CN-121582324-B - High-throughput phenotype measurement method and measurement system for macrobrachium rosenbergii small-size shrimps

CN121582324BCN 121582324 BCN121582324 BCN 121582324BCN-121582324-B

Abstract

The invention relates to the technical field of phenotype measurement in aquatic genetic breeding, in particular to a high-throughput phenotype measurement method and a measurement system for macrobrachium rosenbergii small-size shrimps. In the method, a camera bellows environment is built by a white shooting bottom plate, a dark extinction inner wall and an LED area array light source in a measuring box, and the illumination is adaptively adjusted by combining rough segmentation and local gray histogram analysis of a preview image, so that a stable bimodal gray distribution is formed between a shrimp body and a background. On the basis, a segmentation-key point combined deep learning model is adopted, a shrimp body mask and standard key points are output at the same time, topological consistency loss constraint of the side length proportion and the included angle of a skeleton is introduced in training, and the error marks and drifting of the key points caused by transparent body colors are obviously reduced. The invention realizes the rapid, non-contact and high-precision high-throughput measurement of the phenotypic parameters and the weights of the macrobrachium rosenbergii small-sized individuals, and provides a high-quality structured phenotypic database for genetic evaluation and multi-character joint selection.

Inventors

LI FENGBO
FAN YUNPENG
Ying Haihang
QI YUNHUI
ZHANG HAIQI
GAO QIANG
CHENG HAIHUA
LI BO
YUAN HUWEI
XU YANG
Shen Peijing
PENG FEI

Assignees

浙江省淡水水产研究所(浙江省淡水渔业环境监测站)

Dates

Publication Date: 20260508
Application Date: 20260126

Claims (9)

1. The high-throughput phenotype measurement method for the macrobrachium rosenbergii small-size shrimps is characterized by comprising the following steps of: S1, taking out macrobrachium rosenbergii with single tail body length of 3-5 cm from a culture water body, slightly wiping excessive water on the surface of the macrobrachium rosenbergii, placing the macrobrachium rosenbergii on a white shooting bottom plate of a measuring box body, and providing a camera bellows illumination environment by an LED area array light source arranged in the measuring box body; S2, acquiring a overlooking high-definition image of a target macrobrachium rosenbergii by an industrial camera fixed at the top of the measuring box body, and synchronously acquiring weight data of the macrobrachium rosenbergii by a high-precision electronic scale arranged below the shooting bottom plate to obtain a synchronous measurement result of the image and the weight; S3, inputting the overlooking image obtained in the step S2 into a preloaded macrobrachium rosenbergii segmentation-key point combined deep learning model, wherein the combined deep learning model comprises a shared feature encoder, a shrimp body segmentation branch and a key point detection branch, wherein the shrimp body segmentation branch outputs macrobrachium rosenbergii pixel level shrimp body masks, and the key point detection branch outputs key point thermodynamic diagrams and coordinates which are in one-to-one correspondence with standard key points in an interested area defined by the shrimp body masks: S4, automatically calculating phenotype size parameters at least comprising the following geometric relations of frontal angle length, full length, body length, head chest nail length, abdomen length, head chest nail width and abdomen width based on the standard key point coordinates output in the step S3, and combining the phenotype size parameters with the weight data acquired in the step S2 to form a multi-phenotype parameter set of the macrobrachium rosenbergii individual; In the step S1, a control and processing terminal shoots a preview image with low brightness, performs region automatic segmentation on the preview image to obtain a rough shrimp body contour, calculates a local gray histogram in a shrimp body contour neighborhood, adaptively adjusts the brightness and the duty ratio of an LED area array light source based on the histogram, so that the shrimp body contour gray distribution and the background gray distribution form a spaced double-peak structure in the local histogram, thereby enhancing the contrast of the shrimp body contour and a characteristic point neighborhood under the condition of body color transparency, wherein the region automatic segmentation adopts a lightweight semantic segmentation network or a foreground extraction algorithm based on an adaptive threshold value, is only used for positioning the shrimp body rough contour, and the segmentation result is not directly used for parameter calculation but is only used for limiting the statistical region of the local gray histogram so as to reduce the influence of background noise on illumination adaptive adjustment.
2. The method of claim 1, wherein the standard keypoints in step S3 comprise 10 standard keypoints including a frontal tip front end forhead, a frontal tip root forhead _root, a head-widest uppermost head_up, a head-widest lowermost head_down, a belly-widest uppermost body_up intermediate the second and third sections, a belly-widest lowermost body_down intermediate the second and third sections, a first belly center mind, a tail tip front end tail, a tail tip rearmost tail_tip, and an upper and lower eye socket root intermediate eye; The calculation in the step S4 is as follows, the distance between forhead and forhead _root corresponds to the frontal angle length, the distance between forhead and tail_tip corresponds to the full length, the distance between tail_tip and eye corresponds to the body length, the distance between mind and eye corresponds to the head pectoral length, the distance between mind and tail corresponds to the belly length, the distance between head_up and head_down corresponds to the head pectoral width, and the distance between body_up and body_down corresponds to the belly width.
3. The method of claim 1, wherein the training data set construction of the joint deep learning model comprises collecting top-down images of macrobrachium rosenbergii with body length of 3-5 cm under different families, different culture ponds and different transparency conditions, labeling key points on each image by trained labeling personnel under unified anatomical standard, generating a standard skeleton topological graph by using the labeled key points, and carrying out data enhancement such as rotation, overturning, brightness disturbance, local blurring and the like on the image, so that the model can still stably output key point coordinates meeting skeleton topological constraints under the conditions of transparent body color, local overexposure or reflection.
4. The method of claim 1, wherein in the training stage of the joint deep learning model in step S3, topology consistency loss terms are constructed for the side length proportion and the included angle deviation of the predicted key point coordinates and the standard skeleton by using a macrobrachium rosenbergii standard skeleton topological graph formed by sequentially connecting key points, and the topology consistency loss terms are minimized together with segmentation loss and key point thermodynamic diagram regression loss, so that the dislocation and loss of the key points are restrained by means of the topological constraint of the whole skeleton under the condition that the local edge is blurred due to the transparency of the body color.
5. The method of claim 4, wherein the topology consistency loss comprises at least: The length constraint loss of the difference structure between the side length proportion of the adjacent key points in the standard framework and the reference proportion obtained by statistics of the training sample; And the angle constraint loss of a difference structure between an included angle formed by two adjacent edges in the standard framework and a reference included angle; And the topological consistency loss and the segmentation cross entropy loss and the key point thermodynamic diagram mean square error loss are weighted and summed according to preset weights in the training process to serve as a total loss function of the joint deep learning model.
6. The method of claim 1, further comprising the step of performing a quality check on the measurement result: Simultaneously displaying an original overlooking image, a shrimp body segmentation mask, a predicted skeleton topology, multiple phenotype parameters and weight data on a visual interface, automatically checking whether segmentation connectivity, a skeleton is closed, key points fall in the shrimp body mask and whether the phenotype parameters fall in a reference range according to a preset rule, marking the measurement as invalid when a preset quality standard is not met, prompting to re-execute the steps S2-S4, marking the measurement as valid when the standard is met, and writing the measurement result into a database; And/or, when the macrobrachium rosenbergii is measured in batches, the control and processing terminal automatically generates an individual number and a batch number according to the measurement sequence, and when the multi-phenotype parameter set and the weight data are stored, the individual number is written into the corresponding overlook image file name and the database record at the same time, so that a structured phenotype database capable of being traced according to families, batches and individuals is formed.
7. A macrobrachium rosenbergii small scale shrimp high throughput phenotyping system for implementing the method of any one of claims 1-6, comprising: The measuring box body is internally provided with a white shooting bottom plate and a dark extinction inner wall; the industrial camera is fixedly arranged at the top of the measuring box body, the optical axis is basically vertical to the shooting bottom plate, and the industrial camera is used for collecting overlooking high-definition images of macrobrachium rosenbergii; The LED area array light source is arranged in the measuring box body, surrounds the industrial camera, irradiates towards the shooting bottom plate and is used for forming a camera bellows illumination environment with adjustable illumination in the measuring box body; the high-precision electronic scale is arranged below the shooting bottom plate and is electrically connected with the control and processing terminal, and is used for synchronously collecting the weight data of the macrobrachium rosenbergii; the control and processing terminal comprises a processor, a graphic processing unit GPU and a memory, wherein a computer program and the macrobrachium rosenbergii segmentation-key point combined deep learning model are stored in the memory, and the computer program enables the control and processing terminal to execute the method steps of any one of claims 1-6 when being executed.
8. A computer device comprising a processor, a graphics processing unit GPU and a memory, wherein the memory stores a computer program which, when executed by the processor and the graphics processing unit, causes the computer device to perform the method of any of claims 1 to 6.
9. A computer readable storage medium having stored thereon a computer program which, when executed by a computer device, causes the computer device to perform the high throughput phenotyping method of macrobrachium rosenbergii small scale shrimp of any one of claims 1-6.

Description

High-throughput phenotype measurement method and measurement system for macrobrachium rosenbergii small-size shrimps Technical Field The invention relates to the technical field of phenotype measurement in aquatic genetic breeding, in particular to a high-throughput phenotype measurement method and a measurement system for macrobrachium rosenbergii small-size shrimps. Background In aquaculture genetic breeding, accurate determination of individual phenotype traits is a fundamental link of constructing selection indexes, developing genetic evaluation and family breeding. For economic shrimps such as macrobrachium rosenbergii, the indexes such as body length, body width, head and chest armor size, abdomen width, weight and the like not only reflect individual growth performance and body type structure, but also directly relate to key economic characters such as commodity specification, processing yield and the like. However, for a long time, the breeding and production practice still mainly relies on manual contact measurement, namely, the gesture of the shrimp body is fixed by an operator, and parameters of each shrimp body are read item by means of a vernier caliper, a ruler and an electronic scale and recorded manually. The mode has the problems of high labor intensity, low efficiency, strong subjectivity of reading, poor repeatability, error recording, strong stress of prawn bodies, and even easy mechanical damage, and the like, is difficult to meet the high-throughput measurement requirement of thousands of individual movable and small, and becomes an important bottleneck for restricting the genetic breeding efficiency of the prawns. With the development of computer vision and deep learning technology, attempts are beginning to realize automatic determination of the phenotype of the prawns by using machine vision at home and abroad. For example, patent CN112861872a proposes a method for measuring the phenotype data of penaeus vannamei, by collecting top view and side view of penaeus vannamei, firstly intercepting the shrimp body area by using a target detection network such as YOLO, then extracting a plurality of characteristic point coordinates based on a key point detection network such as HourglassNet, HRNet, and further calculating multiple phenotype data such as body length and body width by combining three-dimensional space transformation, thereby realizing non-contact multipoint phenotype measurement of penaeus vannamei in water environment. For another example, patent CN114612495B discloses a method for determining the phenotype data of litopenaeus vannamei, the body of the litopenaeus vannamei is divided into different areas such as frontal angle, brachytherapy, abdomen by using a GCN image segmentation algorithm, the length and area of each part are calculated by combining the pixel scale factors of a reference object in a container, the abdomen length is corrected by using a three-dimensional space correction model, and meanwhile, a neural network and an SVR model are constructed by taking the pixel characteristics of different parts as inputs, so that the prediction of the body length and the body weight is realized, and the automation level of the phenotype determination of the litopenaeus vannamei is improved to a certain extent. In a processing and grading scene, patent CN114998565B proposes a semi-finished shrimp size detection device and method, an image acquisition module is arranged on a conveying device to automatically acquire an image of a semi-finished argentina red shrimp, a length detection model is built through parameters such as profile characteristics, skeleton lines and the like, and the lengths of a large number of shrimps on a processing line are intelligently detected and graded, so that the size detection efficiency and consistency are remarkably improved. The method has the advantages that the method has a positive pushing effect on the development of the field, namely, a deep convolution network, an image segmentation and a key point detection network are introduced, the shrimp body contour extraction and characteristic point identification are changed from manual experience to data-driven automatic identification, the automation and objectivity of phenotype measurement are greatly improved, the background color, the illumination condition and the shooting visual angle are standardized to a certain extent by means of a special shooting box body, a water box or an illumination box, the interference of environmental changes on a measurement result is reduced, and on the third, the online detection and automatic classification of the lengths of commodity shrimps or semi-finished shrimps are realized through the combination of a conveyor belt and visual detection in the processing classification direction, and the detection efficiency of an industrial link is improved. However, from the technical content and the analysis of application objects, the prior art still has a pl