Detailed Implementation

[0023] First, those skilled in the art should understand that these embodiments are merely used to explain the technical principles of the embodiments of this application and are not intended to limit the scope of protection of the embodiments of this application. Those skilled in the art can make adjustments as needed to adapt to specific application scenarios.

[0024] This invention aims to solve the technical problems existing in current underwater target detection technology, such as the inaccuracy of traditional horizontal bounding boxes in representing slender rotating targets, the difficulty in network convergence caused by the rigid introduction of angle parameters in existing diffusion detection models, and the serious uneven distribution of computing power caused by static evaluation sampling mechanisms on underwater edge devices.

[0025] This invention aims to achieve the following objectives: First, to construct a multi-dimensional state space diffusion and decoupling scaling mechanism for slender rotating targets: For slender underwater targets, especially eels and other objects exhibiting extremely high aspect ratios and arbitrary directional distributions in three-dimensional water, this invention overcomes the limitations of traditional horizontal bounding boxes, which contain significant background noise, leading to severe IoU mismatches and accidental deletions due to non-maximum suppression. It provides a rotating box diffusion strategy that smoothly expands the state space of the diffusion network from the traditional 4 degrees of freedom to 5 degrees of freedom including angle parameters. A decoupling signal scaling mechanism for coordinates, width, height, and angle parameters accurately encloses the target, solves the problem of non-compact morphological representation, and overcomes the signal-to-noise ratio imbalance caused by drastic angle perturbations, ensuring stable convergence in complex underwater acoustic environments.

[0026] Second, an adaptive dynamic inference and computing power scheduling architecture based on scene density awareness is implemented: To achieve high frame rate processing of continuous sonar video streams on an underwater edge computing platform with extremely limited computing power and power consumption, a scene-adaptive dynamic inference engine is designed. Addressing the characteristic of drastic fluctuations in target density over time in real underwater scenes, the traditional DiffusionDet engine breaks away from its rigid, static configuration of fixed evaluation boxes and iteration steps. A lightweight spatial distribution density estimation module is introduced to proactively perceive the target density of the current frame during the inference phase and dynamically look up tables based on density levels to adaptively adjust the number of candidate rotated bounding boxes generated in real time. N eval The number of iterations for reverse denoising S This architecture enables on-demand allocation of computing power at the edge: in sparse scenarios, the number of steps is reduced to significantly reduce power consumption and increase speed, while in dense scenarios, the number of steps and frames is increased to ensure high recall, perfectly achieving a dynamic balance between speed and accuracy.

[0027] Third, this paper addresses the challenges of high-fidelity feature extraction and weak target protection under complex underwater acoustic characteristics: To address the unavoidable strong reverberation, large-size multiplicative speckle noise, and extremely blurred far-field target edges in forward-looking sonar images, a rotational region of interest alignment and dynamic bounding box update mechanism is defined, deeply adapted to the acoustic characteristics of sonar. By accurately extracting directional sonar image features during the iterative denoising inference process and dynamically filtering and resetting low-confidence redundant prediction boxes, the model can effectively distinguish between clean background noise and weak target edges in extremely high background noise. This rotational region of interest alignment and dynamic bounding box update mechanism avoids the serious missed detection phenomenon of traditional lightweight target detection algorithms on sonar images, providing high-fidelity sensing input with an extremely low false negative rate for underwater intelligent equipment.

[0028] Specifically, the present invention provides an underwater rotating target detection method, such as... Figure 1 As shown, the implementation is based on an adaptive rotating box diffusion model, and includes the following steps: Step S1, Sonar Image Acquisition and High-Dimensional Physical Feature Encoding: Acquire the continuous dynamic video stream collected by the forward-looking sonar of the underwater edge device. After performing basic preprocessing on the input current sonar image frame, input it into a deep convolutional neural network or Transformer backbone network to extract the physical feature pyramid containing multi-scale spatial information and construct a physical feature map containing the multi-scale physical feature pyramid. Step S2, Scene Target Distribution Density Perception and Inference Parameter Adaptive Scheduling: A lightweight density estimation strategy is used to assess the density of physical echo targets within the current sonar image frame. Based on the assessed scene density level, a preset resource scheduling table is dynamically searched to determine in real time the number of candidate rotated bounding boxes to be generated for the current sonar image frame. N eval and the number of iterations for reverse denoising. S ; Specifically, the scene density level (also known as the scene target distribution density level) perception is calculated using an acoustic feature proportion model, as follows: For the effective mask region of the input sonar image, set an absolute threshold for acoustic noise floor determination. t bg Calculate the density parameter of the acoustic activity of the entire image. r Based on the calculated density parameters r The scene is divided into sparse, normal, and crowded levels, and dynamically mapped to predefined configuration pairs. N eval , S This allows for the use of fewer frames and steps in sparse scenes to reduce edge chip power consumption, while more frames and steps are used in crowded scenes to improve the recall rate of weak and slender targets.

[0029] in, , I ( x , y ) represents coordinates ( x , y Pixel acoustic echo intensity at () This is an indicator function that takes the value 1 when the condition is met, and 0 otherwise. Area Mask The total pixel area of ​​the effective acoustic region in the sector. Mask For effective masking.

[0030] For any point on the image ( x , y Its effective mask Mask ( x,y ) is defined as ; R min and R max These are the physical detection blind zone radius and the maximum range radius of the sonar, respectively. i min and i max These represent the left and right limit opening angles of the sonar beam. Different sonars have different left and right limit opening angles; the opening angle of the BHQ forward-looking sonar is generally -65° to 65°. Effective mask. Mask ( x,y The region with an output of 1 is the acoustic foreground, and the decoder extracts the rotational region of interest features within the acoustic foreground region; the region with an output of 0 is the shielded region.

[0031] Step S3, Initialization of the five-DOF rotating target state space: Within the physical detection space of the current sonar image frame, generate the state space in a purely random manner. N eval A random candidate rotating bounding box that follows a standard Gaussian distribution; the four-dimensional state of the traditional horizontal box is extended to a five-degree-of-freedom state, forming an initial set of rotating bounding boxes (also known as the initial random target candidate rotating bounding box set) containing the physical center coordinates, width, height and rotation angle. The data structure for the state space of a five-degree-of-freedom rotating target is defined as follows: Regarding the first i A physical target, its state vector b i Represented as ; in,( c i x , c iy ) as the target i The physical center coordinates in the two-dimensional physical coordinate system of the image, ( w i , h i ) for this goal i The physical width and height, For the goal i The physical deflection angle of the central vertical axis relative to the horizontal axis. This high-dimensional state space ensures a high degree of fit between the predicted bounding box and the physical boundary of the slender target, eliminating the interference of more than 80% useless background on the IoU calculation.

[0032] Step S4, Decoupling Scaling and Multi-Step Progressive Inverse Denoising and Fine Filtering: The initial rotated bounding box set and the physical feature map extracted in step S1 are input into the cascaded detection decoder, and the iteration steps are determined in step S2. S Perform Markov inverse denoising process; in each denoising loop, use rotation of region of interest to precisely extract directional features, regress the noiseless rotated bounding box of the current step, and perform dynamic bounding box update according to preset confidence threshold, replacing falsely detected background boxes with random Gaussian noise. The forward noise addition process during model training of the cascaded detection decoder employs a decoupled signal scaling mechanism, and its mathematical model is as follows: During the forward physical diffusion process, the true target reference box is... z 0 Add Gaussian noise to generate the first t Moment-by-moment noise-adding boxes z t : ; in, For the Markov forward process in the following t The cumulative noise variance scheduling coefficient is monotonically decreasing, and ò is standard Gaussian noise, ò∈(0,1). Scaling the diagonal matrix to decouple the signal ;in, s pos , s size , s angle Independent signal scaling factors corresponding to physical center coordinates, aspect ratio, and rotation angle, respectively. It is a diagonal matrix. Since small angular perturbations in physical space can cause a drastic decrease in the intersection-union ratio of two targets, it is set as follows: s angle > s pos This is to ensure that the angle parameter maintains a high signal-to-noise ratio in the deep diffusion denoising network, guaranteeing stable convergence. Among these, spos , s size The preferred value is 2.0; s angle The preferred range is between 3.0 and 4.0. The stable convergence condition of the cascaded detection decoder is that the rate of change of the loss value of the loss function over 10 epochs is less than 0.5%.

[0033] The dynamic bounding box update includes the following physical logic judgment mechanism: In the s After the denoising iteration, the system evaluation output is generated. N eval The physical classification confidence score C of each temporary bounding box. A filter is built to remove boxes below a local adaptive threshold. t conf Remove the poorly bounding boxes and retain M high-quality target boxes with clearly defined physical features. M ≤ N eval To ensure that the continuous denoising process conforms to the physical diffusion law, instead of rigidly truncating discarded bad frames, a forced introduction of... N eval - M A newly sampled, randomly rotated bounding box from a standard Gaussian distribution is concatenated with the retained high-quality bounding boxes along the tensor dimension and then input into the first... s + 1 iteration in progress. This avoids the network being misled by spurious target features in the sonar noise floor.

[0034] The physical classification confidence score C is the core metric for evaluating whether a predicted bounding box contains a real target. First, feature aggregation is performed: the cascaded detection decoder uses the current rotated box coordinates to extract a local acoustic feature map from the image feature pyramid through the Rotated RoI Align operation. Then, forward mapping is performed: this local acoustic feature map is flattened and input into the fully connected classification branch at the head of the cascaded detection decoder. Activation output is then performed: the branch outputs a Logits vector of length K, where K is the number of target classes. This vector is mapped to a probability distribution between 0 and 1 using an activation function. Finally, extreme value extraction is performed: the highest probability value corresponding to the foreground class is extracted, which is the physical classification confidence score C of the predicted bounding box. Local adaptive thresholding is also performed. t conf Typically set to 0.4 to 0.5, the physical classification confidence score C is less than the local adaptive threshold. t conf The predicted bounding box is the bad bounding box.

[0035] Step S5, Rotating frame fusion and outputting realistic physical targets: After... SAfter a complete denoising iteration, the maximum suppression of the rotation intersection-union ratio is applied to the final set of rotated bounding boxes output, eliminating highly overlapping redundant rotated bounding boxes in physical space, and outputting the physical coordinates and classification labels of the denoised and reconstructed target real rotated bounding boxes.

[0036] The rotation-intersection-union (ROU) threshold is typically set to 0.6. Since organisms such as eels often overlap and intertwine, the ROU threshold should not be too low; conversely, an excessively high ROU threshold will result in multiple duplicate bounding boxes on the same target.

[0037] Specifically, the application of the maximum value of the rotation intersection-union ratio to suppress and eliminate redundant rotated bounding boxes is as follows: A1. Descending order: Sort all the final rotated bounding boxes output by the cascaded detection decoder from high to low according to the physical classification confidence score C, forming a list.

[0038] A2. Baseline Selection: The final rotated bounding box with the highest score in the pop-up list is used as the physical baseline box and stored in the target real rotated bounding box set.

[0039] A3. Intersection over Union (IoU) Calculation: Using a polygon intersection algorithm, accurately calculate the rotation IoU (SkewIoU) of the physical reference box with all other final rotated bounding boxes in the remaining list.

[0040] A4. Redundancy Removal: Final rotated bounding boxes with SkewIoU > 60% are identified as redundant rotated bounding boxes with overlapping spatial heights and are directly deleted from the list.

[0041] A5. Iteration: Repeat steps A2-A4 until the list is empty.

[0042] The output of a true physical target specifically includes: Physical coordinates: 5 floating-point numbers after denoising and reconstruction ( c x ,c y , w, h,θ ), representing the precise pixel physical center point of the real physical target in the sonar image coordinate system ( c x ,c y Minimum Enclosing Width w Minimum Enclosure Length h and the specific rotation angle relative to the horizontal axis i .

[0043] Category label: An integer ID or string representing the specific type of the real physical target, for example, Class ID: 0, which maps to the dictionary for eel.

[0044] Physical classification confidence score C: A floating-point number between 0.0 and 1.0, representing the confidence probability of the detection result.

[0045] An underwater target detection system, based on an adaptive rotating frame diffusion model, specifically includes: Multidimensional physical feature encoding module: used to receive continuous dynamic video streams (also known as image streams) from underwater photoelectric or sonar sensors in real time, and construct a physical feature map (also known as a multi-scale acoustic feature base map) containing a multi-scale physical feature pyramid through the built-in convolutional computing unit.

[0046] Adaptive Inference Control Unit: Used to perform pixel-level activity calculations, dynamically issue control commands based on scene density level classification results, and adjust the number of parallel computing channels and the cycle clock period of subsequent processing modules.

[0047] Random Rotation Bounding Box Generator: Used in response to control commands to generate a corresponding number of high-dimensional floating-point random matrices that follow a Gaussian distribution within a specific physical memory region.

[0048] Cascaded Denoising Decoding Accelerator: Deployed with rotation region of interest alignment operator and candidate box update (BoxRenewal) logic gate, it uses the weights of the decoding network to perform inverse Markov denoising process, gradually restores the physical morphological parameters of the slender underwater target and completes the localization.

[0049] A storage medium storing a computer-readable underwater target detection program, which, when executed by an edge-side processor of an underwater intelligent device, implements the aforementioned underwater rotating target detection method.

[0050] An underwater intelligent device, such as Figure 2 As shown, the underwater intelligent equipment includes an acoustic detection system and an edge computing device. The edge computing device deploys the underwater target detection system. This underwater intelligent equipment directly reads data streams from the main control bus and implements an ultra-low power sleep mode in sparsely populated target waters or a full-load computing power operation mode in densely populated target waters through a dynamic peak-shaving configuration module. It outputs rotating frame detection results for precise underwater capture or fish biomass statistics. When the underwater intelligent equipment is an underwater intelligent detection wearable device, the communication interface is used to connect to AR smart swimming goggles; or when the underwater intelligent equipment is an unmanned underwater vehicle, the communication interface is used to connect to a display screen; the communication interface can also be used for data export.

[0051] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for detecting underwater rotating targets, characterized in that, The steps include the following: S1. Sonar Image Acquisition and High-Dimensional Physical Feature Encoding: Acquire the continuous dynamic video stream acquired by the forward-looking sonar, preprocess the current sonar image frame and input it into a deep convolutional neural network or Transformer backbone network to construct a physical feature map containing a multi-scale physical feature pyramid. S2. Scene Target Distribution Density Perception and Inference Parameter Adaptive Scheduling: A lightweight density estimation strategy is used to assess the density of targets within the current sonar image frame. A pre-defined resource scheduling table is consulted based on the scene density level to determine the number of candidate rotated bounding boxes to be generated for the current sonar image frame. N eval and the number of iterations for reverse denoising. S ; S3. Initialization of the five-degree-of-freedom rotating target state space: Generated in a purely random manner within the physical detection space of the current sonar image frame. N eval A set of candidate rotated bounding boxes that follow a standard Gaussian distribution is used to form an initial set of rotated bounding boxes with five degrees of freedom, including the physical center coordinates, width, height, and rotation angle. S4. Decoupled Scaling and Multi-Step Progressive Inverse Denoising Refinement: The initial rotated bounding box set and physical feature map are input into the cascaded detection decoder, and the filtering is performed according to the number of iterations. S Perform Markov inverse denoising; S5. Rotated Bounding Box Fusion and Real Physical Target Output: After iterative inverse denoising, the final set of rotated bounding boxes is processed by applying the maximum value suppression of the rotation intersection-union ratio to remove redundant rotated bounding boxes, and the physical coordinates and classification labels of the real rotated bounding boxes of the targets are output.

2. The underwater rotating target detection method according to claim 1, characterized in that, The lightweight density estimation strategy in step S2 is calculated using an acoustic feature proportion model, specifically as follows: For the effective foreground mask region of the input current sonar image frame, set an absolute threshold for acoustic noise floor determination. τ bg Calculate the density parameter of the acoustic activity of the entire image. ρ Based on density parameters ρ The scene density level of the current sonar image frame is divided and dynamically mapped to a predefined configuration pair for that scene density level. N eval , S ); in, , I ( x , y ) represents coordinates ( x , y Pixel acoustic echo intensity at () This is an indicator function that takes the value 1 when the condition is met, and 0 otherwise. Area Mask The total pixel area of ​​the effective acoustic region in the sector. Mask For effective masking.

3. The underwater rotating target detection method according to claim 1, characterized in that, The data structure of the five-degree-of-freedom rotating target state space in step S3 is defined as follows: Regarding the first i The state vector of each target b i Represented as ; in,( c i x , c i y ) as the target i The physical center coordinates in the two-dimensional physical coordinate system of the image, ( w i , h i ) as the target i The physical width and height, For the goal i The physical deflection angle of the central vertical axis relative to the horizontal axis.

4. The underwater rotating target detection method according to claim 1, characterized in that, In step S4, the cascaded detection decoder employs a decoupled signal scaling mechanism during the forward noise addition process in model training, as detailed below: During the forward physical diffusion process, the true target reference box is... z 0 Add Gaussian noise to generate the first t Moment-by-moment noise-adding boxes z t , ; in, For the Markov forward process in the following t The cumulative noise variance scheduling coefficient, which is monotonically decreasing, is ò, representing standard Gaussian noise. Scaling the diagonal matrix to decouple the signal , s pos , s size , s angle These are independent signal scaling factors for the physical center coordinates, aspect ratio, and rotation angle, respectively. s angle > s pos .

5. The underwater rotating target detection method according to claim 4, characterized in that, In each denoising loop of the Markov inverse denoising step S4, directional features are precisely extracted by rotating the region of interest to regress the noiseless rotated bounding box of the current step, and dynamic bounding box update is performed according to the preset confidence threshold to replace the falsely detected background box with random Gaussian noise.

6. The underwater rotating target detection method according to claim 5, characterized in that, The dynamic bounding box update includes the following physical logic judgment mechanism: For the s After the denoising iteration output, evaluate the generated N eval The physical classification confidence score C of each temporary bounding box is used by the filter to remove boxes that are below the local adaptive threshold. τ conf bad bounding boxes, retain M Each physical feature points to a well-defined, high-quality target bounding box. M ≤ N eval ; Introduction N eval - M A newly sampled, randomly rotated bounding box from a standard Gaussian distribution is concatenated with the retained high-quality bounding boxes along the tensor dimension and then input into the first... s + 1 step of noise reduction iteration in progress.

7. An underwater target detection system for implementing the underwater rotating target detection method as described in any one of claims 1 to 6, characterized in that, include: Multidimensional physical feature encoding module: used to receive continuous dynamic video streams from underwater forward-looking sonar sensors in real time, and extract physical feature maps of multi-scale physical feature pyramids through built-in convolution calculation units; Adaptive inference control unit: used to perform pixel-level activity calculations, dynamically issue control commands based on scene density level division results, and adjust the number of parallel computing channels and the cycle clock period of subsequent processing modules; Random Rotation Bounding Box Generator: Used in response to control commands to generate a corresponding number of high-dimensional floating-point random matrices that follow a Gaussian distribution within a specific physical memory region; Cascaded Denoising Decoding Accelerator: Deployed with rotation region of interest alignment operator and candidate box update logic gate, it uses the weights of the decoding network to perform inverse Markov denoising process, gradually restores the physical morphological parameters of the slender underwater target and completes the localization.

8. A storage medium storing a computer-readable underwater target detection program thereon, characterized in that: When the underwater target detection program is executed by the edge-side processor of the underwater intelligent equipment, it implements the underwater rotating target detection method as described in any one of claims 1 to 6.

9. An underwater intelligent equipment, characterized in that: It includes a forward-looking sonar hardware system and an airborne edge computing device, wherein the airborne edge computing device is equipped with the underwater target detection system as described in claim 7; the underwater intelligent equipment directly reads the data stream from the main control bus, and implements an ultra-low power sleep mode in sparse target waters or a full-load computing power operation mode in dense target waters through a dynamic peak-shaving configuration module, and outputs rotating frame detection results for underwater fine grasping or fish biomass statistics.

10. The underwater intelligent equipment according to claim 9, characterized in that, The underwater intelligent equipment refers to underwater intelligent detection wearable equipment or unmanned underwater vehicles.

Images