A method and system for high-throughput phenotyping of small-sized macrobrachium rosenbergii shrimps

Abstract

The present application relates to the technical field of phenotype measurement in aquatic genetic breeding, and particularly relates to a high-throughput phenotype measurement method and system for small-sized Macrobrachium rosenbergii. In the measurement box, a dark box environment is constructed by a white shooting bottom plate, dark light-absorbing inner walls and an LED surface array light source, combined with preview graph rough segmentation and local gray histogram analysis, to adaptively adjust the light to make the shrimp body and the background form a stable bimodal gray distribution. On this basis, a segmentation-key point joint deep learning model is used to simultaneously output the shrimp body mask and standard key points, and the topological consistency loss constraint of the skeleton side length ratio and the included angle is introduced in the training to significantly reduce the key point mislabeling and drift caused by transparent body color. The present application realizes the rapid, non-contact, high-precision and high-throughput determination of the phenotype parameters and weight of small-sized Macrobrachium rosenbergii individuals, and provides a high-quality structured phenotype database for genetic evaluation and multi-trait joint selection.

Description

Technical Field

[0001] This invention relates to the field of phenotypic measurement technology in aquatic genetic breeding, and particularly to a high-throughput phenotypic measurement method and system for small-sized giant freshwater prawns. Background Technology

[0002] In aquaculture genetic breeding, accurate determination of individual phenotypic traits is a fundamental step in constructing selection indices, conducting genetic assessments, and selecting families. For economically important shrimp species such as the giant freshwater prawn (Macrobrachium rosenbergii), indicators such as body length, body width, carapace size, abdominal width, and weight not only reflect individual growth performance and body structure but also directly relate to key economic traits such as marketable size and processing yield. However, for a long time, breeding and production practices have mainly relied on manual contact measurements: operators hold the shrimp in position by hand and use calipers, rulers, and electronic scales to read and record each body size parameter manually. This method suffers from problems such as high labor intensity, low efficiency, strong subjectivity in readings, poor repeatability, easy recording errors, strong stress on shrimp, and even mechanical damage. It is difficult to meet the high-throughput measurement needs of tens of thousands of individuals and has become a major bottleneck restricting the efficiency of shrimp genetic breeding.

[0003] With the development of computer vision and deep learning technologies, both domestic and international researchers have begun to explore the use of machine vision to automate the determination of shrimp phenotypic traits. For example, patent CN112861872A proposes a method for determining the phenotypic data of Litopenaeus vannamei. By collecting top and side views of the shrimp, the method first uses object detection networks such as YOLO to extract the shrimp body region, then extracts the coordinates of several feature points based on keypoint detection networks such as HourglassNet and HRNet, and finally calculates multiple phenotypic data such as body length and body width by combining three-dimensional spatial transformation. This enables non-contact multi-point phenotypic measurement of Litopenaeus vannamei in an aquatic environment. For example, patent CN114612495B discloses a method for determining the phenotypic data of Litopenaeus vannamei. It uses the GCN image segmentation algorithm to divide the shrimp body into different regions such as the rostrum, carapace, and abdomen. Combining this with the pixel ratio factor of a reference object in the container, it calculates the length and area of ​​each part and uses a stereoscopic correction model to correct the abdominal length. Simultaneously, it constructs a neural network and an SVR model using pixel features from different parts as input to predict body length and weight, which improves the automation level of shrimp phenotypic determination to a certain extent. In processing and grading scenarios, patent CN114998565B proposes a device and method for detecting the size of semi-finished shrimp. By arranging an image acquisition module on a conveyor device, it automatically acquires images of semi-finished Argentine red shrimp and constructs a length detection model using parameters such as contour features and skeletal lines. This enables intelligent detection and grading of the length of large batches of shrimp on the processing line, significantly improving the efficiency and consistency of size detection.

[0004] Overall, the aforementioned existing technologies have played a positive role in promoting the development of this field in the following aspects: First, the introduction of deep convolutional networks, image segmentation, and key point detection networks has transformed shrimp contour extraction and feature point recognition from manual experience to data-driven automatic recognition, significantly improving the automation and objectivity of phenotypic determination; Second, the widespread use of dedicated shooting boxes, water boxes, or light boxes has standardized background color, lighting conditions, and shooting angles to a certain extent, reducing the interference of environmental changes on measurement results; Third, in the processing and grading direction, the combination of conveyor belts and visual inspection has enabled online detection and automatic grading of the length of commercial or semi-finished shrimp, improving the detection efficiency of industrial processes.

[0005] However, from the perspective of technical content and application targets, the aforementioned existing technologies still have several incompatibilities with early phenotypic measurements of small-sized Macrobrachium rosenbergii breeding. First, existing methods are mostly geared towards Litopenaeus vannamei or Litopenaeus vannamei, and the samples are mostly commercial-sized or close to adult size; there are almost no publicly available measurement methods specifically for small-sized Macrobrachium rosenbergii individuals with a body length of 3-5 cm. Small-sized Macrobrachium rosenbergii have robust chelicerae, short and thick abdomens, and more transparent body color. In terms of behavior, they are more prone to out-of-water stress and jumping, and it is difficult for them to maintain a stable standard posture for a short period of time. This makes it difficult to directly transfer the traditional "free swimming in water + three-dimensional skeleton fitting" approach to the short-term out-of-water measurement scenario; existing shrimp phenotyping methods focus more on adapting to bending postures and multiple perspectives in water, rather than on stable, repeatable, high-throughput measurements for small-sized, transparent-colored, and short-term out-of-water conditions.

[0006] Secondly, regarding image features and model design, while existing deep learning phenotyping methods incorporate keypoint detection networks, they primarily treat keypoint detection as a single-point regression problem. Evaluation metrics focus on the average error of keypoints or the overall phenotypic parameter error, failing to address the practical problems of "low contrast between local edges and feature points due to body transparency, and susceptibility to errors in manual / network annotation." They lack a systematic solution strategy encompassing both optical imaging and model structure. For example, in CN112861872A and CN114612495B, illumination control is mostly fixed or simply adjusted, with no adaptive brightness adjustment strategy based on the local grayscale histogram of the preview image. Furthermore, the keypoint detection network lacks topological consistency loss based on the standard skeleton side length ratio and angular relationship, making it difficult to correct single-point misalignment and loss through overall skeleton constraints when transparent body color weakens local texture and blurs edges.

[0007] Furthermore, regarding the measurement indicators and data fusion methods, existing technologies mostly focus on a few body size indicators and visual weight estimation, especially favoring "weight prediction from images," which is suitable for estimating the size of commercial shrimp and grading them for processing. However, for breeding scenarios that require obtaining multiple key body size parameters and actual weight data simultaneously to construct a multi-trait joint selection index, existing patent solutions generally do not provide sufficient support. For example, CN114612495B uses SVR model prediction of weight rather than actual weight, making it difficult to use as a high-precision target trait in genetic evaluation. CN114998565B only targets the length and size detection of semi-finished shrimp, without considering body structure parameters such as carapace width and abdominal width, and does not involve automatically constructing a structured phenotypic database by family and individual number.

[0008] However, after in-depth analysis of existing technologies and their application scenarios, we found that when dealing with the specific target of "early-stage breeding individuals of giant freshwater prawns with a body length of 3-5 cm," existing technologies generally exhibit inadequacies and technological gaps, as specifically manifested in the following ways:

[0009] First, existing technologies primarily measure species such as Litopenaeus vannamei (whiteleg shrimp), focusing on commercially viable sizes or adult stages. Specific measurement techniques for Macrobrachium rosenbergii, especially for its smaller (3-5 cm) breeding individuals, are extremely rare. However, small-sized Macrobrachium rosenbergii exhibit distinct biological characteristics: 1) their highly transparent body color results in low contrast between the shrimp and background in images, blurring their outlines and internal structural features; 2) they exhibit a strong stress response after being removed from water, readily jumping and struggling to maintain a stable posture for extended periods; 3) their small size makes traditional contact-based measurement methods difficult and prone to damage. These unique characteristics mean that directly applying existing techniques used for other species or larger shrimp faces significant challenges in terms of imaging stability, feature recognition accuracy, and measurement success rate.

[0010] Secondly, at the deep learning model level, existing technologies are typically optimized for general poses or background complexity. Their training datasets and model loss function designs do not explicitly consider the core challenge of "local overexposure, reflection, and blurring caused by transparent body color and water stress." Therefore, when applied to small-sized giant freshwater prawns, mislabeling, omission, or severe drift of feature points easily occurs, leading to errors in the calculation of body size parameters.

[0011] Furthermore, at the level of breeding applications, existing solutions mostly focus on estimating a single or a few body size parameters (such as body length) or predicting body weight through models. However, they are clearly insufficient to support the need for early breeding selection to simultaneously obtain more than seven detailed body size parameters, such as frontal horn length, cephalothorax length / width, and abdominal length / width, and accurately correlate them with actual body weight to construct a structured phenotypic database.

[0012] Therefore, existing technologies have not yet provided a complete solution to the series of interconnected technical problems faced by small-sized giant freshwater prawns, namely, the difficulty in imaging transparent body color, the difficulty in fixing stress posture, and the difficulty in simultaneously acquiring multiple traits. There is an urgent need for a high-throughput phenotyping method and system that integrates targeted imaging technology, robust identification algorithms, and unified data management. Summary of the Invention

[0013] The technical objective of this invention is to address the problems encountered in early phenotypic measurements of small-sized giant freshwater prawns (3-5 cm in body length) during breeding, such as high error rates in phenotypic annotation due to transparent body color, difficulties in image acquisition due to strong out-of-water stress and a tendency to jump, and the difficulty in simultaneously acquiring multiple body size parameters and true weight. This invention provides a high-throughput phenotypic measurement method and system based on dark-box imaging and deep learning instance segmentation. This method enables rapid, accurate, non-contact, and integrated measurement of multiple phenotypic parameters, including body length, body width, cephalothorax size, abdominal width, and rostrum length, as well as weight, in small-sized giant freshwater prawns, providing high-quality early phenotypic data support for giant freshwater prawn genetic breeding.

[0014] To achieve the objectives of this invention, the following technical solution is adopted:

[0015] A high-throughput phenotypic measurement method for small-sized Macrobrachium rosenbergii, the method comprising the following steps:

[0016] S1. Take out a single giant freshwater prawn with a body length of 3-5cm from the aquaculture water, wipe off the excess water on its surface, and place it on the white shooting base of the measuring box. The dark box lighting environment is provided by the LED array light source arranged inside the measuring box.

[0017] S2. An industrial camera fixed to the top of the measuring box captures a high-definition top-view image of the target giant freshwater prawn, while a high-precision electronic scale arranged below the shooting base simultaneously captures the weight data of the giant freshwater prawn, obtaining the synchronous measurement results of image and weight.

[0018] S3. Input the top view image obtained in step S2 into a pre-loaded deep learning model for segmentation and keypoint detection of giant freshwater prawns. The joint deep learning model includes a shared feature encoder, a shrimp body segmentation branch, and a keypoint detection branch. The shrimp body segmentation branch outputs a pixel-level shrimp body mask of the giant freshwater prawn, and the keypoint detection branch outputs a keypoint heatmap and coordinates corresponding one-to-one with standard keypoints within the region of interest defined by the shrimp body mask.

[0019] S4. Based on the standard key point coordinates output in step S3, automatically calculate at least the following phenotypic size parameters according to the following geometric relationships: rostrum length, total length, body length, carapace length, abdomen length, carapace width, and abdomen width; and combine the phenotypic size parameters with the weight data collected in step S2 to form a multiphenotypic parameter set for the individual giant freshwater prawn.

[0020] Preferably, in step S1, the control and processing terminal first captures a preview image at low brightness, performs automatic region segmentation on the preview image to obtain a rough outline of the shrimp body, calculates a local grayscale histogram in the neighborhood of the shrimp body outline, and adaptively adjusts the brightness and duty cycle of the LED array light source based on the histogram, so that the grayscale distribution of the shrimp body outline and the grayscale distribution of the background form a bimodal structure in the local histogram, thereby enhancing the contrast of the shrimp body outline and the neighborhood of the characteristic points under the condition of transparent body color.

[0021] The automatic region segmentation uses a lightweight semantic segmentation network or an adaptive threshold-based foreground extraction algorithm, which is only used to locate the rough outline of the shrimp body. The segmentation results are not directly used for parameter calculation, but only used to limit the statistical area of ​​the local gray-level histogram to reduce the impact of background noise on the adaptive adjustment of illumination.

[0022] As a preferred option, the standard key points in step S3 include the following 10 standard key points: frontal tip forhead, frontal tip root forhead_root, head_up, head_down, body_up, body_down, body_down, mind, tail tip, tail_tip, and eye, the midpoint between the roots of the upper and lower eye sockets.

[0023] In step S4, the following calculations are performed: the distance between forhead and forhead_root corresponds to the length of the frontal horn; the distance between forhead and tail_tip corresponds to the total length; the distance between tail_tip and eye corresponds to the body length; the distance between mind and eye corresponds to the length of the cephalothorax; the distance between mind and tail corresponds to the abdominal length; the distance between head_up and head_down corresponds to the width of the cephalothorax; and the distance between body_up and body_down corresponds to the abdominal width.

[0024] Preferably, the training dataset construction for the joint deep learning model includes: collecting top-view images of giant freshwater prawns with a body length of 3-5 cm under different family lines, different breeding ponds, and different transparency conditions; having trained annotators annotate key points for each image under a unified anatomical standard; generating a standard skeleton topology map using the annotated key points; and performing data enhancements such as rotation, flipping, brightness perturbation, and local blurring on the images to ensure that the model can still stably output key point coordinates that satisfy the skeleton topology constraints even under transparent body color, local overexposure, or reflective conditions.

[0025] As a preferred embodiment, in step S3, during the training phase of the joint deep learning model, the topological map of the standard skeleton of the giant freshwater prawn, formed by connecting key points in sequence, is used to construct a topological consistency loss term based on the ratio of the predicted key point coordinates to the side length and the included angle deviation of the standard skeleton. This term is then minimized together with the segmentation loss and the key point heatmap regression loss. Thus, in the case of local edge blurring caused by transparent body color, the overall skeleton topological constraints are used to suppress key point misalignment and loss.

[0026] Preferably, the topology consistency loss includes at least:

[0027] The length constraint loss is constructed by comparing the side length ratio of adjacent key points in the standard skeleton with the reference ratio obtained from the training samples.

[0028] And the angle constraint loss constructed by the difference between the included angle between two adjacent sides in the standard skeleton and the reference included angle;

[0029] The topology consistency loss is weighted and summed with the segmentation cross-entropy loss and the key point heatmap mean square error loss according to preset weights during the training process, and used as the total loss function of the joint deep learning model.

[0030] Preferably, the method further includes a step of quality verification of the measurement results:

[0031] The visualization interface simultaneously displays the original top view image, shrimp body segmentation mask, predicted skeleton topology, and multiple phenotypic parameters and weight data. Based on preset rules, it automatically verifies the segmentation connectivity, whether the skeleton is closed, whether key points fall within the shrimp body mask, and whether each phenotypic parameter falls within the reference range. If the preset quality standards are not met, the measurement is marked as invalid and prompts to re-execute steps S2 to S4. If the standards are met, the measurement result is marked as valid and written to the database.

[0032] Preferably, the method further includes, when performing batch testing on multiple giant freshwater prawns, the control and processing terminal automatically generates individual numbers and batch numbers according to the testing order, and when storing the multi-phenotypic parameter set and weight data, the individual numbers are simultaneously written into the corresponding top-view image file name and database record, so as to form a structured phenotypic database that can be traced by family, batch and individual.

[0033] Furthermore, the present invention also provides a high-throughput phenotypic measurement system for small-sized giant freshwater prawns to implement the method, comprising:

[0034] The measuring chamber has a white imaging base and dark matte inner walls.

[0035] An industrial camera is fixedly installed on the top of the measuring box, with its optical axis basically perpendicular to the shooting base plate, for acquiring high-definition top-view images of giant freshwater prawns;

[0036] An LED array light source is arranged inside the measuring chamber and surrounds the industrial camera, illuminating the shooting base plate to create an adjustable darkroom lighting environment inside the measuring chamber.

[0037] A high-precision electronic scale is installed below the shooting base plate and electrically connected to the control and processing terminal for synchronously collecting the weight data of giant freshwater prawns.

[0038] The control and processing terminal includes a processor, a graphics processing unit (GPU), and a memory. The memory stores a computer program and the aforementioned deep learning model for segmentation and key points of the giant freshwater prawn. When the computer program is executed, the control and processing terminal performs the steps of the method described above.

[0039] Furthermore, the present invention also provides a computer device, the computer device including a processor, a graphics processing unit and a memory, the memory storing a computer program, the computer program being executed by the processor and the graphics processing unit to cause the computer device to perform the method described thereon.

[0040] Furthermore, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a computer device, the computer device performs the high-throughput phenotypic measurement method for small-sized giant freshwater prawns.

[0041] This invention achieves significant improvements in several dimensions of technical performance by tightly coupling dark-box adaptive imaging with a segmentation-keypoint joint deep learning model for small individuals of the giant freshwater prawn (3-5 cm in length and highly transparent body color). First, preview-driven local grayscale histogram analysis and adaptive brightness adjustment of the LED array light source ensure clear and stable boundaries and contrast for the prawn's body contour even under transparent body color conditions, reducing the fundamental error of "unclear and inaccurate" feature points from an optical perspective. Second, through the joint network output of "prawn body mask + 10 standard anatomical keypoints," and by introducing topological consistency loss based on the skeleton side length ratio and included angle during the training phase, keypoint prediction is simultaneously influenced by local texture information and... The overall skeletal constraint can correct single-point drift and misalignment even in cases of local overexposure, reflection, or blurred edges, significantly reducing the error rate of trait feature point labeling and improving the repeatability and robustness of multiple body size parameters such as body length, cephalothorax length and width, and abdominal width. Furthermore, the method completes keypoint detection and multi-index geometric calculation in a single inference, and automatically correlates with the actual body weight obtained synchronously from an electronic scale. Single-tail measurement time is controlled to the order of seconds, significantly improving the throughput of massive individual phenotypic collection in the early stages of breeding. Finally, the structured phenotypic database formed around the unified keypoint system provides a high-quality, traceable quantitative basis for subsequent genetic evaluation and multi-trait joint selection, achieving a comprehensive improvement in measurement accuracy, measurement efficiency, and data utilization value. Attached Figure Description

[0042] Figure 1 Schematic diagram of the hardware structure of a high-throughput phenotyping system for small-sized giant freshwater prawns.

[0043] Figure 2 Schematic diagram of software system structure and functional modules.

[0044] Figure 3 Flowchart of a high-throughput phenotypic determination method for small-sized giant freshwater prawns.

[0045] Figure 4 Top view of Macrobrachium rosenbergii: 10 standard key points and skeletal topology diagram.

[0046] Figure 5 Schematic diagram of the joint segmentation-keypoint deep learning model structure.

[0047] Figure 6 This is a schematic diagram of the calculation results of the present invention. Detailed Implementation

[0048] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of the present invention.

[0049] I. Terminology and Implementation Environment

[0050] To facilitate understanding of this invention, some of the terms involved will be explained first:

[0051] Small-sized giant freshwater prawns: referring to prawns with a body length of... Individual giant freshwater prawns within the range exhibit relatively transparent body color and strong stress response when out of water.

[0052] Dark box lighting environment: refers to the imaging environment consisting of the box, white shooting base, dark matte inner wall and LED array light source, in which external ambient light is basically isolated.

[0053] Key points: These refer to 10 standard points on a top-view view of a shrimp that have clear anatomical significance and are used for geometric measurements.

[0054] orhead: the anterior point of the forehead tip;

[0055] forhead_root: the point at the root of the frontal apex;

[0056] head_up: The uppermost point of the widest part of the head;

[0057] head_down: The lowest point of the widest part of the head;

[0058] body_up: The uppermost point of the widest part of the abdomen, between the second and third abdominal segments;

[0059] body_down: The lowest point of the widest part of the abdomen, between the second and third abdominal segments;

[0060] mind: the central point of the first abdominal segment;

[0061] tail: the front end of the tail fan;

[0062] tail_tip: The furthest point at the end of the tail fan;

[0063] eye: The midpoint between the roots of the left and right eye sockets.

[0064] Standard skeleton topology diagram: A shrimp skeleton wireframe formed by connecting 10 key points as nodes in a predetermined order, used to constrain the relative positional relationships between key points.

[0065] Joint Segmentation-Keypoint Deep Learning Model: This refers to a neural network model that shares a feature encoder and simultaneously outputs a shrimp body segmentation mask and heatmaps and coordinates of 10 keypoints.

[0066] The implementation environment of this invention can be a giant freshwater prawn breeding farm or research laboratory, where the indoor temperature is generally controlled at... The measurement system is installed in a relatively fixed, unobstructed location, close to the breeding pond or temporary holding pond, to facilitate continuous sampling by the operator.

[0067] II. System Structure

[0068] 2.1 Hardware Structure (see...) Figure 1 )

[0069] Figure 1 A schematic diagram of the hardware structure of the high-throughput phenotyping system for small-sized giant freshwater prawns of the present invention is shown. The system mainly includes:

[0070] 2.1.1 Measuring box 100

[0071] The enclosure is preferably made of aluminum alloy or stainless steel, with a dark matte coating applied to the inner wall.

[0072] A transparent observation window and an openable lid are provided on the top of the enclosure. The inside of the lid can be covered with a matte material to prevent reflection.

[0073] A white high-diffuse-reflection imaging base plate 110 is installed at the bottom of the box. The base plate material can be frosted acrylic board or metal plate sprayed with white matte paint.

[0074] 2.1.2 Industrial Camera 120

[0075] It is fixedly installed at the top center of the measuring box, with the optical axis basically perpendicular to the shooting base plate.

[0076] Typical configuration is A color area-array camera with a pixel count, equipped with a 12-16mm fixed-focus lens.

[0077] 2.1.3 LED area array light source 130

[0078] Multiple LED array panels are arranged around the industrial camera, illuminating the shooting base plate evenly.

[0079] The light source is electrically connected to the control and processing terminal, and its brightness, current, and duty cycle can be adjusted to achieve precise brightness control.

[0080] 2.1.4 High-precision electronic scale 140

[0081] Installed below the shooting base plate, it transfers weight to the scale body through openings or a thin plate structure.

[0082] The range can be The resolution is no less than .

[0083] 2.1.5 Control and Processing Terminal 200

[0084] It consists of industrial control computers or high-performance workstations, equipped with multi-core CPUs, GPUs (such as NVIDIA RTX series), ample memory, and solid-state drives.

[0085] It communicates with industrial cameras via USB or GigE, with electronic scales via serial port or USB, and with LED light source controllers via digital I / O or serial port.

[0086] 2.1.6 Input / Output Devices

[0087] It includes a monitor, keyboard, and mouse, used for human-computer interaction, data viewing, and operation control.

[0088] 2.2 Software Structure (see...) Figure 2 )

[0089] Figure 2 The functional modules and data flow relationships of the software system of the present invention are shown, mainly including:

[0090] Equipment control and initialization module 210: responsible for connecting to industrial cameras, LED light sources, and electronic scales, setting parameters, and monitoring status.

[0091] Darkbox adaptive imaging module 220: performs preview shooting, coarse segmentation, local grayscale histogram statistics, and adaptive adjustment of LED brightness.

[0092] Image and weight synchronous acquisition module 230: After determining the lighting conditions, it acquires a single frame of high-definition image and a stable weight value at one time.

[0093] Joint Segmentation-Keypoint Inference Module 240: Calls a pre-trained deep learning model to segment the shrimp body and predict the coordinates of 10 keypoints in the image.

[0094] Multiphenotypic parameter analysis module 250: Calculates phenotypic parameters such as frontal angle length and body length based on the geometric relationship of key points.

[0095] Quality verification and retest control module 260: Automatically checks the segmentation mask, skeleton topology and parameter range, and controls whether to retest.

[0096] Data Management and Visualization Module 270: Responsible for result display, querying, statistics, and database storage.

[0097] III. Overall Method Flow

[0098] Figure 3This is a flowchart of the high-throughput phenotypic determination method for small-sized giant freshwater prawns according to the present invention, including steps S1 to S4 and quality verification and data storage steps. The following is in conjunction with… Figures 3-5 The specific implementation of each step is explained in detail.

[0099] IV. S1: Sample preparation and targeted dark-box imaging with transparent body color

[0100] 4.1 Sample Preparation

[0101] Operators randomly selected animals with a length of approximately [missing information] from the breeding pond or temporary holding pond. Individual giant freshwater prawns were placed in a shallow basin of water for a few minutes to alleviate stress. Before measurement, excess water was gently absorbed from the prawn's surface with a soft absorbent cloth to prevent water droplets from forming bright reflective spots on the photographic substrate. Then, a single giant freshwater prawn was gently placed in the center area of ​​the white photographic substrate.

[0102] 4.2 Preview Image Acquisition and Coarse Segmentation

[0103] After the device control and initialization module completes the camera power-on, exposure time, and initial gain settings, the darkroom adaptive imaging module starts the preview function, continuously capturing preview images using low LED brightness. For each preview image, the module performs the following steps:

[0104] 1) Scaling and Denoising: Scaling the image to a smaller resolution (e.g., Noise can be eliminated by using simple median filtering or Gaussian filtering.

[0105] 2) Coarse segmentation: The shrimp foreground is extracted from the background using Otsu thresholding or adaptive thresholding methods to obtain a coarse binary image. Isolated small objects are further removed using morphological opening operations, and small holes are filled using closing operations.

[0106] 3) Contour extraction: Extract the largest connected region from the binary image as the rough contour of the shrimp body.

[0107] This coarse segmentation step involves relatively little computation, does not rely on complex networks, and only provides an approximate foreground location for subsequent grayscale histogram calculations.

[0108] 4.3 Local Gray-Level Histogram Statistics and Bimodal Target Setting

[0109] After obtaining the rough outline of the shrimp body, the dark box adaptive imaging module expands the shrimp body outline by a certain number of pixels (e.g., 20 pixels) to form a local region. The gray-level distribution of the foreground (shrimp body) and background pixels in this region is statistically analyzed to obtain two gray-level histogram curves.

[0110] Ideally, after enhancing the transparency of the body color, the grayscale histogram of the shrimp body region and the background region should exhibit a "separated bimodal structure," meaning there is a significant grayscale difference between the shrimp body grayscale peak and the background grayscale peak. To simplify implementation, the following objectives can be set: the background grayscale peak is close to medium-high grayscale (e.g., 180–220); the shrimp body grayscale peak is lower than the background grayscale peak by a certain amount (e.g., 30–60 grayscale levels); and the difference between the two peaks is greater than a preset threshold (e.g., 40).

[0111] A lighting evaluation function can be defined using soft constraints, for example:

[0112] ;

[0113] in, The average gray level of the background. The average grayscale value of the shrimp body. The target background grayscale center, For the maximum allowable deviation, , These are the weighting coefficients. A higher value indicates a better distinction between the shrimp and the background under the current lighting conditions.

[0114] 4.4 LED Brightness Adaptive Adjustment

[0115] The darkroom adaptive imaging module is based on the above evaluation function. Evaluate preview images at different brightness levels, and use a simple search strategy (e.g., start with medium brightness and gradually increase or decrease the brightness) to find what... Maximum brightness setting. The specific steps are as follows:

[0116] 1) Set a median brightness when the system is powered on. Take a preview image and calculate ;

[0117] 2) Using step size Adjust the brightness up and down respectively. ,calculate , ;

[0118] 3) If If it is the largest, then let Continue searching in the same direction, or in the opposite direction, until... No further increase or reaching the brightness boundary;

[0119] 4) Will obtain the maximum The brightness value is denoted as This refers to the LED brightness during the formal data acquisition phase.

[0120] Through the above process, the system can automatically adjust the light source output under different batches, different transparency levels, or different ambient temperatures, so that the shrimp body outline maintains a stable contrast and grayscale range in the image, reducing the blurring and mislabeling of feature points caused by the transparency of the body color.

[0121] V. S2: Synchronous acquisition of image and weight

[0122] When the LED brightness is locked to Then, the image and weight synchronization acquisition module begins execution. This module mainly completes camera exposure settings, stable weight value reading, and synchronization triggering.

[0123] 5.1 Camera Exposure and White Balance Adjustment

[0124] Based on the average brightness of the preview image, the system can automatically adjust the camera exposure time and gain appropriately to ensure that the final captured image is neither overexposed nor underexposed. In a darkroom environment, the background is relatively uniform, so a fixed white balance is generally sufficient; however, if necessary, the center area of ​​the shooting substrate can be selected as a white balance reference.

[0125] 5.2 Weight stability judgment

[0126] After the shrimp is placed on the imaging base, the reading on the high-precision electronic scale will fluctuate slightly for a short period of time. The system continuously reads the weight value several times via serial port. The difference between five (e.g., five) readings is less than a preset threshold (e.g., ... When the weight has stabilized, its average value is taken as the final weight.

[0127] 5.3 Synchronous Acquisition Process

[0128] Regarding image acquisition timing, the system can adopt one of two methods:

[0129] Image capture followed by weight: First, trigger the camera to capture a high-resolution image, then read the stable weight;

[0130] Parallel acquisition: While waiting for the weight to stabilize, the camera is triggered as soon as the shrimp's posture is determined to be suitable for shooting, ultimately obtaining images and weights with similar timestamps.

[0131] This embodiment preferably uses a parallel acquisition method to reduce waiting time. Both image and weight data are timestamped and associated with the current individual ID, providing a basis for subsequent data management.

[0132] VI. S3: Joint Segmentation of Transparent Small-sized Giant Freshwater Prawns – Key Point Detection

[0133] This step is one of the core contributions of this invention. By designing a joint network structure and topological constraint loss, it is possible to output stable and reliable key point coordinates even under conditions of transparent body color, local reflection, and partial edge blurring.

[0134] 6.1 Definition of 10 Standard Key Points (see...) Figure 4 )

[0135] Figure 4 The locations of 10 standard key points of the present invention on a top view of the giant freshwater prawn are shown and labeled with numbers 1 to 10:

[0136] forhead: the anterior point of the forehead;

[0137] forhead_root: the point at the root of the frontal apex;

[0138] head_up: The uppermost point of the widest part of the head;

[0139] head_down: The lowest point of the widest part of the head;

[0140] body_up: The uppermost point of the widest part of the abdomen, between the second and third abdominal segments;

[0141] body_down: The lowest point of the widest part of the abdomen, between the second and third abdominal segments;

[0142] mind: the central point of the first abdominal segment;

[0143] tail: the front end of the tail fan;

[0144] tail_tip: The furthest point at the end of the tail fan;

[0145] eye: The midpoint between the roots of the left and right eye sockets.

[0146] Regardless of whether the shrimp is facing up or down, or slightly tilted to the left or right in the top view, the above anatomical position is used as the labeling and prediction standard to ensure that the meaning of key points is consistent across different samples.

[0147] 6.2 Joint Segmentation – Keypoint Network Structure (see...) Figure 5 )

[0148] Figure 5 A schematic diagram of the joint deep learning model of the present invention is shown. This model can be constructed as follows:

[0149] 1) Shared Feature Encoder

[0150] The ResNet or ResNet+FPN residual network structure is used to extract multi-scale features from the input image through multiple convolutions and pooling. The output is a set of high-dimensional feature maps, which are shared by the segmentation branch and the keypoint branch.

[0151] 2) Shrimp body segmentation

[0152] Based on the feature map, upsampling and convolutional layers are applied to output a single-channel segmentation probability map, which, after sigmoid activation, represents the probability that each pixel belongs to a shrimp body. A binary segmentation mask is obtained by applying a threshold (e.g., 0.5).

[0153] 3) Keypoint Detection Branch

[0154] Using shared feature maps and segmentation masks as input, the masks can be used as attention weights to effectively suppress background noise. A 10-channel keypoint heatmap is output through multi-layer convolution, with each channel corresponding to one keypoint. The location of the maximum response in each heatmap is taken as the keypoint coordinates; alternatively, a weighted average can be used to obtain sub-pixel level coordinates.

[0155] 4) Standard skeleton topology diagram

[0156] like Figure 6 As shown, 10 key points are used as nodes and connected in a preset order to obtain a main line such as "forhead→forhead_root→eye→mind→tail→tail_tip", as well as side branches connecting head_up, head_down, body_up, and body_down respectively. The average value of the side length ratio of each skeleton in the training samples and the branch angle are pre-calculated to construct topological constraints.

[0157] 6.3 Loss Function Design and Topological Constraints

[0158] During the model training phase, the total loss function can be written as:

[0159] ;

[0160] in, To divide the loss, For the regression loss of the key point heatmap, For skeleton topology consistency loss, The weights for each loss. Binary cross-entropy or Dice loss can be used to constrain the consistency between the shrimp body mask and the manually labeled mask. Mean squared error is used to compare the differences between the predicted heatmap and the Gaussian heatmap centered on the actual key points. Topology consistency loss is also considered. Including length constraints and angle constraints, for example:

[0161] ;

[0162] in, For the skeleton edge set, For the edges in the current sample The length ratio (relative to the length of the main skeleton). The reference length ratio obtained from statistics in the training set; For the set of included angles of the skeleton, The included angle in the current sample Size, For reference angle; and This represents the weighting coefficient. Through this loss, if a keypoint deviates in position due to local blurring, it will cause one or more side length ratios or angles to deviate from the reference value, thus being... As a penalty, the network automatically adjusts the coordinates of key points during training to make the overall skeletal shape more closely resemble the structure of a real shrimp.

[0163] 6.4 Model Training and Deployment

[0164] 1) Dataset Construction

[0165] Thousands of top-view images of giant freshwater prawns (3–5 cm in length) were collected from multiple families and farming batches, ensuring coverage of different water qualities, transparency levels, and postures. Trained annotators used annotation tools to annotate each image with a prawn segmentation mask and 10 key points, with key point locations based on anatomical standards. Data augmentation was performed on the original images using rotation, flipping, brightness perturbation, local blurring, and noise addition to improve model robustness.

[0166] 2) Training process

[0167] The network is trained using stochastic gradient descent or the Adam optimizer in mini-batch mode. The segmentation loss weights can be increased in the early stages of training. To ensure mask accuracy; gradually increase the weights of the topology loss in the later stages of training. The rationality of the skeleton structure is emphasized. By monitoring the average error of key points and the consistency of the skeleton using the validation set, the optimal model parameters are selected.

[0168] 3) Deployment and Reasoning

[0169] The trained model weights are imported into the control and processing terminal and run on the GPU. Performing a forward inference once for each acquired top-view image simultaneously outputs the shrimp mask and the coordinates of 10 key points. The inference time is typically in the tens of milliseconds, meeting high-throughput requirements.

[0170] VII. S4: Multiphenotypic Parameter Analysis and Data Fusion

[0171] After obtaining the coordinates of 10 key points, the multi-phenotypic parameter parsing module calculates various body size parameters according to the pre-set geometric relationships and combines them with the weight data to form a multi-phenotypic parameter set.

[0172] 7.1 pixels – Length calibration

[0173] During the system installation and debugging phase, operators place a graduated standard ruler on the imaging base and capture calibration images in a darkroom environment. The pixel width between adjacent graduation lines is then calibrated automatically or manually. Corresponding to the actual length (like Thus, the pixel-to-length conversion factor is obtained:

[0174] ;

[0175] in, The unit is All subsequent body size parameters will be converted using this coefficient.

[0176] 7.2 Calculation of Phenotypic Parameters

[0177] Let the pixel coordinates of forhead, forhead_root, head_up, head_down, body_up, body_down, mind, tail, tail_tip, and eye be respectively... ~ The parameters can be calculated as follows:

[0178] horn length : ;

[0179] in The distance between keypoint 1 and keypoint 2 is the Euclidean distance.

[0180] full length : ;

[0181] Body length : ;

[0182] Cephalothorax : ;

[0183] Abdomen : ;

[0184] Cephalothorax : ;

[0185] Wide abdomen : ;

[0186] In the formula, Indicate key points With key points pixel distance, , , , , These represent the length of the brow horn, total length, body length, cephalothorax length, and abdomen length, respectively. , These represent the width of the cephalothorax and the width of the abdomen, respectively.

[0187] The above calculations yield multiple body size parameters for this giant freshwater prawn. The system can further calculate derived parameters such as body length / body width ratio and carapace length / abdomen length ratio as needed, for more detailed body shape evaluation.

[0188] 7.3 Fusion of Multiphenotypic Parameter Sets and Weight Data

[0189] The stable weight value obtained during the image acquisition phase is denoted as The multiphenotypic parameter parsing module combines all body size parameters and weight into a multidimensional vector, for example:

[0190] ;

[0191] in, This represents the set of multiphenotypic parameters for the individual. The system can also record the family pedigree number, batch number, collection timestamp, and image file path, and package them into a complete record for storage in the database.

[0192] VIII. Quality Verification and Visualization Interface

[0193] A schematic diagram of the human-computer interaction interface of the measurement software, including the original image display area, the segmentation result display area, the skeleton topology display area, the data list, and the status prompt bar.

[0194] The quality verification and retest control module performs the following checks on the results after each measurement:

[0195] 1) Segment connectivity check: Is the number of connected regions of the shrimp body mask 1, and is the total area within a reasonable range?

[0196] 2) Check the legality of key point locations: whether all 10 key points fall within the shrimp body mask and are not near the edge of the image.

[0197] 3) Skeleton topology check: Check whether the deviation of each side length ratio and included angle from the reference value is less than the set threshold.

[0198] 4) Body size and weight reasonableness check: for example, whether the body length is within the acceptable range. Whether body length and weight are within the empirical regression range.

[0199] If any check fails, the system will highlight the individual in red on the interface and display the message "Measurement quality is unqualified, remeasurement is recommended." The operator can then reposition the individual and repeat steps S2 through S4. If all checks pass, the system will mark the results as valid and save them automatically.

[0200] IX. Batch Testing and Phenotypic Database Construction

[0201] For each batch of testing, the operator first enters or selects the family number, batch number, and rearing pond number in the software interface. The system automatically generates an individual ID for each giant freshwater prawn, starting from 1, and writes this number into:

[0202] Image file name, such as "A01_batch1_ID0001.png";

[0203] Primary key field of a database record;

[0204] The first column of the exported CSV file.

[0205] Through the above mechanism, any database record can be traced back to the original image and the corresponding individual, thus providing a reliable data foundation for subsequent genetic evaluation and multi-trait joint selection.

[0206] 10. Optional Embodiments and Variations

[0207] 1) Regarding camera and light source configuration

[0208] Without affecting the principle of darkroom imaging, the industrial camera can have a higher or slightly lower resolution, and the light source can adopt different structures such as ring lights or strip lights, as long as it can achieve uniform illumination and controllable brightness of the shooting area. This invention does not limit the specific model.

[0209] 2) Regarding the joint network structure

[0210] The shared feature encoder can adopt other mainstream backbone networks, such as EfficientNet and HRNet; the keypoint detection branch can also be implemented using a Transformer structure. As long as the shrimp body mask and the coordinates of 10 keypoints are output simultaneously, and skeleton topological constraints are introduced during the training phase, it constitutes an equivalent substitution of this invention.

[0211] 3) Regarding the form of topological loss

[0212] The specific expressions of angle and length constraints can vary, for example, by using cosine distance or Huber loss; or by adding weights to certain edges or included angles to strengthen the constraints on the frontal or tail regions. These modifications do not alter the essence of the invention.

[0213] 4) Regarding the expansion of application scope

[0214] Although this invention is mainly aimed at giant freshwater prawns with a body length of 3-5 cm, it is equally applicable to early-stage giant freshwater prawns with slightly larger or smaller body lengths, simply by adding corresponding samples during the training phase; for other economically important shrimp species with similar body structure, the overall technical approach of this invention can also be adopted by adjusting the key point definition and training dataset.

[0215] As can be seen from the above embodiments, this invention forms a complete technical route for small-sized, transparent-body-colored giant freshwater prawns in terms of dark-box adaptive imaging, segmentation-keypoint joint networks, and skeleton topological constraints. It achieves rapid, non-contact, and high-precision integrated measurement of multiple phenotypic parameters and true weight, providing high-quality early phenotypic data support for the genetic breeding of giant freshwater prawns. Those skilled in the art, after reading this specification, can make various modifications or improvements without inventive effort, and all such modifications or improvements should fall within the protection scope of this invention.

[0216] Application Example: Verification of the effect of high-throughput determination of phenotypic characteristics in giant freshwater prawns with a body length of 3-5 cm

[0217] The following is a complete application example to quantitatively demonstrate the technical effectiveness of the present invention in the "phenotyping of small-sized transparent body-color giant freshwater prawns".

[0218] I. Experimental Objectives and Subjects

[0219] This embodiment involves a phenotypic comparison experiment conducted on small-sized individuals (3-5 cm in length) of *Macrobrachium rosenbergii* in the early growth stage at a *Macrobrachium rosenbergii* breeding farm. The purpose is to:

[0220] 1) Verify the accuracy and repeatability of the method of the present invention in identifying key points on small-sized, transparent giant freshwater prawns;

[0221] 2) Compare the errors and correlations between traditional manual contact measurement and the system of this invention in key phenotypic parameters such as body length, cephalothorax length, and abdominal width;

[0222] 3) Compare the differences between the two methods in terms of measurement efficiency and operational intensity;

[0223] 4) Demonstrate the comprehensive technical effectiveness of this invention in reducing the error rate of phenotypic feature point labeling and improving the quality of simultaneous measurement of multiple phenotypic parameters and weight.

[0224] The experimental subjects were 300 giant freshwater prawns randomly selected from the same breeding population. Their body length was controlled between 30 and 50 mm (3 to 5 cm) after pre-screening. Their body color was generally transparent and they had typical morphological characteristics of small-sized giant freshwater prawns.

[0225] II. Experimental Grouping and Measurement Methods

[0226] To eliminate interference from individual differences, this embodiment adopts the method of "measuring the same shrimp using two different methods" without grouping individuals. Each shrimp is first measured using a manual contact method, and then measured using the method of this invention, keeping the measurement order consistent.

[0227] 1. Manual contact measurement procedure (comparative method)

[0228] For each giant freshwater prawn, the following traditional process is used:

[0229] 1) After wiping the surface of the shrimp with a damp cloth, place it on a flat measuring board;

[0230] 2) One operator holds a vernier caliper to fix the shrimp's posture, making it as straight as possible;

[0231] 3) Another recorder reads the body length, cephalothorax length, and abdominal width based on the visually identifiable positions of the rostrum endpoint, tail fan endpoint, etc.

[0232] 4) Weigh the items using an electronic table scale and record the weight manually;

[0233] 5) Each shrimp requires at least two people to work together, and the measurement time for a single shrimp is about 30 to 60 seconds.

[0234] Because small-sized shrimp are transparent and small in size, it is difficult to accurately identify phenotypic features such as the front of the rostrum and the widest part of the abdomen with the naked eye in artificial methods, which can easily lead to reading differences and poor repeatability.

[0235] 2. Measurement procedure of the system of the present invention

[0236] The high-throughput phenotyping system for small-sized giant freshwater prawns of this invention was used to measure 300 giant freshwater prawns from the same batch. Steps S1 to S4 were performed on each prawn:

[0237] 1) Adaptive imaging with dark box

[0238] A single shrimp is placed in the center of the white base plate of the measuring chamber. The adaptive imaging module of the dark chamber automatically completes the preview shooting, coarse segmentation, and local grayscale histogram statistics, and adjusts the brightness of the LED array light source until the grayscale of the shrimp body and the background form a bimodal structure in the histogram. This process takes about 0.5 to 1 second and requires no manual intervention.

[0239] 2) Image and weight acquisition are synchronized.

[0240] Once the lighting conditions are determined, the system automatically triggers the industrial camera to acquire a high-definition top-down image and reads a stable weight value from the high-precision electronic scale within 0.5 seconds. The acquisition time for a single tail image plus weight is less than 1.5 seconds.

[0241] 3) Joint Segmentation – Key Point Reasoning

[0242] The acquired top-down image is immediately fed into a joint deep learning model deployed on the GPU. The system outputs the shrimp body mask and the coordinates of 10 standard key points within approximately 30–50 milliseconds.

[0243] 4) Multiphenotypic parameter parsing and quality verification

[0244] The system automatically calculates parameters such as frontal horn length, total length, body length, cephalothorax length, abdomen length, cephalothorax width, and abdomen width based on the coordinates of forhead, forhead_root, head_up, head_down, body_up, body_down, mind, tail, tail_tip, and eye. It also automatically verifies segmentation connectivity, skeleton topology rationality, and parameter range. If preset standards are not met, the system automatically prompts for remeasurement; if the verification passes, the results are automatically saved to the database.

[0245] In actual testing, the total time from placing a single shrimp on the bottom plate to completing the calculation of multiple phenotypic parameters and weight is about 2 to 3 seconds, and one operator can complete the continuous testing.

[0246] III. Referencing "True Values" and Evaluation Indicators

[0247] To objectively evaluate the measurement accuracy and key point error of this invention, this embodiment introduces a third type of "high-precision benchmark measurement" as a reference true value:

[0248] After the experiment, 80 samples with high transparency and clear key points were randomly selected from 300 samples. Using a high-magnification magnifying glass and a high-resolution scale plate, experienced technicians magnified and marked 10 key points under static laboratory conditions. The body size parameters were calculated based on the distance between the key points and used as the approximate "true value".

[0249] Based on the true value, the absolute error, relative error, and correlation coefficient (R²) of the manual contact measurement and the measurement of the present invention system on indicators such as body length, cephalothorax length, and abdominal width are statistically analyzed.

[0250] Simultaneously, the average measurement time per tail, number of operators, and proportion of non-compliant measurements were recorded for both manual measurement and the method of this invention on 300 tail samples.

[0251] IV. Experimental Results

[0252] 1. Comparison of the accuracy of key phenotypic parameters (80 tails)

[0253] Using the high-precision true value as a reference, the mean absolute error, mean relative error, and correlation coefficient R² of three representative indicators (body length, carapace length, and abdominal width) in the 80 tail samples were statistically analyzed. The results are illustrated in Table 1.

[0254] Table 1. Comparison of measurement accuracy of different methods and true values ​​(n=80)

[0255]

[0256] As can be seen from Table 1:

[0257] In small-sized, transparent-bodied giant freshwater prawns, the method of this invention showed an average absolute error of approximately 40%–50% compared to manual methods for three key indicators: body length, carapace length, and abdominal width. The relative error decreased from 6.8%–9.1% for manual methods to approximately 3%. The correlation coefficient R² with the true values ​​improved from 0.82–0.88 to 0.92–0.96, indicating a high degree of consistency between the measurement results of this invention and the high-precision true values. This demonstrates that through adaptive dark-box imaging, joint segmentation-keypoint network, and skeleton topological constraints, this invention effectively improves the accuracy of keypoint localization and body size calculation for small, transparent-bodied prawns.

[0258] 2. Key point annotation error rate and repeatability

[0259] To evaluate the improvement in the "feature point annotation error rate" caused by body color transparency, this embodiment defines the keypoint error exceedance rate: when the distance deviation of a keypoint from the true value exceeds 1.5mm (approximately 3% to 5% of a body length of 3-5cm), it is recorded as a keypoint error exceedance. Statistics were collected from 80 tail samples and 10 keypoints, totaling 800 keypoints, yielding the following data (Table 2).

[0260] Table 2 Comparison of Key Point Error Exceedance Rates (n=80, Total Number of Key Points=800)

[0261]

[0262] As shown in Table 2, the keypoint error over-limit rate of the method of the present invention is only 3.5%, which is significantly lower than the 18.5% of manual contact annotation. This demonstrates that: adaptive dark-box imaging reduces boundary blurring caused by transparent body color; and the use of shrimp body mask + keypoint joint network and skeleton topological constraints suppress single-point drift, resulting in a significant improvement in the overall consistency of keypoints. Furthermore, the standard deviation of the measurements was calculated by repeatedly measuring the same shrimp three times. The results show that the coefficient of variation (CV) of the repeatability of the method of the present invention is generally less than 2% for indicators such as body length, carapace length, and abdominal width, while that of the manual method is generally above 5%, further indicating that the measurement repeatability of the present invention is better.

[0263] 3. Measurement efficiency and operating load

[0264] In the full batch testing of 300 giant freshwater prawn samples, the average single-tail testing time and the number of operators required by the manual method and the method of the present invention were statistically analyzed, and the results are shown in Table 3.

[0265] Table 3 Comparison of measurement efficiency and operating load (n=300)

[0266]

[0267] It can be seen that the single-tail measurement time of the method of the present invention is about 3 seconds, which is only about 1 / 15 of that of the manual method. For the same 300 small-sized giant freshwater prawns, the manual method requires about 3.75 hours (with 2 people working together), while the method of the present invention takes about 15 minutes (which can be completed by 1 person). In the early stage of breeding, when it is necessary to perform phenotypic screening on thousands or even tens of thousands of small prawns, the traditional manual method is almost infeasible, while the present invention can complete large-scale high-throughput measurement within a reasonable time, which significantly reduces labor costs and operational intensity.

[0268] V. Technical Effect Analysis and Summary

[0269] As can be seen from this embodiment, the present invention has the following comprehensive technical advantages compared with traditional manual contact measurement:

[0270] 1) Significantly reduces the error rate of feature point annotation caused by transparent body color.

[0271] By using dark-box adaptive imaging, the grayscale values ​​of the shrimp body and the background are made to form a bimodal structure in the local histogram, which enhances the contrast between the edge of the transparent shrimp and the neighborhood of key points. The joint segmentation-key point network further suppresses the key point shift when local blurring is constrained by the skeleton topology loss, reducing the key point error over-limit rate from 18.5% to 3.5%.

[0272] 2) Significantly improved accuracy and repeatability in body size measurement

[0273] On 80 samples, the average relative error of body length, carapace length and abdominal width by the method of the present invention was reduced to about 3%, and the correlation coefficient R² with the high-precision true value was increased to 0.92 to 0.96; the coefficient of variation of repeated measurements was generally less than 2%, which is significantly better than the manual method.

[0274] 3) The efficiency of measurement has been improved by orders of magnitude, meeting the needs of high-throughput breeding.

[0275] The single-tail determination time has been reduced from 45 seconds in manual methods to about 3 seconds, the number of operators has been reduced from 2 to 1, and the total time for 300 samples has been shortened from about 3.75 hours to about 15 minutes. This method can be extended to continuous high-throughput screening of tens of thousands of small-sized giant freshwater prawns.

[0276] 4) Multiple phenotypic parameters and actual weight are output in one integrated manner, with a high degree of data structuring.

[0277] This invention simultaneously obtains more than 7 body size parameters and actual weight in a single reasoning process, and automatically associates them with family number, batch number and timestamp to form a structured phenotypic database, providing a stable and traceable data foundation for subsequent genetic evaluation and multi-trait joint selection.

[0278] In summary, this embodiment fully demonstrates that the present invention can significantly improve measurement efficiency, reduce manual labeling error rate and operational intensity in the phenotypic determination of small-sized transparent body-colored giant freshwater prawns while ensuring high accuracy, and has significant practical application value and promotion significance.

[0279] The foregoing description of embodiments of the present invention, through which those skilled in the art are able to implement or use the present invention, will be readily apparent to those skilled in the art. Various modifications to these embodiments will be readily apparent to those skilled in the art. The general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novelty disclosed herein.

Claims

1. A high-throughput phenotypic measurement method for small-sized Macrobrachium rosenbergii, characterized in that, The method includes the following steps: S1. Take out a single giant freshwater prawn with a body length of 3-5 cm from the aquaculture water, wipe off the excess water on its surface and place it on the white photographic base plate of the measuring box. The dark box lighting environment is provided by the LED array light source arranged inside the measuring box. S2. An industrial camera fixed to the top of the measuring box captures a high-definition top-view image of the target giant freshwater prawn, while a high-precision electronic scale arranged below the shooting base simultaneously captures the weight data of the giant freshwater prawn, obtaining the synchronous measurement results of image and weight. S3. Input the top-view image obtained in step S2 into a pre-loaded deep learning model for segmentation and keypoint detection of the giant freshwater prawn. The joint deep learning model includes a shared feature encoder, a shrimp body segmentation branch, and a keypoint detection branch. The shrimp body segmentation branch outputs a pixel-level shrimp body mask of the giant freshwater prawn, and the keypoint detection branch outputs a keypoint heatmap and coordinates corresponding one-to-one with the standard keypoints within the region of interest defined by the shrimp body mask. S4. Based on the standard key point coordinates output in step S3, automatically calculate at least the following phenotypic size parameters according to the following geometric relationships: rostrum length, total length, body length, carapace length, abdomen length, carapace width, and abdomen width; and combine the phenotypic size parameters with the weight data collected in step S2 to form a multiphenotypic parameter set for the individual giant freshwater prawn. In step S1, the control and processing terminal first captures a preview image at low brightness, performs automatic region segmentation on the preview image to obtain a rough outline of the shrimp body, calculates a local gray-level histogram in the neighborhood of the shrimp body outline, and adaptively adjusts the brightness and duty cycle of the LED array light source based on the histogram, so that the gray-level distribution of the shrimp body outline and the gray-level distribution of the background form a bimodal structure in the local histogram, thereby enhancing the contrast of the shrimp body outline and the neighborhood of the characteristic points under the condition of transparent body color; the automatic region segmentation adopts a lightweight semantic segmentation network or a foreground extraction algorithm based on adaptive threshold, which is only used to locate the rough outline of the shrimp body. The segmentation result is not directly used for parameter calculation, but only used to limit the statistical area of ​​the local gray-level histogram to reduce the influence of background noise on the adaptive adjustment of illumination.

2. The method according to claim 1, characterized in that, The standard key points in step S3 include the following 10 standard key points: frontal tip forhead, frontal tip root forhead_root, head_up, head_down, body_up (the widest part of the head), body_down (the widest part of the abdomen between the second and third abdominal segments), body_down (the widest part of the abdomen between the second and third abdominal segments), mind (the center point of the first abdominal segment), tail tip, tail_tip, and eye (the midpoint between the roots of the upper and lower eye sockets). In step S4, the following calculations are performed: the distance between forhead and forhead_root corresponds to the frontal horn length, the distance between forhead and tail_tip corresponds to the total length, the distance between tail_tip and eye corresponds to the body length, the distance between mind and eye corresponds to the cephalothorax length, the distance between mind and tail corresponds to the abdomen length, the distance between head_up and head_down corresponds to the cephalothorax width, and the distance between body_up and body_down corresponds to the abdomen width.

3. The method according to claim 1, characterized in that, The training dataset construction for the joint deep learning model includes: collecting top-view images of giant freshwater prawns with body lengths of 3–5 cm from different families, different culture ponds, and different transparency conditions; having trained annotators annotate key points for each image under a unified anatomical standard; using the annotated key points to generate a standard skeleton topology map; and performing data enhancements such as rotation, flipping, brightness perturbation, and local blurring on the images to ensure that the model can still stably output key point coordinates that satisfy the skeleton topology constraints even under transparent body color, local overexposure, or reflective conditions.

4. The method according to claim 1, characterized in that, In step S3, during the training phase of the joint deep learning model, the topological map of the standard skeleton of the giant freshwater prawn, formed by connecting key points in sequence, is used to construct a topological consistency loss term based on the ratio of the predicted key point coordinates to the side length and the included angle deviation of the standard skeleton. This term is minimized together with the segmentation loss and the key point heatmap regression loss. Thus, in the case of local edge blurring caused by transparent body color, the overall skeleton topological constraints are used to suppress key point misalignment and loss.

5. The method according to claim 4, characterized in that, The topology consistency loss includes at least: The length constraint loss is constructed by comparing the side length ratio of adjacent key points in the standard skeleton with the reference ratio obtained from the training samples. And the angle constraint loss constructed by the difference between the included angle between two adjacent sides in the standard skeleton and the reference included angle; The topology consistency loss is weighted and summed with the segmentation cross-entropy loss and the key point heatmap mean square error loss according to preset weights during the training process, and used as the total loss function of the joint deep learning model.

6. The method according to claim 1, characterized in that, The method also includes a step of quality verification of the measurement results: The visualization interface simultaneously displays the original top view image, shrimp body segmentation mask, predicted skeleton topology, and multiple phenotypic parameters and weight data. Based on preset rules, it automatically verifies the segmentation connectivity, whether the skeleton is closed, whether the key points fall within the shrimp body mask, and whether each phenotypic parameter falls within the reference range. If the preset quality standard is not met, the measurement is marked as invalid and prompts to re-execute steps S2 to S4. If the standard is met, the measurement result is marked as valid and written to the database. And / or, the method further includes, when performing batch measurements on multiple giant freshwater prawns, the control and processing terminal automatically generates individual numbers and batch numbers according to the measurement sequence, and when storing the multi-phenotypic parameter set and weight data, the individual numbers are simultaneously written into the corresponding top-view image file name and database record, so as to form a structured phenotypic database that can be traced by family, batch and individual.

7. A high-throughput phenotypic measurement system for small-sized giant freshwater prawns to implement the method described in any one of claims 1 to 6, characterized in that, include: The measuring chamber has a white imaging base and dark matte inner walls. An industrial camera is fixedly installed on the top of the measuring box, with its optical axis basically perpendicular to the shooting base plate, for acquiring high-definition top-view images of giant freshwater prawns; An LED array light source is arranged inside the measuring chamber and surrounds the industrial camera, illuminating the shooting base plate to create an adjustable darkroom lighting environment inside the measuring chamber. A high-precision electronic scale is installed below the shooting base plate and electrically connected to the control and processing terminal for synchronously collecting the weight data of giant freshwater prawns. A control and processing terminal includes a processor, a graphics processing unit (GPU), and a memory. The memory stores a computer program and the aforementioned deep learning model for segmentation and key points of the giant freshwater prawn. When the computer program is executed, it causes the control and processing terminal to perform the steps of the method described in any one of claims 1 to 6.

8. A computer device, characterized in that, The computer device includes a processor, a graphics processing unit (GPU), and a memory, wherein the memory stores a computer program that, when executed by the processor and the GPU, causes the computer device to perform the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, the computer program, when executed by a computer device, causing the computer device to perform the high-throughput phenotyping method for small-sized giant freshwater prawns as described in any one of claims 1 to 6.

Images