A method for evaluating the growth of cobia based on image analysis

Abstract

The application relates to the technical field of image analysis, and discloses a method for evaluating the growth of Cobia based on image analysis, which comprises the following steps: in an underwater environment, synchronously collecting morphological images of Cobia, environmental spectrum data and data of the environment of a breeding water body; processing the images, dynamically compensating optical distortion at different water depths; extracting morphological characteristic parameters of the Cobia from the images, and establishing individual identity identification; fusing the morphological characteristic parameters and the environmental spectrum data, and outputting a body weight prediction value; predicting a future growth trend and evaluating a health condition; according to the future growth trend, feeding records and the health evaluation result, generating a dynamically adjusted feeding scheme and executing feeding management. The method can significantly improve the accuracy and efficiency of the evaluation of the growth and health condition of fish bodies, improve the overall efficiency of Cobia breeding management, and comprehensively improve the intelligent level of breeding.

Description

Technical Field

[0001] This invention relates to the field of image analysis technology, and in particular to a method for evaluating the growth of cobia based on image analysis. Background Technology

[0002] my country ranks first in the world in aquaculture production. In recent years, the construction of marine ranches in my country has flourished. Cobia, with its rapid growth, delicious taste, and high economic value, has become one of the main fish species farmed in marine cages in southern my country, and is particularly suitable for large-scale aquaculture in deep-sea cages.

[0003] However, traditional assessment methods for evaluating the growth status of cobia in deep-sea cage aquaculture environments primarily rely on manual measurement. This approach is not only inefficient but also causes significant stress to the fish during measurement, affecting their normal growth and health. More importantly, the feasibility of manual measurement is severely limited in the complex environment of the open ocean, making it difficult to meet the actual needs of aquaculture assessment. Against this backdrop, underwater image acquisition technology offers a new approach and direction for solving these problems.

[0004] However, current acquisition technologies face numerous challenges in application. Due to the scattering and absorption of light by water, as well as dynamic blurring that may occur during the shooting process, the quality of the acquired images deteriorates, making it difficult to accurately identify the morphological characteristics of cobia, thus affecting the accuracy and reliability of image-based fish growth assessment.

[0005] Currently, although some image analysis-based aquaculture monitoring technologies exist, most of these technologies are designed for terrestrial animals or shallow-water aquaculture environments, making them difficult to directly apply to the low-light, high-scattering environments of deep-sea cages. Furthermore, existing technologies also have limitations in image processing algorithms, making it difficult to effectively eliminate underwater optical distortions, leading to inaccurate morphological feature extraction and consequently affecting the accuracy of growth assessment.

[0006] Furthermore, existing growth assessment methods often lack dynamic updating and adaptive adjustment mechanisms, failing to adjust assessment strategies based on real-time data, leading to discrepancies between assessment results and actual conditions. Therefore, this invention proposes an image analysis-based growth assessment method for cobia, aiming to solve the above problems and achieve non-contact, high-precision growth assessment. Summary of the Invention

[0007] The purpose of this invention is to propose an image analysis-based method for assessing the growth of cobia, in order to address the problems of insufficient accuracy and poor real-time performance of existing non-contact assessment methods.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: a method for evaluating the growth of cobia based on image analysis, comprising the following steps:

[0009] Step S1: In the underwater environment, morphological images of cobia, environmental spectral data and aquatic environment data of the culture water are collected simultaneously using a multispectral imaging device.

[0010] Step S2: The image is processed using a scattering-absorption separation dehazing algorithm and a feature enhancement method based on the retinal mechanism to dynamically compensate for optical distortion at different water depths.

[0011] Step S3: Extract morphological feature parameters of cobia from the image based on a deep learning model and establish individual identification.

[0012] Step S4: Construct a deep neural network prediction model, integrate morphological feature parameters and environmental spectral data, and output the predicted weight value;

[0013] Step S5: By analyzing the synergistic change patterns between historical growth curves and current morphological characteristic parameters, and combining this with environmental parameter compensation mechanisms, predict future growth trends and assess health status.

[0014] Step S6: Based on future growth trends, feeding records, and health assessment results, generate a dynamically adjusted feeding plan and implement feeding management.

[0015] Furthermore, step S1 also includes the following sub-steps:

[0016] S1-1, acquire morphological images of cobia using a multispectral imaging device. The morphological images of cobia include visible light images and near-infrared images. The visible light images reflect the markings and color features of the cobia's body surface, and the near-infrared images can penetrate water to obtain the outline information of the cobia.

[0017] S1-2, using an integrated environmental sensor in a multispectral imaging device to collect environmental spectral data of the water body and environmental data of the aquaculture water body. The environmental spectral data includes spectral irradiance data, optical characteristic parameters of the water body, water quality-related optical indicators and environmental background light information. The aquaculture water body environmental data includes water temperature, dissolved oxygen, pH, salinity and ammonia nitrogen concentration.

[0018] S1-3 performs spatiotemporal synchronization processing on the collected morphological images, environmental spectral data, and aquaculture water environment data. The acquisition time synchronization is achieved through hardware trigger signals, and a feature matching algorithm is used to establish the spatial mapping relationship of multi-source data.

[0019] Furthermore, step S2 also includes the following sub-steps:

[0020] S2-1 uses a scattering-absorption separation dehazing algorithm to process underwater cobia morphological images. By independently compensating for the light scattering effect of suspended particles and the light absorption effect of water molecules, the degradation of cobia morphological images is corrected.

[0021] S2-2, Perform feature enhancement based on retinal mechanisms, simulate the nonlinear response of retinal ganglion cells to light and dark boundaries, and use multi-scale filtering and adaptive contrast enhancement algorithms to enhance the morphological features of the cobia image, including the edges of body surface markings, fin contours and feature point regions.

[0022] S2-3 integrates the dehazed morphological image of the cobia with the enhanced morphological features through spatial registration and feature fusion algorithms, and outputs a cobia feature image with complete morphological information.

[0023] Furthermore, step S3 also includes the following sub-steps:

[0024] S3-1, Morphological feature points of cobia are located by deformable convolutional neural network. The deformable convolutional neural network adapts to different cobia postures by convolutional kernels with learnable offsets. The morphological feature points include the snout tip A, the dorsal edge end of the skull B, the origin of the first dorsal fin C, the origin of the second dorsal fin D, the base end of the second dorsal fin E, the dorsal origin of the caudal fin F, the base origin of the pectoral fin G, the origin of the pelvic fin H, the origin of the anal fin I, the base end of the anal fin J, and the ventral origin of the caudal fin K.

[0025] S3-2, Based on path analysis, morphological feature parameters are selected from feature points. The morphological feature parameters include DI from the origin of the second dorsal fin to the origin of the anal fin and CI from the origin of the first dorsal fin to the origin of the anal fin.

[0026] S3-3, an individual morphological framework is constructed based on the selected morphological feature parameters, and global geometric features are extracted. The topological relationship of the individual morphological framework is fused with the global geometric features to generate a biometric code with spatiotemporal invariance. The biometric code is used to uniquely identify an individual cobia.

[0027] S3-4 establishes a mapping between morphological feature parameters and individual identification, binding and storing measurement results with acquisition time and spatial location information to construct a traceable three-dimensional individual profile containing a complete growth history.

[0028] Furthermore, step S4 also includes the following sub-steps:

[0029] S4-1, Construct a deep neural network prediction model, including a fully connected branch for morphological features, a 1D convolutional branch for environmental spectra, an attention mechanism fusion layer, and an output layer. The fully connected branch for morphological features includes a batch normalization layer and a Dropout layer. The 1D convolutional branch for environmental spectra has a kernel size of 3 and a filter count of 8. The attention mechanism fusion layer adopts a query-key value calculation mode. The output layer is a single-neuron linear regression unit.

[0030] S4-2, input the morphological feature parameters and environmental spectral data into the deep neural network prediction model, and obtain the weight prediction value in grams through forward propagation calculation. The morphological feature parameters are weighted according to the path analysis results, and the environmental spectral data are spectral data of five characteristic bands: 450nm, 550nm, 650nm, 750nm and 850nm.

[0031] S4-3, when the prediction error exceeds 5% for three consecutive times, the network weights are updated using an incremental learning algorithm. The incremental learning algorithm uses an elastic weight solidification method and determines the importance weights of the parameters by calculating the Fisher information matrix.

[0032] Furthermore, step S5 also includes the following sub-steps:

[0033] S5-1, based on the historical growth curve database and current morphological feature parameters, constructs a time series-morphological feature coupling model, quantifies the correlation weight between growth rate and changes in morphological feature parameters through LSTM neural network, and outputs a weight gain prediction curve for the next 7-30 days.

[0034] S5-2 takes the deviation between the current morphological feature parameters and the predicted growth curve as input, and uses an anomaly detection algorithm based on Mahalanobis distance to calculate the health risk index. When the health risk index exceeds the health threshold, a health warning is triggered; otherwise, it is judged to be in a healthy state.

[0035] S5-3, Establish a compensation function for the water temperature-light intensity-growth response to dynamically correct the prediction results. The specific formula is as follows:

[0036] ;

[0037] in, It is the environmental compensation amount for the predicted weight. This is the water temperature influence coefficient, where T is the real-time measured water temperature. This is the optimal water temperature for the growth of cobia. This is the illumination influence coefficient, where L is the real-time illumination intensity. It is the reference intensity of light.

[0038] Furthermore, step S6 also includes the following sub-steps:

[0039] S6-1, based on the weight gain prediction curve and historical feeding records, a benchmark feeding model for cobia was established using multiple regression analysis. The specific formula is as follows:

[0040] ;

[0041] in, This is the baseline feeding amount. This is a predicted weight gain for the next 7 days. This represents the average feeding amount over the past 7 days. , The coefficient was determined through aquaculture experiments;

[0042] S6-2, The feeding plan is dynamically adjusted based on the health risk index. When the health risk index does not exceed the health threshold, the baseline feeding amount is used. When the health risk index exceeds the health threshold, the feeding amount is adjusted according to a preset rule. The specific formula for the preset rule is as follows:

[0043] ;

[0044] in, To adjust the feeding amount, The baseline feeding amount is given, and S represents the health risk index.

[0045] S6-3 converts the adjusted feeding scheme into control commands, which are then transmitted to the execution device via the communication network to achieve feeding management.

[0046] The beneficial effects of the technical solution provided by this invention include at least the following:

[0047] This invention enables non-contact assessment of the growth status of cobia by simultaneously acquiring multi-angle images and environmental information of cobia in an underwater environment. This significantly reduces stress on the fish and improves assessment efficiency and accuracy.

[0048] This invention employs a scattering-absorption separation dehazing algorithm and a retinal mechanism-based feature enhancement method to process images, dynamically compensating for optical distortions at different water depths. This significantly improves image quality, making morphological feature recognition more accurate and thus enhancing the reliability of growth assessment.

[0049] The deep neural network prediction model constructed in this invention integrates morphological feature parameters and environmental spectral data to output predicted body weight. It can comprehensively consider multiple influencing factors, improve the accuracy and robustness of growth assessment, and provide a scientific basis for aquaculture management.

[0050] This invention analyzes the synergistic change patterns between historical growth curves and current morphological characteristic parameters to predict future growth trends and assess health status. It enables dynamic monitoring and timely adjustment of the growth status of cobia, thereby improving aquaculture efficiency and fish health. Attached Figure Description

[0051] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 A flowchart of the method provided in an embodiment of the present invention;

[0053] Figure 2 This is a feature point marking diagram of the cobra provided in an embodiment of the present invention. Detailed Implementation

[0054] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of an image analysis-based cobia growth assessment method proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0055] 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 invention pertains.

[0056] The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0057] The following description, in conjunction with the accompanying drawings, details a specific scheme for an image analysis-based method for assessing the growth of cobia provided by this invention.

[0058] Please see Figure 1 The diagram illustrates a flowchart of an image analysis-based method for assessing the growth of cobia, according to an embodiment of the present invention. The method includes the following steps:

[0059] Step S1: In the underwater environment, morphological images of cobia, environmental spectral data and aquatic environment data of the culture water are collected simultaneously using a multispectral imaging device.

[0060] Step S1 further includes the following sub-steps:

[0061] S1-1, acquires morphological images of cobia using a multispectral imaging device. The morphological images of cobia include visible light images and near-infrared images. Visible light images reflect the markings and color features of the cobia's body surface, while near-infrared images can penetrate water to obtain the outline information of the cobia.

[0062] S1-2, using an integrated environmental sensor in a multispectral imaging device to collect environmental spectral data of the water body and environmental data of the aquaculture water body. The environmental spectral data includes spectral irradiance data, optical characteristic parameters of the water body, water quality-related optical indicators and environmental background light information. The aquaculture water body environmental data includes water temperature, dissolved oxygen, pH, salinity and ammonia nitrogen concentration.

[0063] S1-3 performs spatiotemporal synchronization processing on the collected morphological images, environmental spectral data, and aquaculture water environment data. The acquisition time synchronization is achieved through hardware trigger signals, and a feature matching algorithm is used to establish the spatial mapping relationship of multi-source data.

[0064] It should be noted that the multispectral imaging device is a camera device capable of simultaneously acquiring optical data in multiple specific wavelength bands, including a visible light sensor (400-700nm) and a near-infrared sensor (700-1100nm), and integrating an environmental monitoring module.

[0065] Specific parameters of the multispectral imaging device:

[0066] 1. Visible light camera: SONY IMX415 sensor, resolution 3840×2160, spectral response range 400-700nm; visible light lens equipped with polarizing filter (to eliminate reflections on water surface).

[0067] 2. Near-infrared camera: FLIR Boson 640, with a response band of 700-1100nm and equipped with an 850nm bandpass filter; near-infrared imaging adopts a coaxial illumination design (wavelength 850nm, power 10W LED array).

[0068] 3. Environmental sensors: Spectroradiometer: ASD FieldSpec4, 350-2500nm, resolution 3nm@700nm; Temperature sensor: Seabird SBE39, accuracy ±0.002℃.

[0069] Serpent fish morphological images: These are images that reflect the body structure and appearance characteristics of serpent fish, used for morphological feature extraction and growth assessment.

[0070] Environmental spectral data includes spectral irradiance of light in the water, optical characteristic parameters (scattering and absorption coefficients), water quality-related optical indicators (chlorophyll concentration, suspended matter concentration), and ambient background light information, which helps to understand the impact of the aquatic environment on image acquisition and cobia growth.

[0071] Aquaculture water environment data: The aquaculture water environment is one of the important factors affecting the growth of cobia. Aquaculture water environment data can provide information about the aquaculture environment conditions and is crucial to the accuracy of growth assessment models.

[0072] Visible light images: Images captured using the visible spectrum (the range of light visible to the human eye, 400-700 nanometers), primarily used to reflect the markings and color features on the body surface of the cobia.

[0073] Near-infrared images: Images captured using the near-infrared spectrum (wavelength slightly longer than visible light, 700-900 nanometers) can penetrate water and obtain the outline information of the cobia, which is particularly useful for identifying fish in turbid water.

[0074] Image acquisition specifications include: Shooting distance: 1.5±0.3m (ensuring that the cobia occupies ≥30% of the frame); Exposure control: automatic exposure for visible light images (target grayscale value 120-180), and fixed exposure for near-infrared images for 50ms; Calibration board usage: deploying a 24-color X-Rite ColorChecker in the pool for white balance calibration.

[0075] Environmental sensors: Sensors integrated into multispectral imaging devices to measure environmental parameters of water bodies, including spectral irradiance and temperature.

[0076] The environmental data synchronous acquisition specifications include: spectral sampling: one spectral scan is triggered for each frame of image (integration time 100ms); temperature sampling: the PT1000 sensor continuously acquires data at a frequency of 10Hz, and the average value is taken 5ms before and after the image exposure time.

[0077] Spatiotemporal synchronization processing: Ensuring that morphological images, environmental spectral data, and aquaculture water environment data are synchronized in time and space is crucial for subsequent data analysis and processing.

[0078] Hardware trigger signal: Use signals generated by hardware devices (timers or synchronization modules) to synchronize the data acquisition process and ensure that all data is captured at the same time.

[0079] Feature matching algorithm: A computational method used to establish spatial mapping relationships between different types of data (images and environmental data) to ensure consistency and comparability between data.

[0080] Feature matching algorithm code:

[0081] def align_data(img_features, sensor_features):

[0082] # Establish mapping relationships using SIFT feature matching

[0083] sift = cv2.SIFT_create()

[0084] kp1, des1 = sift.detectAndCompute(img_features, None)

[0085] kp2, des2 = sift.detectAndCompute(sensor_features, None)

[0086] bf = cv2.BFMatcher()

[0087] matches = bf.knnMatch(des1, des2, k=2)

[0088] # Apply RANSAC to remove mismatches

[0089] good = []

[0090] For m,n in matches:

[0091] if m.distance < 0.75*n.distance:

[0092] good.append([m])

[0093] H, _ = cv2.findHomography(src_pts, dst_pts, cv2.RANSAC, 5.0)

[0094] return H

[0095] Step S2: The image is processed using a scattering-absorption separation dehazing algorithm and a feature enhancement method based on the retinal mechanism to dynamically compensate for optical distortion at different water depths.

[0096] Step S2 further includes the following sub-steps:

[0097] S2-1 uses a scattering-absorption separation dehazing algorithm to process underwater cobia morphological images. By independently compensating for the light scattering effect of suspended particles and the light absorption effect of water molecules, the degradation of cobia morphological images is corrected.

[0098] S2-2 performs feature enhancement based on retinal mechanisms, simulating the nonlinear response of retinal ganglion cells to light and dark boundaries, and uses multi-scale filtering and adaptive contrast enhancement algorithms to enhance the morphological features of cobia images, including the edges of body surface markings, fin contours and feature point regions.

[0099] S2-3 integrates the dehazed morphological image of the cobia with the enhanced morphological features through spatial registration and feature fusion algorithms, and outputs a cobia feature image with complete morphological information.

[0100] It should be noted that the scattering-absorption separation dehazing algorithm is an advanced image processing technique used to eliminate the fogging effect in underwater images. It restores the original clarity of the image by separating the scattering (caused by suspended particles) and absorption (caused by water molecules and other dissolved substances) effects in the image.

[0101] Implementation process of dehazing algorithm based on scattering-absorption separation:

[0102] 1. Light scattering compensation: An improved dark channel prior algorithm is adopted, wherein the dark channel window size is set to 15×15 pixels; the transmittance map optimization uses guided filtering (the guide image is a grayscale image of the original image).

[0103] 2. Optical absorption compensation: A wavelength-dependent attenuation model was established based on the Beer-Lambert law: Attenuation coefficients α = 0.2 / m, 0.15 / m, and 0.25 / m were set for the R (650nm), G (550nm), and B (450nm) channels, respectively; the attenuation coefficient for the near-infrared band (850nm) was fixed at 0.08 / m.

[0104] Feature enhancement methods based on retinal mechanisms: mimicking the human retina's response to light, enhancing key visual features in images, particularly suitable for enhancing morphological features in images, and crucial for the growth assessment of cobia.

[0105] Dynamic compensation for optical distortion at different water depths: Since the optical conditions (light scattering and absorption) are different at different water depths, this step ensures that high-quality images can be obtained under various water depth conditions by dynamically adjusting the image processing parameters.

[0106] Light scattering effect: refers to the scattering phenomenon that occurs when light passes through water containing suspended particles, which can cause images to become blurry and unclear.

[0107] Light absorption effect: refers to the absorption of light by molecules and particles in water, which can cause image color distortion and reduced brightness.

[0108] Correcting morphological image degradation of cobia: restoring the original morphological features of the image, including the outline and texture of the fish body, through a dehazing algorithm.

[0109] Retinal ganglion cells: A type of cell in the human retina that is responsible for converting light signals into nerve signals and transmitting them to the brain for visual processing. These cells are particularly sensitive to the boundaries between light and dark areas and can enhance visual contrast.

[0110] Multiscale filtering: an image processing technique that enhances image features by applying filters at different scales, which helps to extract local details in an image.

[0111] Adaptive contrast enhancement algorithm: an intelligent algorithm that can automatically adjust the contrast based on the local characteristics of an image, making the details in the image stand out more.

[0112] Spatial registration: an image processing technique used to align images captured from different sources or under different conditions to ensure their spatial consistency.

[0113] The technical specifications for spatial registration include:

[0114] 1. Use SURF feature point matching, with the following requirements: number of matching point pairs ≥ 50; reprojection error RMS < 0.5 pixels.

[0115] 2. Registration transformation model: Affine transformation (6 degrees of freedom).

[0116] Feature fusion algorithm: a technique that combines multiple image features to generate a single image containing complete information.

[0117] The specific parameters of the feature fusion algorithm include:

[0118] 1. Fusion weights of dehazed image and texture enhancement image: Low frequency component: 70% dehazed image + 30% enhancement image; High frequency component: 20% dehazed image + 80% enhancement image.

[0119] 2. Image quality evaluation criteria after fusion: Edge Preservation Index (EPI) > 0.9; Information Entropy Increment ≥ 0.8 bits / pixel.

[0120] Output a cobia feature image with complete morphological information: The final output image not only contains the basic morphological information of the cobia, but also includes enhanced morphological features, providing rich data support for subsequent growth assessment.

[0121] Step S3: Extract morphological feature parameters of cobia from the image based on a deep learning model and establish individual identification.

[0122] Step S3 further includes the following sub-steps:

[0123] S3-1, the morphological feature points of cobia are located by deformable convolutional neural network. The deformable convolutional neural network adapts to different cobia postures through convolutional kernels with learnable offsets. The morphological feature points include the snout tip A, the end of the dorsal margin of the skull B, the origin of the first dorsal fin C, the origin of the second dorsal fin D, the end of the base of the second dorsal fin E, the dorsal origin of the caudal fin F, the upper origin of the pectoral fin G, the origin of the pelvic fin H, the origin of the anal fin I, the end of the base of the anal fin J, and the ventral origin of the caudal fin K.

[0124] S3-2, Based on path analysis, morphological feature parameters are selected from feature points. The morphological feature parameters include DI from the origin of the second dorsal fin to the origin of the anal fin and CI from the origin of the first dorsal fin to the origin of the anal fin.

[0125] S3-3, an individual morphological framework is constructed based on the selected morphological feature parameters, and global geometric features are extracted. The topological relationship of the individual morphological framework is fused with the global geometric features to generate a biometric code with spatiotemporal invariance. The biometric code is used to uniquely identify an individual cobia.

[0126] S3-4 establishes a mapping between morphological feature parameters and individual identification, binding and storing measurement results with acquisition time and spatial location information to construct a traceable three-dimensional individual profile containing a complete growth history.

[0127] Please see Figure 2 This is a feature point marking diagram of the cobra provided in an embodiment of the present invention.

[0128] It should be noted that deep learning models are machine learning models based on artificial neural networks, capable of learning complex feature representations from large amounts of data. In this invention, they are used to extract morphological features of cobia from images.

[0129] Morphological characteristics parameters: These are quantifiable parameters that reflect the body shape and appearance of cobia. These parameters are very important for assessing the growth status of cobia.

[0130] Individual identification: Each cobia is assigned a unique identifier to identify and track the growth changes of the same fish in multiple observations.

[0131] Deformable convolutional neural network: a deep learning model, a special type of convolutional neural network whose convolutional kernels can adaptively adjust their shape according to the features of the input data, thereby better adapting to different poses of the cobia.

[0132] The network architecture details of deformable convolutional neural networks include:

[0133] 1. Backbone network: Improved version of ResNet-50 (replacing the last 3 regular convolutional layers with deformable convolutional layers).

[0134] 2. Deformable convolution parameters: Offset learning rate: 0.1 times the base learning rate; Kernel size: 3×3, Number of offset channels per group = 2 × kernel area.

[0135] 3. Feature point localization head: Output heatmap resolution: 1 / 4 of the input image; Loss function: Improved Focal Loss (α=0.8, γ=2).

[0136] Morphological feature points: These refer to specific points on the body of the cobia. The positions of these points can be used to calculate various morphological feature parameters.

[0137] Path analysis: a statistical method used to evaluate the direct impact of different variables (morphological parameters) on the outcome variable (weight). In this invention, it is used to screen the morphological features that are most important for weight prediction.

[0138] The steps for performing a path analysis include:

[0139] 1. Input parameters: 39 possible combinations of lengths (morphological parameters) for 11 feature points.

[0140] 2. Statistical Methods: First, a linear correlation analysis was performed between morphological parameters and body mass to calculate the correlation coefficient (r) and path coefficient (Pi). Then, based on the correlation analysis and path coefficient, the direct determinant coefficient of each morphological parameter on body mass was calculated. d i ) and indirect determination coefficient ( d ij This demonstrates the degree to which various morphological traits affect body mass.

[0141] 3. Results verification: DI has the greatest direct effect on body weight (1.982), followed by CI (1.087), DI from the origin of the second dorsal fin to the origin of the anal fin and CI from the origin of the first dorsal fin to the origin of the anal fin.

[0142] Method for constructing the individual morphological framework: The Delaunay triangulation algorithm is used to construct a triangular mesh model of the cobia's body surface with 11 morphological feature points as vertices; the side length ratio and included angle of specific triangles in the mesh are calculated as framework features.

[0143] Selection of global geometric features: Select the two morphological parameters with the highest weight in the path analysis and their derived ratios.

[0144] Feature encoding generation and fusion: The frame features (a set of proportional values) and global geometric features (a set of measurements) are normalized and then concatenated to form the original feature vector; Principal component analysis (PCA) or an autoencoder is used to reduce the dimensionality and encode the original feature vector to obtain 128-dimensional or 256-dimensional biometric codes.

[0145] Similarity measurement methods: Use cosine similarity or Euclidean distance to compare two biometric codes, and set a threshold (such as 0.95) to determine whether they belong to the same person.

[0146] Biometric coding: A coding method that converts specific characteristics of an organism into digital representations that can be used for identification and classification.

[0147] Association mapping: The process of associating morphological feature parameters with individual identity identifiers to ensure that the morphological feature data of each fish corresponds correctly with its identity identifier.

[0148] Three-dimensional individual profile: A detailed record for each cobia, including its morphological characteristics, identity markers, and growth history, which can be used to track individual growth changes.

[0149] Step S4: Construct a deep neural network prediction model, integrate morphological feature parameters and environmental spectral data, and output the predicted weight value;

[0150] Step S4 further includes the following sub-steps:

[0151] S4-1, Construct a deep neural network prediction model, including a fully connected branch for morphological features, a 1D convolutional branch for environmental spectra, an attention mechanism fusion layer, and an output layer. The fully connected branch for morphological features includes a batch normalization layer and a Dropout layer. The 1D convolutional branch for environmental spectra has a kernel size of 3 and a filter number of 8. The attention mechanism fusion layer adopts a query-key value calculation mode. The output layer is a single neuron linear regression unit.

[0152] S4-2, input the morphological feature parameters and environmental spectral data into the deep neural network prediction model, and obtain the weight prediction value in grams through forward propagation calculation. The morphological feature parameters are weighted according to the path analysis results, and the environmental spectral data are spectral data of five characteristic bands: 450nm, 550nm, 650nm, 750nm and 850nm.

[0153] S4-3 When the prediction error exceeds 5% for three consecutive times, the incremental learning algorithm is used to update the network weights. The incremental learning algorithm adopts the elastic weight solidification method and determines the importance weights of the parameters by calculating the Fisher information matrix.

[0154] It should be noted that the deep neural network prediction model is a multi-layered neural network model that can learn the complex relationship between input data (morphological features and environmental spectrum) and output (weight).

[0155] Output weight prediction: The model predicts the weight of the cobia based on the input morphological features and environmental spectral data.

[0156] Morphological feature fully connected branch: a part of a neural network where morphological feature parameters are processed through fully connected layers (each neuron is connected to all neurons in the previous layer).

[0157] Batch normalization layer: A type of network layer used to normalize the output of the previous layer in order to accelerate the training process and improve the stability of the model.

[0158] Dropout layer: A regularization technique that prevents overfitting by randomly dropping some neurons in the network.

[0159] Environmental Spectrum 1D Convolution Branch: A network branch specifically for processing one-dimensional environmental spectral data, using convolution operations to extract features.

[0160] A kernel size of 3 and a filter count of 8 define the parameters of the convolutional layer, where the kernel size determines the receptive field and the filter count determines the depth of the output feature map.

[0161] Attention fusion layer: A network layer that learns the importance of different input features, thereby improving model performance; The attention fusion layer implementation includes: Query source: 64-dimensional feature vector of morphological branch; Key-Value source: 8-dimensional feature vector of spectral branch; Number of attention heads: 4; Temperature coefficient τ: 0.5 (Softmax scaling factor).

[0162] Query-key value calculation pattern: an implementation of the attention mechanism that determines the weight of a value by calculating the similarity between the query and the key.

[0163] The output layer is a single-neuron linear regression unit: the last layer of the network, which uses a single neuron to perform linear regression and outputs the final predicted value.

[0164] Forward propagation: The training process of a neural network, in which input data is passed forward through the network to calculate the output.

[0165] Weight prediction in grams: The weight prediction output by the model is in grams.

[0166] Weighted by path analysis results: The direct impact of each morphological characteristic parameter determined by path analysis on body weight is weighted.

[0167] Morphological parameter weighting method: DI weight: 1.446 (direct path coefficient); 2L-2R weight: 0.663; EG weight: 0.635; normalization: Min-Max scaling to the [0,1] interval.

[0168] Spectral data in 5 characteristic bands: spectral data at specific wavelengths (450nm, 550nm, 650nm, 750nm and 850nm), which reflect the optical properties of the water body.

[0169] Spectral data calibration includes dark current subtraction: recording dark reference values ​​before each acquisition; whiteboard correction: using a Labsphere Spectralon standard reflector; and band extraction: integrated intensity within the center wavelength ±5 nm range.

[0170] Incremental learning algorithm: A learning strategy that allows a model to update its weights as it receives new data without having to retrain the entire network.

[0171] Elastic weight fixation method: an incremental learning technique that determines the importance of network parameters by calculating the Fisher information matrix and adjusts the update strategy accordingly.

[0172] Fisher information matrix: a statistic used to measure the uncertainty of parameters, and in this case, to determine which parameters should be retained more during updates.

[0173] Step S5: By analyzing the synergistic change patterns between historical growth curves and current morphological characteristic parameters, and combining this with environmental parameter compensation mechanisms, predict future growth trends and assess health status.

[0174] Step S5 further includes the following sub-steps:

[0175] S5-1, based on the historical growth curve database and current morphological feature parameters, constructs a time series-morphological feature coupling model, quantifies the correlation weight between growth rate and changes in morphological feature parameters through LSTM neural network, and outputs a weight gain prediction curve for the next 7-30 days.

[0176] S5-2 takes the deviation between the current morphological feature parameters and the predicted growth curve as input, and uses an anomaly detection algorithm based on Mahalanobis distance to calculate the health risk index. When the health risk index exceeds the health threshold, a health warning is triggered; otherwise, it is judged to be in a healthy state.

[0177] S5-3, Establish a compensation function for the water temperature-light intensity-growth response to dynamically correct the prediction results. The specific formula is as follows:

[0178] ;

[0179] in, It is the environmental compensation amount for the predicted weight. This is the water temperature influence coefficient, where T is the real-time measured water temperature. This is the optimal water temperature for the growth of cobia. This is the illumination influence coefficient, where L is the real-time illumination intensity. It is the reference intensity of light.

[0180] It should be noted that the historical growth curve refers to the growth data of cobia over a period of time, which can reflect the general trend and pattern of its growth.

[0181] Environmental parameter compensation mechanism: a mechanism to adjust the prediction results considering the impact of environmental factors (water temperature and light) on the growth of cobia.

[0182] Predicting future growth trends: Based on current and historical data, predict the growth of cobia in the near future.

[0183] Assess health status: Assess the health status of cobia based on growth trends and deviations from morphological parameters.

[0184] Time-series-morphological feature coupling model: A model that combines time-series analysis and morphological feature analysis to predict the growth of cobia.

[0185] LSTM Neural Network: Long Short-Term Memory Network, a special type of recurrent neural network capable of learning and predicting time series data.

[0186] The LSTM neural network architecture details include: input layer: 7-dimensional time series (historical weight + DI + 2L-2R + EG); hidden layer: two-layer LSTM (64 neurons / layer); output layer: daily weight prediction for the next 30 days (linear activation); sequence processing: sliding window length = 60 days.

[0187] The correlation weight between growth rate and changes in morphological feature parameters: the relationship between growth rate and changes in morphological feature parameters learned through the LSTM network.

[0188] Deviation: The difference between the current morphological feature parameters and the predicted growth curve.

[0189] Mahalanobis distance anomaly detection algorithm: An algorithm based on Mahalanobis distance is used to detect whether data points are outliers.

[0190] Health Risk Index: An indicator calculated based on deviation, used to assess the health status of cobia.

[0191] Health threshold: A preset value used to determine whether the health risk index indicates a healthy state; in this embodiment, it is 0.6.

[0192] Water temperature-light-growth response compensation function: A function used to adjust growth prediction results according to water temperature and light conditions.

[0193] Environmental compensation for predicted weight: Predicted weight adjusted for environmental conditions.

[0194] Water temperature influence coefficient: a quantitative indicator of the effect of water temperature on the growth of cobia, fitted through temperature control experiments.

[0195] Code for calibrating the water temperature influence coefficient:

[0196] # Alpha coefficient fitting code (based on least squares method)

[0197] import numpy as np

[0198] from scipy.optimize import curve_fit

[0199] # Experimental Data: Water Temperature T vs. Relative Growth Rate ΔW_observed

[0200] T_data = np.array([24, 26, 28, 30, 32])

[0201] ΔW_data = np.array([-2.1, -0.8, 0, -1.2, -3.0])

[0202] def temp_model(T, alpha):

[0203] return alpha * (T - 28)**2

[0204] alpha_opt, _ = curve_fit(temp_model, T_data, ΔW_data)

[0205] print(f"The fitted result is α={alpha_opt[0]:.3f}") # Output α≈0.152

[0206] Optimal growth water temperature: The most suitable water temperature for the growth of cobia is 28℃ in this example.

[0207] Light intensity effect coefficient: a quantitative index of the effect of light intensity on the growth of cobia, fitted by the light cycle control experiment.

[0208] Light reference intensity: The light intensity value used for reference.

[0209] Step S6: Based on future growth trends, feeding records, and health assessment results, generate a dynamically adjusted feeding plan and implement feeding management.

[0210] Step S6 further includes the following sub-steps:

[0211] S6-1, based on the weight gain prediction curve and historical feeding records, a benchmark feeding model for cobia was established using multiple regression analysis. The specific formula is as follows:

[0212] ;

[0213] in, This is the baseline feeding amount. This is a predicted weight gain for the next 7 days. This represents the average feeding amount over the past 7 days. , The coefficient was determined through aquaculture experiments;

[0214] S6-2, The feeding plan is dynamically adjusted based on the health risk index. When the health risk index does not exceed the health threshold, the baseline feeding amount is used. When the health risk index exceeds the health threshold, the feeding amount is adjusted according to a preset rule. The specific formula for the preset rule is as follows:

[0215] ;

[0216] in, To adjust the feeding amount, The baseline feeding amount is given, and S represents the health risk index.

[0217] S6-3 converts the adjusted feeding scheme into control commands, which are then transmitted to the execution device via the communication network to achieve feeding management.

[0218] It should be noted that the dynamically adjusted feeding plan is an automatically adjusted feeding plan based on real-time data and predictive model results, designed to meet the growth needs of cobia and prevent overfeeding or underfeeding.

[0219] Feeding management: Implement the adjusted feeding plan through an automated system to ensure the accuracy and timeliness of feeding.

[0220] Benchmark feeding model: A mathematical model based on predicted weight gain and historical feeding data of cobia is used to calculate the recommended feeding amount.

[0221] Multiple regression analysis: a statistical method used to analyze the influence of multiple independent variables (predicted weight gain and historical feeding amount) on a single dependent variable (baseline feeding amount).

[0222] Coefficients: In a multiple regression model, the parameters associated with each independent variable represent the degree of influence of that variable on the baseline feed amount. These coefficients are calibrated through aquaculture experiments.

[0223] Health Risk Index: A quantitative indicator used to assess the health status of cobia, calculated based on the deviation of morphological characteristic parameters from the predicted growth curve.

[0224] Health threshold: A preset value used to determine whether the health risk index indicates that the feeding plan needs to be adjusted.

[0225] Adjusted feeding amount: When the health risk index exceeds the health threshold, the feeding amount is adjusted according to preset rules.

[0226] Preset rules: Rules that dynamically adjust the amount of feed based on the health risk index, ensuring that the amount of feed is reduced when the fish are in poor health, thus avoiding further health problems.

[0227] Control commands: Convert the adjusted feeding plan into commands that can be executed by automated equipment.

[0228] Communication network: A network used to transmit control commands from a central control system to the execution devices.

[0229] Execution equipment: Automatic feeder, used to perform feeding operations according to control commands.

[0230] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for evaluating the growth of cobia based on image analysis, characterized in that, Includes the following steps: Step S1: In the underwater environment, morphological images of cobia, environmental spectral data and aquatic environment data of the culture water are collected simultaneously using a multispectral imaging device. Step S2: The image is processed using a scattering-absorption separation dehazing algorithm and a feature enhancement method based on the retinal mechanism to dynamically compensate for optical distortion at different water depths. Step S3: Extract morphological feature parameters of cobia from the image based on a deep learning model and establish individual identification. Step S4: Construct a deep neural network prediction model, integrate morphological feature parameters and environmental spectral data, and output the predicted weight value; Step S5: By analyzing the synergistic change patterns between historical growth curves and current morphological characteristic parameters, and combining this with environmental parameter compensation mechanisms, predict future growth trends and assess health status. Step S6: Based on future growth trends, feeding records, and health assessment results, generate a dynamically adjusted feeding plan and implement feeding management.

2. The image analysis-based growth assessment method for cobia according to claim 1, characterized in that: Step S1 further includes the following sub-steps: S1-1, acquire morphological images of cobia using a multispectral imaging device. The morphological images of cobia include visible light images and near-infrared images. The visible light images reflect the markings and color features of the cobia's body surface, and the near-infrared images can penetrate water to obtain the outline information of the cobia. S1-2, using an integrated environmental sensor in a multispectral imaging device to collect environmental spectral data of the water body and environmental data of the aquaculture water body. The environmental spectral data includes spectral irradiance data, optical characteristic parameters of the water body, water quality-related optical indicators and environmental background light information. The aquaculture water body environmental data includes water temperature, dissolved oxygen, pH, salinity and ammonia nitrogen concentration. S1-3 performs spatiotemporal synchronization processing on the collected morphological images, environmental spectral data, and aquaculture water environment data. The acquisition time synchronization is achieved through hardware trigger signals, and a feature matching algorithm is used to establish the spatial mapping relationship of multi-source data.

3. The image analysis-based growth assessment method for cobia according to claim 1, characterized in that: Step S2 further includes the following sub-steps: S2-1 uses a scattering-absorption separation dehazing algorithm to process underwater cobia morphological images. By independently compensating for the light scattering effect of suspended particles and the light absorption effect of water molecules, the degradation of cobia morphological images is corrected. S2-2, Perform feature enhancement based on retinal mechanisms, simulate the nonlinear response of retinal ganglion cells to light and dark boundaries, and use multi-scale filtering and adaptive contrast enhancement algorithms to enhance the morphological features of the cobia image, including the edges of body surface markings, fin contours and feature point regions. S2-3 integrates the dehazed morphological image of the cobia with the enhanced morphological features through spatial registration and feature fusion algorithms, and outputs a cobia feature image with complete morphological information.

4. The image analysis-based growth assessment method for cobia according to claim 1, characterized in that: Step S3 further includes the following sub-steps: S3-1, Morphological feature points of cobia are located by deformable convolutional neural network. The deformable convolutional neural network adapts to different cobia postures by convolutional kernels with learnable offsets. The morphological feature points include the snout tip A, the dorsal edge end of the skull B, the origin of the first dorsal fin C, the origin of the second dorsal fin D, the base end of the second dorsal fin E, the dorsal origin of the caudal fin F, the base origin of the pectoral fin G, the origin of the pelvic fin H, the origin of the anal fin I, the base end of the anal fin J, and the ventral origin of the caudal fin K. S3-2, Based on path analysis, key morphological feature parameters are selected from feature points. The morphological feature parameters include DI from the origin of the second dorsal fin to the origin of the anal fin and CI from the origin of the first dorsal fin to the origin of the anal fin. S3-3, an individual morphological framework is constructed based on the selected morphological feature parameters, and global geometric features are extracted. The topological relationship of the individual morphological framework is fused with the global geometric features to generate a biometric code with spatiotemporal invariance. The biometric code is used to uniquely identify an individual cobia. S3-4 establishes a mapping between morphological feature parameters and individual identification, binding and storing measurement results with acquisition time and spatial location information to construct a traceable three-dimensional individual profile containing a complete growth history.

5. The image analysis-based growth assessment method for cobia according to claim 1, characterized in that: Step S4 further includes the following sub-steps: S4-1, Construct a deep neural network prediction model, including a fully connected branch for morphological features, a 1D convolutional branch for environmental spectra, an attention mechanism fusion layer, and an output layer. The fully connected branch for morphological features includes a batch normalization layer and a Dropout layer. The 1D convolutional branch for environmental spectra has a kernel size of 3 and a filter count of 8. The attention mechanism fusion layer adopts a query-key value calculation mode. The output layer is a single-neuron linear regression unit. S4-2, input the morphological feature parameters and environmental spectral data into the deep neural network prediction model, and obtain the weight prediction value in grams through forward propagation calculation. The morphological feature parameters are weighted according to the path analysis results, and the environmental spectral data are spectral data of five characteristic bands: 450nm, 550nm, 650nm, 750nm and 850nm. S4-3, when the prediction error exceeds 5% for three consecutive times, the network weights are updated using an incremental learning algorithm. The incremental learning algorithm uses an elastic weight solidification method and determines the importance weights of the parameters by calculating the Fisher information matrix.

6. The image analysis-based growth assessment method for cobia according to claim 1, characterized in that: Step S5 further includes the following sub-steps: S5-1, based on the historical growth curve database and current morphological feature parameters, constructs a time series-morphological feature coupling model, quantifies the correlation weight between growth rate and changes in morphological feature parameters through LSTM neural network, and outputs a weight gain prediction curve for the next 7-30 days. S5-2 takes the deviation between the current morphological feature parameters and the predicted growth curve as input, and uses an anomaly detection algorithm based on Mahalanobis distance to calculate the health risk index. When the health risk index exceeds the health threshold, a health warning is triggered; otherwise, it is judged to be in a healthy state. S5-3, Establish a compensation function for the water temperature-light intensity-growth response to dynamically correct the prediction results. The specific formula is as follows: ； in, It is the environmental compensation amount for the predicted weight. This is the water temperature influence coefficient, where T is the real-time measured water temperature. This is the optimal water temperature for the growth of cobia. This is the illumination influence coefficient, where L is the real-time illumination intensity. It is the reference intensity of light.

7. The image analysis-based growth assessment method for cobia according to claim 1, characterized in that: Step S6 further includes the following sub-steps: S6-1, based on the weight gain prediction curve and historical feeding records, a benchmark feeding model for cobia was established using multiple regression analysis. The specific formula is as follows: ； in, This is the baseline feeding amount. This is a predicted weight gain for the next 7 days. This represents the average feeding amount over the past 7 days. , The coefficient was determined through aquaculture experiments; S6-2, The feeding plan is dynamically adjusted based on the health risk index. When the health risk index does not exceed the health threshold, the baseline feeding amount is used. When the health risk index exceeds the health threshold, the feeding amount is adjusted according to a preset rule. The specific formula for the preset rule is as follows: ； in, To adjust the feeding amount, The baseline feeding amount is given, and S represents the health risk index. S6-3 converts the adjusted feeding scheme into control commands, which are then transmitted to the execution device via the communication network to achieve feeding management.

Images