A sea urchin instance segmentation and high-throughput growth trait measurement method and system
By constructing a weight prediction model based on the improved SUYOLO-seg model and sea urchin growth patterns, the problems of low efficiency and large error in sea urchin growth trait measurement were solved. This enabled high-throughput, non-contact, and accurate measurement of sea urchin growth traits, making it suitable for rapid phenotypic analysis of large batches of sea urchin samples.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- INST OF OCEANOLOGY - CHINESE ACAD OF SCI
- Filing Date
- 2026-04-23
- Publication Date
- 2026-06-09
Smart Images
Figure CN122176037A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of sea urchin aquaculture technology, and in particular to a method and system for segmenting sea urchin specimens and measuring high-throughput growth traits. Background Technology
[0002] Sea urchins are an important marine economic species. Their gonads (commonly known as sea urchin roe) are rich in protein, unsaturated fatty acids, and various trace elements, possessing high nutritional value and broad market demand. In the process of industrialized sea urchin farming and germplasm resource development, growth traits such as size and weight are key indicators for germplasm evaluation, aquaculture selection, growth monitoring, and phenotypic analysis. Regular and accurate monitoring of these growth traits is a core step in assessing sea urchin health, optimizing feeding strategies, and selecting high-quality strains.
[0003] Currently, obtaining sea urchin growth traits, such as shell diameter and weight, mainly relies on traditional manual measurement methods, namely measuring shell diameter with a ruler and weighing with an electronic scale. However, these traditional methods are inefficient and cumbersome in practical applications, and cannot meet the needs of high-throughput, rapid phenotypic collection of large batches of sea urchin samples. Furthermore, manual measurement is highly subjective, leading to errors in judging the complex boundaries and estimating the size of sea urchins. More seriously, the weighing and measurement process usually requires removing the sea urchins from the water, which not only causes a strong stress response and affects their subsequent growth, but also allows water to easily accumulate on the surface of the sea urchin and between its dense spines, resulting in significant weighing errors. Therefore, developing a non-contact, automated, high-throughput measurement process has become a practical requirement for industry development.
[0004] To achieve high-precision non-contact measurement, it is necessary to accurately segment the outline of each sea urchin from a complex background. However, to date, there is no dedicated segmentation model for sea urchins in industry. Using existing general-purpose segmentation models in industry (such as YOLOv8) faces insurmountable technical bottlenecks when processing sea urchin images, including: the sea urchin's body color blends closely with the background, easily losing high-frequency edge details; the sea urchin's spines exhibit irregular dynamic morphological changes, making it impossible for the model to extract deformation features; sea urchins vary in size, and underwater environments often contain suspended matter and other background interference, making the model susceptible to multi-scale and complex background interference; the sea urchin's peripheral spines have a slender and continuously tortuous structure, making it easy for the model's mask to break when processing slender and curved targets.
[0005] Furthermore, existing methods typically only obtain the pixel area of an image, lacking a stable and high-precision physical scale calibration mechanism, and thus failing to achieve accurate conversion from pixel scale to actual physical scale. Moreover, when constructing the area-weight mapping relationship, existing methods are insufficiently adaptable to abnormal samples and complex backgrounds, lacking anomaly data cleaning mechanisms based on biophysical principles. This results in poor generalization ability and low accuracy in the final trained weight prediction model, making it difficult to apply robustly in real-world aquaculture environments. Summary of the Invention
[0006] To address the problems in existing technologies where weighing and measuring sea urchins requires removing them from the water, causing severe stress and affecting their subsequent growth, and resulting in significant weighing errors due to water retention on the sea urchin's surface and between its dense spines, the first objective of this invention is to provide a method for sea urchin instance segmentation and high-throughput measurement of growth traits, comprising:
[0007] Acquire calibration images containing sea urchins;
[0008] Use a pre-trained SUYOLO-seg model to extract sea urchin pixel boundaries;
[0009] The pixel area of the sea urchin is calculated using the polygon area formula, and the real physical area of the sea urchin is calculated based on the pre-calibrated real physical area of the unit pixel.
[0010] The true weight of sea urchins was estimated using a sea urchin weight trait prediction model based on the mapping relationship between sea urchin area and weight.
[0011] Specifically, the SUYOLO-seg model, based on the original YOLO11-seg visual recognition model, replaces the conventional downsampling layer Conv in the backbone and neck networks with Haar wavelet downsampling modules, replaces the feature extraction module C3k2 in the backbone and neck networks with deformable convolution C2f_DCNv4 modules, replaces the C2PSA module in the feature extraction and fusion stage with a heterogeneous multi-scale attention network, and replaces the original segmentation head with a dynamic snake-shaped convolutional segmentation head.
[0012] Specifically, the Haar wavelet downsampling module uses a two-dimensional Haar discrete wavelet transform to transform the input feature map. Decomposed into low-frequency components Horizontal detail components Vertical detail components and diagonal detail components ;
[0013] The four components are concatenated along the channel dimension to generate an intermediate feature map. ;
[0014] The output features are downsampled by mapping to the target number of output channels C' through 1×1 pointwise convolution, followed by batch normalization and ReLU activation. .
[0015] Specifically, the C2f_DCNv4 module will input the feature map as... After adjusting the number of channels using 1×1 convolution, the channels are evenly divided into reserved branches. and processing branches The processing branch For n Bottleneck structures containing DCNv4, branches will be preserved. Initial processing branch and the output of each Bottleneck The concatenation is performed along the channel dimension, followed by 1×1 convolutional fusion and dimensionality reduction to output the feature map. .
[0016] Specifically, the Bottleneck structure first extracts local features through a 3×3 standard convolution. Then, the offset is dynamically predicted using DCNv4. and modulation coefficient For the reference sampling point corresponding to the output position Adaptive sampling formula for:
[0017] ,
[0018] in, This represents the modulation coefficient of the k-th sampling point in group G. This indicates the reference sampling point at the current output position. This indicates a predefined sampling point offset. This represents the learned dynamic offset, and K represents the total number of sampling points;
[0019] Predicting offsets using additional convolutional branches and modulation scalar :
[0020] .
[0021] Specifically, The heterogeneous multi-scale attention network module will input feature maps First, the number of channels is adjusted using 1×1 convolution, and the features are then divided into g subgroups along the channel dimension. Each subgroup passes through a multi-scale convolutional module in parallel. This module contains at least three parallel branches, each extracting features from a different receptive field, and compressing the output channel count to [number missing]. The features output from each convolutional branch are concatenated along the channel dimension and then fused through a 1×1 convolution to obtain the multi-scale features corresponding to that subgroup. The features of each subgroup are input into the channel attention module and the spatial attention module respectively, and all subgroups are attention-enhanced features. The data is concatenated and connected to the input feature map via residual connections. Add them together to output the feature map. .
[0022] Specifically, the channel attention module focuses on the features of the Jth subgroup. Perform global average pooling and global max pooling respectively. Two were obtained. The two descriptors are fed into a shared multilayer perceptron (MLP), the outputs are summed, and then activated by a sigmoid function to obtain the channel weights. ;
[0023] The spatial attention module employs a directional decomposition mechanism, encoding features in both the height and width directions, and then processing the subgroup features. Perform max pooling and average pooling along the channel axis respectively to obtain two results. The feature maps are concatenated and then subjected to 7×7 convolution and sigmoid activation to obtain spatial weights. .
[0024] Specifically, the dynamic snake-shaped convolution segmentation head will input the feature map Predict the initial offset using a 3×3 convolution. Where K is the number of sampling points, and is scaled using the tanh function to... The offset prediction formula is:
[0025] ,
[0026] in, This represents the initial coordinate offset matrix for prediction, α is the scaling factor, and BN represents the batch normalization operation. This represents a standard convolution with a kernel size of 3×3;
[0027] Coordinates of the original sampling point Based on this, and according to the shape and orientation of the target to be segmented, the horizontal offset component of the initial prediction is adjusted along the X-axis or Y-axis. or vertical offset The process involves iterative accumulation, followed by bilinear interpolation to extract the final interpolated features corresponding to consecutive sampling points p. :
[0028] ,
[0029] In the formula, p represents continuous sampling points. Let p be the four integer pixels surrounding p, and q be the set. Integer pixels in the data. The original feature value corresponding to the integer pixel q. For distance-based interpolation weights, interpolation weights The calculation is as follows:
[0030] ;
[0031] The sampled features are weighted and summed according to the convolution kernel weights to obtain the final output features.
[0032] Specifically, the sea urchin pixel area The calculation formula is:
[0033] ,
[0034] in, The total number of vertices. , Let a be the coordinates of the a-th vertex. , Let the coordinates be those of the (a+1)th vertex.
[0035] The sea urchin weight trait prediction model includes:
[0036] Based on the growth pattern of sea urchins, a physical proportional characteristic variable P is constructed.
[0037] ,
[0038] In the formula, E is the allometric growth parameter for the transformation of the two-dimensional projected area of the sea urchin into its three-dimensional equivalent volume;
[0039] The interquartile range algorithm is used to remove extreme outliers from the physical ratio P. Let the first quartile be Q1 and the third quartile be Q3, then:
[0040] ,
[0041] Samples that meet the following criteria are retained as high-quality training samples:
[0042] ,
[0043] Samples that do not meet the above conditions will be considered abnormal samples and removed.
[0044] Multiple machine learning regression models were built for fitting, and 2-3 mean absolute percentage error (MAPE) models were selected from them. And the coefficient of determination The regression model with a value greater than 0.9 was used to average and fuse the prediction results to obtain the final predicted value of sea urchin weight.
[0045] A second objective of this invention is to provide a sea urchin instance segmentation and high-throughput growth trait measurement system, using the aforementioned sea urchin instance segmentation and high-throughput growth trait measurement method, comprising:
[0046] The image acquisition module acquires calibrated images of sea urchins in the target sea area and images taken at the vertical shooting height from the seabed surface;
[0047] The sea urchin instance segmentation module uses a pre-trained SUYOLO-seg model to extract the pixel boundaries of sea urchins in the underwater calibration image;
[0048] The area calculation module calculates the actual physical area of the sea urchin based on the pixel boundaries of the sea urchin and the pre-calibrated actual physical area per unit pixel.
[0049] The weight prediction module estimates the true weight of sea urchins based on a sea urchin weight trait prediction model that maps the area to the weight of the sea urchin.
[0050] The output module provides a visualization of the sea urchin's area and weight.
[0051] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0052] 1. This invention addresses the underwater characteristics of sea urchins by constructing a dedicated SUYOLO-seg instance segmentation model based on the YOLO architecture. It replaces the conventional downsampling layer with a Haar wavelet downsampling module (HDown) to preserve high-frequency details of sea urchin spines and edges while reducing feature map resolution. It replaces the conventional feature extraction module with a C2f_DCNv4 module to enhance the model's adaptive modeling ability for sea urchin spine bending, posture changes, and local deformation. The heterogeneous multi-scale attention network (HSAN) module enhances the model's representation ability for sea urchins of different sizes through multi-scale convolution and channel-space dual attention mechanisms, while suppressing interference from complex background noise such as mud, attachments, and suspended particles. The dynamic serpentine convolutional segmentation head (DSCseg) allows the convolutional sampling path to continuously extend along the curved boundaries of the sea urchin, thereby improving the accuracy of the segmentation mask in fitting the complex contours and slender spine structures of the sea urchin. SUYOLO-seg achieves a mAP50 index of 0.89382 for instance segmentation, a 2.69% improvement compared to YOLO11-seg.
[0053] 2. This invention achieves completely non-contact measurement of growth traits by acquiring images, segmenting instances, and automatically converting physical dimensions, without ever removing sea urchins from the water. This avoids the severe stress caused by removing sea urchins from the water, ensuring their normal growth and development. It also completely solves the weighing error problem caused by water retention on the sea urchin surface and between its spines, improving the authenticity and accuracy of growth trait data. Furthermore, it can simultaneously extract the boundaries, calculate the area, and estimate the weight of multiple sea urchins in a single frame image, significantly improving the efficiency of sea urchin phenotypic data acquisition. This fully meets the needs of rapid phenotypic analysis of large batches of samples in factory farming and breeding, significantly reducing the labor and time costs of aquaculture production and germplasm selection. Attached Figure Description
[0054] Figure 1This is an overall flowchart of the method of the present invention;
[0055] Figure 2 The overall network structure diagram of the SUYOLO-seg instance segmentation model;
[0056] Figure 3 This is a structural diagram of the Haar wavelet downsampling (HDown) module;
[0057] Figure 4 This is a structural diagram of the deformable convolution (C2f_DCNv4) module;
[0058] Figure 5 This is a diagram of the Heterogeneous Multiscale Attention Network (HSAN) module structure.
[0059] Figure 6 This is a diagram showing the overall structure of the Dynamic Snake Convolution Segment (DSCsegment).
[0060] Figure 7 This is a diagram of the Dynamic Snake Convolution (DSCConv) module structure.
[0061] Figure 8 This is a comparison chart showing the segmentation results of sea urchin instances using different segmentation models.
[0062] Where a is the test result graph of the SUYOLO-seg model, and b is the test result graph of the YOLO11-seg model;
[0063] Figure 9 A comparison of the fitting relationship between sea urchin weight and area under different models;
[0064] Figure 10 This is a schematic diagram showing the segmentation and area measurement results of a sea urchin instance. Detailed Implementation
[0065] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.
[0066] This invention provides a method for sea urchin instance segmentation and high-throughput growth trait measurement, which can automatically obtain the precise boundary of sea urchins, automatically complete the pixel-to-actual scale conversion, and has anti-interference and anomaly removal mechanisms, thereby realizing intelligent non-contact high-throughput measurement of sea urchin growth properties with accurate estimation.
[0067] S1: Acquisition and calibration of sea urchin image set.
[0068] Industrial cameras with polarizing filters or high-definition imaging equipment were used to acquire images in the sea urchin farming area. Before acquisition, the camera's intrinsic parameters were calibrated and distortion corrected using Zhang's camera calibration method. Under fixed shooting height, fixed shooting angle, and fixed focal length, the mapping relationship between image pixels and actual physical scale was established to obtain the true physical area corresponding to each pixel.
[0069] Preferably, image samples of no fewer than 300 sea urchins are collected under the same aquaculture area and the same species conditions, with the sea urchins covering different sizes, ages, and growth stages as much as possible. Simultaneously, each sea urchin is manually weighed to obtain its corresponding actual weight data, constructing a dataset of sea urchin image instances and their corresponding weights.
[0070] S2: Extracting sea urchin pixel boundaries based on a pre-trained accurate instance segmentation model.
[0071] To address the challenges of sea urchins in complex aquaculture environments, such as strong background interference, fragmented boundaries, irregular spine morphology, and large scale variations, a precise instance segmentation model, SUYOLO-seg, suitable for sea urchin targets, is constructed based on YOLO11-seg. Its structure is as follows: Figure 2 As shown.
[0072] S2.1: This invention replaces the conventional downsampling layer Conv in the backbone and neck networks with the Haar wavelet downsampling (HDown) module to preserve high-frequency details of sea urchin spines and edges while reducing feature map resolution.
[0073] Because sea urchins are covered with irregular spines and their body color often blends well with complex backgrounds such as mud and sand, traditional stride convolution downsampling results in a significant loss of high-frequency edge details. This invention utilizes the HDown module, which, through the characteristics of Haar wavelet transform, preserves the spine outlines and high-frequency boundary components of the sea urchin while reducing spatial resolution, solving the segmentation problem of blurred edges and easy background confusion in complex underwater environments. Furthermore, Haar wavelets are computationally simple and provide accurate edge localization, making them suitable for real-time segmentation tasks.
[0074] To reduce the loss of sea urchin edge texture information during downsampling, the conventional downsampling layers in the backbone and neck networks were replaced with Haar wavelet downsampling (HDown) modules, the structure of which is shown in the diagram below. Figure 3 .
[0075] Input feature map ,
[0076] Where C represents the number of channels; H and W represent the height and width of the feature map, respectively.
[0077] Two-dimensional Haar discrete wavelet transform (DWT) is used to decompose the feature into four sub-bands with half the spatial resolution. Low-frequency components are then extracted. And three high-frequency detail components, including the horizontal detail component. Vertical detail components Diagonal detail components For any spatial coordinates The formulas for calculating each component are as follows:
[0078] ,
[0079] ,
[0080] ,
[0081] ,
[0082] All component dimensions are .
[0083] Subsequently, the four components are concatenated along the channel dimension to generate an intermediate feature map. This helps preserve edge detail information:
[0084] ,
[0085] Finally, the downsampled features are obtained by mapping the target output channel number C' to a 1×1 pointwise convolution, followed by batch normalization and ReLU activation. :
[0086] ,
[0087] This downsampling module reduces spatial resolution while preserving high-frequency edge information, thus mitigating the loss of sea urchin boundary details caused by downsampling.
[0088] S2.2: This invention replaces the feature extraction module C3k2 in the backbone and neck networks with the deformable convolutional module C2f_DCNv4 to enhance the model's adaptive modeling ability for sea urchin spine bending, posture changes and local deformation.
[0089] In seawater, sea urchin spines exhibit irregular morphological changes. Fixed square convolutional kernels cannot effectively extract such dynamic features. The C2f_DCNv4 module introduced in this invention, through adaptive prediction of spatial offset, enables the sampling points of the convolutional kernel to dynamically fit the actual spine morphology of the sea urchin, improving the model's generalization ability to the varied postures of sea urchins and the completeness of feature extraction.
[0090] To improve the network's ability to model sea urchin posture changes, this invention replaces all feature extraction modules C3k2 in the backbone and neck networks with deformable convolutional modules C2f_DCNv4, the structure of which is shown in the diagram below. Figure 4 .
[0091] The input feature map is After adjusting the number of channels using 1×1 convolution, the channels are evenly divided into reserved branches. and processing branches , The system sequentially enters n Bottleneck structures containing DCNv4. Each Bottleneck structure first extracts local features through a standard 3×3 convolution. Then, the offset is dynamically predicted using DCNv4. And modulation coefficient (modulation scalar) For the reference sampling point corresponding to the output position Adaptive sampling formula for:
[0092] ,
[0093] in, This represents the modulation coefficient of the k-th sampling point in group G. This indicates the reference sampling point at the current output position. This indicates a predefined sampling point offset. This represents the dynamic offset obtained through learning, and K represents the total number of sampling points.
[0094] offset and modulation scalar Predicted through additional convolutional branches:
[0095] .
[0096] Let the output of the nth Bottleneck be denoted as Finally, the branch will be preserved. Initial processing branch and the output of each Bottleneck By concatenating the data along the channel dimension and then performing 1×1 convolutional fusion for dimensionality reduction, a feature map is obtained that enhances the ability to represent target deformation. .
[0097] ,
[0098] The convolution kernels in this C2f_DCNv4 module can adaptively adjust the sampling position according to the target shape, thereby enhancing the ability to represent the irregular curved contours of sea urchins.
[0099] S2.3: This invention replaces the C2PSA module in the feature extraction and fusion stage with a heterogeneous multi-scale attention network (HSAN), which improves the model's ability to represent sea urchins of different sizes through multi-scale convolution and channel-space dual attention mechanism, and suppresses the interference of complex background noise such as mud, attached substances, and suspended particles.
[0100] Due to differences in shooting distance and individual sea urchins, the sea urchins in the images vary in size, and background interference such as suspended objects may occur. This invention employs a heterogeneous multi-scale attention network (HSAN). On one hand, it accurately captures sea urchins of different sizes through multi-scale convolutional modules. On the other hand, it uses a dual attention mechanism to coordinate spatial and channel attention, focusing on the key directional distribution of the sea urchin body and suppressing underwater environmental noise and false target interference.
[0101] To accurately locate the spatial position of deformation features, the C2PSA module in the feature extraction and fusion stage is replaced with a heterogeneous multi-scale attention network (HSAN), thereby improving the model's ability to jointly model multi-scale local structures, directional spatial distributions, and channel importance. Its structure diagram is shown below. Figure 5 .
[0102] Input feature map After adjusting the number of channels using 1×1 convolution, the features are divided into g subgroups along the channel dimension. Each subgroup passes through a multi-scale convolutional module (MSCM) in parallel. The MSCM contains at least three parallel branches, employing at least 1×1, 3×3, and 5×5 convolutional kernels respectively, to extract features from different receptive fields, compressing the output channel count to [number missing]. The features output from each convolutional branch are concatenated along the channel dimension and then fused using a 1×1 convolution to obtain the multi-scale features corresponding to that subgroup. The features of each subgroup are then input into the channel attention module and the spatial attention module respectively, and the attention weights are calculated in parallel.
[0103] Channel attention module for the J-th subgroup features Perform global average pooling and global max pooling respectively. Two were obtained. The two descriptors are fed into a shared multilayer perceptron (MLP), the outputs are summed, and then activated by a sigmoid function to obtain the channel weights. :
[0104] ;
[0105] The spatial attention module employs a direction decomposition mechanism, encoding features in both the height and width directions, and then processing the subgroup features. Perform max pooling and average pooling along the channel axis respectively to obtain two results. The feature maps are concatenated and then subjected to 7×7 convolution and sigmoid activation to obtain spatial weights. :
[0106] .
[0107] The attention-enhanced features are then summed to obtain the attention-enhanced features of the Jth subgroup. :
[0108] ,
[0109] In the formula, This indicates element-wise multiplication.
[0110] Finally, attention enhancement features for all subgroups The data is concatenated and connected to the input feature map via residual connections. Adding them together yields the global output feature map of the HSAN module. ,
[0111] .
[0112] This module jointly models multi-scale local structures and directional spatial distributions, effectively suppressing underwater noise and background interference.
[0113] S2.4: This invention replaces the original segmentation head with a dynamic serpentine convolution segmentation head (DSCsegment), enabling the convolution sampling path to extend continuously along the curved boundary of the sea urchin, thereby improving the fitting accuracy of the segmentation mask to the complex contours and slender spiky structures of the sea urchin.
[0114] The peripheral spines of sea urchins exhibit a continuous, tortuous structure morphologically. Conventional segmentation heads are prone to mask breakage or poor edge fit when processing such slender and curved targets. This invention employs a dynamic serpentine convolution segmentation head (DSCsegment), which continuously iterates and accumulates the coordinates of sampling points, ensuring that the receptive field strictly follows the tortuous path of the sea urchin in space, thereby generating an instance segmentation mask with high edge fit and structural continuity.
[0115] To enhance the detection head's ability to fit the complex boundaries of sea urchins, a dynamic serpentine convolution (DSCConv) is introduced in the mask coefficient prediction branch of the final segmentation detection head to replace the standard convolution, constructing a dynamic serpentine convolution segmentation detection head (DSCsegment). This improves the modeling ability for curved, slender, and complex boundary targets. Its structure diagram is shown below. Figure 6 and Figure 7 .
[0116] Input feature map First, the initial offset is predicted using a 3×3 convolution. , Where K is the number of sampling points, and is scaled using the tanh function to... The offset prediction formula is:
[0117] ,
[0118] in, This represents the initial coordinate offset matrix for prediction; α is the scale factor used to control the maximum magnitude of the offset, with a default value of 2; BN indicates batch normalization operation. This represents a standard convolution with a kernel size of 3×3.
[0119] Coordinates of the original sampling point Based on this, and according to the shape and orientation of the target to be segmented, the horizontal offset component of the initial prediction is adjusted along the X-axis or Y-axis. or vertical offset Iterative accumulation is performed to generate a continuously curved sampling path that fits the target contour.
[0120] Let the center index of the convolution kernel be... Taking the horizontal direction as an example, the cumulative offset of the center sampling point in the initial state , ,
[0121] Cumulative offset of the vth sampling point to the right of the center point for:
[0122] ,
[0123] Cumulative offset of the v-th sampling point to the left of the center point for:
[0124] ,
[0125] Cumulative offset of the vth sampling point above the center point for:
[0126] ,
[0127] Cumulative offset of the vth sampling point below the center point for:
[0128] ,
[0129] Coordinates of the original sampling point The horizontal offset component is accumulated in the previous iteration. and vertical offset Obtain the final sampling point coordinates .
[0130] Since the cumulative coordinates may be non-integer, bilinear interpolation is used to extract the final interpolation features corresponding to consecutive sampling points p. :
[0131] ,
[0132] In the formula, p represents continuous sampling points. Let p be the four integer pixels surrounding p, and q be the set. Integer pixels in the data. The original feature value corresponding to the integer pixel q. For distance-based interpolation weights, interpolation weights The calculation is as follows:
[0133] .
[0134] The sampled features are weighted and summed according to the convolution kernel weights to obtain the final output features. Finally, the extracted deformation features are processed through 1×1 convolution to generate bounding box parameters (Box), classification scores (Cls), and mask coefficients (MaskCoef), which are then combined with the prototype mask to obtain the instance segmentation result. This module ensures that the receptive field extends continuously along the curved path of the target, thereby avoiding mask breaks.
[0135] refer to Figure 8Using the same training and validation sets and under the same hardware configuration, both the SUYOLO-seg model of this invention and the basic YOLO11-seg model were trained using the minimum parameter version (n-series), with 100 training epochs for each model. Experimental results show that SUYOLO-seg significantly outperforms YOLO11-seg in all performance metrics for object detection and instance segmentation tasks. In terms of object detection performance, SUYOLO-seg achieves Precision, Recall, mAP50, and mAP50-95 of 0.93730, 0.81798, 0.89878, and 0.65148, respectively, which are improvements of 0.03190, 0.03585, 0.02930, and 0.06601 compared to YOLO11-seg. In terms of instance segmentation performance, SUYOLO-seg achieves Precision, Recall, mAP50, and mAP50-95 of 0.93788, 0.81881, 0.89382, and 0.58474, respectively, which are improvements of 0.02662, 0.03798, 0.02687, and 0.02289 compared to YOLO11-seg.
[0136] Among them, SUYOLO-seg achieved an mAP50 score of 0.89382 for instance segmentation, which is 2.69% higher than YOLO11-seg, indicating that it has better localization and mask fitting capabilities. This shows that SUYOLO-seg has stronger feature representation and boundary recovery capabilities for targets with complex boundaries, such as sea urchins, and is more suitable for sea urchin instance segmentation tasks.
[0137] Step S3: Construct a sea urchin growth trait measurement model.
[0138] The sea urchin pixel boundaries are extracted by inputting the image to be tested into the SUYOLO-seg model, and the true physical area of the sea urchin is calculated based on the Shoelace polygon theorem. Multidimensional outlier data is cleaned using the interquartile range (IQR) algorithm based on allometric growth, and then input into a network that integrates a multi-model regression algorithm to output the predicted live weight of the sea urchin.
[0139] S3.1: Sea urchin area extraction and calculation based on the SUYOLO-seg model.
[0140] The image to be tested is input into the pre-trained SUYOLO-seg instance segmentation model, which outputs the set of vertex coordinates of the boundary polygon corresponding to each sea urchin.
[0141] Let the sea urchin dividing boundary be... The sea urchin is composed of vertices, then the pixel area is... The area is calculated using the polygon area formula:
[0142] ,
[0143] in, The total number of vertices. , Let a be the coordinates of the a-th vertex. , Let the coordinates be the (a+1)th vertex;
[0144] Based on the actual physical area per pixel obtained from previous camera calibration Further calculation of the true physical area of sea urchins :
[0145] ,
[0146] in, The preferred unit is cm²;
[0147] The above method enables automatic conversion from sea urchin boundary segmentation results to the actual area characteristics of sea urchins.
[0148] S3.2: Construct a prediction model for sea urchin weight traits.
[0149] To establish the mapping relationship between sea urchin area and weight, at least 300 sea urchin samples were collected from the same aquaculture area and under the same species conditions, covering different growth stages and individual sizes as much as possible. For each sea urchin, the SUYOLO-seg model was used for instance segmentation, and its actual physical area (Sactual) was calculated according to step 3.1. Simultaneously, the actual weight (W) of each sea urchin was obtained by manual weighing, forming a sample set with a one-to-one correspondence between area and weight.
[0150] Based on the growth pattern of sea urchins, a physical proportion characteristic variable P is constructed:
[0151] ,
[0152] In the formula, E is the allometric growth parameter for the transformation of the two-dimensional projected area of the sea urchin into the three-dimensional equivalent volume, which is preferably 1.5.
[0153] To improve the stability of the weight prediction model, the interquartile range (IQR) algorithm is used to remove extreme outliers from the physical proportion P.
[0154] Let the first quartile be Q1 and the third quartile be Q3, then we have:
[0155] ,
[0156] Samples that meet the following criteria are retained as high-quality training samples:
[0157] ,
[0158] Samples that do not meet the above conditions will be considered abnormal samples and removed to reduce measurement errors or interference caused by abnormal individuals.
[0159] After outlier cleaning, the high-quality sample set is divided into training and testing sets, and a feature matrix (sea urchin area) is constructed. ) and target vector (true weight of sea urchin) ).
[0160] Based on this, various machine learning regression models were established for fitting, referencing... Figure 9 The regression model includes at least one or more of the following: power function regression model, quadratic polynomial regression model, Huber robust polynomial regression model, support vector regression model, random forest regression model, generalized additive model, etc.
[0161] To evaluate the performance of each regression model, the mean absolute percentage error (MAPE) and the coefficient of determination were used. As an evaluation indicator:
[0162] ,
[0163] ,
[0164] in, Indicates the actual weight. Indicates the predicted weight. This represents the average actual weight.
[0165] Preferably, the constructed model satisfies: MAPE < 8%. >0.9.
[0166] Two to three regression models with the best performance can be selected based on the evaluation results of the test set. Their prediction results can be averaged and fused to obtain the final predicted value of sea urchin weight, thereby improving the model's generalization ability and robustness.
[0167] In practical applications, the sea urchin image to be measured is input into the SUYOLO-seg model to achieve accurate segmentation of each sea urchin instance. Subsequently, the true area of the sea urchin is calculated based on the segmentation boundaries. Finally, the true area is input into the weight trait prediction model to obtain the predicted weight value of the sea urchin, thereby achieving high-throughput, non-contact, and automated measurement of growth traits such as sea urchin area and weight.
[0168] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention. The above are merely preferred embodiments of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the protection scope of the present invention.
Claims
1. A method for segmenting sea urchin specimens and measuring high-throughput growth traits, characterized in that, include: Acquire calibration images containing sea urchins; Use a pre-trained SUYOLO-seg model to extract sea urchin pixel boundaries; The pixel area of the sea urchin is calculated using the polygon area formula, and the real physical area of the sea urchin is calculated based on the pre-calibrated real physical area of the unit pixel. The true weight of sea urchins was estimated using a sea urchin weight trait prediction model based on the mapping relationship between sea urchin area and weight.
2. The method for segmenting sea urchin instances and measuring high-throughput growth traits according to claim 1, characterized in that, The SUYOLO-seg model, based on the original YOLO11-seg visual recognition model, replaces the conventional downsampling layer Conv in the backbone and neck networks with Haar wavelet downsampling modules, replaces the feature extraction module C3k2 in the backbone and neck networks with deformable convolutional C2f_DCNv4 modules, replaces the C2PSA module in the feature extraction and fusion stage with a heterogeneous multi-scale attention network, and replaces the original segmentation head with a dynamic snake-shaped convolutional segmentation head.
3. The method for segmenting sea urchin instances and measuring high-throughput growth traits according to claim 2, characterized in that, The Haar wavelet downsampling module uses a two-dimensional Haar discrete wavelet transform to process the input feature map. Decomposed into low-frequency components Horizontal detail components Vertical detail components and diagonal detail components ; The four components are concatenated along the channel dimension to generate an intermediate feature map. ; The output features are downsampled by mapping to the target number of output channels C' through 1×1 pointwise convolution, followed by batch normalization and ReLU activation. .
4. The method for segmenting sea urchin instances and measuring high-throughput growth traits according to claim 2, characterized in that, The C2f_DCNv4 module will input the feature map as... After adjusting the number of channels using 1×1 convolution, the channels are evenly divided into reserved branches. and processing branches The processing branch For n Bottleneck structures containing DCNv4, branches will be preserved. Initial processing branch and the output of each Bottleneck The concatenation is performed along the channel dimension, followed by 1×1 convolutional fusion and dimensionality reduction to output the feature map. .
5. The method for segmenting sea urchin instances and measuring high-throughput growth traits according to claim 4, characterized in that, The Bottleneck structure first extracts local features through a 3×3 standard convolution. Then, the offset is dynamically predicted using DCNv4. and modulation coefficient For the reference sampling point corresponding to the output position Adaptive sampling formula for: , in, This represents the modulation coefficient of the k-th sampling point in group G. This indicates the reference sampling point at the current output position. This indicates a predefined sampling point offset. This represents the learned dynamic offset, and K represents the total number of sampling points; Predicting offsets using additional convolutional branches and modulation scalar : 。 6. The method for segmenting sea urchin instances and measuring high-throughput growth traits according to claim 2, characterized in that, The heterogeneous multi-scale attention network module will input the feature map First, the number of channels is adjusted using 1×1 convolution, and the features are then divided into g subgroups along the channel dimension. Each subgroup passes through a multi-scale convolutional module in parallel. This module contains at least three parallel branches, each extracting features from a different receptive field, and compressing the output channel count to [number missing]. The features output from each convolutional branch are concatenated along the channel dimension and then fused through a 1×1 convolution to obtain the multi-scale features corresponding to that subgroup. The features of each subgroup are input into the channel attention module and the spatial attention module respectively, and all subgroups are attention-enhanced features. The data is concatenated and connected to the input feature map via residual connections. Add them together to output the feature map. .
7. The method for segmenting sea urchin instances and measuring high-throughput growth traits according to claim 6, characterized in that, The channel attention module focuses on the features of the Jth subgroup. Perform global average pooling and global max pooling respectively. Two were obtained. The two descriptors are fed into a shared multilayer perceptron (MLP), the outputs are summed, and then activated by a sigmoid function to obtain the channel weights. ; The spatial attention module employs a directional decomposition mechanism, encoding features in both the height and width directions, and then processing the subgroup features. Perform max pooling and average pooling along the channel axis respectively to obtain two results. The feature maps are concatenated and then subjected to 7×7 convolution and sigmoid activation to obtain spatial weights. .
8. The method for segmenting sea urchin instances and measuring high-throughput growth traits according to claim 2, characterized in that, The dynamic snake-shaped convolution segmentation head will input the feature map Predict the initial offset using a 3×3 convolution. Where K is the number of sampling points, and is scaled using the tanh function to... The offset prediction formula is: ， in, This represents the initial coordinate offset matrix for prediction, α is the scaling factor, and BN represents the batch normalization operation. This represents a standard convolution with a kernel size of 3×3; Coordinates of the original sampling point Based on this, and according to the shape and orientation of the target to be segmented, the horizontal offset component of the initial prediction is adjusted along the X-axis or Y-axis. or vertical offset The process involves iterative accumulation, followed by bilinear interpolation to extract the final interpolated features corresponding to consecutive sampling points p. : ， In the formula, p represents continuous sampling points. Let p be the four integer pixels surrounding p, and q be the set. Integer pixels in the data. The original feature value corresponding to the integer pixel q. For distance-based interpolation weights, interpolation weights The calculation is as follows: ； The sampled features are weighted and summed according to the convolution kernel weights to obtain the final output features.
9. The method for segmenting sea urchin instances and measuring high-throughput growth traits according to claim 1, characterized in that, The sea urchin pixel area The calculation formula is: ， in, The total number of vertices. , Let a be the coordinates of the a-th vertex. , Let the coordinates be those of the (a+1)th vertex. The sea urchin weight trait prediction model includes: Based on the growth pattern of sea urchins, a physical proportional characteristic variable P is constructed. ， In the formula, E is the allometric growth parameter for the transformation of the two-dimensional projected area of the sea urchin into its three-dimensional equivalent volume; The interquartile range algorithm is used to remove extreme outliers from the physical ratio P. Let the first quartile be Q1 and the third quartile be Q3, then: , Samples that meet the following criteria are retained as high-quality training samples: , Samples that do not meet the above conditions will be considered abnormal samples and removed. Multiple machine learning regression models were built for fitting, and 2-3 mean absolute percentage error (MAPE) models were selected from them. And the coefficient of determination The regression model with a value greater than 0.9 was used to average and fuse the prediction results to obtain the final predicted value of sea urchin weight.
10. A system for segmenting sea urchin instances and measuring high-throughput growth traits, using the method for segmenting sea urchin instances and measuring high-throughput growth traits as described in any one of claims 1-9, characterized in that, include: The image acquisition module acquires calibrated images of sea urchins in the target sea area and images taken at the vertical shooting height from the seabed surface; The sea urchin instance segmentation module uses a pre-trained SUYOLO-seg model to extract the pixel boundaries of sea urchins in the underwater calibration image; The area calculation module calculates the actual physical area of the sea urchin based on the pixel boundaries of the sea urchin and the pre-calibrated actual physical area per unit pixel. The weight prediction module estimates the true weight of sea urchins based on a sea urchin weight trait prediction model that maps the area to the weight of the sea urchin. The output module provides a visualization of the sea urchin's area and weight.