A large model-based method and system for evaluating the growth stage feeding amount of large yellow croaker

Abstract

The application discloses a large model-based evaluation method and system for growth stage feeding amount of large yellow croaker, and belongs to the field of aquaculture. The method comprises the following steps: simultaneously collecting underwater images of a fish pond in multiple shooting directions, identifying and segmenting fish body images corresponding to each entity fish in each underwater image; determining the distance between the fish body image corresponding entity fish and the camera collecting the fish body image, and extracting the external features of the fish body image; selecting one fish body image from each shooting direction to form a fish body image associated combination, and screening a second fish body image associated combination from all possible fish body image associated combinations according to the distance and the external features of the fish body image; constructing three-dimensional point cloud data of the entity fish according to the second fish body image associated combination, fitting the three-dimensional body shape parameters of the entity fish through the three-dimensional point cloud data, and determining the final feeding amount of the fish group according to the three-dimensional body shape parameters of each entity fish. The application can solve the problem of low precision of real-time feeding amount of the fish group.

Description

Technical Field

[0001] This application relates to the field of aquaculture, and in particular to a method and system for evaluating the feeding amount of large yellow croaker during its growth stage based on a large model. Background Technology

[0002] In the field of aquaculture, the scientific feeding of large yellow croaker is a crucial factor affecting farming efficiency and fish health, therefore precise feeding of large yellow croaker is necessary.

[0003] Traditional large yellow croaker farming methods rely heavily on visual assessment of fish activity, historical feeding records, or water quality parameters to roughly estimate feeding amounts. However, these methods suffer from insufficient accuracy, significant susceptibility to human subjectivity, and inaccurate identification of fish growth stages. To address this issue, existing technologies utilize image recognition to periodically capture live large yellow croaker samples. These samples are then placed in a laboratory environment to accurately identify the current growth stage of the fish, thus enabling precise feeding. However, this method requires capturing live large yellow croaker samples and a large sample size to accurately estimate the overall feeding amount for the entire fish population, making it unsuitable for practical large yellow croaker farming.

[0004] Therefore, how to accurately estimate the overall growth stage of a fish population in real time using image recognition technology without capturing live samples of large yellow croaker, thereby improving the accuracy of feeding, is a technical problem that needs to be solved. Summary of the Invention

[0005] This application provides a method and system for evaluating the feeding amount of large yellow croaker during its growth stages based on a large model. This method can solve the problem in the prior art that the overall growth stage of the fish population cannot be accurately estimated in real time through image recognition technology, resulting in low accuracy of the feeding amount.

[0006] One embodiment of this application provides a method for evaluating the feeding amount of large yellow croaker during its growth stage based on a large model, including:

[0007] Simultaneously, underwater images of the fishpond are acquired from multiple shooting positions, and the fish body images corresponding to each physical fish in each of the underwater images are identified and segmented; the cameras corresponding to each of the multiple shooting positions are respectively placed at the vertices of a preset regular polygon; the fish body image includes: the pixel value and depth value of each pixel in the fish body image;

[0008] Based on the depth value, the distance between the actual fish corresponding to the current fish image and the camera that acquired the fish image is determined, and the external features of the fish corresponding to the fish image are extracted by a preset convolutional neural network based on the pixel value of each pixel in the fish image.

[0009] From each shooting position, a fish image is randomly selected to form a fish image association combination, and a first fish image association combination is selected from all the fish image association combinations; wherein, the distance between each fish image in the first fish image association combination is less than the diameter of the circumcircle of the regular polygon;

[0010] Select a second fish image association combination from all the first fish image association combinations; wherein the first Euclidean distance between the fish images corresponding to the fish external features in the second fish image association combination is less than a preset value;

[0011] Based on the association and combination of the second fish body image, construct the three-dimensional point cloud data corresponding to the actual fish, and obtain the three-dimensional body shape parameters corresponding to the actual fish by fitting the three-dimensional point cloud data;

[0012] Based on the three-dimensional body shape parameters, the probability distribution of the fish population at each growth stage is determined using a pre-trained classification model. Then, based on the probability distribution and the first feeding amount for the fish population at each growth stage, the final feeding amount is determined.

[0013] Compared with existing technologies, the above embodiments have the following beneficial effects: By setting cameras at different shooting positions underwater in the fishpond, underwater images collected from different positions can be obtained. Without capturing fish samples, the overall visual information of the fish school can be obtained, thereby improving the real-time performance of fish growth stage estimation. Furthermore, underwater images collected from different positions provide body shape information of the fish in different locations, effectively solving the problem of difficulty in fish identification caused by underwater fish occlusion, and providing a more comprehensive basis for accurately extracting fish external features. Furthermore, a target recognition model is used to accurately identify and segment each fish in the image, and the distance between the actual fish and the camera is calculated by combining depth information, ensuring… The spatial localization accuracy of fish images is improved; simultaneously, fish external features are extracted through convolutional neural networks, and the spatial localization of fish images is combined with fish external features to effectively improve the matching accuracy of fish images of the same entity across different angles. This provides accurate input data for the subsequent construction of 3D point cloud data of the entity fish, while also avoiding the problem of inaccurate fish size estimation due to mutual occlusion. Finally, based on the accurately obtained 3D point cloud data of each entity fish, the 3D body shape parameters of each entity fish are precisely fitted, thereby providing a large amount of effective sample data for estimating the growth stage of the fish population, improving the accuracy of the estimation of the growth stage of the fish population, and thus improving the accuracy of the final fish feeding amount.

[0014] Furthermore, determining the probability distribution of the fish population at each growth stage based on the three-dimensional body shape parameters using a pre-trained classification model includes:

[0015] The first top-down image of the fishpond was captured during the last feeding. Based on the first top-down image, several images of fish schools clustered together were obtained by using a threshold segmentation method.

[0016] Based on the images of fish gathering and the feeding coordinates of the last feeding, the feeding activity density of the fish is estimated.

[0017] The three-dimensional body size parameters of each of the described fish and the feeding activity density of the fish group are input into the classification model to determine the probability distribution of the fish group at each growth stage.

[0018] Compared with existing technologies, the above embodiments have the following beneficial effects: Based on the original three-dimensional body shape parameters, auxiliary features of fish feeding activity density are introduced. By obtaining the top view image at the time of the last feeding, the gathering point of the fish at the time of feeding is segmented, and combined with the feeding coordinate point information, the overall feeding activity of the fish can be reflected, thereby improving the adaptability of the model under different environmental, seasonal and lighting conditions, thus providing more reliable basic data for subsequent feeding amount calculation.

[0019] Further, the step of segmenting the first top-view image into several fish-gathering images using a threshold segmentation method includes:

[0020] The first top view is converted into a grayscale image, and the grayscale image is filtered and denoised to obtain the third top view;

[0021] The threshold segmentation method is used to convert the third top view into a binary image, and connected component analysis is performed on the binary image to obtain several connected components.

[0022] The contour information of each connected component is extracted, and the corresponding fish aggregation image is obtained by segmenting from the third top view based on the contour information.

[0023] Compared with existing technologies, the above embodiments have the following beneficial effects: When feeding fish, the fish will gather on the surface of the pond. The higher the degree of gathering, the lower the transparency of the water. Therefore, when performing image recognition processing, the fish gathering area can be effectively segmented by simply converting the top view image to grayscale, filtering and reducing noise, and then performing threshold segmentation and connected component analysis, thus improving the computational efficiency of real-time monitoring. Furthermore, by performing connected component analysis and extracting the edge contours of each connected component, the boundaries of the fish distribution are clearly distinguishable, thereby accurately obtaining the pixel range of each gathering area and providing accurate input for subsequent calculations of fish density and activity.

[0024] Further, estimating the feeding activity density of the fish school based on the fish aggregation image and the feeding coordinates of the last feeding includes:

[0025] The fish aggregation image includes: a third top view slice corresponding to each of the connected components;

[0026] Calculate the centroid coordinates of each of the connected domains, and map each of the centroid coordinates from the camera coordinate system to the world coordinate system;

[0027] Calculate the average gray value of all pixels in the third top view slice corresponding to each of the connected components, and evaluate the density of fish in the corresponding connected component based on the average gray value;

[0028] The feeding activity density of the fish is estimated based on the density of the fish population corresponding to each connected region and the second Euclidean distance between the centroid coordinates of the connected region and the feeding coordinate point.

[0029] Compared with existing technologies, the above embodiments have the following beneficial effects: By calculating the centroid coordinates of each fish school gathering area and mapping them from the camera coordinate system to the world coordinate system, the precise location of the fish school in physical space is achieved; furthermore, the fish school density is evaluated by combining the average gray value, and the level of the gray value can reflect the strength of the number and activity of the fish school in a local area, which can effectively reduce the deviation in activity calculation caused by the short-term movement or dispersion of the fish school; then, the fish school gathering degree of each area is used as a weight, and the Euclidean distance between the centroid of the area and the feeding coordinate point is weighted for evaluation, which can reflect the fish school's response to the feeding location, thereby more accurately reflecting the feeding enthusiasm and providing a more scientific reference for subsequent growth stage prediction and feeding amount decision-making.

[0030] Further, the step of identifying and segmenting the fish body image corresponding to each entity fish in each of the underwater images includes:

[0031] A second top-down view image currently captured above the fishpond is obtained; the water color features of the fishpond are extracted from the second top-down view image; and the underwater images are color-corrected based on the water color features.

[0032] The color-corrected underwater images are input into the pre-trained target recognition model to obtain the detection box and segmentation mask of each fish in the underwater images.

[0033] By removing noise points in the underwater image corresponding to the detection frame using the segmentation mask, the fish body image corresponding to each physical fish in each underwater image is obtained.

[0034] Compared with existing technologies, the above embodiments have the following beneficial effects: By using the top-down view and extracting water color features through image recognition technology, the underwater image is uniformly adjusted, effectively eliminating color distortion caused by differences in water quality, lighting, and shooting angle, thereby improving the accuracy of target recognition; furthermore, the color-corrected image is input into the pre-trained target recognition model, which can obtain a more stable detection box and segmentation mask, reducing missed detections and false detections; finally, combined with mask removal of noise points, it can ensure that the extracted fish body image has a clear outline and complete details, providing high-quality input data for 3D point cloud reconstruction and body shape parameter fitting, and improving the stability and accuracy of the entire evaluation process.

[0035] Further, the step of constructing three-dimensional point cloud data corresponding to the actual fish based on the association and combination of the second fish body image, and obtaining three-dimensional body shape parameters corresponding to the actual fish by fitting the three-dimensional point cloud data, includes:

[0036] The three-dimensional body shape parameters include: fish length, fish width, and fish thickness;

[0037] The key point detection model is used to detect key points in each of the fish images in the second fish image association combination; wherein each key point corresponds to a fish body part.

[0038] By using triangulation, the coordinates of key points in each of the fish images in the second fish image association combination are converted into three-dimensional point cloud coordinates to obtain the three-dimensional point cloud data;

[0039] The pose of the 3D point cloud data is adjusted to a unified coordinate system using principal component analysis.

[0040] Based on the three-dimensional point cloud data, a preset parametric three-dimensional model of the fish body is fitted using the least squares method, and the three-dimensional body shape parameters are extracted based on the parametric three-dimensional model of the fish body.

[0041] Compared to existing technologies, the above embodiments have the following advantages: By acquiring fish body feature parts through a keypoint detection model and converting multi-view keypoint coordinates into a 3D point cloud using triangulation, accurate reconstruction of the fish's geometric morphology is achieved. Furthermore, principal component analysis is used to unify the point cloud pose, avoiding model inconsistencies caused by different shooting angles. Finally, by fitting a pre-defined parametric 3D model of the fish body using the least squares method, core parameters such as fish length, width, and thickness can be stably extracted. This method balances geometric accuracy and computational efficiency, enabling the rapid acquisition of stable 3D body shape parameters in dynamic swimming environments, providing high-precision input features for growth stage classification models, and significantly improving the reliability of evaluation results.

[0042] Further, determining the final feeding amount based on the probability distribution and the initial feeding amount of the fish population at each growth stage includes:

[0043] The probability distribution includes the probability that the fish population is in each of the growth stages.

[0044] The second feeding amount is obtained by multiplying each probability by the first feeding amount corresponding to the growth stage and summing the results.

[0045] The current water temperature of the fishpond is collected, and the second feeding amount is adjusted according to the water temperature to obtain the final feeding amount.

[0046] Compared with the prior art, the above embodiments have the following beneficial effects: Based on obtaining the probability distribution of each growth stage of the fish population, the second feeding amount after comprehensive weighting is obtained by multiplying it with the first feeding amount of the corresponding stage and accumulating it, ensuring that the feeding amount matches the actual structure of the fish population; further, by combining the real-time collected water temperature information to correct the feeding amount, the difference in feeding amount caused by water temperature changes can be compensated, preventing overfeeding or underfeeding, so that the final feeding amount is more in line with the actual needs of the fish population.

[0047] Another embodiment of this application provides a feeding amount assessment system for the growth stage of large yellow croaker based on a large model, including: a fish body image segmentation module, a fish body external feature extraction module, a first fish body image association module, a second fish body image association module, a point cloud fitting module, and a feeding amount calculation module.

[0048] The fish image segmentation module is used to simultaneously acquire underwater images of the fishpond from multiple shooting positions, identify and segment the fish images corresponding to each actual fish in each underwater image; the cameras corresponding to each of the multiple shooting positions are respectively placed at the vertices of a preset regular polygon; the fish image includes: the pixel value and depth value of each pixel in the fish image.

[0049] The fish body exterior feature extraction module is used to determine the distance between the actual fish corresponding to the current fish body image and the camera that acquired the fish body image based on the depth value, and to extract the fish body exterior features corresponding to the fish body image based on the pixel value of each pixel in the fish body image through a preset convolutional neural network.

[0050] The first fish image association module is used to arbitrarily select a fish image from each shooting position to form a fish image association combination, and to filter the first fish image association combination from all the fish image association combinations; wherein, the distance between each fish image in the first fish image association combination is less than the diameter of the circumcircle of the regular polygon.

[0051] The second fish image association module is used to filter a second fish image association combination from all the first fish image association combinations; wherein, the first Euclidean distance between the fish images corresponding to the fish external features in the second fish image association combination is less than a preset value;

[0052] The point cloud fitting module is used to construct three-dimensional point cloud data corresponding to the actual fish based on the association and combination of the second fish body image, and to obtain three-dimensional body shape parameters corresponding to the actual fish based on the fitting of the three-dimensional point cloud data.

[0053] The feeding amount calculation module is used to determine the probability distribution of the fish population at each growth stage based on the three-dimensional body shape parameters and a pre-trained classification model, and to determine the final feeding amount based on the probability distribution and the first feeding amount of the fish population at each growth stage.

[0054] Furthermore, the feeding amount calculation module includes: a threshold segmentation unit, a fish feeding activity density estimation unit, and a classification unit; the feeding amount calculation module is used to determine the probability distribution of the fish population at each growth stage based on the three-dimensional body shape parameters and a pre-trained classification model, including:

[0055] The threshold segmentation unit is used to acquire a first top-down view image taken above the fishpond during the last feeding, and to segment and obtain several fish aggregation images based on the first top-down view image using a threshold segmentation method.

[0056] The fish feeding activity density estimation unit is used to estimate the fish feeding activity density based on the images of the fish agglomeration and the feeding coordinates at the time of the last feeding.

[0057] The classification unit is used to input the three-dimensional body shape parameters of each of the fish entities and the feeding activity density of the fish group into the classification model to determine the probability distribution of the fish group at each growth stage.

[0058] Further, the point cloud fitting module includes: a key point detection unit, a 3D point cloud acquisition unit, a coordinate unification unit, and a 3D model parameter fitting unit; the point cloud fitting module is used to construct 3D point cloud data corresponding to the actual fish based on the association and combination of the second fish body image, and to obtain 3D body shape parameters corresponding to the actual fish based on the 3D point cloud data, including:

[0059] The three-dimensional body shape parameters include: fish length, fish width, and fish thickness;

[0060] The key point detection unit is used to detect key points of each fish image in the second fish image association combination through a key point detection model; wherein each key point corresponds to a fish body part.

[0061] The three-dimensional point cloud acquisition unit is used to convert the coordinates of key points of each fish image in the second fish image association combination into three-dimensional point cloud coordinates through triangulation, and obtain the three-dimensional point cloud data.

[0062] The coordinate unification unit is used to adjust the pose of the three-dimensional point cloud data to a unified coordinate system using principal component analysis.

[0063] The three-dimensional model parameter fitting unit is used to fit a preset fish body parameterized three-dimensional model based on the three-dimensional point cloud data using the least squares method, and to extract the three-dimensional body shape parameters based on the fish body parameterized three-dimensional model. Attached Figure Description

[0064] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0065] Figure 1 This is a flowchart illustrating a method for evaluating the feeding amount of large yellow croaker during its growth stage based on a large model, provided in some embodiments of this application.

[0066] Figure 2 This is a schematic diagram of the structure of a feeding amount assessment system for large yellow croaker growth stages based on a large model, provided in some embodiments of this application. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0068] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0069] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0070] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0071] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0072] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0073] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0074] Traditional large yellow croaker farming methods rely heavily on visual assessment of fish activity, historical feeding records, or water quality parameters to roughly estimate feeding amounts. However, these methods suffer from insufficient accuracy, significant susceptibility to human subjectivity, and inaccurate identification of fish growth stages. To address this issue, existing technologies utilize image recognition to periodically capture live large yellow croaker samples. These samples are then placed in a laboratory environment to accurately identify the current growth stage of the fish, thus enabling precise feeding. However, this method requires capturing live large yellow croaker samples and a large sample size to accurately estimate the overall feeding amount for the entire fish population, making it unsuitable for practical large yellow croaker farming.

[0075] Please refer to Figure 1 To address the issue of low accuracy in real-time feeding of fish in existing technologies, this application provides a method for evaluating feeding amounts for large yellow croaker during its growth stage based on a large model, comprising steps S101 to S106, specifically:

[0076] S101: Simultaneously acquire underwater images of the fishpond from multiple shooting positions, identify and segment the fish body images corresponding to each entity fish in each of the underwater images; the cameras corresponding to each of the multiple shooting positions are respectively placed at the vertices of a preset regular polygon; the fish body image includes: the pixel value and depth value of each pixel in the fish body image.

[0077] Furthermore, in some embodiments of this application, the step of identifying and segmenting the fish body image corresponding to each entity fish in each of the underwater images includes:

[0078] A second top-down view image currently captured above the fishpond is obtained; the water color features of the fishpond are extracted from the second top-down view image; and the underwater images are color-corrected based on the water color features.

[0079] The color-corrected underwater images are input into the pre-trained target recognition model to obtain the detection box and segmentation mask of each fish in the underwater images.

[0080] By removing noise points in the underwater image corresponding to the detection frame using the segmentation mask, the fish body image corresponding to each physical fish in each underwater image is obtained.

[0081] Preferably, in some embodiments of this application, the step of extracting the water color features of the fishpond from the second top-view image includes: calculating the average value of each channel (R / G / B) in the second top-view image to obtain water color reference values ​​(R_w, G_w, B_w), and using the water color reference values ​​as water color features.

[0082] Preferably, in some embodiments of this application, the step of color correction of each underwater image based on the water color characteristics includes: taking neutral gray as the target, obtaining the gain coefficient of each channel based on the water color reference value; multiplying each pixel in the underwater image by the corresponding channel gain coefficient to eliminate the overall color shift; additionally increasing the red channel intensity of each pixel in each underwater image to address red light attenuation; and finally using Retinex or white balance algorithms to enhance the local contrast and color naturalness of each underwater image.

[0083] Preferably, in some embodiments of this application, the target recognition model employs target recognition models such as Mask R-CNN, YOLOv8-seg, or Detectron2. This application does not limit the model used. Based on the input underwater image, the target recognition model outputs a detection box and segmentation mask for each physical fish in the underwater image.

[0084] As can be seen from the above embodiments, this application utilizes the top-down view and extracts water color features through image recognition technology to uniformly adjust underwater images, effectively eliminating color distortion caused by differences in water quality, lighting, and shooting angle, thereby improving the accuracy of target recognition. Furthermore, the color-corrected image is input into the pre-trained target recognition model, which can obtain more stable detection boxes and segmentation masks, reducing missed detections and false detections. Finally, combined with mask removal of noise points, it can ensure that the extracted fish body image has clear contours and complete details, providing high-quality input data for 3D point cloud reconstruction and body shape parameter fitting, and improving the stability and accuracy of the entire evaluation process.

[0085] S102: Based on the depth value, determine the distance between the actual fish corresponding to the current fish image and the camera that acquired the fish image, and extract the fish body appearance features corresponding to the fish image through a preset convolutional neural network based on the pixel value of each pixel in the fish image.

[0086] Preferably, in some embodiments of this application, determining the distance between the actual fish corresponding to the current fish image and the camera that acquired the fish image based on the depth value includes: calculating the depth value of each pixel in the fish image, adding the depth values ​​of all pixels in the fish image, averaging them to obtain the average depth value between the actual fish corresponding to the fish image and the camera, and calculating the distance between the actual fish and the camera that acquired the fish image based on the average depth value. The method of converting the depth value into a specific distance is an existing method, and this application embodiment does not further limit or explain this process.

[0087] Preferably, in some embodiments of this application, the step of extracting the fish body exterior features corresponding to the fish body image based on the pixel value of each pixel in the fish body image through a preset convolutional neural network includes: scaling the fish body image to a resolution of 256×256 and normalizing the pixel values ​​of the fish body image so that the pixel values ​​of each channel are distributed within the standard range during model training; performing multi-layer convolution operations on the normalized fish body image, with each convolutional kernel automatically learning to extract local features of different scales and directions. In shallow convolutions, the network captures low-level features such as the fish body outline, color patterns, and fin edges. In deeper convolutions, the network learns more complex shape combinations and texture patterns. Between convolutional layers, the spatial resolution of the feature map is gradually reduced through max pooling to retain significant information and filter local noise, thereby enabling the network to integrate exterior information within a larger receptive field. The multi-channel feature map output by the last convolutional layer is then subjected to global average pooling to obtain the fish body exterior features.

[0088] S103: Randomly select a fish image from each shooting position to form a fish image association combination, and filter the first fish image association combination from all the fish image association combinations; wherein, the distance between each fish image in the first fish image association combination is less than the diameter of the circumcircle of the regular polygon.

[0089] Preferably, in some embodiments of this application, the step of selecting a first fish image association combination from all the fish image association combinations includes: since multiple cameras corresponding to multiple shooting positions are distributed on the vertices of a regular polygon, the maximum distance from any vertex in the regular polygon to its interior point will not exceed the diameter of its circumcircle. When the distance between the actual fish corresponding to a certain fish image and the camera exceeds this diameter, it means that at least one camera at a shooting position cannot capture the actual fish, and a valid fish image association combination cannot be formed. By selecting a first fish image association combination from all fish image association combinations, preliminary data screening is performed, improving the efficiency of subsequent calculation of fish appearance feature similarity.

[0090] S104: Select a second fish image association combination from all the first fish image association combinations; wherein, the first Euclidean distance between the fish images corresponding to the fish external features in the second fish image association combination is less than a preset value.

[0091] Preferably, in some embodiments of this application, the step of selecting a second fish image association combination from all the first fish image association combinations includes: for each first fish image association combination, calculating the first Euclidean distance between the fish body external features of each corresponding fish image; if there is a first fish image association combination where each first Euclidean distance is less than a preset value, it means that the fish images in the first fish image association combination all point to the same entity fish, and the first fish image association combination is used as the second fish image association combination.

[0092] S105: Based on the association and combination of the second fish body image, construct the three-dimensional point cloud data corresponding to the actual fish, and obtain the three-dimensional body shape parameters corresponding to the actual fish by fitting the three-dimensional point cloud data.

[0093] Furthermore, in some embodiments of this application, the step of constructing three-dimensional point cloud data corresponding to the actual fish based on the association and combination of the second fish body image, and obtaining three-dimensional body shape parameters corresponding to the actual fish by fitting the three-dimensional point cloud data, includes:

[0094] The three-dimensional body shape parameters include: fish length, fish width, and fish thickness;

[0095] The key point detection model is used to detect key points in each of the fish images in the second fish image association combination; wherein each key point corresponds to a fish body part.

[0096] By using triangulation, the coordinates of key points in each of the fish images in the second fish image association combination are converted into three-dimensional point cloud coordinates to obtain the three-dimensional point cloud data;

[0097] The pose of the 3D point cloud data is adjusted to a unified coordinate system using principal component analysis.

[0098] Based on the three-dimensional point cloud data, a preset parametric three-dimensional model of the fish body is fitted using the least squares method, and the three-dimensional body shape parameters are extracted based on the parametric three-dimensional model of the fish body.

[0099] Preferably, in some embodiments of this application, the step of detecting key points of each fish image in the second fish image association combination using a key point detection model includes: wherein the key point detection model adopts Lite-HRNet to meet the needs of real-time underwater detection, and the key point detection model extracts feature information of each fish image through a convolutional neural network, including local features of the fish outline, fins, head and tail, etc.; further, based on the extracted feature information, the key point detection model predicts the position coordinates of several key points in each fish image, such as typical positions like the head, tail, dorsal fin, and pelvic fin.

[0100] Preferably, in some embodiments of this application, the step of converting the coordinates of key points of each fish image in the second fish image association combination into three-dimensional point cloud coordinates through triangulation to obtain the three-dimensional point cloud data includes: wherein the coordinates corresponding to the key points are two-dimensional coordinates; acquiring the intrinsic and extrinsic parameters of the camera used to acquire the fish images, including: focal length, principal point position, camera position and attitude. It should be noted that the above intrinsic and extrinsic parameters are parameter data after the camera calibration is completed; matching the key points based on the corresponding parts of each key point; constructing the projection matrix of the corresponding camera according to the camera intrinsic and extrinsic parameters, and substituting the projection matrix of each camera and the coordinates of each key point into the triangulation algorithm to calculate the coordinates of the key points in three-dimensional space, thereby obtaining the three-dimensional point cloud coordinates.

[0101] Preferably, in some embodiments of this application, adjusting the pose of the three-dimensional point cloud data to a unified coordinate system using principal component analysis includes: calculating the centroid position of the three-dimensional point cloud data and translating all three-dimensional point cloud coordinates so that the centroid is located at the origin; calculating the covariance matrix of the three-dimensional coordinates based on the decentralized point cloud coordinates; performing eigenvalue decomposition on the covariance matrix to obtain the principal component directions; and rotating the three-dimensional point cloud data along the principal component directions so that the principal axis of the fish body is aligned with the X / Y / Z axes of the unified coordinate system, thereby obtaining the pose-adjusted three-dimensional point cloud data.

[0102] Preferably, in some embodiments of this application, the step of fitting a preset parametric 3D model of a fish body using the least squares method based on the 3D point cloud data, and extracting the 3D body shape parameters based on the parametric 3D model of the fish body, includes: selecting a preset parametric 3D model of a fish body (such as an ellipsoid, a multi-segment cross-sectional surface, or a statistical shape model), whose parameters correspond to the body shape features such as the length, width, and thickness of the fish body; defining the sum of squared distances from the 3D point cloud surface to the model surface as the fitting error; adjusting the model parameters using the least squares optimization algorithm to minimize the fitting error defined above; and directly extracting the 3D body shape parameters such as the length, height, and thickness of the fish body from the fitted model.

[0103] As can be seen from the above embodiments, this application obtains fish body feature parts through a key point detection model and uses triangulation to convert the coordinates of key points from multiple perspectives into a three-dimensional point cloud, achieving accurate reconstruction of the fish's geometric shape. Furthermore, principal component analysis is used to unify the point cloud pose, avoiding inconsistencies caused by different shooting angles. Finally, the least squares method is used to fit a pre-set parametric three-dimensional model of the fish, enabling stable extraction of core parameters such as length, width, and thickness. This method balances geometric accuracy and computational efficiency, quickly obtaining stable three-dimensional body shape parameters in the dynamic swimming environment of fish schools, providing high-precision input features for growth stage classification models, and significantly improving the reliability of evaluation results.

[0104] S106: Based on the three-dimensional body shape parameters, the probability distribution of the fish population at each growth stage is determined using the pre-trained classification model, and the final feeding amount is determined based on the probability distribution and the first feeding amount of the fish population at each growth stage.

[0105] Furthermore, in some embodiments of this application, determining the probability distribution of the fish population at each growth stage based on the three-dimensional body shape parameters using a pre-trained classification model includes:

[0106] The first top-down image of the fishpond was captured during the last feeding. Based on the first top-down image, several images of fish schools clustered together were obtained by using a threshold segmentation method.

[0107] Based on the images of fish gathering and the feeding coordinates of the last feeding, the feeding activity density of the fish is estimated.

[0108] The three-dimensional body size parameters of each of the described fish and the feeding activity density of the fish group are input into the classification model to determine the probability distribution of the fish group at each growth stage.

[0109] Preferably, in some embodiments of this application, the method for determining the feeding activity density of fish schools is as follows: each fish school aggregation image corresponds to a region; based on the segmented fish school aggregation images, the number of fish in the region corresponding to each fish school aggregation image is determined by image recognition technology; further, regions closer to the feeding point are given higher weights, and the number of fish in each fish school aggregation image within the region is weighted and summed to obtain the feeding activity density of fish schools.

[0110] Preferably, in some embodiments of this application, the classification model is based on collecting a large amount of target fish data labeled with known growth stages. The three-dimensional body shape parameters of each fish are obtained, and the feeding activity density of the fish population at the corresponding time period is recorded simultaneously. The three-dimensional body shape parameters of all fish are averaged, and the averaged three-dimensional body shape parameters and the feeding activity density of the fish population are further normalized and then feature-stitched to obtain sample data. A multilayer perceptron is used as the basic classification model, and the multilayer perceptron is trained based on the obtained sample data to obtain a pre-trained classification model.

[0111] As can be seen from the above embodiments, this application introduces auxiliary features of fish feeding activity density on the basis of the original three-dimensional body shape parameters. By obtaining the top view image at the time of the last feeding, the gathering point of the fish at the time of feeding is segmented, and combined with the feeding coordinate point information, the overall feeding activity of the fish can be reflected, thereby improving the adaptability of the model under different environmental, seasonal and lighting conditions, thus providing more reliable basic data for subsequent feeding amount calculation.

[0112] Furthermore, in some embodiments of this application, the step of segmenting and obtaining several fish-gathering images based on the first top-view image using a threshold segmentation method includes:

[0113] The first top view is converted into a grayscale image, and the grayscale image is filtered and denoised to obtain the third top view;

[0114] The threshold segmentation method is used to convert the third top view into a binary image, and connected component analysis is performed on the binary image to obtain several connected components.

[0115] The contour information of each connected component is extracted, and the corresponding fish aggregation image is obtained by segmenting from the third top view based on the contour information.

[0116] Preferably, in some embodiments of this application, by converting the first top view into a grayscale image, the amount of computation is reduced and color interference is eliminated. At the same time, based on the shooting conditions when the first top view was acquired, the grayscale image is filtered and noise-reduced to improve the subsequent segmentation effect.

[0117] Preferably, in some embodiments of this application, the threshold segmentation method is used to convert the third top view into a binary image, and connected component analysis is performed on the binary image to obtain several connected components, including: areas where fish are densely clustered, which are usually darker in the top view because the fish absorb light or overlap and occlude, while sparse areas are brighter. Therefore, a global threshold or adaptive threshold is used to distinguish the fish from the background based on pixel brightness to obtain a binary image, where the fish pixel is 1 and the water background pixel is 0; further, connected component analysis is performed on the binary image to obtain several connected components, and contour information of each connected component is obtained through contour detection.

[0118] When feeding the fish, they gather at the surface of the pond. The higher the concentration, the lower the water transparency. Therefore, during image recognition processing, the fish gathering areas can be effectively segmented simply by converting the top-view image to grayscale, filtering and reducing noise, and then performing threshold segmentation and connected component analysis. This improves the computational efficiency of real-time monitoring. Furthermore, by analyzing connected components and extracting the edge contours of each connected component, the boundaries of the fish distribution become clearly discernible. This allows for the accurate acquisition of the pixel range of each gathering area, providing precise input for subsequent calculations of fish density and activity.

[0119] Furthermore, in some embodiments of this application, estimating the feeding activity density of the fish school based on the fish aggregation image and the feeding coordinates at the time of the last feeding includes:

[0120] The fish aggregation image includes: a third top view slice corresponding to each of the connected components;

[0121] Calculate the centroid coordinates of each of the connected domains, and map each of the centroid coordinates from the camera coordinate system to the world coordinate system;

[0122] Calculate the average gray value of all pixels in the third top view slice corresponding to each of the connected components, and evaluate the density of fish in the corresponding connected component based on the average gray value;

[0123] The feeding activity density of the fish is estimated based on the density of the fish population corresponding to each connected region and the second Euclidean distance between the centroid coordinates of the connected region and the feeding coordinate point.

[0124] Preferably, in some embodiments of this application, the step of evaluating the density of fish in the corresponding connected domain based on the average gray value includes: for average gray values ​​less than 50, it is a high density; for average gray values ​​greater than or equal to 50 and less than 100, it is a medium density; and for average gray values ​​greater than or equal to 100, it is a low density.

[0125] Preferably, in some embodiments of this application, the step of determining the fish density corresponding to each connected component and the second Euclidean distance between the centroid coordinates of the connected component and the feeding coordinate point includes: setting corresponding weights for high density, medium density, and low density, wherein a higher density indicates a higher fish activity level, and therefore a larger weight value; multiplying the reciprocal of the second Euclidean distance by the weight of the fish density of the corresponding connected component and then adding them together to obtain the fish feeding activity density.

[0126] As can be seen from the above embodiments, this application achieves precise positioning of the fish schools in physical space by calculating the centroid coordinates of each fish school gathering area and mapping them from the camera coordinate system to the world coordinate system. Furthermore, the fish school density is assessed by combining the average gray value. The level of the gray value can reflect the strength of the fish school quantity and activity in a local area, which can effectively reduce the deviation in activity calculation caused by short-term movement or dispersion of the fish school. Then, the fish school gathering degree of each area is used as a weight, and the Euclidean distance between the centroid of the area and the feeding coordinate point is weighted for evaluation. This can reflect the fish school's response to the feeding location, thereby more accurately reflecting the feeding enthusiasm and providing a more scientific reference for subsequent growth stage prediction and feeding amount decision-making.

[0127] Furthermore, in some embodiments of this application, determining the final feeding amount based on the probability distribution and the first feeding amount of the fish population at each growth stage includes:

[0128] The probability distribution includes the probability that the fish population is in each of the growth stages.

[0129] The second feeding amount is obtained by multiplying each probability by the first feeding amount corresponding to the growth stage and summing the results.

[0130] The current water temperature of the fishpond is collected, and the second feeding amount is adjusted according to the water temperature to obtain the final feeding amount.

[0131] Preferably, in some embodiments of this application, the sum of the probabilities of the fish population being in each of the growth stages is 1.

[0132] Preferably, in some embodiments of this application, the step of correcting the second feeding amount based on the water temperature to obtain the final feeding amount includes: determining the optimal temperature for the growth of large yellow croaker and calculating the difference between the current water temperature and the optimal temperature; determining a feeding amount correction coefficient based on the difference, which is obtained through experiments; and multiplying the second feeding amount by the feeding amount correction coefficient to obtain the final feeding amount. The feeding amount correction coefficient is obtained through experiments, including: selecting multiple large yellow croakers with normal growth and no disease risks and placing them in a water environment where the water temperature can be regulated; first, recording the average daily feeding amount when the large yellow croakers are at the optimal temperature (preferably 23 degrees Celsius); then adjusting the water temperature every other day by one degree Celsius, thereby recording the average daily feeding amount of the large yellow croakers at other degrees Celsius; and determining the feeding amount correction coefficient corresponding to the difference between the average daily feeding amount at other degrees Celsius and the average daily feeding amount at the optimal temperature.

[0133] As can be seen from the above embodiments, this application, based on obtaining the probability distribution of each growth stage of the fish population, multiplies and sums the first feeding amount for the corresponding stage to obtain a second feeding amount with comprehensive weighting, ensuring that the feeding amount matches the actual structure of the fish population; further, by combining the real-time collected water temperature information to correct the feeding amount, it can compensate for the difference in feeding amount caused by water temperature changes, prevent overfeeding or underfeeding, and make the final feeding amount more in line with the actual needs of the fish population.

[0134] In summary, the feeding amount assessment method for large yellow croaker growth stages based on a large model provided in this application has the following advantages compared to existing technologies: By setting up cameras at different shooting positions underwater in the fishpond to acquire underwater images collected from different positions, the overall visual information of the fish school can be obtained without capturing fish samples, thereby improving the real-time performance of fish growth stage estimation. Moreover, the underwater images collected from different positions provide body shape information of the fish in different positions, effectively solving the problem of fish identification difficulties caused by fish occlusion, and providing a more comprehensive basis for accurately identifying fish external features; furthermore, the target recognition model performs accurate image recognition and segmentation of each fish in the image, combined with depth information... The system calculates the distance between the fish and the camera to ensure the spatial positioning accuracy of the fish images. Simultaneously, it extracts the fish's external features using a convolutional neural network, combining the spatial positioning with these features to effectively improve the matching accuracy of the same fish image across different angles. This provides accurate input data for the subsequent construction of 3D point cloud data for the fish, while also avoiding inaccurate fish size estimation due to mutual occlusion. Finally, based on the accurately obtained 3D point cloud data of each fish, the system precisely fits and obtains the 3D body shape parameters of each fish, thus providing a large amount of effective sample data for estimating the fish's growth stage, improving the accuracy of the growth stage estimation, and ultimately increasing the precision of the final fish feeding amount.

[0135] like Figure 2 As shown, based on the above-mentioned method embodiments, an embodiment of this application provides a feeding amount assessment system for the growth stage of large yellow croaker based on a large model, including: a fish body image segmentation module 201, a fish body external feature extraction module 202, a first fish body image association module 203, a second fish body image association module 204, a point cloud fitting module 205, and a feeding amount calculation module 206.

[0136] Further, in some embodiments of this application, the fish image segmentation module 201 is used to simultaneously acquire underwater images of a fishpond from multiple shooting positions, identify and segment the fish images corresponding to each entity fish in each of the underwater images; the cameras corresponding to each of the multiple shooting positions are respectively placed at the vertices of a preset regular polygon; the fish image includes: the pixel value and depth value of each pixel in the fish image; the fish external feature extraction module 202 is used to determine the distance between the entity fish corresponding to the current fish image and the camera that acquired the fish image based on the depth value, and extract the fish external features corresponding to the fish image through a preset convolutional neural network based on the pixel value of each pixel in the fish image; the first fish image association module 203 is used to arbitrarily select a fish image from each shooting position to form a fish image association combination, and filter the first fish image from all the fish image association combinations. The first fish image association module 204 is used to filter second fish image association combinations from all the first fish image association combinations; wherein the first Euclidean distance between the fish images corresponding to the external features of the fish in the second fish image association combination is less than a preset value; the second fish image association module 205 is used to construct three-dimensional point cloud data corresponding to the actual fish based on the second fish image association combination, and to obtain three-dimensional body shape parameters corresponding to the actual fish based on the three-dimensional point cloud data; the third feeding amount calculation module 206 is used to determine the probability distribution of the fish group in each growth stage based on the three-dimensional body shape parameters and a pre-trained classification model, and to determine the final feeding amount based on the probability distribution and the first feeding amount of the fish group in each growth stage.

[0137] Further, in some embodiments of this application, the feeding amount calculation module 206 includes: a threshold segmentation unit, a fish feeding activity density estimation unit, and a classification unit; the feeding amount calculation module is used to determine the probability distribution of the fish population at each growth stage based on the three-dimensional body shape parameters and a pre-trained classification model, including: the threshold segmentation unit is used to acquire a first top-view image taken above the fishpond during the last feeding, and to segment several fish population images based on the first top-view image using a threshold segmentation method; the fish feeding activity density estimation unit is used to estimate the fish feeding activity density based on the several fish population images and the feeding coordinate points during the last feeding; the classification unit is used to input the three-dimensional body shape parameters of each fish and the fish feeding activity density into the classification model to determine the probability distribution of the fish population at each growth stage.

[0138] Further, in some embodiments of this application, the point cloud fitting module 205 includes: a key point detection unit, a 3D point cloud acquisition unit, a coordinate unification unit, and a 3D model parameter fitting unit; the point cloud fitting module is used to construct 3D point cloud data corresponding to the entity fish based on the second fish body image association combination, and to obtain 3D body shape parameters corresponding to the entity fish based on the 3D point cloud data fitting, including: wherein the 3D body shape parameters include: fish body length, fish body width, and fish body thickness; the key point detection unit is used to detect each key point in the second fish body image association combination through a key point detection model. The key points of the fish body image are described; wherein each key point corresponds to a fish body part; the three-dimensional point cloud acquisition unit is used to convert the coordinates of the key points of each fish body image in the second fish body image association combination into three-dimensional point cloud coordinates through triangulation, thereby obtaining the three-dimensional point cloud data; the coordinate unification unit is used to adjust the pose of the three-dimensional point cloud data to a unified coordinate system through principal component analysis; the three-dimensional model parameter fitting unit is used to fit a preset fish body parameterized three-dimensional model according to the three-dimensional point cloud data through the least squares method, and extract the three-dimensional body shape parameters according to the fish body parameterized three-dimensional model.

[0139] It is understood that the above system embodiments correspond to the method embodiments of this application, and can realize the feeding amount assessment method for large yellow croaker growth stage based on a large model provided by any of the above method embodiments of this application.

[0140] In summary, the feeding amount assessment device for large yellow croaker growth stages provided in this application has the following advantages compared to the prior art: By setting cameras at different shooting positions underwater in the fishpond, underwater images collected from different positions can be obtained without capturing fish samples, thus improving the real-time performance of fish growth stage estimation. Furthermore, the underwater images collected from different positions provide body shape information of the fish in different positions, effectively solving the problem of difficulty in fish identification caused by underwater fish occlusion, and providing a more comprehensive basis for accurately extracting fish external features. Moreover, the target recognition model performs precise image recognition and segmentation of each fish in the image, combined with depth... The system calculates the distance between the fish and the camera to ensure the spatial positioning accuracy of the fish images. Simultaneously, it extracts external features of the fish using a convolutional neural network, combining the spatial positioning with these features to effectively improve the matching accuracy of the same fish image across different angles. This provides accurate input data for the subsequent construction of 3D point cloud data for the fish, while also avoiding inaccurate fish size estimation due to mutual occlusion. Finally, based on the accurately obtained 3D point cloud data of each fish, the system precisely fits and obtains the 3D body shape parameters of each fish, thus providing a large amount of effective sample data for estimating the fish population's growth stage, improving the accuracy of the growth stage estimation, and ultimately increasing the precision of the final fish feeding amount.

[0141] It should be noted that the system embodiments described above are merely illustrative, and some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided in this application, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0142] Based on the above embodiments of the large-model-based method for evaluating the feeding amount of large yellow croaker during its growth stage, another embodiment of this application provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the large-model-based method for evaluating the feeding amount of large yellow croaker during its growth stage according to any embodiment of this application.

[0143] For example, in this embodiment, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete this application. The one or more module units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the terminal device.

[0144] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0145] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0146] Based on the above-described method embodiments, another embodiment of this application provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the feeding amount assessment method for large yellow croaker growth stage based on a large model as described in any of the above-described method embodiments of this application.

[0147] The modules / units integrated in the device / terminal equipment, if implemented as software functional units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

Claims

1. A method for evaluating feeding amounts during the growth stages of large yellow croaker based on a large model, characterized in that, include: Simultaneously, underwater images of the fishpond are captured from multiple shooting positions, and the fish body images corresponding to each physical fish in each of the underwater images are identified and segmented. The cameras corresponding to the multiple shooting positions are respectively placed at the vertices of a preset regular polygon; the fish image includes: the pixel value and depth value of each pixel in the fish image; Based on the depth value, the distance between the actual fish corresponding to the current fish image and the camera that acquired the fish image is determined, and the external features of the fish corresponding to the fish image are extracted by a preset convolutional neural network based on the pixel value of each pixel in the fish image. From each shooting position, a fish image is randomly selected to form a fish image association combination, and a first fish image association combination is selected from all the fish image association combinations; wherein, the distance between each fish image in the first fish image association combination is less than the diameter of the circumcircle of the regular polygon; Select a second fish image association combination from all the first fish image association combinations; wherein the first Euclidean distance between the fish images corresponding to the fish external features in the second fish image association combination is less than a preset value; Based on the association and combination of the second fish body image, construct the three-dimensional point cloud data corresponding to the actual fish, and obtain the three-dimensional body shape parameters corresponding to the actual fish by fitting the three-dimensional point cloud data; Based on the three-dimensional body shape parameters, the probability distribution of the fish population at each growth stage is determined using a pre-trained classification model. Then, based on the probability distribution and the first feeding amount for the fish population at each growth stage, the final feeding amount is determined.

2. The method for evaluating the feeding amount of large yellow croaker during its growth stage based on a large model as described in claim 1, characterized in that, The step of determining the probability distribution of the fish population at each growth stage based on the three-dimensional body shape parameters and a pre-trained classification model includes: The first top-down image of the fishpond was captured during the last feeding. Based on the first top-down image, several images of fish schools clustered together were obtained by using a threshold segmentation method. Based on the images of fish gathering and the feeding coordinates of the last feeding, the feeding activity density of the fish is estimated. The three-dimensional body size parameters of each of the described fish and the feeding activity density of the fish group are input into the classification model to determine the probability distribution of the fish group at each growth stage.

3. The method for evaluating the feeding amount of large yellow croaker during its growth stage based on a large model as described in claim 2, characterized in that, The step of segmenting the first top-view image into several fish-gathering images using a threshold segmentation method includes: The first top view is converted into a grayscale image, and the grayscale image is filtered and denoised to obtain the third top view; The threshold segmentation method is used to convert the third top view into a binary image, and connected component analysis is performed on the binary image to obtain several connected components. The contour information of each connected component is extracted, and the corresponding fish aggregation image is obtained by segmenting from the third top view based on the contour information.

4. The method for evaluating the feeding amount of large yellow croaker during its growth stage based on a large model as described in claim 3, characterized in that, The feeding activity density of the fish is estimated based on the images of the fish gathering and the feeding coordinates of the last feeding. include: The fish aggregation image includes: a third top view slice corresponding to each of the connected components; Calculate the centroid coordinates of each of the connected domains, and map each of the centroid coordinates from the camera coordinate system to the world coordinate system; Calculate the average gray value of all pixels in the third top view slice corresponding to each of the connected components, and evaluate the density of fish in the corresponding connected component based on the average gray value; The feeding activity density of the fish is estimated based on the density of the fish population corresponding to each connected region and the second Euclidean distance between the centroid coordinates of the connected region and the feeding coordinate point.

5. The method for evaluating the feeding amount of large yellow croaker during its growth stage based on a large model as described in claim 1, characterized in that, The process of identifying and segmenting the fish body images corresponding to each entity fish in each of the underwater images includes: A second top-down view image currently captured above the fishpond is obtained; the water color features of the fishpond are extracted from the second top-down view image; and the underwater images are color-corrected based on the water color features. The color-corrected underwater images are input into the pre-trained target recognition model to obtain the detection box and segmentation mask of each fish in the underwater images. By removing noise points in the underwater image corresponding to the detection frame using the segmentation mask, the fish body image corresponding to each physical fish in each underwater image is obtained.

6. The method for evaluating the feeding amount of large yellow croaker during its growth stage based on a large model as described in claim 1, characterized in that, The process involves constructing three-dimensional point cloud data corresponding to the actual fish based on the association and combination of the second fish body image, and then fitting the three-dimensional point cloud data to obtain the three-dimensional body shape parameters corresponding to the actual fish. include: The three-dimensional body shape parameters include: fish length, fish width, and fish thickness; The key point detection model is used to detect key points in each of the fish images in the second fish image association combination; wherein each key point corresponds to a fish body part. By using triangulation, the coordinates of key points in each of the fish images in the second fish image association combination are converted into three-dimensional point cloud coordinates to obtain the three-dimensional point cloud data; The pose of the 3D point cloud data is adjusted to a unified coordinate system using principal component analysis. Based on the three-dimensional point cloud data, a preset parametric three-dimensional model of the fish body is fitted using the least squares method, and the three-dimensional body shape parameters are extracted based on the parametric three-dimensional model of the fish body.

7. A method for evaluating the feeding amount of large yellow croaker during its growth stage based on a large model, as described in any one of claims 1 to 6, characterized in that... The final feeding amount is determined based on the probability distribution and the initial feeding amount of the fish at each growth stage. include: The probability distribution includes the probability that the fish population is in each of the growth stages. The second feeding amount is obtained by multiplying each probability by the first feeding amount corresponding to the growth stage and summing the results. The current water temperature of the fishpond is collected, and the second feeding amount is adjusted according to the water temperature to obtain the final feeding amount.

8. A feeding amount assessment system for large yellow croaker during its growth stages based on a large model, characterized in that, include: The system includes a fish body image segmentation module, a fish body external feature extraction module, a first fish body image association module, a second fish body image association module, a point cloud fitting module, and a feeding amount calculation module. The fish image segmentation module is used to simultaneously acquire underwater images of the fishpond from multiple shooting positions, identify and segment the fish images corresponding to each actual fish in each underwater image; the cameras corresponding to each of the multiple shooting positions are respectively placed at the vertices of a preset regular polygon; the fish image includes: the pixel value and depth value of each pixel in the fish image. The fish body exterior feature extraction module is used to determine the distance between the actual fish corresponding to the current fish body image and the camera that acquired the fish body image based on the depth value, and to extract the fish body exterior features corresponding to the fish body image based on the pixel value of each pixel in the fish body image through a preset convolutional neural network. The first fish image association module is used to arbitrarily select a fish image from each shooting position to form a fish image association combination, and to filter the first fish image association combination from all the fish image association combinations; wherein, the distance between each fish image in the first fish image association combination is less than the diameter of the circumcircle of the regular polygon. The second fish image association module is used to filter a second fish image association combination from all the first fish image association combinations; wherein, the first Euclidean distance between the fish images corresponding to the fish external features in the second fish image association combination is less than a preset value; The point cloud fitting module is used to construct three-dimensional point cloud data corresponding to the actual fish based on the association and combination of the second fish body image, and to obtain three-dimensional body shape parameters corresponding to the actual fish based on the fitting of the three-dimensional point cloud data. The feeding amount calculation module is used to determine the probability distribution of the fish population at each growth stage based on the three-dimensional body shape parameters and a pre-trained classification model, and to determine the final feeding amount based on the probability distribution and the first feeding amount of the fish population at each growth stage.

9. The feeding amount assessment system for large yellow croaker growth stages based on a large model as described in claim 8, characterized in that, The feeding amount calculation module includes: a threshold segmentation unit, a fish feeding activity density estimation unit, and a classification unit; the feeding amount calculation module is used to determine the probability distribution of the fish population at each growth stage based on the three-dimensional body shape parameters and a pre-trained classification model, including: The threshold segmentation unit is used to acquire a first top-down view image taken above the fishpond during the last feeding, and to segment and obtain several fish aggregation images based on the first top-down view image using a threshold segmentation method. The fish feeding activity density estimation unit is used to estimate the fish feeding activity density based on the images of the fish agglomeration and the feeding coordinates at the time of the last feeding. The classification unit is used to input the three-dimensional body shape parameters of each of the fish entities and the feeding activity density of the fish group into the classification model to determine the probability distribution of the fish group at each growth stage.

10. The feeding amount assessment system for large yellow croaker growth stages based on a large model as described in claim 8, characterized in that, The point cloud fitting module includes: a key point detection unit, a 3D point cloud acquisition unit, a coordinate unification unit, and a 3D model parameter fitting unit; the point cloud fitting module is used to construct 3D point cloud data corresponding to the actual fish based on the association and combination of the second fish body image, and to obtain the 3D body shape parameters corresponding to the actual fish based on the 3D point cloud data, including: The three-dimensional body shape parameters include: fish length, fish width, and fish thickness; The key point detection unit is used to detect key points of each fish image in the second fish image association combination through a key point detection model; wherein each key point corresponds to a fish body part. The three-dimensional point cloud acquisition unit is used to convert the coordinates of key points of each fish image in the second fish image association combination into three-dimensional point cloud coordinates through triangulation, and obtain the three-dimensional point cloud data. The coordinate unification unit is used to adjust the pose of the three-dimensional point cloud data to a unified coordinate system using principal component analysis. The three-dimensional model parameter fitting unit is used to fit a preset fish body parameterized three-dimensional model based on the three-dimensional point cloud data using the least squares method, and to extract the three-dimensional body shape parameters based on the fish body parameterized three-dimensional model.

Images