A method and system for predicting the concentration of artificial sweetener acesulfame-k in offshore aquaculture seawater based on the behavior characteristics of juvenile black seabream

Abstract

The application discloses a kind of based on the prediction method and system of artificial sweetener acesulfame concentration in offshore aquaculture seawater based on black seabream juvenile behavior characteristics, comprising: collecting the motion video data of black seabream juvenile under different acesulfame concentration environments and pretreatment;Based on the improved YOLOv8 network model training gets the behavior recognition model of juvenile fish, and the behavior trajectory data set corresponding to the black seabream juvenile motion video data set;Based on GNN and GRU network constructs initial space-time correlation network model, utilizes the behavior trajectory data set to the initial space-time correlation network model training, by analyzing the behavior trajectory of black seabream juvenile, acesulfame concentration is predicted, and the space-time correlation network model of training is complete is obtained.The application can predict the artificial sweetener acesulfame concentration in the environment where black seabream juvenile is located by the behavior characteristics of black seabream juvenile, can provide strong support for offshore aquaculture seawater environment monitoring and early warning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of offshore marine environment monitoring, and in particular to a method and system for predicting the concentration of artificial sweetener acesulfame in offshore aquaculture seawater based on the behavior characteristics of juvenile black seabream. BACKGROUND

[0002] Sweeteners are substances that impart sweetness to food and belong to a category of food additives. Artificial sweeteners occupy half of the sweetener market in China due to their high sweetness, low usage, low cost, and strong market competitiveness. In recent years, artificial sweeteners found in the environment have been considered a new pollutant, and their potential biological toxicity, ecological risk, and health risk are of concern. Among them, acesulfame is not only favored in the field of food additives due to its simple production process, low price, and excellent performance, but also widely used in pharmaceuticals, personal care products, and feed. However, acesulfame cannot be metabolized by the human body, is difficult to completely remove through wastewater treatment, and has relatively high stability and a long half-life, making it easy to accumulate in the environment. Currently, acesulfame has been detected in the marine environment, with concentrations ranging from 1.9 x 10 -4 ~ 9.9 μg / L, posing an undeniable potential hazard to the marine ecosystem environment. Therefore, accurately predicting the concentration of artificial sweetener acesulfame in offshore seawater is of great significance for seawater environmental monitoring and early warning.

[0003] As a widely used artificial synthetic sweetener, acesulfame detection techniques currently include high-performance liquid chromatography, ion chromatography, and liquid chromatography-mass spectrometry. The detection limit of high-performance liquid chromatography for extremely low concentrations of acesulfame may still be high, making it impossible to accurately detect low concentrations of acesulfame in seawater. The results of high-performance liquid chromatography are affected by the selection and stability of the chromatographic column, and different brands and models of chromatographic columns may have different effects on the separation of acesulfame. In addition, the separation effect may decrease due to column efficiency decline or contamination. Ion chromatography is mainly suitable for compounds that can be easily ionized. Although acesulfame exists in the form of anions under acidic conditions, it may not be completely ionized under certain specific conditions, affecting the accuracy of the detection results. Liquid chromatography-mass spectrometry combines the techniques of high-performance liquid chromatography and mass spectrometry, which is more difficult to operate and maintain, and the equipment cost is higher.

[0004] In order to accurately establish an offshore marine ecological early warning monitoring system for acesulfame, the present application proposes a method and system for predicting the concentration of artificial sweetener acesulfame in offshore aquaculture seawater based on the behavior characteristics of juvenile black seabream. By using the behavioral response of aquatic organisms, the concentration of acesulfame in seawater is predicted, and the ecological risk of acesulfame as a new pollutant is evaluated from the perspective of ecological behavior. SUMMARY

[0005] Therefore, the application provides a method for predicting the concentration of artificial sweetener acesulfame in offshore aquaculture seawater based on the behavior characteristics of juvenile black seabream, to solve the technical problems of insufficient detection sensitivity, complex operation and high equipment cost of low concentration acesulfame in offshore aquaculture seawater.

[0006] To achieve the above technical purpose, the application adopts the following technical scheme:

[0007] On the one hand, the application provides a method for predicting the concentration of artificial sweetener acesulfame in offshore aquaculture seawater based on the behavior characteristics of juvenile black seabream, comprising:

[0008] Collecting the motion video data of juvenile black seabream in different acesulfame concentration environments and preprocessing to obtain the juvenile black seabream motion video data set;

[0009] Building an initial juvenile behavior recognition model based on an improved YOLOv8 network model, training the initial juvenile behavior recognition model using the juvenile black seabream motion video data set to obtain a trained juvenile behavior recognition model and the corresponding behavior trajectory data set of the juvenile black seabream motion video data set;

[0010] Building an initial spatio-temporal correlation network model based on GNN and GRU networks, training the initial spatio-temporal correlation network model using the behavior trajectory data set, and predicting the acesulfame concentration by analyzing the behavior trajectory of juvenile black seabream to obtain a trained spatio-temporal correlation network model;

[0011] Obtaining the actual motion video of juvenile black seabream in offshore aquaculture seawater, and predicting the acesulfame concentration in the offshore aquaculture seawater according to the actual motion video using the juvenile behavior recognition model and the spatio-temporal correlation model.

[0012] Further, collecting the motion video of juvenile black seabream in different acesulfame concentration environments and preprocessing, comprising:

[0013] Collecting the motion video of juvenile black seabream in different acesulfame concentration environments based on the preset shooting angle, resolution and frame rate to obtain the original video data;

[0014] Segmenting the original video data into segments of equal length and labeling the acesulfame concentration for each video to obtain the motion video data set of juvenile black seabream;

[0015] Expanding the video data set using data enhancement technology to obtain an expanded data set;

[0016] Labeling the juvenile black seabream in each video in the expanded data set and converting the labeled video to obtain the preprocessed motion video data set;

[0017] The preprocessed motion video dataset is divided into a training set, a validation set and a test set.

[0018] Further, the improved YOLOv8 network model comprises:

[0019] The C2f module in the YOLOv8 backbone network is replaced by a C2f-MSADCN module stacked by a C2f module and an MSADCN module;

[0020] The C2f module in the YOLOv8 neck network is integrated into an MSSimAM attention module;

[0021] The CIoU loss function in the YOLOv8 head network is replaced by an SEIoU loss function.

[0022] Further, the MSADCN module introduces a multi-scale and attention mechanism, and the convolution operation of the module is represented by a formula:

[0023]

[0024] Where y(p) is the feature value of position p in the output feature map, S represents the number of multi-scales, α s is a learnable weight on scale s, used to adjust the contribution of features of different scales, K is the number of sampling points of the convolution kernel, is the convolution weight of the kth sampling position under the s th scale, is the offset reference of the kth sampling position, is a learnable offset, is the modulation coefficient of the kth sampling point, used to adjust the importance of the feature at this position.

[0025] Further, the MSSimAM attention module comprises a multi-scale convolution layer, a similarity adjustment layer, an attention layer and a feature fusion layer, which is used to capture local and global information of an image using a multi-scale receptive field;

[0026] The multi-scale convolution layer is used to extract features of different scales using convolution kernels of different sizes;

[0027] The similarity measurement layer is used to measure the similarity between features of different scales;

[0028] The attention layer is used to generate an attention map according to the similarity measurement, and to weight the multi-scale features to highlight key features;

[0029] The feature fusion layer is used to fuse the weighted multi-scale features into the final output of the module.

[0030] Further, the SEIoU loss function is represented by a formula:

[0031]

[0032] Among them, 1-IoU represents the intersection-over-union loss, which is used to measure the overlap difference between the predicted box and the real box, ρ 2 (b,b gt ) represents the square of the Euclidean distance between the center point of the predicted box and the center point of the real box, w and w gt The widths of the predicted box and the real box, h and h respectively gt Represents the height of the predicted box and the real box, w c and h c are the width and height of the minimum bounding rectangle that can wrap the predicted box and the true box, respectively.

[0033] Furthermore, the initial juvenile behavior recognition model was trained using the black sea bream juvenile motion video dataset to obtain a fully trained juvenile behavior recognition model, including:

[0034] The initial juvenile behavior recognition model is trained using the training set and validation set of the black sea bream juvenile motion video dataset to obtain a trained juvenile behavior recognition model.

[0035] The trained juvenile fish behavior recognition model was tested using the test set of the black sea bream juvenile motion video dataset;

[0036] When the prediction results of the model meet the training expectations, a fully trained juvenile fish behavior recognition model is obtained, and the coordinate time series of all black sea bream juveniles in the training set, validation set and test set videos are extracted respectively.

[0037] Furthermore, the initial spatiotemporal correlation network model based on GNN-GRU is constructed, including:

[0038] The initial spatiotemporal correlation model includes a GNN module, a multi-layer GRU module and a fully connected module;

[0039] The GNN module is used to obtain spatial features based on the position feature input of the black sea bream juveniles at different times, so as to capture the spatial correlation between the black sea bream juveniles;

[0040] The multi-layer GRU module includes multiple GRU network models connected in series, and calculates the hidden state of each GRU network model according to the spatial characteristics of each moment; it is used to capture the temporal correlation of the behavior of black sea bream juveniles;

[0041] The fully connected network model predicts the acesulfame potassium concentration based on the hidden state output by the multi-layer GRU module.

[0042] Further, the initial spatio-temporal correlation network model is trained by using the behavior trajectory dataset, the acesulfame-K concentration is predicted by analyzing the behavior trajectory of the juvenile red drum, and a trained spatio-temporal correlation network model is obtained, including:

[0043] The initial spatio-temporal correlation network model is supervised trained by using the coordinate time series of the juvenile red drum in the training set and the validation set and the corresponding acesulfame-K concentration as labels.

[0044] The trained spatio-temporal correlation network model is tested by using the coordinate time series of the juvenile red drum in the test set and the corresponding acesulfame-K concentration as labels, and when the accuracy of the prediction result reaches the expectation, the trained spatio-temporal correlation network model is obtained.

[0045] On the other hand, the application also provides a system for predicting the concentration of artificial sweetener acesulfame-K in offshore aquaculture seawater based on the behavior characteristics of juvenile red drum, including:

[0046] A data acquisition module is configured to collect and preprocess the motion video data of juvenile red drum in different acesulfame-K concentration environments to obtain a juvenile red drum motion video dataset.

[0047] A behavior recognition model construction module is configured to construct an initial juvenile fish behavior recognition model based on an improved YOLOv8 network model, train the initial juvenile fish behavior recognition model by using the juvenile red drum motion video dataset, obtain a trained juvenile fish behavior recognition model, and obtain a behavior trajectory dataset corresponding to the juvenile red drum motion video dataset.

[0048] A spatio-temporal correlation network model construction module is configured to construct an initial spatio-temporal correlation network model based on GNN-GRU, train the initial spatio-temporal correlation network model by using the behavior trajectory dataset, predict the acesulfame-K concentration by analyzing the behavior trajectory of the juvenile red drum, and obtain a trained spatio-temporal correlation network model.

[0049] A concentration prediction module is configured to obtain the actual motion video of juvenile red drum in offshore aquaculture seawater, and predict the acesulfame-K concentration of the offshore aquaculture seawater according to the actual motion video by using the juvenile fish behavior recognition model and the spatio-temporal correlation model.

[0050] Compared with the prior art, the application provides a prediction method for the concentration of artificial sweetener Acesulfame-K based on the behavior characteristics of juvenile black seabream, which improves the classification accuracy of different concentrations of Acesulfame-K according to the behavior characteristics of juvenile black seabream by constructing a behavior recognition model and a spatio-temporal joint model based on improved YOLOv8 and GNN-GRU respectively, thereby providing a scientific basis for evaluating the pollution of Acesulfame-K to the marine ecological environment. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 A flowchart of the prediction method for the concentration of artificial sweetener Acesulfame-K in offshore aquaculture seawater based on the behavior characteristics of juvenile black seabream is provided.

[0052] Figure 2 A structure diagram of an initial juvenile behavior recognition model constructed based on the improved YOLOv8 network model is provided.

[0053] Figure 3 A structure diagram of the MSADCN module is provided.

[0054] Figure 4 A structure diagram of the MSSimAM attention module is provided.

[0055] Figure 5 A structure diagram of the initial spatio-temporal correlation network model is provided.

[0056] Figure 6 An mAP result diagram of the improved YOLOv8 model combined with different modules for training is provided.

[0057] Figure 7 A training time consumption result diagram of the improved YOLOv8 model combined with different modules for training is provided.

[0058] Figure 8 A comparison diagram of the prediction accuracy of the spatio-temporal correlation model and the random forest algorithm is provided. DETAILED DESCRIPTION

[0059] The preferred embodiments of the application will be specifically described below with reference to the accompanying drawings, which form a part of this application, and are used to explain the principles of the application together with the embodiments of the application, but are not used to limit the scope of the application.

[0060] Before introducing the embodiments of the application, the inventive concept of the present application is described.

[0061] Behavioral changes are the most direct and sensitive responses of organisms to external environmental stimuli and internal environmental changes. The behavioral responses of aquatic organisms have been used as a method for environmental monitoring, but they have received much less attention than other fields such as animal toxicology. Among the organisms used for behavioral assessment, fish are particularly noteworthy. Due to their wide distribution in aquatic ecosystems and their key ecological position in the food web, fish play a crucial role in ecological toxicology monitoring. Fish have longer iterative cycles compared to other aquatic organisms, and thus play a crucial role in ecological toxicology monitoring.

[0062] Etelis ocellatus is a typical offshore fish widely distributed in the coastal waters of China and other Asian countries. In these areas, the important position of Etelis ocellatus in the food chain makes it a key species in the aquatic ecosystem. As a benthic fish, Etelis ocellatus larvae are highly sensitive to changes in the physical and chemical properties of seawater, especially when the concentration of pollutants such as dissolved oxygen, temperature, and pollutants changes. By monitoring the behavioral changes of Etelis ocellatus larvae, not only can the effects of acesulfame-K concentration on this species be reflected, but also the effects of pollutants on the entire ecosystem can be indirectly reflected, thereby providing data support for ecological risk assessment.

[0063] Artificial sweeteners can have toxic effects on fish behavior. Existing studies have shown that artificial sweeteners can cause abnormal behavior in fish, such as social behavior, anxiety, and memory and cognitive behavior. For example, adult zebrafish exposed to aspartame and acesulfame-K showed increased anxiety behavior and reduced activity; high-dose aspartame increased the phototaxis and preference for blue in zebrafish, and reduced their exploration ability and social behavior. Zebrafish fed with aspartame and high-fat diet also showed acute decline in swimming ability. Similarly, zebrafish exposed to saccharin also showed anxiety-like behavior, reduced exploration ability and sociality, and interference with color-based associative learning and memory ability. In summary, fish behavior has the advantages of high sensitivity and easy observation, and plays an indispensable role in the ecological risk assessment of pollutants. It is an effective entry point for establishing an ecological early warning monitoring system.

[0064] Therefore, from the perspective of ecological behavior, assessing the ecological risk of acesulfame-K, a new type of pollutant, can help establish an early warning monitoring system for acesulfame-K in offshore marine ecosystems, and provide scientific support for its ecological risk and pollution control.

[0065] Based on the potential link between fish behavior and artificial sweeteners, the present application uses deep learning technology to identify the trajectory of Etelis ocellatus larvae in an offshore aquaculture seawater environment contaminated with artificial sweetener acesulfame-K. According to the behavioral characteristics of Etelis ocellatus larvae, the concentration of artificial sweetener acesulfame-K in their environment is classified to predict the concentration of artificial sweetener acesulfame-K in seawater, thereby achieving monitoring of the offshore aquaculture seawater environment.

[0066] Please refer toFigure 1 The embodiment provides a method for predicting the concentration of acesulfame in offshore aquaculture seawater based on the behavior characteristics of juvenile black seabreams, comprising the following steps:

[0067] Step S101: Collecting motion video data of juvenile black seabreams in different acesulfame concentration environments and performing preprocessing to obtain a juvenile black seabream motion video data set;

[0068] Step S102: Constructing an initial juvenile fish behavior recognition model based on an improved YOLOv8 network model, training the initial juvenile fish behavior recognition model by using the juvenile black seabream motion video data set to obtain a trained juvenile fish behavior recognition model and a behavior trajectory data set corresponding to the juvenile black seabream motion video data set;

[0069] Step S103: Constructing an initial spatiotemporal correlation network model based on a GNN and a GRU network, training the initial spatiotemporal correlation network model by using the behavior trajectory data set, predicting the acesulfame concentration by analyzing the behavior trajectory of the juvenile black seabream to obtain a trained spatiotemporal correlation network model;

[0070] Step S104: Obtaining actual motion video of juvenile black seabreams in offshore aquaculture seawater, and predicting the acesulfame concentration in the offshore aquaculture seawater according to the actual motion video by using the juvenile fish behavior recognition model and the spatiotemporal correlation model.

[0071] The method of the embodiment uses the improved YOLOv8 network model to recognize the behavior of juvenile black seabreams, and the behavior recognition model obtained by training can accurately extract the motion behavior data of juvenile black seabreams. The spatiotemporal correlation analysis of the behavior trajectory by combining the GNN (graph neural network) and the GRU (gated recurrent unit) network can more accurately predict the acesulfame concentration in seawater. By continuously collecting behavior data of juvenile fish and performing real-time analysis, real-time monitoring and prediction of the acesulfame concentration in the offshore aquaculture environment can be realized. This method avoids the cumbersome process of traditional manual sampling and laboratory analysis, improves the prediction efficiency and timeliness, and can accurately predict the acesulfame concentration in the offshore aquaculture environment in real time, efficiently and accurately. By predicting the acesulfame concentration in the offshore aquaculture water area, water quality changes can be detected in advance, and timely measures can be taken for adjustment. This helps to improve the efficiency of aquaculture water area management and ensure the health of the aquaculture environment, thereby improving the growth efficiency of black seabream and the economic benefits of aquaculture. In addition, the method can reduce the need for manual operation, reduce labor costs, and reduce the influence of human factors on the prediction results. For large-scale farms, it can effectively reduce the workload of management personnel and improve the efficiency of aquaculture management.

[0072] As a preferred embodiment, in step S101, the motion video of juvenile black seabreams in different acesulfame concentration environments is collected and preprocessed, comprising:

[0073] Based on the preset shooting angle, resolution, frame rate, the juvenile black seabream movement videos in different acesulfame-K concentration environments were collected to obtain the original video data;

[0074] The original video data was segmented into segments with equal time length, and each video was labeled with acesulfame-K concentration to obtain the juvenile black seabream movement video dataset;

[0075] Data augmentation techniques were used to expand the video dataset to obtain the expanded dataset;

[0076] The juvenile black seabream in each video in the expanded dataset was labeled, and the labeled video was format-converted to obtain the preprocessed movement video dataset;

[0077] The preprocessed movement video dataset was divided into a training set, a validation set, and a test set.

[0078] As a specific example, in order to ensure the quality and diversity of the video data, so as to better analyze the movement behavior of juvenile black seabream, four different acesulfame-K concentration environments (0 μg / L, 10 μg / L, 100 μg / L, and 1000 μg / L) were constructed, and 10 juvenile black seabreams were placed in each environment. A Hikvision camera was used to shoot from a top-down angle, with a video resolution of 720p and a frame rate of 30 frames per second. All videos were segmented according to a 15-second time length, and each video was labeled with acesulfame-K concentration to form a juvenile black seabream movement video dataset.

[0079] Data augmentation techniques were used to expand the video dataset, specifically: by rotating the video by 90 degrees, 180 degrees, and 270 degrees, as well as performing horizontal and vertical mirror inversion, cropping, mirroring, brightness adjustment, etc., the video dataset was expanded. This can make the training model better generalize, avoid overfitting, increase the number of training samples, and improve the robustness of the model.

[0080] For the expanded video dataset, LabelMe was used to label the juvenile black seabream in the video. The labeled data can help researchers analyze the movement patterns of fish and identify factors that affect movement behavior (such as acesulfame-K concentration). It was converted to YOLOv8 format, and the dataset was divided into a training set, a validation set, and a test set according to different acesulfame-K concentrations.

[0081] YOLOv8 is a new version of the YOLO (You Only Look Once) series of target detection models released after YOLOv5. It inherits the core idea of the YOLO series, which is based on a convolutional neural network (CNN) and directly regresses the class and location of the target in the image. YOLOv8 performs well in target detection, especially in simultaneously detecting multiple targets in an image. The present invention selects YOLOv8 model as the basic framework of the model, and improves and optimizes the model according to the dynamic moving video characteristics of juvenile red sea bream in seawater, so as to meet the actual needs of analyzing the motion video of juvenile red sea bream and identifying its behavior trajectory.

[0082] When identifying juvenile red sea bream in nearshore seawater, YOLOv8 has the following defects:

[0083] 1. Since red sea bream usually appears in a complex nearshore seawater background, factors such as reflection of nearshore seawater, changes in illumination, transparency of fish school, etc. can cause low contrast between fish school and background. Due to the small size of the juvenile fish school and the flow fluctuation of seawater and the refraction of underwater light, the video background is cluttered. YOLOv8, as a general target detection network, lacks the ability to extract key features such as fast swimming or dense aggregation, and cannot well distinguish between fish school and background, thus failing to complete the identification task of fish school behavior.

[0084] 2. The red sea bream school is often a dynamic group with small distances between individuals, and even overlaps. The size, speed and direction of motion of the red sea bream school vary greatly, so the motion trajectory may appear at different spatial scales. YOLOv8 has difficulty in identifying target features of different sizes, which affects the effect of fish school behavior identification.

[0085] 3. The loss function of the YOLOv8 base model is CIoU, which is a loss function commonly used to improve the accuracy of bounding box regression in target detection tasks. It combines center distance, aspect ratio and overlapping area to optimize the regression of target bounding box. In a dynamic environment, the behavior of the fish school causes the bounding box to change constantly, and the motion trajectory of the red sea bream school may involve the interaction of multiple individuals, with large changes in target shape and scale. At this time, the CIoU loss function cannot accurately reflect the real-time position and changes of the fish school.

[0086] In order to solve the above problems, we optimize and improve the YOLOv8 network model to better identify the behavior of juvenile red sea bream in nearshore seawater scenes. As a preferred embodiment, in step S102, the specific improvement method of the YOLOv8 network model includes:

[0087] Replace the C2f module in the YOLOv8 backbone network with a C2f-MSADCN module stacked by the C2f module and the MSADCN module;

[0088] Integrate the C2f module in the YOLOv8 neck network into the MSSimAM attention module;

[0089] Replace the CIoU loss function in the YOLOv8 head network with the SEIoU loss function.

[0090] The following is a detailed introduction to the specific improvement methods:

[0091] (1) Replace the C2f module in the YOLOv8 backbone network with a C2f-MSADCN module stacked by the C2f module and the MSADCN module:

[0092] The strategy of MSADCN mainly includes multi-scale feature extraction, attention mechanism and dynamic context learning.

[0093] As a specific example, Figure 2 As shown, Figure 2 This is a schematic diagram of the structure of an initial juvenile fish behavior recognition model based on the improved YOLOv8 network model. In this embodiment, the C2f module in the original YOLOv8 backbone network is replaced with a C2f-MSADCN module formed by stacking the C2f module and the MSADCN module.

[0094] The MSADCN module (Multiscale Attention Deformable Convolutional Network) is an improvement to the variable convolutional network DCNv2. The network introduces multi-scale and attention mechanisms based on DCNv2, enabling it to more effectively handle irregular deformation and multi-scale features. The MSADCN module structure is as follows: Figure 3 As shown. The MSADCN core convolution operation can be expressed as the following formula:

[0095]

[0096] Among them, y(p) is the eigenvalue of position p in the output feature map, S represents the number of multi-scales, α s is the learnable weight on scale s, which is used to adjust the feature contribution of different scales. K is the number of sampling points of the convolution kernel. is the convolution weight of the kth sampling position at the sth scale, is the offset reference of the k-th sampling position, is the learnable offset, is the modulation coefficient of the kth sampling point, used to adjust the importance of the feature at this position.

[0097] By replacing the C2f module with a C2f-MSADCN module formed by stacking a C2f module and a MSADCN module, more accurate identification of irregular deformations exhibited by a school of black seabream when swimming, as well as the rapid swimming of the school of fish, blurring caused by water flow, and a constantly changing background, can be achieved in underwater video monitoring, thereby improving the overall recognition effect of the behavior of juvenile black seabream. By focusing on the fish behavior in the complex background of the video, which is obscured by interference factors in the surrounding environment (such as water flow fluctuations, plants, sand, etc.), and by automatically adjusting the position and weight of the convolution sampling point through dynamic context learning, the performance of YOLOv8 in black seabream school behavior trajectory recognition can be better improved.

[0098] (2) integrate the C2f module in the YOLOv8 neck network into the MSSimAM attention module:

[0099] MSSimAM (Multi-Scale Simulated Attention Module) is an improvement of the SimAM attention module. This module can better capture local and global information of an image using a multi-scale receptive field. The structure of the MSSimAM attention module is shown in Figure 4 .

[0100] For an input feature map where C represents the number of channels, H represents the height of the input feature map, and W represents the width of the input feature map. In some embodiments, C = 256, H = 1280, and W = 720. The energy function of MSSimAM at position (h, w) under scale s can be represented as:

[0101]

[0102] where y j represents the output of the jth neuron, is the mean of all neuron outputs under scale s, is the variance of neuron outputs under scale s, λ is a parameter that controls the balance between the mean term and the variance term, and M is the number of neurons in the neighborhood.

[0103] The attention weight of MSSimAM at position (h, w) under different scales s can be represented as:

[0104]

[0105] where ∈ is a constant to prevent division by zero.

[0106] The attention weight of the multi-scale fusion post-position (h, w) of MSSimAM can be represented as:

[0107]

[0108] where S represents the number of different scales, w s represents the fusion weight of scale s, used to control the contribution of each scale to the final feature.

[0109] The feature value of the enhanced post-position (h, w) of MSSimAM can be represented as:

[0110] y h,w = A(h, w)x h,w

[0111] where x h,w represents the feature value of position (h, w) in the input feature map, y h,w represents the feature value of the enhanced post-position (h, w).

[0112] Through the above improvements, MSSimAM can identify the movement of juvenile black seabream in fast swimming, turning, or between different water layers, thereby capturing dynamic behavior information at different scales. For example, when the black seabream moves in a long distance or a small background, MSSimAM can identify the overall movement pattern of the fish through a large receptive field; when the black seabream swims quickly in a short distance or a complex background, a small-scale receptive field helps the model focus on local details and rapidly changing features. At the same time, MSSimAM can balance the influence of local and global information by calculating the energy function at different scales, avoiding missing important movement information due to the overweight of a certain scale, to help the model identify subtle changes in group interaction. Through the combination of local and global information, the ability to capture complex behaviors is enhanced. For example, when a school of fish moves collectively, their global movement trend needs to be captured; when individual black seabream performs detailed actions (such as jumping, pausing, fast swimming, etc.), local changes need to be focused on. By calculating the attention weight of each scale, local and global information is organically integrated, thereby enhancing the ability to capture these complex behaviors.

[0113] Therefore, integrating the C2f module in the YOLOv8 neck network with the MSSimAM attention module can significantly improve the behavior recognition accuracy in black seabream movement videos, especially in dynamic scenes, complex backgrounds, and small object detection. Through the optimization of multi-scale feature fusion and attention mechanism, MSSimAM not only can capture more rich feature information, but also can improve the robustness and generalization ability of the model, thereby enhancing the behavior classification, positioning, and tracking ability of black seabream.

[0114] (3) Replace the CIoU loss function in the YOLOv8 head network with the SEIoU loss function:

[0115] The SEIoU loss function is called Simplified Efficient Intersection over Union, which is an improvement of the EIoU loss function. The SEIoU loss function can be expressed as the following formula:

[0116]

[0117] where 1-IoU represents the intersection over union loss, which is used to measure the overlap difference between the predicted box and the real box, and p 2 (b, b gt ) represents the square of the Euclidean distance between the center points of the predicted box and the real box, w and w gt are the width of the predicted box and the real box, h and h gt are the height of the predicted box and the real box, w c and h c are the width and height of the smallest bounding rectangle that can wrap the predicted box and the real box.

[0118] Using SEIoU as the loss function of the model not only makes the model pay attention to the overlap between the boxes (IoU), but also introduces the Euclidean distance of the center points and the scale adjustment term, which can accurately locate the behavior patterns of the fast-moving or small-range changing black seabream. SEIoU increases the constraint on the position of the box center, making the predicted box closer to the real box and reducing the error caused by the deviation of the predicted box position. Especially in the fast-moving or overlapping seabream motion scene, it effectively improves the positioning accuracy of the model. In addition, the scale adjustment term in SEIoU, especially by calculating the proportion of the width and height difference relative to the smallest bounding rectangle, can improve the accuracy of behavior recognition affected by the size difference of the fish body. For example, small black seabream at a distance or black seabream blocked by other fish often produce boundary boxes of different sizes. Under the influence of long distance or light conditions, it appears as a relatively small black seabream juvenile. The scale adjustment part in the SEIoU loss function can more effectively handle small object detection, especially in situations far from the camera or blocked by other objects, maintaining high-precision detection, ultimately enabling the YOLOv8-based black seabream behavior recognition model to more accurately and reliably identify and track multiple behavior patterns of black seabream.

[0119] As a preferred embodiment, the initial juvenile behavior recognition model is trained using the juvenile seabream motion video dataset to obtain a trained juvenile behavior recognition model, including:

[0120] The initial juvenile fish behavior recognition model is trained using the training set and the validation set of the juvenile fish motion video data set of black seabream, and a trained juvenile fish behavior recognition model is obtained;

[0121] The trained juvenile fish behavior recognition model is tested using the test set of the juvenile fish motion video data set of black seabream;

[0122] When the prediction result of the model meets the training expectation, a trained juvenile fish behavior recognition model is obtained, and the coordinate time series of all juvenile fish of black seabream in the training set, the validation set and the test set videos are extracted respectively.

[0123] Specifically, the training expectation is measured by mAP (mean Average Precision, average precision mean). If the mAP does not meet the expectation (usually set to 0.95), it may be because there are some problems in the model training or the hyperparameters are not properly adjusted, and the model needs to be further trained until the mAP meets the preset standard threshold. A trained juvenile fish behavior recognition model is obtained, and the coordinate time series of all juvenile fish of black seabream in the training set, the validation set and the test set videos are extracted respectively. These data are position time series, that is, the position information of juvenile fish of black seabream at each time, and their positions (coordinates) and possible labels (such as behavior categories) are output.

[0124] As a specific example, the coordinate time series of all juvenile fish of black seabream in the training set, the validation set and the test set videos are extracted respectively, the position coordinates are extracted from the first frame and the 16th frame of each second, and the coordinate time series of each video is represented as a three-dimensional matrix

[0125] Wherein, T represents the time length, N represents the number of juvenile fish of black seabream in the video, 2 represents the x and y coordinates of each juvenile fish of black seabream, and in this embodiment, T and N take values of 30 and 10 respectively.

[0126] As a preferred embodiment, in step S103, the initial spatio-temporal correlation network model based on GNN-GRU is constructed, comprising:

[0127] The initial spatio-temporal correlation model comprises a GNN module, a multi-layer GRU module and a full connection module;

[0128] The GNN module is used to input the position features of juvenile fish of black seabream at different times to obtain spatial features, which are used to capture the spatial correlation between juvenile fish of black seabream;

[0129] The multi-layer GRU module comprises a plurality of serially connected GRU network models, which calculate the hidden states of each GRU network model according to the spatial features at each time, and are used to capture the temporal correlation of the behavior of juvenile fish of black seabream;

[0130] The fully connected network model predicts the acesulfame-K concentration according to the hidden state output by the multi-layer GRU module.

[0131] As a specific embodiment, as shown in Figure 5 , Figure 5 The initial spatio-temporal correlation network model structure is shown. The spatio-temporal correlation network model GNN-GRU is composed of a GNN network model, a multi-layer GRU network model and a fully connected network model. The GNN network model is used to process the location feature data X t of each time t and output a spatial feature H t with a dimension of Nxd. The multi-layer GRU network model is composed of T GRU network models connected in series, which sequentially accept the spatial feature H t of each time t, gradually calculate the hidden state of each GRU network model, and finally output the hidden state h T of the Tth GRU network model as the output result of the entire multi-layer GRU network model. The fully connected network model uses the hidden state h T to predict the acesulfame-K concentration.

[0132] In the spatio-temporal correlation model, the GNN network model is used to capture the spatial correlation between juvenile black seabreams. For input data The N x 2 location feature at time t (1≤t≤T) represents the node position matrix at this time, and the GNN is used to capture the spatial relationship between nodes at time t. The GNN network structure includes input node structure, adjacency matrix calculation and graph convolution module. Among them, the GNN input node structure is the N x 2 location feature at time t, which can be represented as a matrix Each row represents the node coordinates x, y; the adjacency matrix is represented as a matrix It reflects the spatial relationship between nodes, which is generated according to the Euclidean distance between nodes, that is, The graph convolution module convolves the features of each node to capture the spatial correlation between targets, and its update rule is:

[0133]

[0134] Among them, is the node feature of the lth layer, and the initial value is X t , A t is the normalized adjacency matrix, D is the degree matrix of the adjacency matrix, W (l) is the learned weight matrix, and σ is the nonlinear activation function ReLU. The output result of the GNN is an Nxd feature matrix H t , where d represents the node feature embedding dimension.

[0135] The GRU network model is used to capture the time correlation of the juvenile fish behavior. t The GRU network model is used to capture the time dependence of the nodes within T time. t The input node structure is an Nxd-dimensional feature matrix H t―1 The GRU update rule is represented by the following formula:

[0136]

[0137] where z t represents the update gate, h t―1 represents the hidden state at the previous time, represents the candidate hidden state.

[0138] As a preferred embodiment, the initial spatiotemporal correlation network model is trained using the behavior trajectory dataset, and the acesulfame-K concentration is predicted by analyzing the behavior trajectory of the juvenile fish, to obtain a trained spatiotemporal correlation network model, including:

[0139] The initial spatiotemporal correlation network model is supervised trained using the coordinate time series of the juvenile fish in the training set and the validation set and the corresponding acesulfame-K concentration as labels;

[0140] The trained spatiotemporal correlation network model is tested using the coordinate time series of the juvenile fish in the test set and the corresponding acesulfame-K concentration as labels, and when the accuracy of the prediction result reaches the expectation, the trained spatiotemporal correlation network model is obtained.

[0141] In order to test the actual effect of the present application, the following experiments were carried out: the YOLOv8 basic model, the YOLOv8 basic model respectively combined with the MSADCN module, the MSSimAM module and the SEIoU loss function were used to train the juvenile fish dataset, and the mAP and training time of each model were respectively shown in Figure 6 and Figure 7 .

[0142] Model A is the basic YOLOv8 model, with an mAP of 87.6% and a training time of 11.04 hours.

[0143] Model B, which combines MSADCN in the basic model, has an mAP increase of 1.6% compared to Model A, and a training time increase of 1.42 hours. Although the training time is slightly increased, MSADCN enhances the feature extraction and representation ability of the model in low-light water environment, improving the detection accuracy of the model.

[0144] Model C integrates MSSimAM into the base model. Similar to Model B, MSSimAM further improves mAP with a small increase in training time. This is mainly because MSSimAM enhances local detailed features, thus improving the accuracy of juvenile red drum behavior recognition tasks.

[0145] Model D improves the loss function of the base model, resulting in a slight increase in mAP and a decrease in training time compared to Model A. This is mainly due to SEIoU, which accelerates model convergence and reduces training time.

[0146] Model E integrates MSADCN and MSSimAM into the base model, resulting in a significant improvement in accuracy. The integration of these three significantly improves the model's ability to describe features.

[0147] Model F integrates MSADCN in the base model and uses SEIoU, with an mAP of 89.9% and a training time of 12.39 hours. The integration of these technologies enables the model to more accurately identify targets and focus on high-quality priors.

[0148] Model G uses MSSimAM and SEIoU technologies to improve the accuracy of identifying juvenile red drum behavior while reducing training time. Model H integrates MSADCN, MSSimAM, and SEIoU, with an mAP of 95.7% and a slight increase in training time to 13.04 hours. By supplementing different modules, the newly constructed YOLOv8 improved model achieves better performance improvement and ensures a good trade-off between the recognition accuracy and computational speed of juvenile red drum behavior detection.

[0149] For the improved YOLOv8 model extracted juvenile red drum position time series, GNN-GRU spatiotemporal correlation model and traditional classification method are used to predict acesulfame-K concentration, respectively. The traditional classification method calculates the group movement speed, nearest neighbor distance, individual swimming speed, individual swimming speed synchronicity, and polarity of juvenile red drum, and realizes the prediction of acesulfame-K concentration through the random forest algorithm. The prediction results of the two methods are shown in Figure 8 . Among them, the prediction accuracy of the random forest is 89.97%, while the prediction accuracy of GNN-GRU is 95.46%, which is much higher than that of the traditional prediction method.

[0150] From the above experimental results, it can be seen that the method of the present application has excellent performance in juvenile red drum behavior recognition and acesulfame-K concentration prediction tasks. The comprehensive effect of target recognition and classification prediction is better than that of existing methods, and the adaptability and detection accuracy in complex environments have also been further improved. The method not only has significant theoretical value, but also provides strong support for the practical application of aquatic ecological environment.

[0151] The embodiment also provides a prediction system for the concentration of artificial sweetener acesulfame in offshore aquaculture seawater based on the behavior characteristics of juvenile black seabream, comprising:

[0152] A data acquisition module is configured to collect motion video data of juvenile black seabream in different acesulfame concentrations and perform preprocessing to obtain a juvenile black seabream motion video dataset;

[0153] A behavior recognition model construction module is configured to construct an initial juvenile fish behavior recognition model based on an improved YOLOv8 network model, train the initial juvenile fish behavior recognition model using the juvenile black seabream motion video dataset, obtain a trained juvenile fish behavior recognition model, and obtain a behavior trajectory dataset corresponding to the juvenile black seabream motion video dataset;

[0154] A spatio-temporal correlation network model construction module is configured to construct an initial spatio-temporal correlation network model based on GNN-GRU, train the initial spatio-temporal correlation network model using the behavior trajectory dataset, analyze the behavior trajectory of juvenile black seabream to predict the acesulfame concentration, and obtain a trained spatio-temporal correlation network model;

[0155] A concentration prediction module is configured to obtain actual motion video of juvenile black seabream in offshore aquaculture seawater, and predict the acesulfame concentration in the offshore aquaculture seawater according to the actual motion video using the juvenile fish behavior recognition model and the spatio-temporal correlation model.

[0156] The prediction method for the concentration of artificial sweetener acesulfame in offshore aquaculture seawater based on the behavior characteristics of juvenile black seabream has the core advantage that the juvenile fish behavior recognition model and the spatio-temporal correlation network model are respectively constructed by improved YOLOv8 and GNN-GRU models. The two models are optimized for the recognition task of the behavior characteristics of juvenile black seabream, and the model performance is improved. By introducing a multi-scale attention deformable convolution network (MSADCN) in YOLOv8, more accurate recognition can be performed on the irregular deformation of juvenile black seabream in the underwater video monitoring when swimming, the rapid swimming of the fish school, the blur caused by the water flow, and the constantly changing background and multi-scale features. In addition, a multi-scale simulation attention module (MSSimAM) is added to improve the feature fusion capability. At the same time, a simple and efficient intersection over union (SEIoU) loss function is used to improve the accuracy of the boundary box regression, thereby further optimizing the detection effect and calculation speed.

[0157] On the other hand, the GNN-GRU spatio-temporal joint model models the spatial features in each time step through GNN, captures the complex relationship between targets, and gradually updates the hidden state on the time series through GRU to capture long and short term dependencies, forming a joint modeling of spatio-temporal dependencies. Combined with the spatial feature extraction of GNN and the time dynamic modeling of GRU, the classification accuracy of different concentrations of acesulfame according to the behavior characteristics of juvenile red drum is improved, thereby providing a scientific basis for evaluating the pollution of acesulfame to the marine ecological environment.

[0158] In summary, the present application has wide application potential, such as fish ecological behavior research, environmental monitoring, underwater robot navigation, etc., and has important application and theoretical value in promoting the technical progress in the field of marine environmental protection and ecological monitoring.

[0159] The above description is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for predicting the concentration of an artificial sweetener, Acesulfame-K, in offshore aquaculture seawater based on the behavioral characteristics of juvenile black seabream, Sparus macrocephalus, characterized by, The application relates to a method for predicting the concentration of acesulfame in seawater based on a juvenile fish behavior recognition model and a spatiotemporal correlation model. The application comprises the following steps: Video data of juvenile fish of different acesulfame concentrations is collected and preprocessed to obtain a juvenile fish motion video data set; An initial juvenile fish behavior recognition model is constructed based on an improved YOLOv8 network model, the initial juvenile fish behavior recognition model is trained by using the juvenile fish motion video data set, a trained juvenile fish behavior recognition model is obtained, and a behavior trajectory data set corresponding to the juvenile fish motion video data set is obtained; the improved YOLOv8 network model comprises the following steps: replacing a C2f module in a YOLOv8 backbone network with a C2f-MSADCN module formed by stacking a C2f module and an MSADCN module; integrating a C2f module in a YOLOv8 neck network into an MSSimAM attention module; the MSSimAM attention module comprises a multi-scale convolution layer, a similarity adjustment layer, an attention layer and a feature fusion layer, is used for capturing local and global information of an image by using a multi-scale receptive field; the multi-scale convolution layer is used for extracting features of different scales by using convolution kernels of different sizes; the similarity adjustment layer is used for measuring the similarity between features of different scales; the attention layer is used for generating an attention map according to the similarity measurement, weighting the multi-scale features to highlight key features; and the feature fusion layer is used for fusing the weighted multi-scale features into final outputs of the module; A CIoU loss function in a YOLOv8 head network is replaced with an SEIoU loss function; An initial spatiotemporal correlation network model is constructed based on a GNN and a GRU network, the initial spatiotemporal correlation network model is trained by using the behavior trajectory data set, the acesulfame concentration is predicted by analyzing the behavior trajectory of the juvenile fish, and a trained spatiotemporal correlation network model is obtained; the initial spatiotemporal correlation network model is constructed based on a GNN and a GRU network, and comprises the following steps: The initial spatiotemporal correlation network model comprises a GNN module, a multi-layer GRU module and a full connection module; The GNN module is used for inputting position features of the juvenile fish at different time points to obtain spatial features, and is used for capturing spatial correlation between the juvenile fish; The multi-layer GRU module comprises a plurality of serially connected GRU network models, the hidden states of the GRU network models are calculated according to the spatial features at each time point, and the multi-layer GRU module is used for capturing the time correlation of the behavior of the juvenile fish; The full connection module is used for predicting the acesulfame concentration according to the hidden states output by the multi-layer GRU module; 2. The method for predicting the concentration of artificial sweetener acesulfame-K in offshore farming seawater based on the behavioral characteristics of juvenile black seabream according to claim 1, characterized in that, Actual motion video of juvenile fish in offshore aquaculture seawater is obtained, and the acesulfame concentration of the offshore aquaculture seawater is predicted according to the actual motion video by using the juvenile fish behavior recognition model and the spatiotemporal correlation model. The juvenile fish motion video under different acesulfame concentrations is collected and preprocessed, and the method comprises the following steps: The juvenile fish motion video under different acesulfame concentrations is collected based on a preset shooting angle, resolution and frame rate to obtain original video data; The original video data is segmented into segments with equal time lengths, and each video is labeled with an acesulfame concentration to obtain a juvenile fish motion video data set; The video dataset is expanded by using a data enhancement technology to obtain an expanded dataset; Each video in the expanded dataset is labeled with juvenile snapper, and the labeled video is converted to obtain a preprocessed motion video dataset; The preprocessed motion video dataset is divided into a training set, a validation set and a test set.

3. The method for predicting the concentration of artificial sweetener acesulfame-K in offshore farming seawater based on the behavioral characteristics of juvenile black seabream according to claim 1, characterized in that, The MSADCN module introduces a multi-scale and attention mechanism, and the convolution operation of the module is represented by a formula: where y (i, j) is the feature value at position (i, j) in the output feature map, p p S represents the number of multi-scales, is the learnable weight at scale s, s for adjusting the feature contribution of different scales, K is the number of sampling points of the convolution kernel, is the convolution weight at the s th sampling position at the k th scale, is the offset reference of the k th sampling position, is the learnable offset, is the modulation coefficient of the k th sampling point for adjusting the feature importance of the position.​​ 4. The method for predicting the concentration of artificial sweetener acesulfame-K in offshore farming seawater based on the behavioral characteristics of juvenile black seabream according to claim 1, characterized in that, The SEIoU loss function is represented by a formula: wherein, represents the intersection over union loss, which measures the difference in overlap between the predicted and the ground truth bounding boxes, represents the square of the Euclidean distance between the center points of the predicted and the ground truth bounding boxes, w and w gt are the width of the predicted and the ground truth bounding boxes, respectively, h and h gt are the height of the predicted and the ground truth bounding boxes, respectively, w c and h c are the width and the height of the smallest enclosing rectangle that can enclose the predicted and the ground truth bounding boxes, respectively.

5. The method for predicting the concentration of artificial sweetener acesulfame-K in offshore farming seawater based on the behavioral characteristics of juvenile black seabream according to claim 2, characterized in that, The initial juvenile fish behavior recognition model is trained using the juvenile fish motion video dataset to obtain a trained juvenile fish behavior recognition model, including: The initial juvenile fish behavior recognition model is trained using the training set and the validation set of the juvenile fish motion video dataset to obtain a trained juvenile fish behavior recognition model; The trained juvenile fish behavior recognition model is tested using the test set of the juvenile fish motion video dataset. When the prediction result of the model reaches the training expectation, a trained juvenile fish behavior recognition model is obtained, and the coordinate time series of all juvenile snappers in the training set, the validation set and the test set are extracted.

6. The method for predicting the concentration of artificial sweetener acesulfame-K in offshore farming seawater based on the behavioral characteristics of juvenile black seabream according to claim 2, characterized in that, The initial spatiotemporal correlation network model is trained using the behavior trajectory dataset, and the behavior trajectory of the juvenile snapper is analyzed to predict the acesulfame-K concentration to obtain a trained spatiotemporal correlation network model, including: The coordinate time series of the juvenile snapper in the training set and the validation set and the corresponding acesulfame-K concentration are used as labels to supervise the training of the initial spatiotemporal correlation network model; The coordinate time series of the juvenile snapper in the test set and the corresponding acesulfame-K concentration are used as labels to test the trained spatiotemporal correlation network model, and when the accuracy of the prediction result reaches the expectation, a trained spatiotemporal correlation network model is obtained.

7. A system for predicting the concentration of an artificial sweetener, Acesulfame-K, in offshore seawater aquaculture based on the behavioral characteristics of juvenile black seabream, S. chrysurus, characterized by, It includes: A data acquisition module is configured to collect motion video data of juvenile snappers in different acesulfame-K concentration environments and preprocess the motion video data to obtain a juvenile snapper motion video dataset. The behavior recognition model construction module is configured to construct an initial juvenile fish behavior recognition model based on an improved YOLOv8 network model, train the initial juvenile fish behavior recognition model by using a juvenile fish motion video dataset, obtain a trained juvenile fish behavior recognition model, and obtain a behavior trajectory dataset corresponding to the juvenile fish motion video dataset; the improved YOLOv8 network model comprises: replacing a C2f module in a YOLOv8 backbone network with a C2f-MSADCN module stacked by a C2f module and an MSADCN module; integrating a C2f module in a YOLOv8 neck network into an MSSimAM attention module; the MSSimAM attention module comprises a multi-scale convolution layer, a similarity adjustment layer, an attention layer, and a feature fusion layer, and is configured to capture local and global information of an image by using a multi-scale receptive field; the multi-scale convolution layer is configured to extract features of different scales by using convolution kernels of different sizes; the similarity adjustment layer is configured to measure the similarity between features of different scales; the attention layer is configured to generate an attention map according to the similarity measurement, and weight the multi-scale features to highlight key features; and the feature fusion layer is configured to fuse the weighted multi-scale features into final outputs of the module; replacing a CIoU loss function in a YOLOv8 head network with an SEIoU loss function; The spatio-temporal correlation network model construction module is configured to construct an initial spatio-temporal correlation network model based on a GNN and a GRU network, train the initial spatio-temporal correlation network model by using the behavior trajectory dataset, analyze the behavior trajectory of the juvenile fish, and predict the acesulfame-K concentration, and obtain a trained spatio-temporal correlation network model; the initial spatio-temporal correlation network model comprises a GNN module, a multi-layer GRU module, and a full connection module; the GNN module is configured to input position features of the juvenile fish at different times, obtain spatial features, and capture spatial correlations between the juvenile fish; the multi-layer GRU module comprises a plurality of serially connected GRU network models, and is configured to calculate hidden states of the GRU network models according to the spatial features at each time, capture temporal correlations of behaviors of the juvenile fish, and predict the acesulfame-K concentration according to the hidden states output by the multi-layer GRU module; The concentration prediction module is configured to obtain actual motion videos of the juvenile fish in offshore aquaculture seawater, and predict the acesulfame-K concentration of the offshore aquaculture seawater by using the juvenile fish behavior recognition model and the spatio-temporal correlation model according to the actual motion videos.

Images