A fish feeding intensity evaluation method based on binocular vision

Abstract

The application provides a fish feeding intensity evaluation method based on binocular vision. The method acquires fish feeding images through a binocular camera, and calibrates and corrects the images; then extracts fish targets and fish feature points by using an image segmentation technology, and acquires fish two-dimensional space features; after restoring and matching the missing feature points, the method calculates fish three-dimensional space features by using a weighted triangular partition algorithm, realizes deep perception of behavior details; then performs feature fusion by using a self-encoder, and generates fusion features highly related to fish feeding intensity; finally, the method performs multi-scale decomposition on the fusion features by using an empirical mode decomposition algorithm to acquire feeding intensity change trends, and performs uniform division to realize fine evaluation of fish feeding intensity. By introducing binocular vision technology, the application effectively solves the problem that two-dimensional images are not complete in representing longitudinal behavior, realizes effective compensation of evaluation errors caused by lack of depth information, and significantly improves the precision of feeding intensity evaluation.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and more specifically to a method for assessing fish feeding intensity based on binocular vision. Background Technology

[0002] In large-scale aquaculture, fish obtain the nutrients and energy needed for growth primarily through artificial feeding, making feed cost one of the main expenses. With the continuous expansion of aquaculture scale, feed cost has become a significant factor limiting profit margins. Therefore, optimizing the feed feeding process is crucial for reducing aquaculture costs and improving economic efficiency. In traditional aquaculture, farmers often rely on personal experience to develop feeding plans, neglecting the influence of fish feeding needs and environmental changes. This can easily lead to insufficient or excessive feeding, resulting in water pollution and resource waste.

[0003] With the rapid development of computer vision technology, computer vision-based methods for assessing fish feeding intensity have emerged. Due to their advantages such as being non-contact, easy to use, and highly efficient, they have become an important tool for studying feeding intensity assessment. However, in existing technologies, fish feeding intensity assessment mostly relies on monocular vision technology. Since fish feeding behavior is inherently three-dimensional, monocular vision-based assessment can only acquire two-dimensional information on the projection plane, making it difficult to capture depth information in three-dimensional space. This results in incomplete behavioral modeling, thus limiting the accuracy of feeding intensity assessment. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a method for assessing fish feeding intensity based on binocular vision and spatial feature fusion to achieve fine segmentation.

[0005] Technical solution: A method for assessing fish feeding intensity based on binocular vision, comprising the following steps:

[0006] S1. Use a binocular camera to acquire images of fish feeding, calibrate and correct the images, preprocess the corrected images, and extract fish outlines and feature points.

[0007] S2. Calculate the two-dimensional spatial features of fish based on fish outlines and fish feature points;

[0008] S3. Perform feature point matching on fish feature points in fish feeding images, reconstruct the three-dimensional coordinates of fish feature points, and calculate the three-dimensional spatial features of fish based on the three-dimensional coordinates.

[0009] S4. The autoencoder algorithm is used to fuse the two-dimensional and three-dimensional spatial features of fish to obtain the fused features of fish. The empirical mode decomposition algorithm is used to decompose the fused features of fish to obtain the corresponding intrinsic mode function components. The range of the intrinsic mode function components is calculated. An adaptive equal division strategy based on the range is adopted for the intrinsic mode function components to divide the intrinsic mode function components into multiple intervals. Each interval corresponds to a fish feeding intensity level, thus completing the fish feeding intensity assessment.

[0010] Specifically, the calibration and correction of fish feeding images includes: performing camera calibration, obtaining the camera's intrinsic and extrinsic parameters and distortion parameters, using the intrinsic and extrinsic parameters and distortion parameters to correct the fish feeding images, and calculating the binocular image disparity after stereo matching of the corrected fish feeding images to generate the corresponding disparity map.

[0011] Specifically, preprocessing of the corrected fish feeding images includes:

[0012] The target fish body is extracted from the fish feeding image by image segmentation. The image is enhanced by linear transformation and histogram equalization to improve the clarity of the target fish body. Then the image is binarized, and the fish contour is extracted and drawn by contour detection function. The centroid of the fish contour is extracted as the fish feature point.

[0013] Specifically, the two-dimensional spatial characteristics of fish include the average distance between an individual and the water surface, the angle of inclination of the fish body, the area ratio of the aggregation region, and the proportion of the number of individuals in the aggregation region.

[0014] Specifically, the formula for calculating the average distance between an individual and the water surface is:

[0015]

[0016] In the formula: This represents the average distance between an individual and the water surface. This represents the distance from the i-th individual to the water surface. The number of individual fish;

[0017] The formula for calculating the fish's body tilt angle is:

[0018]

[0019] In the formula: The angle of the fish's body tilt. Let x be the x-coordinate of the fish head. Let be the ordinate of the fish head. Let x be the x-coordinate of the fish tail. The vertical coordinate of the fish tail;

[0020] The formula for calculating the area ratio of clustered areas is:

[0021]

[0022] In the formula: This represents the percentage of the area within the cluster. The area of ​​the cluster. This represents the total area of ​​the fish habitat.

[0023] The formula for calculating the percentage of individuals in a cluster is:

[0024]

[0025] In the formula: The percentage of individuals in a cluster. The number of individuals in the cluster.

[0026] Specifically, feature point matching for fish feature points in fish feeding images includes:

[0027] The fish feature points in the left and right images of a fish feeding image are ordered spatially. Using the fish feature points in the left image as a reference, the coordinates of matching points in the right image are predicted using the disparity values ​​of the corresponding positions. Fish feature points are then searched for in the neighborhood around these matching point coordinates. If a fish feature point exists, a feature point matching pair is established; otherwise, the corresponding position in the right image is marked as a missing feature point. The above steps are repeated, using the fish feature points in the right image as a reference, to detect missing feature points in the left image. For missing feature points, the position coordinates are calculated in reverse using the disparity value and the corresponding fish feature point to obtain the restored feature point. The spatial distance and distribution consistency between the restored feature point and neighboring fish feature points are calculated for validity verification. The verified restored feature points and the original fish feature points are then input into the ORB feature point matching algorithm to complete feature point matching. Finally, the three-dimensional coordinates of the fish feature points are reconstructed based on the principle of binocular vision.

[0028] Specifically, the three-dimensional spatial characteristics of fish include the dispersion of fish schools, the average distance between individuals, and the relationship between individuals and feeding points.

[0029] Specifically, the fish swarm dispersion is calculated as follows:

[0030] Using fish feature points as vertices of the Delaunay triangle, spatial triangulation is performed to obtain multiple tetrahedral structures containing fish feature points. The triangular faces containing fish feature points within each tetrahedral structure are then divided, with the triangular faces contributing to the formation of multiple tetrahedra designated as shared triangles and the remaining triangular faces designated as boundary triangles. A weighting formula is then established.

[0031]

[0032] In the formula: As weight, The number of boundary triangles, To share the number of triangles;

[0033] A formula for calculating the dispersion of a fish population is established based on weights:

[0034]

[0035] In the formula: The dispersion of the fish school The weight of the i-th fish feature point is... Let be the surface area of ​​the tetrahedron formed by the i-th fish feature point. The number of individual fish;

[0036] The formula for calculating the average spacing between individuals is:

[0037]

[0038] In the formula: The average spacing between individuals Let be the distance between the i-th fish feature point and the j-th fish feature point;

[0039] The formula for calculating the relationship between an individual and a feeding point is:

[0040]

[0041] In the formula: For the relationship between individuals and feeding sites, The distance between an individual fish and the feeding point. The angle of inclination of the line connecting the individual fish and the feeding point.

[0042] Specifically, adaptive equal division strategies based on range include:

[0043] Based on the range of intrinsic mode function components, the intervals are divided into equal parts. The number of equal parts is set to obtain the division schemes corresponding to the number of equal parts. The grading effect of different division schemes is evaluated by using the intraclass minimum difference criterion, and the optimal scheme for fish feeding intensity assessment is determined.

[0044] Specifically, the range of the number of equally divided intervals is: N represents the number of individual fish.

[0045] Beneficial Effects: Compared with existing technologies, the significant advantages of this invention are: This invention acquires fish feeding images using a binocular camera, effectively solving the problem of incomplete longitudinal behavior representation in two-dimensional images. It effectively compensates for evaluation errors caused by missing depth information, significantly improving the accuracy of feeding intensity assessment. Then, it acquires two-dimensional and three-dimensional spatial features related to fish feeding intensity through binocular images. These features are then fused using an autoencoder to obtain fused features highly correlated with fish feeding intensity. Finally, the fused features are decomposed at multiple scales using an empirical mode decomposition algorithm to obtain intrinsic mode function (IMF) components, revealing the trend of feeding intensity changes. Based on the range of the IMF components, an adaptive equal division strategy based on the range is used to divide the IMF components into several intervals corresponding to different fish feeding intensities. This invention differs from the traditional four-class classification method of "none," "weak," "medium," and "strong," extending it to multiple finer levels. It provides an effective solution to the intra-class confusion problem inherent in four-class classification methods, has a wide range of applications, and can achieve refined assessment of fish feeding intensity. Attached Figure Description

[0046] Figure 1 This is a flowchart of the method in Embodiment 1 of the present invention.

[0047] Figure 2 This is a flowchart of fish three-dimensional spatial feature extraction in Embodiment 1 of the present invention.

[0048] Figure 3 This is a flowchart of feature point matching in Embodiment 1 of the present invention.

[0049] Figure 4 This is a schematic diagram of the weighted triangulation principle in Embodiment 1 of the present invention. Detailed Implementation

[0051] A preferred embodiment of the present invention will be further described below with reference to the accompanying drawings.

[0052] Example 1

[0053] Please see Figure 1 As shown, this embodiment provides a method for assessing fish feeding intensity based on binocular vision, including the following steps:

[0054] S1. Use a binocular camera to acquire images of fish feeding, calibrate and correct the images, preprocess the corrected images, and extract the fish outline and feature points.

[0055] This embodiment uses a laboratory setting as an example. First, a fish image acquisition platform is set up in the laboratory to collect images of fish feeding. The main hardware equipment of the fish image acquisition platform includes: one fish tank, two aerators, one binocular camera (model ZED 2, parameters: 63 mm baseline, 5.48 million pixels, 4 mm focal length), and one computer. The binocular camera is fixed to the side of the fish tank using a bracket, 0.5 m away from the tank. The binocular camera is connected to the computer via a data cable, and the acquired fish video data is transmitted to the computer for storage. Video frames are extracted from the acquired fish video data to obtain images of fish feeding.

[0056] After acquiring images of fish feeding using a stereo camera, the images need to undergo stereo image processing, including camera calibration, image correction, and stereo matching. Due to lens distortion, images may exhibit deformation and proportional imbalance. Therefore, camera calibration is performed first to obtain the camera's intrinsic and extrinsic parameters, as well as distortion parameters. Then, these parameters are used to correct the images, restoring their original shape and proportions. The corrected stereo images then undergo stereo matching, where corresponding pixels in the left and right images are matched to calculate the stereo image disparity and generate a disparity map.

[0057] Underwater fish feeding images suffer from uneven lighting, distortion, and blurring due to water ripples, which can easily lead to inaccurate feature extraction and thus affect the assessment of feeding intensity. To improve the accuracy of image feature extraction, this invention performs preprocessing operations on the corrected fish feeding images, including image segmentation, image enhancement, image binarization, and contour extraction. Image segmentation extracts the target fish from the feeding images. Linear transformation and histogram equalization are used for image enhancement to improve the clarity of the target fish and reduce noise. Image binarization then highlights the morphological features and details of the target fish. Finally, the fish contour is extracted and drawn using the findContours() function, and the centroid of the fish contour is extracted as the fish feature point.

[0058] S2. Calculate the two-dimensional spatial features of fish based on fish outlines and fish feature points.

[0059] When the feeding intensity of fish changes, their location, distribution, aggregation, posture, etc. will change accordingly. Therefore, this invention extracts four two-dimensional spatial features: average distance from the individual to the water surface (ADS), fish body tilt (FT), aggregation area ratio (PAA), and aggregation individual ratio (PAI) to evaluate the feeding intensity of fish.

[0060] Average distance from individuals to the water surface (ADS) is crucial in aquaculture. When not feeding, fish typically swim in the lower layers of the water. However, scattering feed forces them to rise to the surface to feed. Generally, the more intense the feeding, the more individuals rise to the surface, and the greater the degree of surfacing, resulting in a smaller average distance from the surface. Therefore, this characteristic describes the feeding intensity of fish. The specific calculation formula is as follows:

[0061]

[0062] In the formula: This represents the average distance between an individual and the water surface. This represents the distance from the i-th individual to the water surface. This represents the number of individual fish.

[0063] Fish tilt (FT) is a characteristic of fish swimming horizontally when not feeding. However, when consuming food at the water's surface, fish exhibit a tilted posture with their heads above water and their bodies submerged. Therefore, the FT value increases significantly during feeding, describing the intensity of the fish's feeding. The specific calculation formula is as follows:

[0064]

[0065] In the formula: The angle of the fish's body tilt. Let x be the x-coordinate of the fish head. Let be the ordinate of the fish head. Let x be the x-coordinate of the fish tail. The vertical coordinate is the fish tail.

[0066] The proportion of aggregation area (PAA) is a feature observed in fish feeding, where they tend to cluster and block each other out of food. Generally, the more intense the feeding, the larger the PAA and its proportion. Therefore, this invention extracts this feature to identify and assess the feeding intensity of fish groups. The specific calculation formula is as follows:

[0067]

[0068] In the formula: This represents the percentage of the area within the cluster. The area of ​​the cluster. This represents the total area of ​​the fish habitat.

[0069] The proportion of aggregated individuals (PAI) is another indicator used to measure the aggregation degree of a fish school. Unlike the proportion of aggregation area, it can correct for identification errors that may occur with area measurement. The aggregation index of the fish school is calculated by statistically analyzing the number of aggregated and non-aggregated individuals, and the PAI is ultimately used as one of the indicators to identify the feeding intensity of the fish school. The specific calculation formula is as follows:

[0070]

[0071] In the formula: The percentage of individuals in a cluster. The number of individuals in the cluster.

[0072] S3. Perform feature point matching on fish feature points in fish feeding images, reconstruct the three-dimensional coordinates of fish feature points, and calculate the three-dimensional spatial features of fish based on the three-dimensional coordinates.

[0073] Because two-dimensional spatial features only provide two-dimensional information in a plane and lack depth information in three-dimensional space, they cannot fully and accurately represent the true spatial behavior and state of fish. Furthermore, aquaculture feeding methods are mostly fixed-point feeding, causing fish to gather around the feed when feeding, whereas when not feeding, fish swim freely and are relatively dispersed. Therefore, the gathering phenomenon when fish are feeding means that the distance between individual fish is significantly reduced compared to when they are not feeding, and the fish school dispersion will also decrease accordingly. Feed will sink after absorbing water, so individual fish often exhibit the behavior of first ingesting the sinking feed below the feeding point and then gradually rising to feed. Based on the above three-dimensional information and behavioral characteristics of fish, in order to obtain fish information more comprehensively and accurately and analyze fish spatial characteristics, this invention designs three three-dimensional spatial features—Gen Dispersion (GD), Average Individual Distance (AID), and Individual-Feeding-Point Relationship (RFPR)—for fish feeding behavior analysis and feeding intensity assessment.

[0074] Please refer to Figure 2 To obtain the three-dimensional spatial features of fish, the method first reads the feature point information of the fish's binocular images. Specifically, it reads the fish's feature points from the binocular images and performs a feature point matching operation based on the reconstruction of missing feature points, restoring the missing feature points and matching them accordingly. Then, it reconstructs the three-dimensional coordinates of the feature points based on the matching results. On this basis, a weighted triangulation algorithm is used to accurately obtain the fish school dispersion, the average distance between individuals, and the relationship between individuals and feeding points. This method solves the matching failure problem caused by missing feature points and enhances the ability to represent the spatial dispersion of fish by introducing a weighted triangulation algorithm, greatly improving the accuracy of fish feeding intensity assessment.

[0075] Feature point matching refers to finding corresponding feature points in two images and establishing their correspondence. Accurate feature point matching is crucial for reconstructing the 3D coordinates of fish feature points, which are essential for extracting the 3D spatial features of fish. However, in fish feeding images acquired using binocular cameras, overlapping and occlusion between individual fish can lead to missed detections, resulting in missing feature points. Traditional feature point matching algorithms are affected by these missing feature points, easily leading to matching failures and failing to reconstruct the true spatial information of the feature points, thus impacting the accuracy of 3D spatial feature extraction. To address this issue, this invention proposes a feature point matching method based on missing feature point reconstruction. This method calculates the coordinates of missing feature points based on the disparity of binocular images, reconstructs the missing feature points, and ultimately achieves accurate matching of feature points in binocular images.

[0076] Please refer to Figure 3 The method involves spatially sorting fish feature points in the left and right images of a fish feeding image. Using the fish feature points in the left image as a reference, the coordinates of matching points in the right image are predicted using the disparity values ​​of the corresponding positions. Fish feature points are then searched for in the neighborhood around these matching point coordinates. If a fish feature point exists, a feature point matching pair is established; otherwise, the corresponding position in the right image is marked as a missing feature point. This process is repeated, using the fish feature points in the right image as a reference, to detect missing feature points in the left image. For missing feature points, the disparity consistency principle is used to calculate the position coordinates inversely using the corresponding fish feature points and disparity values ​​to obtain the restored feature points. The spatial distance and distribution consistency between the restored feature points and neighboring fish feature points are calculated for validity verification. The verified restored feature points and the original fish feature points are then input into the ORB feature point matching algorithm to complete feature point matching. Based on the principle of binocular vision, the three-dimensional coordinates of the fish feature points are reconstructed.

[0077] Based on the three-dimensional coordinates of fish feature points, calculate the three-dimensional spatial features of fish respectively:

[0078] The group dispersity (GD) is calculated as follows:

[0079] To accurately describe the discreteness of fish, this invention proposes a weighted triangulation algorithm suitable for three-dimensional space. Fish feature points are used as vertices of Delaunay triangles for spatial triangulation, resulting in multiple tetrahedral structures containing fish feature points. The triangular faces containing fish feature points within each tetrahedral structure are then divided. The triangular faces contributing to the formation of multiple tetrahedra are designated as shared triangles, while the remaining triangular faces are designated as boundary triangles. (Please refer to...) Figure 4Taking point A as an example, there are five triangular faces including point A, which are triangles ABC, ACD, ACE, ABD and ABE respectively. A triangle that forms two tetrahedrons, such as triangle ABC, is defined as a shared triangle, and other triangles are defined as boundary triangles, and a weight formula is established:

[0080]

[0081] Where: is the weight, is the number of boundary triangles, is the number of shared triangles; when a>b, it means that the number of boundary triangles is relatively large, the feature point is located at the boundary of the fish school, the weight k will increase accordingly, so as to strengthen the weight of boundary individuals in the group and improve the sensitivity to changes in boundary features; when a<b, it means that the number of shared triangles is relatively large, the feature point is located in the fish aggregation area, and the weight will decrease accordingly, so as to balance the abnormal feature value caused by repeated statistics of multiple feature points.

[0082] After obtaining the weights, multiply the area of all triangles containing fish feature points by their weight k, so as to further calculate the average surface area of the tetrahedral mesh, which is used as the feature quantity describing the spatial dispersion of fish. In this way, the sensitivity of the tetrahedral mesh to local feature changes of the fish school can be improved, so as to more accurately reflect the dispersion degree of the fish school in space. The calculation formula of fish school dispersion is established as follows:

[0083]

[0084] Where: is the fish school dispersion, is the weight of the i-th fish feature point, is the surface area of the tetrahedron formed by the participation of the i-th fish feature point, is the number of individual fish.

[0085] Average inter-individual distance (AID), according to the aggregation phenomenon of fish during feeding, the average inter-individual distance is obviously an important indicator for evaluating fish feeding intensity, and the specific calculation formula is:

[0086]

[0087] Where: is the average inter-individual distance, is the distance between the i-th fish feature point and the j-th fish feature point.

[0088] The relationship between fish and the feeding point (RFPR) is used to describe the distance-angle relationship between the fish and the feeding point. Due to the water absorption and sedimentation of feed, fish often exhibit a behavior where they first ingest the sedimented feed from below the feeding point and then gradually rise to the surface to feed. The specific calculation formula is as follows:

[0089]

[0090] In the formula: For the relationship between individuals and feeding sites, The distance between an individual fish and the feeding point. The angle of inclination of the line connecting the individual fish and the feeding point.

[0091] S4. The autoencoder algorithm is used to fuse the two-dimensional and three-dimensional spatial features of fish to obtain the fused features of fish. The empirical mode decomposition algorithm is used to decompose the fused features of fish to obtain the corresponding intrinsic mode function components. The range of the intrinsic mode function components is calculated. An adaptive equal division strategy based on the range is adopted for the intrinsic mode function components to divide the intrinsic mode function components into multiple intervals. Each interval corresponds to a fish feeding intensity level, thus completing the fish feeding intensity assessment.

[0092] This invention uses an unsupervised autoencoder algorithm to reduce and fuse the extracted two-dimensional and three-dimensional spatial features of fish, remove redundant information from the data, and automatically extract important features related to fish feeding behavior. Finally, the fish spatial features are represented as a low-dimensional fusion vector, which contains the core features of the input data and retains as much information as possible related to the intensity of fish feeding.

[0093] After obtaining the fish fusion features, this invention uses the Empirical Mode Decomposition (EMD) algorithm to process these features. EMD is a data-driven time series analysis method specifically designed for analyzing nonlinear and non-stationary signals. It can decompose complex signals or time series into several intrinsic mode functions (IMFs). EMD can decompose data based on its own temporal characteristics, without relying on any predefined basis functions, thus exhibiting significant advantages in processing non-stationary and nonlinear data.

[0094] The main process of the empirical mode decomposition method includes: extracting local extrema, constructing upper and lower envelopes, calculating the mean line, extracting intrinsic mode functions, repeating iterations, and resetting the signal, as detailed below.

[0095] For the input signal First, local maxima and local minima are extracted from the signal. Then, using these local maxima, the upper envelope of the signal is constructed via spline interpolation. and lower envelope The upper envelope is formed by interpolation curves at local maxima, and the lower envelope is formed by interpolation curves at local minima. Based on this, the mean lines of the upper and lower envelopes are calculated:

[0096]

[0097] in, This is the mean line of the upper and lower envelopes.

[0098] After obtaining the mean line Then, calculate its relationship with the original signal. The difference is used to obtain the intrinsic mode functions. .

[0099]

[0100] examine Does it meet the characteristic conditions of the IMF component? If not, then... As a new Repeat the above steps until... Without further changes, we now obtain the first IMF component:

[0101]

[0102]

[0103] Then As a new Repeat the above steps to filter out a series of IMF components. Finally, the original signal... It can be represented as the sum of the various IMFs, plus a residual term. .

[0104]

[0105] After processing the fish fusion features using the Empirical Mode Decomposition (EMD) algorithm, the trend of fish feeding intensity changes can be obtained from the IMF components. At the same time, by calculating the range of the IMF, the fluctuation range of fish feeding intensity can be obtained. Then, based on the range range, the fish feeding behavior is divided into multiple different intensity levels, thereby achieving a refined assessment of feeding intensity.

[0106] The formula for calculating the range is:

[0107]

[0108] In the formula: For the i-th IMF component Extremely poor.

[0109] After obtaining the range of the intrinsic mode function components, in order to achieve a refined assessment of fish feeding intensity, this invention adopts an adaptive equal division strategy based on the range. It uses the criterion of minimizing intra-class differences to evaluate the grading effect under different numbers of equal division points, and determines the optimal scheme for multi-level fish assessment by calculating the inter-class and intra-class differences of each candidate division scheme.

[0110] Boundaries between equal zones It can be represented as:

[0111]

[0112] in, This represents the minimum value of the intrinsic mode function components. To divide into equal parts, Indicates the first Divide the intervals equally. This represents the range of the intrinsic mode function components. The corresponding intervals for each feeding intensity level are represented as follows: .

[0113] To ensure higher resolution for the feeding intensity classification compared to traditional four-category methods, while avoiding a decrease in statistical significance due to an excessive number of classifications, based on the empirical rule of nonparametric modeling, the maximum number of groups should not exceed the square root of the sample size. Therefore, this invention sets the range of values ​​for the number of equal divisions to be [value missing]. .

[0114] Finally, based on the corresponding equal intervals, the feeding intensity of fish was divided into several refined levels.

Claims

1. A method for assessing fish feeding intensity based on binocular vision, characterized in that, Includes the following steps: S1. Use a binocular camera to acquire images of fish feeding, calibrate and correct the images, preprocess the corrected images, and extract fish outlines and feature points. S2. Calculate the two-dimensional spatial features of fish based on fish outlines and fish feature points; S3. Perform feature point matching on fish feature points in fish feeding images, reconstruct the three-dimensional coordinates of fish feature points, and calculate the three-dimensional spatial features of fish based on the three-dimensional coordinates. S4. The autoencoder algorithm is used to fuse the two-dimensional and three-dimensional spatial features of fish to obtain the fused features of fish. The empirical mode decomposition algorithm is used to decompose the fused features of fish to obtain the corresponding intrinsic mode function components. The range of the intrinsic mode function components is calculated. An adaptive equal division strategy based on the range is adopted for the intrinsic mode function components to divide the intrinsic mode function components into multiple intervals. Each interval corresponds to a fish feeding intensity level, thus completing the fish feeding intensity assessment.

2. The method for assessing fish feeding intensity based on binocular vision according to claim 1, characterized in that, The calibration and correction of fish feeding images includes: performing camera calibration, obtaining the camera's intrinsic and extrinsic parameters and distortion parameters, using the intrinsic and extrinsic parameters and distortion parameters to correct the fish feeding images, and calculating the binocular image disparity after stereo matching of the corrected fish feeding images to generate a corresponding disparity map.

3. The method for assessing fish feeding intensity based on binocular vision according to claim 1, characterized in that, The preprocessing of the corrected fish feeding images includes: The target fish body is extracted from the fish feeding image by image segmentation. The image is enhanced by linear transformation and histogram equalization to improve the clarity of the target fish body. Then the image is binarized, and the fish contour is extracted and drawn by contour detection function. The centroid of the fish contour is extracted as the fish feature point.

4. The method for assessing fish feeding intensity based on binocular vision according to claim 1, characterized in that, The two-dimensional spatial characteristics of the fish include the average distance between the individual and the water surface, the tilt angle of the fish body, the area ratio of the aggregation region, and the proportion of the number of individuals in the aggregation region.

5. The method for assessing fish feeding intensity based on binocular vision according to claim 4, characterized in that, The formula for calculating the average distance between the individual and the water surface is: In the formula: This represents the average distance between an individual and the water surface. This represents the distance from the i-th individual to the water surface. The number of individual fish; The formula for calculating the fish's body tilt angle is as follows: In the formula: The angle of the fish's body tilt. Let x be the x-coordinate of the fish head. Let be the ordinate of the fish head. Let x be the x-coordinate of the fish tail. The vertical coordinate is the tail of the fish. The formula for calculating the area ratio of the clustered region is: In the formula: This represents the percentage of the area within the cluster. The area of ​​the cluster. This represents the total area of ​​the fish habitat. The formula for calculating the proportion of the number of clustered individuals is: In the formula: The percentage of individuals in a cluster. The number of individuals in the cluster.

6. The method for assessing fish feeding intensity based on binocular vision according to claim 1, characterized in that, The feature point matching of fish feature points in fish feeding images includes: The fish feature points in the left and right images of a fish feeding image are ordered spatially. Using the fish feature points in the left image as a reference, the coordinates of matching points in the right image are predicted using the disparity values ​​of the corresponding positions. Fish feature points are then searched for in the neighborhood around these matching point coordinates. If a fish feature point exists, a feature point matching pair is established; otherwise, the corresponding position in the right image is marked as a missing feature point. The above steps are repeated, using the fish feature points in the right image as a reference, to detect missing feature points in the left image. For missing feature points, the position coordinates are calculated in reverse using the disparity value and the corresponding fish feature point to obtain the restored feature point. The spatial distance and distribution consistency between the restored feature point and neighboring fish feature points are calculated for validity verification. The verified restored feature points and the original fish feature points are then input into the ORB feature point matching algorithm to complete feature point matching. Finally, the three-dimensional coordinates of the fish feature points are reconstructed based on the principle of binocular vision.

7. The method for assessing fish feeding intensity based on binocular vision according to claim 6, characterized in that, The three-dimensional spatial characteristics of the fish include the dispersion of the fish school, the average distance between individuals, and the relationship between individuals and feeding points.

8. The method for assessing fish feeding intensity based on binocular vision according to claim 7, characterized in that, The fish swarm dispersion is calculated as follows: Using fish feature points as vertices of the Delaunay triangle, spatial triangulation is performed to obtain multiple tetrahedral structures containing fish feature points. The triangular faces containing fish feature points within each tetrahedral structure are then divided, with the triangular faces contributing to the formation of multiple tetrahedra designated as shared triangles and the remaining triangular faces designated as boundary triangles. A weighting formula is then established. In the formula: As weight, The number of boundary triangles, To share the number of triangles; Based on the aforementioned weights, a formula for calculating the dispersion of the fish population is established: In the formula: The dispersion of the fish school The weight of the i-th fish feature point is... Let be the surface area of ​​the tetrahedron formed by the i-th fish feature point. The number of individual fish; The formula for calculating the average spacing between individuals is: In the formula: The average spacing between individuals Let be the distance between the i-th fish feature point and the j-th fish feature point; The formula for calculating the relationship between the individual and the feeding point is: In the formula: For the relationship between individuals and feeding sites, The distance between an individual fish and the feeding point. The angle of inclination of the line connecting the individual fish and the feeding point.

9. The method for assessing fish feeding intensity based on binocular vision according to claim 1, characterized in that, The adaptive equal division strategy based on range includes: Based on the range of intrinsic mode function components, the intervals are divided into equal parts. The number of equal parts is set to obtain the division schemes corresponding to the number of equal parts. The grading effect of different division schemes is evaluated by using the intraclass minimum difference criterion, and the optimal scheme for fish feeding intensity assessment is determined.

10. The method for assessing fish feeding intensity based on binocular vision according to claim 9, characterized in that, The range of the number of equally divided intervals is: N represents the number of individual fish.

Images