Machine Vision-Based 3D Reconstruction and Weight Estimation Method for Cultured Biomass

Abstract

This invention belongs to the field of visual monitoring and processing technology for aquaculture organisms. Specifically, it discloses a machine vision-based method for three-dimensional reconstruction and weight estimation of aquaculture biomass, including: calculating the polarization angle of reflected light from the water surface, synchronously adjusting the direction of the underwater light source and the camera's polarization filter to achieve dynamic polarization suppression, effectively blocking specular reflection from the water surface; improving the positioning accuracy of key feature points of the fish body in complex image backgrounds by locating the fin base connection point and combining it with a re-search mechanism in the adjacent area along the spine extension direction; retrieving the body shape correction constant of the current aquaculture fish species, scaling and correcting the physical weight estimate, correcting the weight deviation of bones / viscera not covered in the physical volume calculation, and thus generating the final biomass weight value. This achieves automated three-dimensional reconstruction and weight estimation of aquaculture biomass, improving the accuracy and reliability of aquaculture biomass monitoring.

Description

Technical Field

[0001] This invention belongs to the field of visual monitoring and processing technology for aquaculture organisms, and relates to a method for three-dimensional reconstruction and weight estimation of aquaculture biomass based on machine vision. Background Technology

[0002] In underwater biological monitoring and aquaculture management, real-time and accurate acquisition of fish morphological characteristics and movement trajectories is crucial for assessing fish condition, optimizing the aquaculture environment, and improving aquaculture efficiency. However, the unique characteristics of the underwater environment, such as light reflection from the water surface, result in poor image quality, blurred fish features, and difficulty in tracking movement trajectories. These problems severely restrict the in-depth development of image-based fish condition research. Therefore, developing a method based on advanced image processing technology that can efficiently and accurately extract fish features and track their movement trajectories is of great significance for promoting the intelligent development of underwater biological behavior research and aquaculture management.

[0003] To address these challenges, existing technologies, such as employing high-sensitivity sensors and optimizing optical lens design, focus on improving the performance of underwater imaging equipment to enhance image quality. Furthermore, some studies have incorporated manual annotation and expert experience to assist the image analysis process. These methods have alleviated the difficulties of underwater image processing to some extent.

[0004] Despite some progress made by traditional methods, the following drawbacks remain: 1. While basic underwater image acquisition and simple analysis functions have been achieved, there are still significant shortcomings in improving image quality and extracting clear features in complex underwater environments, resulting in blurred fish features and difficulty in accurately tracking movement trajectories. 2. Although manual or semi-automatic extraction of fish features has been achieved, there are still significant shortcomings in fully automated and efficient feature recognition and classification. Reliance on manual annotation or complex algorithm parameter tuning makes it difficult to meet the needs of large-scale real-time monitoring, and there is also a certain cost burden in the use of manual equipment. Summary of the Invention

[0005] To overcome the above-mentioned deficiencies of the prior art and to achieve the above objectives, the present invention proposes the following technical solution: a three-dimensional reconstruction and weight estimation method for aquaculture biomass based on machine vision, comprising the following: S1, controlling two sets of industrial cameras to simultaneously capture images of the aquaculture area at different shooting frequencies; the first set of cameras continuously captures images at a high frequency to generate a heat map of fish movement distribution; the second set of cameras captures images at a normal frequency and triggers exposure actions through the partitioning of the heat map.

[0006] S2. Rotate and adjust the angle of the polarizer in front of the second set of camera lenses, and control the polarization direction of the light source in the aquaculture area so that the light received by the camera and the light reflected from the water surface form an orthogonal polarization state.

[0007] S3. Extract the key points of the skeleton of multiple fish bodies from the high-definition images captured after the second set of camera polarization adjustment, and then align the key points of the skeleton with the preset skeleton template to automatically correct the deformed parts of the fish body.

[0008] S4. Using the aligned skeleton nodes as boundaries, divide the fish body into head, torso, and tail segments along the length of the fish body. Based on the spacing between key skeleton points in each segment, deduce the maximum thickness of the fish body and obtain the major and minor axes of the ellipse for each segment of the fish body.

[0009] S5. Calculate the volume of each segment by elliptical rotation integral, and calculate the estimated physical weight of a single fish by combining the density coefficient of the corresponding segment.

[0010] S6. Retrieve the body size correction constant of the current farmed fish species, scale and correct the physical weight estimate, and output the final biomass weight value.

[0011] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention calculates the polarization angle of the reflected light from the water surface and simultaneously adjusts the direction of the underwater light source and the camera polarization filter to achieve dynamic polarization suppression effect, ensuring that the camera only receives the direct light that penetrates the water body, while effectively blocking the mirror reflection light from the water surface, significantly improving the contrast and clarity of the underwater image.

[0012] (2) The present invention significantly improves the positioning accuracy of key feature points of fish in complex image backgrounds by using the fin base connection point positioning method and combining the re-search mechanism of the adjacent area in the direction of the spine extension. It reduces the risk of positioning failure due to local errors and improves the accuracy and efficiency of fish morphological feature extraction.

[0013] (3) This invention realizes automated three-dimensional reconstruction and weight estimation of aquaculture biomass by locating key points of the skeleton, correcting deformation, calculating the major and minor axes of the ellipse and the volume integral of the image, reducing the need for manual annotation and complex algorithm parameter tuning, realizing fully automated and efficient feature recognition and classification, while reducing the cost of manual equipment use.

[0014] (4) This invention retrieves the body shape correction constant of the current farmed fish species, scales and corrects the physical weight estimate, outputs the final biomass weight value, corrects the weight deviation of bones / internal organs not covered in the physical volume calculation, optimizes the weight estimation accuracy, makes the weight estimation result closer to the true value, and improves the accuracy and reliability of farmed biomass monitoring. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments 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.

[0016] Figure 1 This is a schematic diagram of the implementation steps of the method of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 As shown, the machine vision-based three-dimensional reconstruction and weight estimation method for aquaculture biomass proposed in this invention includes the following: S1, controlling two sets of industrial cameras to simultaneously capture images of the aquaculture area at different shooting frequencies; the first set of cameras continuously captures images at a high frequency to generate a heat map of fish movement distribution; the second set of cameras captures images at a normal frequency and triggers exposure actions through the heat map partitioning.

[0019] The two sets of industrial cameras refer to visible light imaging equipment groups with fixed physical locations and consistent optical parameter calibration; synchronous shooting means that all cameras are controlled by the same timing chip to control the shutter action; high frequency is defined as the camera acquisition frame rate being at least D times the maximum swimming speed of the fish, where D is a constant and swimming speed is the longest body length that can be moved per unit time.

[0020] In a preferred embodiment, controlling two sets of industrial cameras to simultaneously capture images of the aquaculture area at different shooting frequencies specifically involves configuring two sets of industrial cameras deployed on the water surface and underwater as image acquisition devices. For example, the first set of cameras is set to continuously acquire dynamic images of the fish school at a shooting rate of more than 500 frames per second, while the second set of cameras is set to standby at a normal shooting rate.

[0021] The continuous image sequence captured by the first set of cameras is input into the motion analysis unit. At the same time, the center position of the fish's eye is used as a specific marker point for the fish body to divide the image into different sections.

[0022] The linear distance of movement of the corresponding fish-specific marker points in each fish-body partition in adjacent images is compared frame by frame. When the linear distance of movement exceeds the preset distance, it is determined that the position of the fish-body partition has been significantly displaced. At this time, the displacement of the fish-body partition is superimposed and counted to form a partition count value matrix based on pixels.

[0023] The frequency of significant displacement of fish body regions in each image per unit time is counted to generate a two-dimensional distribution map reflecting the intensity of fish movement within the aquaculture area, denoted as the fish movement distribution heatmap.

[0024] Based on the displacement count values ​​of each fish body partition in the partition count value matrix, the total displacement count values ​​of different pixel blocks are marked with different color depths in the fish movement distribution heatmap. The lightest color region represents the total displacement count value of the fish body that is lower than a set critical threshold and is recorded as a low movement intensity region.

[0025] The exposure trigger mechanism for the second group of cameras is set to send the coordinates of the center point of the low-motion intensity area in the fish movement distribution heatmap to the exposure controller of the second group of cameras.

[0026] The exposure controller of the second set of cameras generates a rectangular focus frame based on the center point coordinates of the low-motion intensity area. This frame drives the autofocus module of the second set of cameras to lock onto the area and release the shutter, executing the exposure action. In the low-motion intensity area, the fish's movement is relatively slow, and the frequency of position changes is low. This makes it easier for the camera to capture a clear image because the fish will not undergo significant positional changes in a short period, thus reducing the possibility of image blur or distortion.

[0027] For example, when a group of crucian carp in the target aquaculture pond swims at a typical speed of 0.5 m / s, the first set of cameras is set to acquire images at a frame rate of 600 frames / second. Every five consecutive frames received, the linear distance of the fish eye marker point between the first and last frames is detected. If the linear distance exceeds a preset distance, it is determined that the fish body's position has significantly shifted, and the count of the corresponding pixel block in the heat map is incremented by 1. After 10 seconds, a 1920×1080 pixel heat map is formed, where pixel blocks with a count of less than 3 are marked as light blue low-motion-intensity areas. After receiving the signal coordinates of the light blue low-motion-intensity area, the second set of cameras immediately rotates the gimbal to center the area and triggers the exposure action after autofocusing. This process ensures that the second set of cameras only acquires high-definition textures of the fish body in a relatively stationary state, avoiding motion blur caused by swimming.

[0028] S2. Rotate and adjust the angle of the polarizer in front of the second set of camera lenses, and control the polarization direction of the light source in the aquaculture area so that the light received by the camera and the light reflected from the water surface form an orthogonal polarization state, so as to suppress the interference of water surface reflection. The interference of water surface reflection refers to the strong specular reflection generated by the ripples on the surface of the aquaculture water.

[0029] In a preferred embodiment, S2 includes: when the second set of cameras performs an exposure operation, detecting the polarization angle of the current water surface reflected light by a photoelectric sensor located above the water surface, wherein the polarization angle of the current water surface reflected light refers to the angle between the vibration direction of the electric field vector in the water surface reflected light and the horizontal plane.

[0030] Increase the polarization angle of the light reflected from the current water surface. The target rotation angle of the polarizing filter for the camera is then used to drive the polarizing filter in front of the second set of camera lenses to rotate around the optical axis at the target rotation angle via a micro stepper motor. At the same time, a fixed-angle polarizing film is installed in front of each underwater LED light. The polarization direction of the light emitted by the underwater LED array in the aquaculture area is controlled by the central controller to be parallel to the direction of the camera filter, so that the polarization direction of the light entering the camera lens is perpendicular to the polarization direction of the light reflected from the water surface.

[0031] Among them, the polarization direction of the light reflected from the water surface refers to the vibration direction of the electric field vector of the light during the reflection process on the water surface, which is measured by a photoelectric sensor; the polarization direction of the light entering the camera lens refers to the vibration direction of the electric field vector of the light that can pass through the camera lens after passing through the polarization filter.

[0032] The criterion for determining the formation of orthogonal polarization states is: when the polarization angle of the light reflected from the water surface is... At that time, the target rotation angle of the camera polarization filter was set to This ensures that the camera only receives direct light that penetrates the water, while the mirror-reflected light from the water surface is physically blocked due to its orthogonal polarization, thus improving the image's clarity and contrast.

[0033] For example, in a tilapia farming cage scenario, after triggering the second set of cameras to perform an exposure action on a low-motion-intensity area, the photoelectric sensor located above the water surface detects that the polarization angle of the light reflected from the water surface is... The camera controller immediately drives the lens polarization ring to rotate to Position, synchronously switch the underwater LED array to In this polarization direction, the light received by the camera consists of two parts: one part is the direct underwater light matched with the polarizer's direction, which carries the image of the fish; the other part is the water surface reflected light orthogonal to the polarizer, which is blocked by the polarizer. At this time, the brightness of the reflected area in the image decreases, making the fish tail texture, which was originally covered by reflection, clearly visible, and improving the accuracy of fish eye key point recognition.

[0034] This invention achieves dynamic polarization suppression by calculating the polarization angle of reflected light from the water surface and simultaneously adjusting the direction of the underwater LED light source and the camera's polarization filter. This ensures that the camera only receives direct light penetrating the water, while effectively blocking specular reflections from the water surface. This significantly improves the contrast and clarity of underwater images, especially achieving high signal-to-noise ratio imaging in low-motion-intensity areas. It provides high-quality data support for fish behavior research and solves the image blurring problem caused by water surface reflections in traditional underwater imaging techniques.

[0035] S3. Extract multiple key skeletal points of the fish body from the high-definition images captured after the second set of camera polarization adjustment. The key skeletal points are anatomical locations that are not easily deformed when the fish moves, including the center of the fish eye and the connection point at the base of the fin. The connection point at the base of the fin includes the dorsal fin, pelvic fin, caudal fin, pectoral fin, and anal fin. Then, align the key skeletal points with the preset skeletal template to automatically correct the deformed parts of the fish body. The skeletal template includes the size ratio rules between the fish body skeletal nodes. The skeletal nodes include, but are not limited to, key skeletal points, the widest point of the head node, the front end of the skull node, the starting point of the spine node, and the end of the tailbone node.

[0036] Among them, position alignment refers to minimizing the total error of the distance between the center point of the fish eye and the front node of the skull, and the distance between the connection point of the fin base and the starting node of the spine through iterative calculation; automatic correction of deformed parts of the fish body refers to resampling and inserting nodes according to the template ratio when the measured node distance exceeds the threshold, and the position of the new node is obtained by weighted average of the coordinates of adjacent nodes.

[0037] In a preferred embodiment, the extraction of key skeletal points of multiple fish bodies from the high-definition images captured after the second set of camera polarization adjustment includes: enhancing the outline edges of the fish bodies using the gradient difference method; and for each visible fish body, identifying the two most prominent circular dark spots in its head region as candidate fish eye regions.

[0038] Search along the direction of the fish's spine for the junctions between each fin and the trunk, and mark them as fin base connection points. The fins include the dorsal fin, pelvic fin, caudal fin, pectoral fin, and anal fin.

[0039] Obtain the preset fish body size ratio rules. If the ratio of the center distance between the two sides of the fish eye area to the distance of the widest part of the head conforms to the fish body size ratio rules, then the corresponding skeleton key point of the visible fish body is determined to be established. At this time, the candidate fish eye area and each fin part are the corresponding skeleton key points of the visible fish body, and their image coordinates are recorded. Otherwise, the corresponding skeleton key point of the visible fish body is determined to be a positioning error, and the corresponding skeleton key points in the neighboring area are searched again.

[0040] Among them, gradient difference method refers to calculating the sum of the absolute values ​​of the gray level difference between each pixel in the image and its eight neighboring pixels. If the sum exceeds a set threshold, it is marked as a contour edge point; circular dark spot is defined as a connected region in the image where the gray level of a pixel is lower than the average gray level of the surrounding pixels by a certain proportion and the pixel distribution meets the requirements of circular fitting error; fin base connection point refers to the vertex of the angle formed by the junction of the bottom of the fin and the side of the fish body, which is located by detecting the maximum value of the lateral gradient along the extension direction of the spine; fish body size ratio rule refers to the upper and lower limit thresholds of the distance between adjacent bone nodes, such as the distance between two nodes in the head is always 10%-12% of the body length, which is statistically derived based on the anatomical database of common farmed fish, including but not limited to the ratio of the distance between the center of the two eye areas to the distance of the widest part of the head; neighboring region refers to a rectangular, circular or other shaped region that extends outward from the coordinates of the initially located fin base connection point by several pixels.

[0041] For example, when processing high-resolution images of grass carp, two dark circular areas were detected in the head region and marked as candidate eye areas on both sides. The distance between them was measured to be 6 mm. Starting from the candidate eye areas on both sides, the trunk was detected posteriorly along the direction of the spine. A transverse gradient abrupt increase point was found at a certain distance from the center of the eye. Its position coincided with the starting point of the dorsal fin recorded in the anatomical atlas. At the same time, the width of the widest part of the head was measured to be 8.1 mm. The ratio of the distance between the two eyes (6 mm) to the width of the widest part of the head (8.1 mm) was calculated to be 0.74, which is within the range of 0.75-0.85 of the fish size ratio rule. Therefore, the candidate eye areas on both sides and the connection point of the fin base were confirmed as effective skeletal key points.

[0042] In a further preferred embodiment, the step of aligning the key points of the skeleton with a preset skeleton template and automatically correcting the deformed parts of the fish body includes: establishing a mapping relationship between the key points of the skeleton and the skeletal nodes in the skeleton template, including but not limited to matching the center point of the fish eye to the front node of the skull, matching the connection point of the fin base to the starting node of the spine, and generating matching point pairs for each part of the fish body. The skeleton template refers to a pre-stored topological set of fish skeleton key points with no less than 7 nodes, and its fish body size ratio rules are determined based on the natural body shape measurement data of the fish species.

[0043] Calculate the sum of the Euclidean distances for each pair of matching points, and achieve position alignment by minimizing this sum through rotation and scaling transformations, thus obtaining the corrected 3D coordinates of the deformed parts of the fish body.

[0044] If, during the alignment process, the distance between adjacent bone nodes in the middle region of the spine exceeds the maximum allowable deviation of the distance between the corresponding bone nodes of the fish, new nodes are automatically interpolated and generated according to the fish size ratio rule of the distance between the corresponding bone nodes of the fish in the skeleton template, thereby forcibly restoring the biological rationality of the linear structure of the fish.

[0045] Specifically, the formula for minimizing the rotation and scaling transformation is as follows: Its output is the optimal rotation matrix. and scaling factor In the formula, i is the number of the skeleton key point. , Extracting the first from the image Three-dimensional coordinates of key points in the skeleton. For the first in the skeleton template The preset coordinates of key skeletal points were obtained from a fish species anatomy database. Given a rotation matrix, the optimal solution is obtained through singular value decomposition, ensuring that the rotated 3D coordinates spatially match the coordinates of the skeleton keypoints extracted from the image as closely as possible. This is a preset scaling factor used to adjust the size of the skeleton template, minimizing the error between the resized 3D coordinates and the coordinates of the skeleton key points in the image.

[0046] Different biological parts have specific anatomical structures and features, and the ratio of adjacent node spacing is an important parameter reflecting these features. Setting constraints helps to ensure that the deformed skeleton template matches the actual anatomical structure of the organism, guaranteeing that the anatomical features of the organism are not destroyed during spatial registration. Therefore, the constraint on the ratio of adjacent node spacing in the skeleton template is set as follows: In the formula Represents two adjacent nodes in the skeleton template and The Euclidean distance between them This represents the cumulative distance between each node from the second node to the Nth node and its preceding node, reflecting the overall length of the skeleton template. These represent the minimum and maximum percentage thresholds for the distance between adjacent nodes, respectively, derived from anatomical statistics. For example, in the head segment, the following rules apply. If the calculated percentage of the distance between adjacent nodes is less than The distance between adjacent nodes is too small; greater than If the distance between adjacent nodes is too large, this constraint can prevent unreasonable deformation.

[0047] For example, for a 400 mm long bass, the three-dimensional coordinates of the left eye (3,2,50), the right eye (7,2,50), and the base of the dorsal fin (5,10,120) were extracted. A bass skeleton template was retrieved, which specifies that the distance between the eyes should be 0.05-0.06 mm of the body length, and the distance between the dorsal fin and the gill cover should be 0.18-0.2 mm of the body length. Alignment and matching revealed that the measured distance between the eyes (4 mm) was much lower than the lower limit of the body length (which should be 20 mm). Based on this, it was determined that the head size was distorted due to the shooting angle. The three-dimensional coordinates of the left and right eyes were reset to (4,2,55) and (8,2,55) according to the proportional rules, while keeping the position of the base of the dorsal fin unchanged. After correction, the key points of the skeleton were restored to a linear arrangement structure that conforms to biological proportions.

[0048] This invention significantly improves the positioning accuracy of key feature points of fish in complex image backgrounds by using a fin base connection point positioning method combined with a re-search mechanism of the adjacent area in the direction of the spine extension. It reduces the risk of positioning failure due to local errors, provides a reliable basis for subsequent behavioral analysis or morphological measurement, and improves the accuracy and efficiency of fish morphological feature extraction.

[0049] S4. Using the aligned skeleton nodes as boundaries, divide the fish body into head, torso, and tail segments along the length of the fish body. Based on the spacing between key skeleton points in each segment, deduce the maximum thickness of the fish body and obtain the major and minor axes of the ellipse for each segment of the fish body.

[0050] Among them, the spacing between key points of the skeleton in each section refers to the straight-line distance in the width direction between the two farthest key points of the skeleton in the corresponding section; the maximum fish body thickness represents the thickest dimension of the fish body perpendicular to the major axis, and its value must not exceed the maximum thickness ratio of the corresponding section of the fish species in the preset rule library; the major axis of the ellipse is defined as the maximum width of the section, and the minor axis is the maximum thickness.

[0051] In a preferred embodiment, the step of dividing the fish body into head, trunk, and tail segments along the length direction using the aligned skeleton nodes as boundaries includes: obtaining the straight-line distance from the front node of the skull to the end node of the tailbone, obtaining the total length of the fish body, and dividing the length by three to obtain the standard segment length.

[0052] Starting from the head, dividing lines are set every standard segment length along the main axis: the first dividing line is located one standard segment length after the head starting point, marked as the end of the head segment; the second dividing line is located one standard segment length after the end of the head segment, marked as the end of the trunk segment; the remaining part is the tail segment. The head segment includes the area where the fish's eyes are located, the trunk segment covers the fin base to the caudal peduncle starting point, and the tail segment includes the caudal fin base node.

[0053] For example, when processing a carp with a body length of 450 mm, the coordinate sequence from head to tail is output as 9 skeletal nodes. The distance between the head and tail nodes is measured to be 450 mm, and the standard length of each segment is calculated to be 150 mm. A first boundary is set at 150 mm from the starting point of the head, and a second boundary is set at 300 mm from the starting point of the head. The head segment contains the first three skeletal nodes, the torso segment contains the middle three nodes, and the tail segment contains the last three nodes.

[0054] In a further preferred embodiment, obtaining the major and minor axes of the ellipse of each segment of the fish body includes: selecting reference points on the central axis of each segment of the fish body based on the three divided segments. For example, the first third of the central axis of the head segment is selected as the reference point, the midpoint of the central axis of the torso segment is selected as the reference point, and the last third of the central axis of the tail segment is selected as the reference point. The central axis is the central axis of the fish body from the head to the tail.

[0055] At the reference point of each section, a local coordinate system is established perpendicular to the central axis of the fish.

[0056] Analyze the spatial distribution of key skeleton points within the current segment, and measure the spatial straight-line distance in the width direction between the two farthest key skeleton points within the segment, using it as a reference value for the minor axis of the ellipse.

[0057] Based on the pre-set constraint ratio of the relationship between the thickness and width of each segment of the fish body, the reference value of the minor axis of the ellipse is adjusted, and the maximum thickness of the fish body is calculated as the minor axis of the ellipse of the corresponding segment. The maximum thickness of the fish body represents the thickest dimension of the fish body perpendicular to the major axis, and its value must not exceed the maximum thickness ratio of the corresponding segment of the fish species in the pre-set rule library. The constraint ratio refers to the limit threshold of the relationship between the thickness and width of each part of the fish body, which is established by statistically analyzing the anatomical measurement data of various common farmed fish.

[0058] Define the maximum width of the segment as the major axis of the ellipse.

[0059] The major and minor axes of the ellipse in each segment of the fish's body were statistically determined.

[0060] Specifically, the formulas for calculating the major and minor axes of the ellipse in each segment are as follows: , ,in For each section number, These correspond to the head, torso, and tail segments, respectively. For the first Major axis of the segmented ellipse, For the first The maximum width of a section is measured by the spatial distance between key points on the skeleton within the section boundary. For the first The minor axis of the segmented ellipse, For the preset fish body The constraint ratio of the thickness to width of a segment, such as the head segment of a fish. , For the first The maximum allowable thickness of a section is preset by the variety template, such as the carp torso. .

[0061] For example, when processing the trunk segment of a silver carp, the distance between the two outermost skeletal key points of the segment was measured to be 85 mm. The anatomical rule base was consulted to find that the upper limit of the trunk thickness is 0.65 of the width. Therefore, the maximum thickness of 55.25 mm = 85 mm × 0.65 was deduced. At the same time, it was verified that this value does not exceed the preset maximum allowable value of 60 mm. Finally, the major axis was 85 mm and the minor axis was 55 mm.

[0062] This invention generates the major and minor axis parameters of the ellipse at the center of the current small slice by linear interpolation using the major and minor axis parameters of the ellipse in adjacent segments. This avoids geometric discontinuities in parameters caused by missing data or measurement errors during the 3D reconstruction of the fish, thus improving the accuracy and efficiency of the reconstruction model. It simplifies the parameter fitting steps in the 3D reconstruction process while ensuring the geometric accuracy of the reconstructed model, making it suitable for applications such as fish morphology analysis or fluid dynamics simulation.

[0063] S5. Calculate the volume of each segment by elliptical rotation integral, and calculate the estimated physical weight of a single fish by combining the density coefficient of the corresponding segment.

[0064] In a preferred embodiment, S5 includes: dividing the elliptical segments of the fish body into several equal slices using a discrete slicing method, and obtaining the elliptical cross-sectional area at the center of each small slice within each segment by linear interpolation based on the parameters of adjacent segments.

[0065] The discrete slicing method requires at least 10 slices, with the thickness of each slice being the major axis of the ellipse of the corresponding segment divided by the number of slices. The linear interpolation refers to allocating the major and minor axis values ​​between the start and end sections of the segment according to the current slice position. The ellipse area is calculated using the formula "3.1416 × major axis radius × minor axis radius", where the major axis radius = major axis value / 2.

[0066] The volume of each small segment within the corresponding segment is calculated by multiplying the area of ​​the ellipse within the segment by the slice thickness. The total volume of the fish body in that segment is obtained by summing the volumes of each small segment.

[0067] The total volume of the head segment, the total volume of the torso segment, and the total volume of the tail segment are obtained, and each is multiplied by the preset density value of the corresponding segment. Then, the product results of the three segments are added together to obtain the estimated physical weight of a single fish.

[0068] Specifically, the formula for calculating the total volume of the section is as follows: ,in Numbering of each small segment within the fish body section. , The total volume of the section The first The radii of the major and minor axes of the ellipse at the slice location. For the first section The thickness of each slice within a segment is the result of dividing the segment length by the number of smaller slices. This represents the total number of small slices. Based on this, the total volume of each segment of the fish's body can be determined. .

[0069] The formula for estimating the physical weight of a single fish is as follows: ,in, Estimate the physical weight of a single fish. For the first The preset density coefficients for the segments are as follows: the head segment has a higher density coefficient than the tail segment, and the tail segment has a higher density coefficient than the torso segment; the preset density coefficient is the head segment. Trunk segment Tail section .

[0070] This invention achieves automated 3D reconstruction and weight estimation of aquaculture biomass by locating key skeleton points, correcting deformation, and calculating the major and minor axes and volume integrals of ellipses in images. This reduces the need for manual annotation and complex algorithm parameter tuning, enabling fully automated and highly efficient feature recognition and classification, meeting the requirements of large-scale real-time monitoring, and lowering the cost of manual equipment use.

[0071] S6. Retrieve the body size correction constant of the current farmed fish species, scale and correct the physical weight estimate, and output the final biomass weight value.

[0072] Among them, the body shape correction constant refers to the multiplication coefficient preset based on the statistical data of live dissection of farmed fish species, ranging from 0.8 to 1.2, which is used to correct the weight deviation of bones / internal organs not covered in the physical volume calculation.

[0073] In a preferred embodiment, S6 includes: after obtaining the estimated physical weight of a single fish, retrieving the body size correction constant of the current farmed fish species from a preset biological species parameter library.

[0074] It should be noted that the body shape correction constant is set based on the differences in the proportion of internal organs and skeletons of the currently farmed fish species: a constant less than 1 is assigned to fish species with large skeletons or well-developed internal organs, such as 0.86-0.93, and a constant close to 1 is assigned to fish species with full muscles, such as 0.96-1.04, thereby compensating for the differences in physiological structure that are not distinguished by volume calculation; the difference in physiological structure refers to the difference in the mass proportion of non-muscle tissue in the fish body, which is obtained by sampling and weighing the fish body before and after removing the skin and internal organs in the laboratory.

[0075] The specific physiological structure difference compensation mechanism is as follows: if the bone density of farmed fish is significantly higher than that of muscle, the constant is set to be less than 1 to reduce weight; if the muscle content of farmed fish is high, the constant is greater than 1 to increase weight.

[0076] The body size correction constant is multiplied by the estimated physical weight, and the product is output as the final biomass weight value. Specifically, the formula for calculating the final biomass weight value is as follows: ,in, This represents the final biomass weight value. The body shape correction constant for the currently tested fish species is determined by sampling and weighing live fish before and after skinning and eviscerating, such as crucian carp. .

[0077] This invention retrieves the body shape correction constant of the current farmed fish species, scales and corrects the physical weight estimate, and outputs the final biomass weight value. It corrects the weight deviation of bones / viscera not covered in the physical volume calculation, optimizes the weight estimation accuracy, makes the weight estimation result closer to the true value, and improves the accuracy and reliability of farmed biomass monitoring.

[0078] It should be noted that the formulas described above, through the principle of dimensional consistency and mathematical standardization methods (such as normalization, dimensionless parameter conversion, or unit system unification), can translate physical quantities with different properties into unitless standard values ​​or superimposed parameters of the same dimension. This eliminates the interference of different dimensions on the computational logic, allowing the formulas to retain the original data distribution characteristics while possessing mathematical rationality and adaptability to objective laws. The descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the invention.

Claims

1. A method for three-dimensional reconstruction and weight estimation of cultured biomass based on machine vision, characterized in that, include: S1. Control two sets of industrial cameras to synchronously capture images of the aquaculture area at different shooting frequencies; the first set of cameras captures images continuously at a high frequency, and generates a fish movement distribution heatmap reflecting the intensity of fish movement in the aquaculture area by statistically analyzing the frequency of significant displacement of fish positions per unit time, and records areas where the total fish displacement count is lower than a set critical threshold as low movement intensity areas; the second set of cameras captures images at a normal frequency, and generates a rectangular focus frame based on the low movement intensity areas, and performs an exposure operation on the rectangular focus frame; S2. Rotate and adjust the angle of the polarizer in front of the second set of camera lenses, and control the polarization direction of the light source in the aquaculture area so that the light received by the camera and the light reflected from the water surface form an orthogonal polarization state. S3. Extract the key points of the skeleton of multiple fish bodies from the high-definition images captured after the second set of camera polarization adjustment, and then align the key points of the skeleton with the preset skeleton template to automatically correct the deformed parts of the fish body. S4. Using the aligned skeleton nodes as boundaries, divide the fish body into head, torso, and tail segments along the length of the fish body. Based on the spacing between the key skeleton points in each segment, deduce the maximum thickness of the fish body and obtain the major and minor axes of the ellipse for each segment of the fish body. S5. Calculate the volume of each segment by elliptical rotation integral, and calculate the estimated physical weight of a single fish by combining the density coefficient of the corresponding segment. S6. Retrieve the body size correction constant of the current farmed fish species, scale and correct the physical weight estimate, and output the final biomass weight value.

2. The method for three-dimensional reconstruction and weight estimation of cultured biomass based on machine vision according to claim 1, characterized in that, The control of two sets of industrial cameras to synchronously capture images of the aquaculture area at different shooting frequencies is specifically as follows: Two sets of industrial cameras, one deployed on the water surface and the other underwater, are configured as image acquisition devices. The continuous image sequence captured by the first set of cameras is input into the motion analysis unit. At the same time, the center position of the fish's eye is used as a specific marker point for the fish body to divide the image into different sections. The linear distance of movement of the corresponding fish-specific identifier points in each fish-body partition in adjacent images is compared frame by frame. When the linear distance of movement exceeds the preset distance, it is determined that the position of the fish-body partition has been significantly displaced. At this time, the displacement of the fish-body partition is superimposed and counted to form a partition count value matrix based on pixels. The frequency of significant displacement of fish body regions in each image per unit time is counted to generate a two-dimensional distribution map reflecting the intensity of fish movement within the aquaculture area, which is denoted as the fish movement distribution heat map. Based on the displacement count values ​​of each fish body partition in the partition count value matrix, the total displacement count values ​​of different pixel blocks are marked with different color depths in the fish movement distribution heatmap. The lightest color region represents the total displacement count value of the fish body below a set critical threshold and is recorded as a low movement intensity region. The exposure trigger mechanism for the second group of cameras is set to send the coordinates of the center point of the low-motion intensity area of ​​the fish movement distribution heatmap to the exposure controller of the second group of cameras. The exposure controller of the second group of cameras generates a rectangular focus frame based on the center point coordinates of the low motion intensity area. After locking the area, the autofocus module of the second group of cameras releases the shutter and performs the exposure action.

3. The method for three-dimensional reconstruction and weight estimation of cultured biomass based on machine vision according to claim 1, characterized in that, The content of S2 includes: When the second set of cameras performs an exposure operation, the polarization angle of the light reflected from the water surface is detected by a photoelectric sensor located above the water surface. Increase the polarization angle of the light reflected from the current water surface. The target rotation angle of the polarizing filter in the camera is then used to drive the polarizing filter in front of the second set of camera lenses to rotate around the optical axis at the target rotation angle. At the same time, the polarization direction of the light emitted by the underwater LED array in the aquaculture area is controlled to be parallel to the direction of the camera filter, so that the polarization direction of the light entering the camera lens is perpendicular to the polarization direction of the light reflected from the water surface.

4. The method for three-dimensional reconstruction and weight estimation of cultured biomass based on machine vision according to claim 1, characterized in that, The extraction of key skeletal points of multiple fish bodies from the high-definition images captured after polarization adjustment of the second set of cameras includes: Gradient difference method is used to enhance the outline edge of fish body. For each visible fish body, the two most prominent circular dark spots in its head region are identified as candidate fish eye regions. Search along the direction of the fish's spine for the junctions between each fin and the trunk, and mark them as the fin base connection points. Obtain the preset fish body size ratio rules. If the ratio of the distance between the center of the two sides of the fish eye area to the distance of the widest part of the head conforms to the fish body size ratio rules, it is determined that the corresponding skeleton key point of the visible fish body is established. At this time, the candidate fish eye area and each fin part are the corresponding skeleton key points of the visible fish body. Otherwise, the corresponding skeletal key points of the visible fish body are considered to be in the wrong location, and the corresponding skeletal key points in the neighboring area are searched again.

5. The method for three-dimensional reconstruction and weight estimation of cultured biomass based on machine vision according to claim 1, characterized in that, The process of aligning key points of the skeleton with a preset skeleton template and automatically correcting deformed areas of the fish body includes: Establish the mapping relationship between the key points of the skeleton and the bone nodes in the skeleton template, and generate matching point pairs for each part of the fish body; Calculate the sum of the Euclidean distances for each pair of matching points, and achieve position alignment by minimizing this sum through rotation and scaling transformations, thus obtaining the corrected 3D coordinates of the deformed parts of the fish body.

6. The method for three-dimensional reconstruction and weight estimation of cultured biomass based on machine vision according to claim 1, characterized in that, The method of dividing the fish body into three equal segments—head, torso, and tail—along the length of the fish body, using the aligned skeletal nodes as boundaries, includes: Obtain the straight-line distance from the front node of the skull to the end node of the tailbone to get the total length of the fish. Divide this length by three to get the standard segment length. Starting from the head, dividing lines are set every standard segment length along the main axis: the first dividing line is located one standard segment length after the head starting point and is marked as the end of the head segment; the second dividing line is located one standard segment length after the end of the head segment and is marked as the end of the torso segment; the remaining part is the tail segment.

7. The method for three-dimensional reconstruction and weight estimation of cultured biomass based on machine vision according to claim 1, characterized in that, The acquisition of the major and minor axes of the ellipse in each segment of the fish body includes: Based on the division of the fish body into three segments, reference points are selected on the central axis of each segment. At the reference point of each section, a local coordinate system is established perpendicular to the central axis of the fish. Analyze the spatial distribution of key skeleton points within the current segment, and measure the spatial straight-line distance in the width direction between the two farthest key skeleton points within the segment, which serves as a reference value for the minor axis of the ellipse. Based on the pre-defined constraint ratio of the relationship between the thickness and width of each segment of the fish body, the reference value of the minor axis of the ellipse is adjusted, and the maximum thickness of the fish body is calculated as the minor axis of the ellipse of the corresponding segment. Define the maximum width of the segment as the major axis of the ellipse; The major and minor axes of the ellipse in each segment of the fish's body were statistically determined.

8. The method for three-dimensional reconstruction and weight estimation of cultured biomass based on machine vision according to claim 1, characterized in that, The content of S5 includes: The discrete slicing method is used to divide the elliptical segments of the fish body into several equal slices. The elliptical cross-sectional area at the center of each small slice within each segment is obtained by linear interpolation based on the parameters of adjacent segments. The volume of each small segment within the corresponding segment is calculated by multiplying the area of ​​the ellipse within the segment by the slice thickness. The total volume of the fish body in that segment is obtained by summing the volumes of each small segment. The total volume of the head segment, the total volume of the torso segment, and the total volume of the tail segment are obtained, and each is multiplied by the preset density value of the corresponding segment. Then, the product results of the three segments are added together to obtain the estimated physical weight of a single fish.

9. The method for three-dimensional reconstruction and weight estimation of cultured biomass based on machine vision according to claim 1, characterized in that, The content of S6 includes: After obtaining the estimated physical weight of a single fish, the body size correction constant of the current farmed fish species is retrieved from the preset biological species parameter library; Multiply the body size correction constant by the estimated physical weight, and output the product as the final biomass weight value.

Images