Submarine organism detection method, system and equipment based on deep learning and medium

Abstract

The invention discloses a deep learning-based benthic organism detection method, system and device and a medium, and belongs to the technical field of computer vision, and the method comprises the steps: obtaining a benthic organism data set picture, and generating a training data set according to the benthic organism data set picture; on the basis of the training data set, a coordinate attention mechanism and loss function calculation are combined, and a frame of a benthic organism detection model is built; the loss function is used for calculating the positioning loss of the actual target frame and the predicted target frame; training the frame of the benthic organism detection model to obtain the benthic organism detection model; and determining a benthic organism detection result based on the benthic organism detection model. According to the method, the deep learning algorithm is utilized, the morphological characteristics of the benthos and the environmental background information are fused, accurate detection and dynamic tracking of the seabed target in the complex underwater environment can be achieved, the accuracy and efficiency of benthos detection are improved, and powerful support is provided for protection and management of marine resources.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a method, system, device, and medium for detecting submarine organisms based on deep learning. Background Art

[0002] With the development of deep learning technology, deep learning-based detection of seabed organisms has made significant progress. Currently, researchers widely use convolutional neural networks, region proposal networks, and more advanced target detection frameworks to achieve automatic recognition and classification of organisms in seabed images. These models can learn the morphological characteristics, texture information, and background relationships of seabed organisms from a large number of annotated images, and still have strong recognition capabilities in complex environments. At the same time, with the improvement of seabed image acquisition equipment and underwater remote sensing technology, the acquired image data is clearer and more informative, providing better training samples for deep learning models. Through continuous training and optimization of the model, the current detection method has demonstrated strong stability and versatility in multiple practical application scenarios, providing important technical support for marine ecological monitoring and biodiversity research.

[0003] However, traditional methods mainly use manually designed features and classifiers for seabed organism detection. Although they have high computational efficiency, their disadvantages are low detection accuracy and difficulty in adapting to different seabed environments. Deep learning methods use deep neural networks to learn biological features in seabed images. Their advantages are high detection accuracy and the ability to adapt to different seabed environments. They have become the mainstream method of vision-based seabed organism detection. However, in actual seabed organism detection scenarios, due to the complex seabed environment, existing methods often cannot achieve ideal results, especially for the detection of small seabed organisms, there are still challenges. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is: although seabed organism detection is crucial for the investigation and protection of marine resources, traditional methods face problems such as low recognition accuracy and slow processing speed. By using deep learning algorithms and combining the characteristics of seabed organisms and environmental background, accurate detection and tracking of seabed organisms can be achieved, the accuracy and efficiency of seabed organism detection can be improved, and strong support can be provided for the protection and management of marine resources.

[0006] To solve the above technical problems, the present invention provides the following technical solution: a method for detecting submarine organisms based on deep learning, which includes the following steps:

[0007] Acquire a submarine organism dataset image and generate a training dataset based on the submarine organism dataset image; build a submarine organism detection model framework based on the training dataset and in combination with a coordinate attention mechanism and loss function calculation; the loss function is used to calculate the positioning loss between the actual target frame and the predicted target frame; train the submarine organism detection model framework to obtain a submarine organism detection model; and determine the submarine organism detection results based on the submarine organism detection model.

[0008] As a preferred embodiment of the deep learning-based seabed organism detection method described in the present invention, a training dataset is generated based on the seabed organism dataset images, comprising the following steps: classifying objects in the seabed organism dataset images based on target seabed organisms and providing bounding boxes; and formatting the class-labeled seabed organism dataset images to obtain a training dataset.

[0009] As a preferred solution of the deep learning-based seabed biological detection method described in the present invention, the framework for building a seabed biological detection model includes: using the CSPNet network as the backbone network to extract features from the training data set, temporarily storing the obtained feature map and cross-connecting it in the subsequent different stages of the CSPNet network; using the FPN network to construct a feature pyramid, fusing the feature maps of different levels of the CSPNet network and the FPN network to obtain a fused feature map; performing a decomposition attention operation on the fused feature map based on the coordinate attention mechanism, aggregating the global information along the height and width coordinate axes; using the PAN network to perform bidirectional path aggregation on the feature map processed by the coordinate attention mechanism; using the Wise-IoU loss function to calculate the positioning loss of the actual target box and the predicted target box during the training stage. The beneficial effect of this preferred technical solution is to realize multi-level fusion of features and key area enhancement by combining the CSPNet network and the FPN feature pyramid with the coordinate attention mechanism. At the same time, the Wise-IoU loss function optimizes the target box positioning accuracy, thereby improving the accuracy and robustness of seabed biological detection.

[0010] As a preferred solution of the deep learning-based seabed organism detection method described in the present invention, wherein: the framework of the seabed organism detection model is trained to obtain the seabed organism detection model, including: setting the initial learning rate based on the step size of the parameter update of the seabed organism detection model; setting the final value of the learning rate decay according to the initial learning rate; setting the momentum in combination with the parameter update and historical gradient information; setting the number of learning rate preheating rounds based on the training preheating process. The beneficial effect of this preferred technical solution is that by setting a reasonable learning rate strategy and momentum parameters, combined with parameter updates and historical gradient information, the stable convergence of the model training process and the improvement of optimization efficiency are achieved, the overfitting problem is effectively avoided, and the generalization ability of the seabed organism detection model is enhanced.

[0011] As a preferred solution of the deep learning-based seabed organism detection method described in the present invention, the seabed organism detection results are determined based on the seabed organism detection model, including the following steps: deploying the seabed organism detection model into an intelligent device; performing seabed organism detection based on the seabed organism detection model, and using a perception network to dynamically adjust the detection strategy to obtain the category and confidence information of the seabed organisms.

[0012] As a preferred solution of the deep learning-based seabed organism detection method described in the present invention, the dynamic adjustment of the detection strategy using a perception network includes the following steps: based on the fused feature map, a dynamic context perception network is used to analyze the fused feature map to obtain the environmental context around the seabed organism; and the detection strategy is dynamically adjusted according to the environmental context around the seabed organism.

[0013] As a preferred embodiment of the deep learning-based seabed organism detection method described herein, the method includes dynamically adjusting the detection strategy based on the environmental context surrounding the seabed organism, including analyzing the environmental context surrounding the seabed organism to determine the seabed area type in which the seabed organism resides, and adjusting the weights of the perception network based on the seabed area type in which the seabed organism resides. This preferred technical solution has the beneficial effect of analyzing the environmental context surrounding the seabed organism through the perception network and adjusting the detection strategy and network weights, enabling the model to adaptively optimize based on different seabed area types, thereby improving detection accuracy and adaptability in underwater environments.

[0014] Another object of the present invention is to provide a seabed biological detection system based on deep learning.

[0015] To solve the above technical problems, the present invention provides the following technical solutions: a deep learning-based seabed organism detection system, comprising: a data acquisition module for acquiring seabed organism data set images and generating a training data set based on the seabed organism data set images; a framework building module for building a framework of a seabed organism detection model based on the training data set and in combination with a coordinate attention mechanism and loss function calculation; the loss function is used to calculate the positioning loss between the actual target frame and the predicted target frame; a model training module for training the framework of the seabed organism detection model to obtain a seabed organism detection model; and a biological detection module for determining the seabed organism detection results based on the seabed organism detection model.

[0016] The present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the processor executes the computer program, the steps of the deep learning-based seabed organism detection method are implemented.

[0017] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, the steps of the deep learning-based seabed organism detection method are implemented.

[0018] The beneficial effects of the present invention are as follows: by utilizing a deep learning algorithm and integrating the morphological characteristics of seabed organisms with environmental background information, the present invention can realize accurate detection and dynamic tracking of seabed targets in complex underwater environments, thereby improving the accuracy and efficiency of seabed organism detection and providing strong support for the protection and management of marine resources; the coordinate attention mechanism is utilized to enhance the model's perception of small targets and key areas, while the Wise-IoU loss function is introduced to optimize detection accuracy and enhance the robustness of the model under conditions of large target scale differences and imbalanced categories; in the model training stage, a combination of multiple feature fusion strategies and dynamic parameter optimization methods is used to improve the model's convergence efficiency and generalization ability; in the deployment stage, a real-time detection architecture on the intelligent device side is adopted to significantly enhance the practicality and response speed of the detection method; it not only effectively improves the accuracy and automation level of seabed organism detection, but also has good adaptability and scalability, providing efficient and reliable technical support for marine resource surveys, underwater robot operations and intelligent monitoring of the marine ecological environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive effort. Among them:

[0020] Figure 1This is a flow chart of a detection model for a method for detecting submarine organisms according to an embodiment of the present invention.

[0021] Figure 2 This is a coordinate attention structure diagram described in an embodiment of the present invention.

[0022] Figure 3 The figure is a schematic structural diagram of a computer device according to an embodiment of the present invention.

[0023] Figure 4 A schematic diagram of submarine organism detection according to a method for submarine organism detection according to an embodiment of the present invention. DETAILED DESCRIPTION

[0024] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0025] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0026] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0027] Example 1, with reference to Figure 1 , which is the first embodiment of the present invention, provides a method for detecting submarine organisms based on deep learning, including:

[0028] S100: Acquire a submarine organism dataset image, and generate a training dataset based on the submarine organism dataset image.

[0029] S200: Build a framework for a submarine biological detection model based on a training dataset and incorporating a coordinate attention mechanism and loss function calculation. The loss function is used to calculate the positioning loss between the actual target frame and the predicted target frame.

[0030] S300: Training the framework of the submarine organism detection model to obtain the submarine organism detection model.

[0031] S400: Determine a submarine organism detection result based on a submarine organism detection model.

[0032] It should be noted that the seabed environment has characteristics such as weak lighting, low visibility, diverse biological forms, and complex background. Traditional image processing methods based on artificial features or shallow models are difficult to cope with detection difficulties such as blurred target object outlines, variable scales, and dense distribution of small targets in real underwater scenes; in addition, due to the high similarity between seabed organisms and the surrounding environment, false detection and missed detection problems occur frequently, especially in the identification of small targets such as sea urchins, sea cucumbers, and scallops. It is more challenging; seabed operations often use underwater robots for image acquisition and analysis, and the performance of the detection algorithm directly affects the operating efficiency of intelligent equipment and the reliability of resource detection.

[0033] Therefore, in response to the above-mentioned problems of insufficient target recognition accuracy, poor robustness and poor environmental adaptability, a deep learning-based seabed organism detection method was constructed through steps S100-S400, wherein the authenticity and validity of the data source and the scientific nature of the annotation are guaranteed by obtaining the original seabed organism images and generating a training data set; by combining the coordinate attention mechanism and the improved loss function, a more spatially sensitive detection model structure is constructed, thereby improving the model's ability to pay attention to target positions and boundaries, especially for more accurate recognition of small targets in complex backgrounds; by training the model structure and optimizing hyperparameters, it is ensured that the model has good generalization ability and convergence efficiency among multiple types of targets; by applying the trained model to actual images, intelligent detection and recognition of seabed targets are realized, and detection results with categories and confidence are output.

[0034] Example 2, reference Figure 1 and Figure 2 , which is the second embodiment of the present invention, provides a deep learning-based seabed organism detection method based on the above embodiments.

[0035] In the embodiment of the present invention, obtaining the submarine organism dataset images in step S100 refers to obtaining the submarine organism dataset images by photographing the submarine with a camera.

[0036] In an optional embodiment, the seabed biological dataset images obtained in step S100 can also be collected by autonomous cruising using an underwater robot equipped with a multispectral imaging device, thereby being able to capture image information at different wavelengths, effectively overcoming the limitations of the weak light environment on the seabed, and fusing the collected image information through computer vision algorithms to enhance the distinguishability of target features.

[0037] In another optional embodiment, the seabed biological dataset image obtained in step S100 can also be acquired through acoustic imaging technology combined with optical images for data collection, and the acoustic image generated by the sonar system is aligned and fused with the image taken by a traditional optical camera to form a composite image dataset containing depth information and morphological features. It can have stronger penetration ability in turbid water environments and obtain more comprehensive seabed biological morphological information.

[0038] In an embodiment of the present invention, generating a training dataset based on a dataset of submarine organisms in step S100 includes the following steps:

[0039] Use labeling tools such as LabelImg to label the objects in the seabed organism dataset images based on the target seabed organisms and give bounding boxes.

[0040] Save the seabed organism dataset images after category annotation, and convert the seabed organism dataset images after category annotation into a format acceptable to the seabed organism detection model to obtain a training dataset. The seabed organism dataset images after category annotation are usually saved in XML format, and the format acceptable to the seabed organism detection model is COCO format.

[0041] In an optional embodiment, in step S100, a training data set is generated based on the images of the seabed biological data set. A method combining data enhancement and semi-supervised learning can also be used. By performing transformation operations such as rotation, scaling, horizontal flipping, brightness and contrast adjustment, and Gaussian blur on the original image, imaging conditions in different environments are simulated. At the same time, transformations unique to the underwater environment, such as water body fluctuation simulation, light attenuation simulation, and suspended particle addition, are introduced to generate enhanced samples that are many times the amount of original data. For unlabeled seabed images, a consistency regularized semi-supervised learning strategy is applied, and pseudo-labels with high confidence are used to expand the training set to solve the problem of scarce labeled data of seabed organisms.

[0042] In another optional embodiment, in step S100, a training data set is generated based on the images of the seabed biological data set. A cross-domain data generation framework can also be constructed based on the generative adversarial network technology, and style transfer is used to convert the biological images of land or shallow water areas into deep-sea environment image features. At the same time, a method combining instance segmentation and interactive annotation is adopted. The most informative samples are selected through an active learning strategy for manual fine annotation, and a semi-automatic annotation tool combining point-level supervision and area-level supervision is combined to reduce the annotation cost and improve the accuracy of the bounding box and the consistency of the category label. It is particularly suitable for processing seabed organisms with complex morphology and fuzzy boundaries such as sea urchins and sea cucumbers.

[0043] In an embodiment of the present invention, the target seabed organisms may be small seabed organisms such as sea cucumbers, sea urchins or scallops.

[0044] In an embodiment of the present invention, the training data set includes a training set, a validation set and a test set, wherein 70% of the data is used as a training set, 20% as a validation set and 10% as a test set to ensure that the model has good generalization ability.

[0045] It should be noted that in actual seabed environments, due to weak lighting, complex background and high turbidity, traditional public data sets often cannot cover all the changing forms of target organisms in natural scenes, resulting in low model generalization performance and poor adaptability; the present invention utilizes the process of manual fine labeling and format conversion to ensure that the training data not only has a high degree of scene closeness, but also has clear semantic consistency in the labeling, which is better than the traditional method of relying on general image sets or only sampling part of the scene. Especially for small seabed organisms with complex shapes and severe occlusion such as sea cucumbers, sea urchins, and scallops, the data structure after format conversion is more adaptable to the input requirements of deep learning models, thereby improving the stability and recognition efficiency of subsequent model training; secondly, in traditional methods, small targets are easily ignored in the multi-scale feature extraction stage. The present invention effectively alleviates the training deviation caused by missing and mislabeled small target labels by controlling data quality from the source.

[0046] In the embodiment of the present invention, the framework of the submarine biological detection model is constructed in step S200, including:

[0047] The CSPNet network is used as the backbone network to extract features from the training dataset, and the obtained feature maps are temporarily stored and cross-connected in the subsequent different stages of the CSPNet network.

[0048] The FPN network is used to construct a feature pyramid, and the feature maps of different levels of the CSPNet network and the FPN network are fused to obtain the fused feature map, so that the FPN network can obtain high-level semantic information and low-level detail information at the same time.

[0049] Based on the coordinate attention mechanism, the fused feature map is decomposed and the global information is aggregated along the height and width coordinate axes.

[0050] For the feature maps processed by the coordinate attention mechanism, the PAN network is used to perform bidirectional path aggregation and optimize multi-scale information fusion.

[0051] During the training phase, the Wise-IoU loss function is used to calculate the positioning loss between the actual target box and the predicted target box.

[0052] In an embodiment of the present invention, for the feature graph processed by the coordinate attention mechanism, a PAN network is used to perform bidirectional path aggregation, including:

[0053] A top-down refinement network is used to perform downsampling and upsampling operations, starting from high-resolution semantic feature maps, gradually reducing the resolution, and fusing these feature maps with the corresponding low-resolution feature maps, so that the model can simultaneously obtain rich semantic information and detailed information.

[0054] A bottom-up aggregation network is used to aggregate feature maps at different levels, aggregating multi-scale information from coarse to fine to improve the model's ability to detect and segment targets.

[0055] Furthermore, the Wise-IoU loss function (Wise Intersection over Union) optimizes the traditional IoU (Intersection over Union) measurement effect on bounding box regression by introducing a weight mechanism. The core idea is to adjust the weight according to the overlap quality between the detection box and the ground-truth box, focusing on difficult samples that are difficult to accurately predict, thereby improving the positioning accuracy of the model in complex scenarios. The specific formula of the Wise-IoU loss function is as follows:

[0056]

[0057] L IoU = 1-IoU;

[0058] L WIoU =F WIoU *L IoU ;

[0059]

[0060] Among them, B gt and B prd They are the real bounding box of the detected target and the predicted bounding box respectively; IoU is the real bounding box of the detected target B gt And the detection target prediction bounding box B prd The ratio of the intersection and union of IoU is the intersection-over-union loss function; L WIoU is the Wise-IoU loss function; F WIoU is the calculation formula of Wise-IoU loss function, F WIoU ∈[1,e); x, y are the coordinates of the center point of the detection target frame; x gt 、y gt is the center point coordinate of the prediction box; H g and W g are the height and width of the minimum bounding rectangle of the predicted bounding box and the true bounding box respectively; the intersection-over-union loss function L IoU ∈[0,1], when the anchor box coincides with the true value box, the calculation formula of the Wise-IoU loss function of the high-quality anchor box will be significantly reduced. WIoUThe value of (the more the anchor box and the true value box overlap, the more the Wise-IoU loss function is calculated WIoU The smaller the value of ) and its focus on the distance between center points.

[0061] Furthermore, drawing on the idea of focalloss, a monotonic focal coefficient is added to the Wise-IoU loss function In order to prevent the network from converging slowly in the later stage, a normalization factor is used to speed up the convergence of the network in the later stage, and the monotonic focusing coefficient The specific formula is as follows:

[0062]

[0063] in, is a normalization term used to balance targets of different scales; the monotonic focusing coefficient With Wise-IoU loss function L WIoU Multiply them together to get the final Wise-IoU loss value.

[0064] It should be noted that by introducing the monotonic focusing coefficient Small target (H g 、W g The gradient contribution of small objects is suppressed to avoid overfitting of the model due to noisy localization of small objects; when jointly training classification and regression tasks, the fluctuation of regression loss caused by target scale differences is reduced.

[0065] It should be noted that the use of the CSPNet network as the backbone network for feature extraction aims to address the issues of information transfer and feature reuse in deep neural networks. The CSPNet network improves the efficiency of information transfer and feature reuse by dividing the input feature map into two parts and cross-connecting these parts at different stages of the network. This design effectively reduces the number of parameters and computational complexity in the network while improving model performance. Secondly, the FPN network combines high-resolution, semantically rich feature maps with low-resolution, semantically weaker feature maps that contain more detailed information in a top-down manner to form a multi-scale feature pyramid. This design enables the network to effectively detect and locate targets at different scales, thereby improving detection performance and robustness. In addition, Coordinate Attention is an improved attention mechanism used to enhance the feature representation capability of Convolutional Neural Networks (CNNs). Traditional attention mechanisms usually capture global spatial information, while the Coordinate Attention Mechanism aggregates global information along the height and width coordinate axes by decomposing the attention operation. The Coordinate Attention Mechanism not only retains spatial information but also enhances the model's attention to important areas, especially in small object detection and fine-grained image classification tasks. In this way, the Coordinate Attention Mechanism can better capture long-range dependencies in features, thereby improving the overall performance of the model.

[0066] Further, such as Figure 2 The figure shows the coordinate attention structure diagram, in which the coordinate attention mechanism effectively enhances the model's ability to represent the target by integrating spatial coordinate information into channel attention; the input feature map is globally average pooled in the height and width directions respectively to obtain aggregated features in the two directions, which contain global information in the corresponding directions; the features in the height and width directions are nonlinearly transformed through a shared MLP (multi-layer perceptron), so that the network can simultaneously capture the relationship between the two directions; the transformed features are weighted and expanded in the height and width directions respectively to generate two attention weight maps; then, the two attention weight maps are multiplied to obtain the final attention weight map, which is used to weight the original feature map; finally, the attention weight map is element-by-element multiplied with the original feature map to realize the redistribution of the importance of different channels, thereby enhancing the network's attention to key features.

[0067] It should be noted that, unlike FPN networks, PAN networks use a bidirectional top-down and bottom-up path to aggregate feature information at different levels, effectively fusing multi-scale information. Furthermore, the Wise-IoU loss function improves upon the traditional IoU (Intersection over Union) metric by weighting the IoU values of different categories to more accurately reflect the model's performance in multi-class segmentation tasks. While traditional IoU metrics can perform poorly in cases of class imbalance, the Wise-IoU loss function, by considering the importance of each category, aims to provide a more fair and reasonable evaluation. This is particularly true in scenarios with uneven class distribution, enabling the Wise-IoU loss function to better guide model optimization and evaluation.

[0068] In the embodiment of the present invention, step S300 trains the framework of the submarine organism detection model to obtain the submarine organism detection model, including:

[0069] The initial learning rate is set based on the step size of the parameter update of the benthic organism detection model.

[0070] Sets the final value of learning rate decay based on the initial learning rate.

[0071] Momentum is set by combining parameter updates and historical gradient information.

[0072] Set the number of learning rate warmup epochs based on the warmup process of training.

[0073] For example, the framework of the seabed organism detection model is trained as follows:

[0074] At the beginning of the training of the framework of the submarine biological detection model, the initial learning rate is set to 0.01, which determines the step size of the model parameter update. The initial learning rate setting cannot be too large to avoid large oscillations or divergence during the training process.

[0075] The final value of the learning rate decay is set to 0.01, which is a multiple of the initial learning rate. This is a learning rate scheduling strategy that can adaptively adjust the learning rate during training and improve the generalization ability of the model. The setting of the final value of the learning rate decay affects the decay speed of the learning rate and is usually multiplied by the initial learning rate.

[0076] The momentum is set to 0.937, which indicates the proportion of historical gradient information considered by the seabed organism detection model when updating parameters. The larger the momentum, the more attention the seabed organism detection model pays to historical gradient information when updating parameters, which helps stabilize the training process.

[0077] The number of learning rate warmup rounds is set to 3.0. At the beginning of training, the learning rate is gradually increased to the initial learning rate. The warmup process helps stabilize training and prevent gradient explosion or gradient vanishing.

[0078] In an optional embodiment, in step S300, the framework of the submarine biological detection model is trained to obtain the submarine biological detection model. A method combining a cyclic learning rate strategy with mixed precision training can also be used. By periodically adjusting the learning rate during the training process, the model can jump out of the local optimal solution. At the same time, FP16 and FP32 mixed precision calculations are used to accelerate the training process and reduce memory usage. In addition, knowledge distillation technology can be introduced to first train a teacher network with a large number of parameters, and then migrate the feature representation learned to a student network with a smaller number of parameters. This ensures the accuracy of the model and improves the inference speed. It is particularly suitable for deployment on underwater robots with limited computing resources.

[0079] In another optional embodiment, in step S300, the framework of the seabed organism detection model is trained to obtain the seabed organism detection model. A method combining adversarial training and data enhancement strategy can also be used to simulate complex conditions such as illumination changes and turbidity changes in the underwater environment by adding disturbances to the image during the training process. At the same time, online case mining technology is introduced to automatically identify and focus on training samples that are easily misjudged by the model, thereby improving the model's detection ability for small seabed organisms that are difficult to distinguish (such as different types of sea cucumbers with similar morphology). In addition, transfer learning can be combined to use weights pre-trained on a large-scale general data set as initialization parameters, and then fine-tune them on the seabed organism data set to effectively solve the problem of insufficient labeled data for seabed organisms.

[0080] It should be noted that in the training process of traditional deep models, when faced with high-noise environments and data with significant differences in target scale, it is very easy to produce problems such as training oscillation, gradient disappearance or slow convergence. In particular, for small targets, the loss value is often dominated by large targets, resulting in insufficient model learning of small targets, affecting the final detection performance; the present invention reasonably sets the learning rate decay mechanism, momentum factor and warm-up strategy, which not only ensures the stability of the training process, but also improves the adaptability of the model to complex input data, thereby effectively avoiding the loss of accuracy caused by parameter jumps in the early stage of training; more importantly, combined with the improved Wise-IoU loss function, the monotonic focus idea is introduced, which can adjust the contribution ratio of targets of each scale to the loss during training, thereby enhancing the model's attention to small targets and dense areas. This fine-grained gradient control mechanism provides a more balanced and efficient optimization path for deep learning models in complex seabed scenes.

[0081] In an embodiment of the present invention, determining the submarine organism detection result based on the submarine organism detection model in step S400 includes the following steps:

[0082] Deploy seafloor biodetection models to smart devices.

[0083] Seabed organism detection is performed based on the seabed organism detection model, and the perception network is used to dynamically adjust the detection strategy to obtain the category and confidence information of the seabed organisms.

[0084] In the embodiment of the present invention, step S400 adopts the perception network to dynamically adjust the detection strategy, including the following steps:

[0085] According to the fused feature map, a dynamic context-aware network is used to analyze the fused feature map to obtain the environmental context around the seabed organisms.

[0086] Dynamically adjust detection strategies based on the environmental context surrounding seafloor organisms.

[0087] In an optional embodiment, the perception network is used to dynamically adjust the detection strategy in step S400, and it can also be optimized through an adaptive illumination compensation mechanism combined with multi-scale feature fusion. The adaptive illumination compensation mechanism can automatically adjust the contrast and brightness at the pixel level according to the illumination unevenness and scattering effect of the underwater image, and at the same time use a multi-scale feature pyramid to adaptively allocate weights of features at different levels, so that the detection network can maintain stable performance under different water depths, visibility and water turbidity conditions, which is particularly suitable for the identification of small organisms in seabed environments with changing illumination conditions.

[0088] In another optional embodiment, the perception network is used to dynamically adjust the detection strategy in step S400. The motion trajectory information of the target in continuous frame images can be used to construct a spatiotemporal feature map through a method combining spatiotemporal sequence modeling and motion prediction, and a dual-stream network architecture is used to process appearance features and motion features respectively. The two features are then adaptively fused through the attention mechanism. This can effectively handle complex situations such as rapid movement, occlusion, and posture changes of seabed organisms, and predict the next moment position of the target through algorithms such as Kalman filtering or particle filtering, thereby improving the tracking and detection accuracy of fast-moving seabed organisms. It is particularly suitable for real-time monitoring scenarios of small organisms such as sea cucumbers and sea urchins that move in complex benthic environments.

[0089] In an embodiment of the present invention, the step S400 dynamically adjusts the detection strategy according to the environmental context surrounding the submarine organisms, including:

[0090] Analyze the environmental context surrounding the seabed organisms to determine the type of seabed area in which the seabed organisms are located.

[0091] Adjust the weights of the perception network according to the type of seabed area where the seabed organisms are located.

[0092] For example, when combining sensors such as millimeter wave radar, adjusting the weights of the perception network according to the type of seabed area where the seabed organisms are located includes:

[0093] When seabed organisms are detected in dense seabed areas, that is, when more than 30% of the detected seabed organisms are occluded and the distance between seabed organisms is less than 10 meters, the analysis weights of the relationships between seabed organisms are adjusted, and the number of downsampling times in the CSPNet network is reduced to avoid losing occluded target information due to excessive stride and prevent false detection.

[0094] When seabed organisms are detected in sparse seabed areas, that is, when less than 30% of the detected seabed organisms are obscured and the distance between seabed organisms is greater than 10 meters, the original characteristics of the seabed organisms are emphasized, and Mosaic data enhancement is disabled to avoid over-cropping of sparse targets.

[0095] When seabed organisms are detected in the intersection area, the channel attention weight is increased and the C-dimensional scaling factor is adjusted from 1.0 to 1.2, which is equivalent to nonlinearly amplifying the weight, so that the network pays more attention to channels that are strongly related to the biological category (such as texture, color, etc.), ensuring that different types of seabed organisms are correctly distinguished.

[0096] It should be noted that in the actual application of seabed biological detection, the environment changes dynamically (such as light fluctuations, seabed topography, etc.). Due to the static reasoning mechanism and fixed strategy, the traditional model is often insufficient in the adaptability of the detection strategy, resulting in poor real-time performance and large fluctuations in recognition accuracy. The present invention introduces the environmental context perception mechanism and combines the perception network to dynamically adjust the detection strategy. It can intelligently adjust the detection focus according to factors such as seabed area density and target occlusion during the reasoning stage, and realize online adaptive adjustment at the strategy level. Especially in areas with severe target overlap and blurred scenes, the present invention can enhance the model's responsiveness to key features and significantly reduce false detection and missed detection rates. At the same time, the dynamic adjustment mechanism of the present invention breaks through the limitations of traditional rule-based or fixed threshold strategies, and has higher scene generalization capabilities and practical deployment value.

[0097] In summary, the present invention utilizes deep learning algorithms to integrate the morphological characteristics of seabed organisms with environmental background information, and can achieve accurate detection and dynamic tracking of seabed targets in complex underwater environments, thereby improving the accuracy and efficiency of seabed organism detection and providing strong support for the protection and management of marine resources. The coordinate attention mechanism is used to enhance the model's perception of small targets and key areas, while the Wise-IoU loss function is introduced to optimize detection accuracy and enhance the robustness of the model under conditions of large target scale differences and imbalanced categories. In the model training stage, a combination of multiple feature fusion strategies and dynamic parameter optimization methods is used to improve the model's convergence efficiency and generalization ability. In the deployment stage, the real-time detection architecture of the intelligent device end is adopted to significantly enhance the practicality and response speed of the detection method. It not only effectively improves the accuracy and automation level of seabed organism detection, but also has good adaptability and scalability, providing efficient and reliable technical support for marine resource surveys, underwater robot operations, and intelligent monitoring of the marine ecological environment.

[0098] Example 3 is the third embodiment of the present invention. This embodiment provides a seabed organism detection system based on deep learning, including: a data acquisition module, used to obtain seabed organism data set images, and generate a training data set based on the seabed organism data set images; a framework building module, used to build a framework of a seabed organism detection model based on the training data set and combined with a coordinate attention mechanism and loss function calculation; the loss function is used to calculate the positioning loss between the actual target box and the predicted target box; a model training module, used to train the framework of the seabed organism detection model to obtain a seabed organism detection model; a biological detection module, used to determine the seabed organism detection results based on the seabed organism detection model.

[0099] Example 4 is the fourth embodiment of the present invention, which is different from the first three embodiments in that:

[0100] like Figure 3 As shown, if the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0101] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0102] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0103] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one or combination of the following technologies known in the art can be used: a discrete logic circuit having logic gate circuits for implementing logic functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0104] Example 5, with reference to Figure 4 , which is the fifth embodiment of the present invention, provides a seabed organism detection method based on deep learning. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0105] This example trains a dataset on a constructed seabed organism detection model, adjusts model parameters based on the dataset and neural network model, and performs training for a preset number of iterations. First, images of various types of deep-sea organisms, including sea cucumbers, sea urchins, and scallops, are collected. These images are captured using a camera mounted on a submersible to obtain a dataset of deep-sea organism images. Next, the deep-sea organism images are processed and annotated to generate a YOLO-formatted training dataset suitable for deep-sea organism detection. The YOLO-formatted training dataset contains information such as the center point coordinates, length, width, and height of the deep-sea organisms, and is divided into training, validation, and test datasets in a ratio of 7:2:1.

[0106] Secondly, the features in the YOLO format training data set are extracted, and the features of different scales are fused and detected to obtain the feature map. Finally, the feature map is input into the seabed organism detection model to output the detection results.

[0107] In addition, during the training process of the submarine organism detection model, the training was performed according to the preset number of iterations, and the parameters were adjusted according to the data set and the neural network model. At the same time, the performance indicators of the submarine organism detection model, including accuracy, recall rate and loss value, were monitored, and the model performance was evaluated using the validation set. Based on the evaluation results, the learning rate and optimizer parameters of the submarine organism detection model were adjusted to ensure that the model has good generalization ability. Finally, as Figure 4 Shown is a schematic diagram of the seabed organism detection results. By inserting the trained model seabed organism detection model into the submersible, deep-sea organism detection is performed on the images sent by the camera, and a dynamic adjustment strategy of the perception network is adopted to output the category and confidence of the deep-sea organisms in the image. It can be seen that the present invention realizes the accurate detection and tracking of seabed organisms by utilizing deep learning algorithms and combining the characteristics of seabed organisms and environmental background.

[0108] 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 the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A deep learning-based method for detecting submarine organisms, characterized by: include, Obtaining a submarine organism dataset image, and generating a training dataset based on the submarine organism dataset image; Based on the training dataset and in combination with the coordinate attention mechanism and loss function calculation, a framework of a submarine organism detection model is built; the loss function is used to calculate the positioning loss between the actual target frame and the predicted target frame; Training the framework of the submarine organism detection model to obtain a submarine organism detection model; A seabed organism detection result is determined based on the seabed organism detection model.

2. The deep learning-based submarine organism detection method according to claim 1, wherein: Generating a training dataset based on the seabed organism dataset images includes the following steps: Based on the target seabed organisms, the objects in the seabed organism dataset are labeled by category and given a bounding box; The format of the seabed biological dataset images after category annotation is converted to obtain the training dataset.

3. The deep learning-based submarine organism detection method according to claim 2, characterized in that: The framework for building a submarine biological detection model includes: Use the CSPNet network as the backbone network to extract features from the training data set, temporarily store the obtained feature maps and cross-connect them in different subsequent stages of the CSPNet network; The FPN network is used to construct a feature pyramid, and the feature maps of different levels of the CSPNet network and the FPN network are fused to obtain a fused feature map; Performing a decomposition attention operation on the fused feature map based on a coordinate attention mechanism to aggregate global information along the height and width coordinate axes; For the feature graph processed by the coordinate attention mechanism, a PAN network is used to perform bidirectional path aggregation; During the training phase, the Wise-IoU loss function is used to calculate the positioning loss between the actual target box and the predicted target box.

4. The deep learning-based submarine organism detection method according to claim 3, wherein: The framework of the submarine organism detection model is trained to obtain a submarine organism detection model, including: Setting an initial learning rate based on a step size for updating parameters of the seabed organism detection model; Set the final value of the learning rate decay based on the initial learning rate; Setting momentum based on the parameter update and historical gradient information; Set the number of learning rate warmup epochs based on the warmup process of training.

5. The deep learning-based submarine organism detection method according to claim 4, characterized in that: Determining a submarine organism detection result based on the submarine organism detection model includes the following steps: deploying the seabed organism detection model into a smart device; Based on the seabed organism detection model, seabed organism detection is performed, and the detection strategy is dynamically adjusted using a perception network to obtain the category and confidence information of the seabed organism.

6. The deep learning-based submarine organism detection method according to claim 5, characterized in that: The method of dynamically adjusting the detection strategy using the perception network includes the following steps: According to the fused feature map, a dynamic context-aware network is used to analyze the fused feature map to obtain the environmental context surrounding the seabed organisms; The detection strategy is dynamically adjusted according to the environmental context surrounding the seabed organism.

7. The deep learning-based submarine organism detection method according to claim 6, characterized in that: Dynamically adjust the detection strategy based on the environmental context surrounding the seabed organism, including: Analyzing the environmental context surrounding the seabed organism to determine the type of seabed area in which the seabed organism resides; The weight of the perception network is adjusted according to the type of the seabed area where the seabed organism is located.

8. A deep learning-based submarine organism detection system, applying the deep learning-based submarine organism detection method according to any one of claims 1 to 7, characterized in that: include, The data acquisition module is used to obtain the submarine biological dataset images and generate a training dataset based on the submarine biological dataset images; A framework building module, which is used to build a framework for submarine biological detection models based on the training dataset and combined with the coordinate attention mechanism and loss function calculation; The loss function is used to calculate the positioning loss between the actual target box and the predicted target box; A model training module is used to train the framework of the submarine organism detection model to obtain a submarine organism detection model; The biological detection module is used to determine the seabed biological detection results based on the seabed biological detection model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the deep learning-based seabed organism detection method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the deep learning-based seabed organism detection method according to any one of claims 1 to 7 are implemented.

Images