Non-contact method and system for assessing the area of a live pig eye

Abstract

The application provides a non-contact live pig eye muscle area evaluation method and system, relates to the field of computer vision, and the method comprises the following steps: obtaining three-dimensional topographic data of a live pig to be measured and basic data of the live pig to be measured; performing feature extraction on the three-dimensional topographic data by using a feature extraction backbone network of a feature extraction network to obtain pose-invariant features irrelevant to changes in the pig body posture; performing mapping on the pose-invariant features by using a feature mapping subnetwork of the feature extraction network to obtain area-sensitive features sensitive to changes in the eye muscle area; matching the basic data of the live pig to be measured, the area-sensitive features and reference area-sensitive features in a preset feature library based on the similarity between the area-sensitive features and the reference area-sensitive features in the preset feature library to determine candidate reference samples; and determining an eye muscle area estimation value according to an eye muscle area reference value corresponding to the candidate reference samples. The technical problem that the prior art cannot realize efficient, accurate and anti-posture interference live pig eye muscle area evaluation under non-contact conditions is solved.

Description

Technical Field

[0001] This application relates to the field of computer vision, and in particular to a non-contact method and system for assessing the area of ​​live pig eye muscles. Background Technology

[0002] In the field of pig farming and breeding, eye muscle area is a key indicator for evaluating lean meat percentage, meat quality, and growth performance in pigs, and its accurate measurement is of great significance for genetic improvement. Currently, the measurement of live eye muscle area mainly relies on ultrasound or computed tomography (CT) technology. However, ultrasound measurement is a contact-based method; during operation, probe pressure and the pig's struggle can easily cause changes in muscle morphology and stress responses. Furthermore, the measurement results are highly dependent on the operator's professional skills and experience, making it difficult to achieve standardized and high-throughput operations in large-scale farming scenarios. While CT technology offers high accuracy, the equipment is expensive, maintenance costs are high, and it usually requires anesthesia of the pigs, increasing operational risks and health hazards. Therefore, there is an urgent need for a method to solve the technical problem of achieving efficient, accurate, and posture-independent assessment of live pig eye muscle area under non-contact conditions. Summary of the Invention

[0003] This application provides a non-contact method and system for assessing the area of ​​live pig eye muscles, which solves the technical problem that existing technologies cannot achieve efficient, accurate and posture-resistant assessment of the area of ​​live pig eye muscles under non-contact conditions.

[0004] To achieve the above objectives, this application adopts the following technical solution: Firstly, a non-contact method for assessing the eye muscle area of ​​live pigs is provided, comprising: acquiring three-dimensional morphological data and basic data of the live pig to be tested; the basic data includes breed information and age information; extracting features from the three-dimensional morphological data through the feature extraction backbone network of a feature extraction network to obtain posture-invariant features unrelated to changes in pig posture; mapping the posture-invariant features using the feature mapping sub-network of the feature extraction network to obtain area-sensitive features sensitive to changes in eye muscle area; matching the similarity between the basic data of the live pig to be tested, the area-sensitive features, and reference area-sensitive features in a pre-set feature library to determine candidate reference samples; wherein, the data elements in the pre-set feature library include: breed information, age information, reference values ​​for eye muscle area, and reference area-sensitive features; the pre-set feature library is constructed using training samples with labeled reference values ​​for eye muscle area; and determining the estimated value of the eye muscle area of ​​the live pig to be tested based on the reference values ​​for eye muscle area corresponding to the candidate reference samples.

[0005] In conjunction with the first aspect mentioned above, in one possible implementation, the feature extraction network includes: a feature extraction backbone network and a feature mapping subnetwork; wherein, the feature extraction backbone network is used to extract pose-invariant features from 3D topography data; and the feature mapping subnetwork is used to map pose-invariant features to area-sensitive features.

[0006] In conjunction with the first aspect mentioned above, in one possible implementation, the training process of the feature extraction backbone network includes: constructing a first training set, which contains 3D morphological data of multiple different poses corresponding to the same live pig; constructing multiple learnable surrogate vectors to form a surrogate tensor based on the first training set; inputting the 3D morphological data in the first training set into the feature extraction backbone network to obtain pose-invariant features; calculating the angle between the pose-invariant features and each learnable surrogate vector; constructing a first loss function based on the angle; and optimizing the parameters of the feature extraction backbone network using a gradient iterative algorithm based on the first loss function until convergence, thereby obtaining the trained feature extraction backbone network.

[0007] In conjunction with the first aspect mentioned above, in one possible implementation, the first loss function satisfies the following formula:

[0008] in, This represents the angle between the pose-invariant feature extracted from the current input 3D topography data and the i-th surrogate vector. , These are the pose-invariant features extracted from the current sample. For the surrogate vectors corresponding to other live pigs; This represents the cosine similarity between pose-invariant features and their corresponding proxy vectors, with 0.2 representing the preset angle boundary and 30 representing the scaling factor. n This indicates the category number of the live pig to which the currently input 3D topography data belongs; N This represents the total number of live pigs in the first training set, and also corresponds to the number of surrogate vectors in the surrogate tensor; This indicates that the pose-invariant features extracted from the current input 3D topography data are related to the first... n The angle between the proxy vectors; It is an exponential operation function; It is the natural logarithm function.

[0009] In conjunction with the first aspect mentioned above, in one possible implementation, the first loss function increases the cosine similarity between the pose-invariant feature corresponding to the current sample and its corresponding surrogate vector, and decreases the cosine similarity between the pose-invariant feature corresponding to the current sample and other surrogate vectors, thereby constraining the pose-invariant features corresponding to the same live pig to cluster towards their corresponding surrogate vectors and separate them from the surrogate vectors corresponding to other live pigs, so as to obtain pose-invariant features that are unrelated to pose changes.

[0010] In conjunction with the first aspect mentioned above, in one possible implementation, the training process of the feature mapping sub-network includes: constructing a second training set, which includes multi-pose 3D morphological data with labeled eye muscle area reference values; selecting 3D morphological data corresponding to multiple different eye muscle area reference values ​​from the same breed and age in the second training set, and sorting them according to the eye muscle area reference values ​​from smallest to largest to obtain a candidate 3D morphological data sequence; inputting the candidate 3D morphological data sequence into the trained feature extraction backbone network to extract the corresponding pose-invariant features; inputting the pose-invariant features into the feature mapping sub-network to obtain the corresponding area-sensitive features; constructing a second loss function based on the feature distance between area-sensitive features and the difference relationship between the corresponding eye muscle area reference values; and optimizing the parameters of the feature mapping sub-network using a gradient iteration algorithm based on the second loss function until convergence to obtain the trained feature mapping sub-network.

[0011] In conjunction with the first aspect mentioned above, in one possible implementation, candidate reference samples are determined by matching the basic data of the live pig to be tested, the area-sensitive features, and the similarity between these features and reference area-sensitive features in a pre-set feature library. This includes: obtaining the breed information and age information of the live pig to be tested; selecting data elements from the pre-set feature library that are identical to the breed information and age information; calculating the cosine similarity between the area-sensitive features and each reference area-sensitive feature in the data elements; removing data elements whose cosine similarity is lower than a preset similarity threshold, and determining the remaining data elements as candidate reference samples.

[0012] In conjunction with the first aspect mentioned above, in one possible implementation, the process of constructing the pre-set feature library includes: acquiring multi-pose 3D morphological data with labeled eye muscle area reference values; extracting reference area sensitive features from the multi-pose 3D morphological data using a feature extraction network; constructing data elements containing variety information, age information, eye muscle area reference values, and reference area sensitive features; calculating the cosine similarity between data elements corresponding to different eye muscle area reference values ​​under the same variety information and age information; when the proportion of data elements with a cosine similarity greater than a preset similarity threshold is greater than a preset proportion threshold, the corresponding data elements are removed; and constructing the pre-set feature library based on the retained data elements.

[0013] In conjunction with the first aspect above, in one possible implementation, the estimated value of the eye muscle area of ​​the live pig to be tested is determined based on the reference value of the eye muscle area corresponding to the candidate reference sample, including: normalizing the cosine similarity corresponding to the candidate reference sample based on the temperature parameter to determine the weight corresponding to each candidate reference sample; and performing a weighted summation of the reference values ​​of the eye muscle area corresponding to the candidate reference sample based on the weight to obtain the estimated value of the eye muscle area. In conjunction with the first aspect mentioned above, in one possible implementation, the weights corresponding to the candidate reference samples... Satisfy the following formula:

[0014] estimated area of ​​eye muscles Satisfy the following formula:

[0015] in, Indicates the first in the candidate reference sample j The weights of the reference values ​​for eye muscle area corresponding to each cosine similarity. and They represent the first and second candidates in the reference sample, respectively. j The and the first k Cosine similarity, Indicates temperature parameter, Indicates the first in the candidate reference sample k Reference values ​​for eye muscle area corresponding to each cosine similarity. Indicates the first in the candidate reference sample k The weights of the reference values ​​for eye muscle area corresponding to each cosine similarity.

[0016] Secondly, a non-contact live pig eye muscle area assessment system is provided, applicable to any of the methods in the first aspect above. The system includes: a data acquisition unit for acquiring three-dimensional morphological data of the live pig to be tested and basic data of the live pig to be tested; the basic data includes breed information and age information; a feature extraction unit for extracting features from the three-dimensional morphological data through the feature extraction backbone network of the feature extraction network to obtain posture-invariant features unrelated to changes in pig posture; mapping the posture-invariant features using the feature mapping sub-network of the feature extraction network to obtain area-sensitive features sensitive to changes in eye muscle area; a candidate reference sample extraction unit for matching the basic data of the live pig to be tested, the area-sensitive features, and the reference area-sensitive features in a pre-set feature library to determine candidate reference samples; wherein, the data elements in the pre-set feature library include: breed information, age information, eye muscle area reference value, and reference area-sensitive features; the pre-set feature library is constructed using training samples with labeled eye muscle area reference values; and an eye muscle area assessment unit for determining the estimated eye muscle area of ​​the live pig to be tested based on the eye muscle area reference value corresponding to the candidate reference sample.

[0017] This application provides a non-contact live pig eye muscle area assessment method and system. The feature extraction backbone network uses a surrogate vector contrast learning mechanism to force features from the same pig in different postures to cluster towards a specific surrogate vector, thereby extracting posture-invariant features robust to posture changes. Based on this, the feature mapping sub-network further mines area-sensitive features strongly correlated with eye muscle area values. This staged feature decoupling strategy fundamentally eliminates noise interference introduced by posture changes such as standing, walking, or muscle tension in live measurements, ensuring the stability and accuracy of subsequent area estimation. Simultaneously, a similarity-weighted estimation mechanism based on a pre-built feature library is adopted, rather than traditional end-to-end regression prediction. This involves fine-grained matching of the area-sensitive features of the test sample with reference features of known labels in the library, and weighted fusion using temperature-controlled Softmax normalized weights. This ensures that the estimation result does not rely solely on a single nearest neighbor but integrates information from multiple highly similar samples, effectively improving the model's generalization ability and prediction accuracy in complex scenarios. Furthermore, the entire process, based on non-contact acquisition and automated processing of 3D topographic data, avoids the physical contact, stress response, and anesthesia risks associated with traditional ultrasound or CT measurements. This significantly reduces operational barriers and labor costs, meeting the practical needs of rapid, high-frequency, and standardized monitoring of ocular muscle area in large-scale intensive farming scenarios. It solves the technical problem that existing technologies cannot achieve efficient, accurate, and posture-resistant assessment of live pig ocular muscle area under non-contact conditions.

[0018] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0019] Figure 1 A system architecture diagram of a non-contact live pig eye muscle area assessment system provided in this application embodiment; Figure 2 A flowchart illustrating a non-contact live pig eye muscle area assessment method provided in this application embodiment. Figure 1 ; Figure 3 A flowchart illustrating a non-contact live pig eye muscle area assessment method provided in this application embodiment. Figure 2 ; Figure 4 A flowchart illustrating a non-contact live pig eye muscle area assessment method provided in this application embodiment. Figure 3 ; Figure 5 A flowchart illustrating a non-contact live pig eye muscle area assessment method provided in this application embodiment. Figure 4 . Detailed Implementation

[0020] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0021] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0022] The non-contact live pig eye muscle area assessment method provided in this application embodiment can be applied to, for example... Figure 1 The non-contact live pig eye muscle area assessment system shown, such as Figure 1 As shown, the system includes: a data acquisition unit 101, used to acquire the three-dimensional morphological data of the live pig to be tested and the basic data of the live pig to be tested; the basic data includes breed information and age information; a feature extraction unit 102, used to extract features from the three-dimensional morphological data through the feature extraction backbone network of the feature extraction network to obtain posture-invariant features that are independent of changes in pig posture; and to map the posture-invariant features using the feature mapping sub-network of the feature extraction network to obtain area-sensitive features that are sensitive to changes in eye muscle area; a candidate reference sample extraction unit 103, used to match the basic data of the live pig to be tested, the area-sensitive features, and the reference area-sensitive features in the pre-set feature library to determine candidate reference samples; wherein, the data elements in the pre-set feature library include: breed information, age information, reference value of eye muscle area, and reference area-sensitive features; the pre-set feature library is constructed through training samples with labeled reference values ​​of eye muscle area; and an eye muscle area evaluation unit 104, used to determine the estimated value of eye muscle area of ​​the live pig to be tested based on the reference value of eye muscle area corresponding to the candidate reference sample.

[0023] To address the technical problem that existing technologies cannot achieve efficient, accurate, and posture-insensitive assessment of live pig eye muscle area under non-contact conditions, this application provides a non-contact method for assessing live pig eye muscle area. Figure 2 Schematic diagram of the non-contact live pig eye muscle area assessment method provided in this application embodiment Figure 1 ,like Figure 2 As shown, it includes: S201. Obtain the three-dimensional morphological data of the live pig to be tested and the basic data of the live pig to be tested.

[0024] Among them, three-dimensional morphological data refers to information that can characterize the spatial structure of the body surface of a live pig, which is acquired using a three-dimensional camera. Three-dimensional morphological data can be point cloud data, depth map data, three-dimensional mesh data, or a combination thereof; basic data refers to information used to describe the growth stage and genetic background of the live pig to be tested, including breed information and age information.

[0025] In one possible implementation, a 3D camera is deployed above the pigpen passage, weighing passage, or measurement area to continuously collect 3D image data at multiple moments during the natural walking of the live pigs. The collected raw 3D data is then processed to remove backgrounds, filter outliers, normalize coordinates, and unify scales to obtain standardized 3D morphological data. Simultaneously, the electronic ear tag information or the file information in the breeding management system corresponding to the live pig being tested is read to obtain breed information and age information, and a data association relationship is established with the 3D morphological data as input data for subsequent eye muscle area assessment.

[0026] It should be noted that the three-dimensional morphology data does not require the live pig to maintain a fixed posture, and the live pig is allowed to be in different natural states such as standing, walking, turning around or lowering its head; in order to improve the stability of subsequent feature extraction, the data of the main body area of ​​the pig can be retained first, and the interference data caused by the fence, the ground and other animals can be removed.

[0027] As an example, a structured light 3D camera was used to scan Duroc pigs. 3D point cloud data from multiple perspectives were continuously acquired as the pigs passed through the measurement channel. After point cloud filtering and coordinate standardization, 3D morphological data containing the contour information of the pig's back, waist, and buttocks were obtained. At the same time, the breed of the pig was identified as Duroc and its age as 180 days based on the ear tag record.

[0028] Based on the above steps, this step acquires three-dimensional data reflecting the external morphological characteristics of pigs and corresponding growth information in a non-contact manner. This avoids animal stress while ensuring data acquisition efficiency and provides a reliable data source for establishing the correlation between eye muscle area and body surface morphology.

[0029] S202. The feature extraction backbone network of the feature extraction network is used to extract features from the three-dimensional morphology data to obtain posture-invariant features that are independent of changes in pig posture.

[0030] Among them, posture-invariant features refer to information that can stably represent the inherent body shape and structure of a pig under different posture conditions; the feature extraction backbone network is used to extract this type of stable feature from three-dimensional morphology data.

[0031] In one possible implementation, a feature extraction network is first constructed. The backbone network for feature extraction can be implemented using PointNet, PointNet++, DGCNN, or other neural network structures suitable for 3D point cloud processing. This network encodes the input 3D shape data into pose-invariant feature vectors of fixed dimensions. The feature mapping subnetwork consists of at least one fully connected layer and is used to map the pose-invariant features to the eye muscle area feature space. Then, a first training set is constructed, containing 3D shape data collected from the same live pig in multiple different poses, and grouped according to the individual live pigs. For each live pig in the first training set, a corresponding learnable surrogate vector is created, forming a surrogate tensor. The 3D topography data from the first training set is input into the feature extraction backbone network to obtain pose-invariant features, and the angle between the pose-invariant features and each surrogate vector is calculated. Based on the principle that the surrogate vector corresponding to the current sample is closest and the distance to other surrogate vectors is farthest, the first loss function is constructed. The parameters of the feature extraction backbone network are continuously optimized through gradient iteration algorithm, so that the pose-invariant features obtained from the same live pig under different poses are clustered in the same region, while the pose-invariant features corresponding to different live pigs are separated from each other, thus obtaining the trained feature extraction backbone network.

[0032] It should be noted that the first loss function is used to constrain the posture-invariant features corresponding to the same live pig to cluster towards its corresponding surrogate vector and separate them from the surrogate vectors corresponding to other live pigs, so as to obtain posture-invariant features that are independent of posture changes. By introducing a preset angle boundary and scaling factor in the angle space, the posture-invariant features corresponding to the same live pig are made to cluster towards its surrogate vector and remain separated from the surrogate vectors corresponding to other live pigs, thereby eliminating the influence of posture changes; Preferably, the first loss function satisfies the following formula:

[0033] in, This indicates that the pose-invariant features extracted from the current input 3D topography data are related to the first... i The angle between the proxy vectors , These are the pose-invariant features extracted from the current sample. For the surrogate vectors corresponding to other live pigs; This represents the cosine similarity between pose-invariant features and their corresponding proxy vectors, with 0.2 representing the preset angle boundary and 30 representing the scaling factor. n This indicates the category number of the live pig to which the currently input 3D topography data belongs; N This represents the total number of live pigs in the first training set, and also corresponds to the number of surrogate vectors in the surrogate tensor. This indicates that the pose-invariant features extracted from the current input 3D topography data are related to the first...n The angle between the proxy vectors; It is an exponential operation function; It is the natural logarithm function.

[0034] Based on the above steps, this step can effectively eliminate the influence of posture changes on the morphology analysis results, making the subsequent eye muscle area estimation more dependent on the real structural features of the pig, thereby improving the stability and consistency of the evaluation results.

[0035] S203. The feature mapping subnetwork of the feature extraction network is used to map the pose-invariant features to obtain area-sensitive features that are sensitive to changes in eye muscle area.

[0036] Among them, area-sensitive features refer to feature representations that can reflect the degree of difference in the area of ​​different eye muscles; the feature mapping subnetwork is used to establish the correlation between the structural features of pigs and the area of ​​eye muscles.

[0037] In one possible implementation, pose-invariant features are input into a feature mapping sub-network. The feature mapping sub-network of the feature extraction network maps the pose-invariant features to obtain area-sensitive features that are sensitive to changes in eye muscle area. The feature mapping sub-network is also trained, and the training process is as follows: a second training set is constructed, containing multi-pose 3D topographic data with labeled eye muscle area reference values; multiple training samples of the same species and age but different eye muscle area reference values ​​are selected from the second training set and sorted in ascending order of eye muscle area reference values; the trained feature extraction backbone network extracts the pose-invariant features corresponding to each training sample, and then inputs them into the feature mapping sub-network to obtain the corresponding area-sensitive features; three samples with different eye muscle area reference values ​​are sequentially selected from the sorted samples to form training triplets, and a second loss function is constructed based on the feature distance between the area-sensitive features corresponding to the training triplets and the difference between the eye muscle area reference values, such that the distance between the area-sensitive features corresponding to samples with large differences in eye muscle area is greater than the distance between the area-sensitive features corresponding to samples with small differences in eye muscle area; wherein the eye muscle area reference values ​​corresponding to the training triplets satisfy an ascending order. The feature mapping subnetwork is trained using a second loss function until convergence. After training, pose-invariant features are input into the trained feature mapping subnetwork, which outputs the corresponding area-sensitive features.

[0038] It should be noted that the second loss function uses a ranking constraint to establish a monotonic correspondence between eye muscle area and feature distance, ensuring that area-sensitive features not only retain pig body structure information but also characterize the trend of eye muscle area changes. During training, training sample groups are preferably constructed under conditions of the same breed and age to reduce the impact of genetic factors and differences in growth stage on the model training results. The second loss function constrains the magnitude relationship between eye muscle area reference values ​​to maintain consistency with the distance relationship between area-sensitive features, ensuring that the sample distribution in the area-sensitive feature space can reflect the trend of eye muscle area changes; specifically, the larger the difference in eye muscle area, the larger the distance between the corresponding area-sensitive features.

[0039] It should also be noted that the same breed and same age constraints are used to eliminate the influence of genetic differences and growth stage differences on the training results, so that the area-sensitive features mainly reflect the change pattern of eye muscle area; the feature distance can be measured by Euclidean distance, cosine distance or other vector distance methods.

[0040] Preferably, the second loss satisfies the following formula:

[0041] in, This represents the area-sensitive feature vector obtained after processing the three-dimensional morphological data corresponding to three different ocular muscle area reference values ​​selected under the same breed and the same age conditions through the feature extraction backbone network and the feature mapping sub-network.

[0042] As an example, in a sample of Duroc pigs all 180 days old, training samples with eye muscle areas of 35cm², 45cm², and 55cm² were selected for training, such that the distance between the 55cm² sample and the 35cm² sample is greater than the distance between the 45cm² sample and the 35cm² sample.

[0043] As an example, the area-sensitive feature can be represented as a fixed-dimensional feature vector, which can be 64-dimensional, 128-dimensional, 256-dimensional, or other preset dimensions; preferably, the area-sensitive feature is a 128-dimensional feature vector.

[0044] Based on the above steps, this step can further transform the abstract body structure features into feature representations related to the variation of eye muscle area, thereby improving the accuracy of subsequent feature retrieval and area estimation.

[0045] S204. Based on the basic data of the live pig to be tested, the similarity between the area-sensitive features and the reference area-sensitive features in the pre-set feature library is matched to determine the candidate reference samples.

[0046] The data elements in the pre-set feature library include: variety information, age information, reference value of eye muscle area, and reference area sensitive features; the pre-set feature library is constructed using training samples with labeled reference values ​​of eye muscle area.

[0047] In one possible implementation, firstly, corresponding data elements are selected from a pre-set feature library based on the breed and age information of the live pig to be tested; then, the cosine similarity between the area-sensitive features of the live pig to be tested and each reference area-sensitive feature is calculated; data elements with a cosine similarity lower than a preset similarity threshold are removed, and the remaining data elements are used as candidate reference samples; wherein, during the construction of the pre-set feature library, the feature similarity between samples of different eye muscle areas under the same breed and age is further calculated, and data elements with high similarity anomalies are removed to improve the quality of the feature library.

[0048] It should be noted that the similarity threshold can be set based on historical verification results, for example, 0.95. When no candidate reference sample meets the criteria after filtering, an estimation failure message can be output or the similarity threshold can be lowered to perform the search again.

[0049] As an example, when the test sample is a 180-day-old Duroc pig, only the data elements corresponding to the 180-day-old Duroc pig are retrieved from the feature library, and the data elements with a cosine similarity greater than 0.95 are retained as candidate reference samples.

[0050] Based on the above steps, this step can limit the search scope to a set of samples with the same growth background, and further utilize feature similarity to select the most valuable samples, thereby improving the reliability of the estimation results.

[0051] S205. Determine the estimated value of the eye muscle area of ​​the live pig to be tested based on the reference value of the eye muscle area corresponding to the candidate reference sample.

[0052] Among them, the estimated eye muscle area is the prediction result calculated using the corresponding eye muscle area reference value of the candidate reference sample and the similarity weight; the temperature parameter is used to adjust the smoothness of the weight distribution of different candidate reference samples.

[0053] In one possible implementation, the cosine similarity of the candidate reference samples is processed by exponential mapping and normalized based on the temperature parameter to obtain the weight of each candidate reference sample. Then, the reference values ​​of the eye muscle area of ​​each candidate reference sample are fused by weighted summation to obtain the estimated value of the eye muscle area of ​​the live pig to be tested. Among them, the candidate reference sample with higher similarity gets a larger weight, thus having a greater impact on the final estimation result.

[0054] It should be noted that the temperature parameter is used to control the weight distribution. When the temperature parameter is small, the influence of highly similar samples is more prominent; when the temperature parameter is large, multiple candidate reference samples participate in the estimation, thereby improving the robustness of the estimation results.

[0055] Based on the above steps, this step makes full use of the eye muscle area information contained in historical reference samples and obtains continuous prediction results through similarity weighted fusion. Compared with directly using the results of a single sample, this can effectively reduce random errors and improve the accuracy of eye muscle area assessment.

[0056] Compared to traditional methods such as ultrasound measurement, the embodiments of this application can estimate the area of ​​the eye muscles without contact with a live pig, which can reduce animal stress response and human operation error and improve measurement efficiency. At the same time, by combining posture-invariant feature extraction with feature library retrieval, the accuracy and stability of the eye muscle area assessment results under different posture conditions are improved.

[0057] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 3 As shown, the above S204 can be specifically implemented through the following S301 to S304, which are explained in detail below: S301. Obtain the breed and age information of the live pig to be tested.

[0058] Among them, breed information is used to characterize the genetic lineage to which the live pig to be tested belongs, and age information is used to characterize the current growth stage of the live pig to be tested; breed information and age information together constitute the basic constraints used to limit the search scope.

[0059] In one possible implementation, the breed information and age information of the live pig to be tested are read from the basic data, and the reading results are standardized. Specifically, the breed information is uniformly converted into a preset breed code, and the age information is uniformly converted into a preset time unit representation. After standardization, a set of attributes to be retrieved is generated and used as input data for the subsequent feature library screening process.

[0060] It should be noted that the breed information and age information are preferably obtained from the breeding management system, electronic ear tag system or manually entered information; when there is a small range of error in the age information, it can be merged according to the preset age range to improve the success rate of feature database matching.

[0061] Based on the above steps, this step, by introducing variety information and age information to constrain the subsequent search scope, can reduce the interference of samples with different genetic backgrounds and different growth stages on the matching results and improve the comparability of candidate reference samples.

[0062] S302. Select data elements from the preset feature library that are the same as the variety information and age information.

[0063] Among them, the data element is the basic storage unit in the pre-set feature library. Each data element contains at least variety information, age information, reference value of eye muscle area, and reference area sensitive features. The pre-set feature library is a structured feature database constructed using training samples with labeled reference values ​​of eye muscle area.

[0064] In one possible implementation, breed information and age information are used as query conditions to traverse and search all data elements in the preset feature library; for each data element, it is determined whether its breed information is consistent with the live pig to be tested, and at the same time, it is determined whether its age information is consistent with the live pig to be tested or is in the same preset age range; when both conditions are met, the corresponding data element is added to the candidate set; after the traversal is completed, the set of data elements that meet the conditions is output.

[0065] It should be noted that when there are no data elements with exactly the same age in the preset feature library, the search can be expanded according to the preset age tolerance range, for example, data elements within the range of ±20 days are allowed to participate in the screening; however, the screening should prioritize the use of completely consistent variety information and age information to ensure sample consistency.

[0066] Based on the above steps, this step can eliminate incomparable sample data before feature matching, reduce the data scale of subsequent similarity calculations, and improve the relevance of search results.

[0067] S303. Calculate the cosine similarity between the area-sensitive feature and each reference area-sensitive feature in the data element.

[0068] Cosine similarity is used to characterize the degree of orientation consistency between two feature vectors in the feature space; the larger the cosine similarity value, the higher the similarity between the two feature vectors.

[0069] In one possible implementation, the area-sensitive feature corresponding to the live pig to be tested is denoted as q, and the reference area-sensitive features in the data elements are denoted as... For each reference area-sensitive feature, the cosine similarity is calculated using the following formula:

[0070] in, Indicates the first i Cosine similarity between a reference area-sensitive feature and the area-sensitive feature to be tested Represents the vector dot product. These represent the magnitudes of the corresponding vectors; after calculation, the corresponding cosine similarity set is obtained.

[0071] It should be noted that before calculating cosine similarity, area-sensitive features can be normalized to ensure that each feature vector has a uniform scale. At the same time, when the feature vector dimension is high, matrix operations can be used to calculate cosine similarity in batches to improve processing efficiency.

[0072] Based on the above steps, this step uses the similarity between area-sensitive features to quantify the association between the test sample and historical reference samples, thereby achieving refined matching based on feature space.

[0073] S304. Remove data elements with a cosine similarity lower than a preset similarity threshold, and determine the remaining data elements as candidate reference samples.

[0074] The preset similarity threshold is used to limit the minimum similarity requirement between the candidate reference sample and the live pig to be tested; the candidate reference sample refers to the data element that simultaneously meets the basic attribute constraints and feature similarity constraints.

[0075] In one possible implementation, the cosine similarity set obtained in S303 is read, and each cosine similarity is compared with a preset similarity threshold m. When the cosine similarity is greater than or equal to the preset similarity threshold, the corresponding data element is retained. When the cosine similarity is less than the preset similarity threshold, the corresponding data element is deleted. After all comparisons are completed, the retained data elements are output as candidate reference samples. If the number of retained candidate reference samples is zero, a matching failure message is output, or the similarity threshold is lowered according to preset rules and the filtering process is re-executed.

[0076] It should be noted that the preset similarity threshold can be determined based on historical verification results, for example, set to 0.95. Setting the threshold too high may result in an insufficient number of candidate reference samples, while setting the threshold too low may introduce reference samples that are significantly different from the sample to be tested. Therefore, it should be set according to the actual data distribution.

[0077] Based on the above steps, this step retains the reference sample that is closest to the characteristics of the live pig to be tested through a similarity threshold screening mechanism, thereby reducing the impact of low-relevance samples on the subsequent eye muscle area estimation results and improving the accuracy and stability of the evaluation results.

[0078] This embodiment implements a step-by-step screening process from basic attribute matching to feature similarity matching. Compared to directly searching through all samples, this scheme can effectively reduce interference from irrelevant samples, improve the correlation between candidate reference samples and the live pigs to be tested, and thus improve the accuracy and stability of the eye muscle area estimation results.

[0079] Furthermore, in one possible implementation of the embodiments of this application, combined with Figure 2 and Figure 3 ,like Figure 4 As shown, the process of constructing the preset feature library in S204 above can be specifically implemented through the following S401 to S405, which are explained in detail below: S401. Obtain multi-pose three-dimensional topographic data with labeled eye muscle area reference values.

[0080] Among them, the labeled reference value of eye muscle area refers to the actual value of eye muscle area obtained by ultrasound measurement, CT measurement or manual measurement after slaughter; multi-pose three-dimensional morphology data refers to three-dimensional point cloud data, depth map data or three-dimensional mesh data collected from the same live pig under different pose conditions.

[0081] In one possible implementation, live pig samples with completed eye muscle area measurements are obtained from historical measurement data of the farm, and three-dimensional morphological data of the corresponding live pigs under different postures are obtained simultaneously; the obtained data is cleaned, outlier is removed, coordinates are unified, and scale is normalized; at the same time, a one-to-one correspondence between the three-dimensional morphological data and the eye muscle area reference value is established to form a sample dataset for constructing a feature library.

[0082] It should be noted that the same live pig can correspond to multiple posture samples, and these posture samples share the same reference value for eye muscle area; it is preferable to cover live pigs of different breeds and different age stages to improve the applicability of the subsequent feature library.

[0083] Based on the above steps, this step can obtain standardized sample data containing real eye muscle area information, providing a data source for building a high-quality reference feature library.

[0084] S402. Use a feature extraction network to extract reference area sensitive features from multi-pose 3D topography data.

[0085] Among them, the reference area sensitive feature refers to the feature vector generated by the trained feature extraction network that can characterize the pattern of changes in eye muscle area.

[0086] In one possible implementation, each 3D topography data is input into a trained feature extraction backbone network to obtain the corresponding pose-invariant features; then the pose-invariant features are input into a trained feature mapping subnetwork to obtain the corresponding reference area sensitive features; the above process is repeated for all 3D topography data in the sample dataset to finally generate a set of reference area sensitive features corresponding to each sample.

[0087] It should be noted that the parameters of the feature extraction network remain fixed at this stage and no longer participate in training; only forward inference calculations are performed. The obtained reference area-sensitive feature dimension is consistent with the area-sensitive feature dimension output by S203.

[0088] Based on the above steps, this step can convert the original three-dimensional topography data into feature vectors with a unified expression form, thereby improving the efficiency of subsequent sample comparison and retrieval.

[0089] S403. Construct data elements that include variety information, age information, reference values ​​for eye muscle area, and reference area sensitive features.

[0090] In one possible implementation, for each reference area sensitive feature, the corresponding variety information, age information, and eye muscle area reference value are read from the original sample data; then, they are encapsulated according to a unified data structure to form data elements, and then all data elements are stored in the feature set to be screened.

[0091] It should be noted that data elements corresponding to different postures of the same live pig can be retained simultaneously to enhance the feature library's ability to cover posture changes.

[0092] S404. Calculate the cosine similarity between data elements corresponding to different reference values ​​of eye muscle area under the same variety information and age information.

[0093] In one possible implementation, each data element in the set of features to be screened is selected as the target data element in turn; among the remaining data elements, data elements with the same variety information and age information as the target data element but different reference values ​​for eye muscle area are selected; then the cosine similarity between the reference area sensitive features of the target data element and the reference area sensitive features of the above data elements is calculated.

[0094] It should be noted that only data elements with different reference values ​​for eye muscle area are compared; if the reference values ​​for eye muscle area are the same, they are not included in this step of the calculation.

[0095] S405. When the proportion of data elements with a cosine similarity greater than a preset similarity threshold is greater than a preset proportion threshold, the corresponding data elements are removed; a preset feature library is constructed based on the retained data elements.

[0096] Among them, the preset similarity threshold is used to determine whether two samples with different eye muscle areas are too similar; the preset proportion threshold is used to determine whether the target data element belongs to an abnormal sample, and the value range can be set to 5% to 30%.

[0097] In one possible implementation, for a target data element, the number of times its cosine similarity to all data elements corresponding to different eye muscle area reference values ​​exceeds a preset similarity threshold is counted, and the total number of comparisons is also counted. The proportion is then calculated. If the proportion exceeds a preset proportion threshold, and the difference between the eye muscle area reference values ​​corresponding to two data elements exceeds a preset area difference threshold, the target data element is determined to have abnormally similar feature representations to a large number of different eye muscle area samples, and therefore does not meet the feature discrimination requirements. The target data element is then removed from the feature set to be screened; otherwise, the target data element is retained. After traversing all data elements, the set of retained data elements is used to build a preset feature library.

[0098] It should be noted that deleted data elements usually correspond to measurement error samples, abnormal acquisition samples, or samples with weak feature representation capabilities; removing such samples can improve the quality of the feature library.

[0099] This embodiment constructs data elements containing breed information, age information, reference values ​​for eye muscle area, and sensitive features related to the reference area. Furthermore, it utilizes an abnormal similar sample removal mechanism to filter the data elements for quality, obtaining a pre-built feature library with high discriminative power. Compared to directly constructing the feature library using all training samples, this approach reduces the impact of abnormal and noisy samples on the similarity matching process, improves the accuracy of candidate reference sample retrieval, and thus enhances the accuracy and stability of live pig eye muscle area assessment results.

[0100] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 5 As shown, the above S205 can be implemented through the following S501 to S502, which are explained in detail below: S501. Based on the temperature parameter, normalize the cosine similarity of the candidate reference samples to determine the weight of each candidate reference sample.

[0101] The temperature parameter is used to adjust the degree of difference in weight distribution among candidate reference samples; the weight is used to characterize the contribution of the corresponding candidate reference sample to the final eye muscle area estimation result.

[0102] In one possible implementation, all candidate reference samples and their corresponding cosine similarities are obtained, and each cosine similarity is normalized using a preset temperature parameter to convert the similarity of different candidate reference samples into corresponding weights. Among them, the candidate reference sample with the higher the cosine similarity receives a larger weight, and the candidate reference sample with the lower the cosine similarity receives a smaller weight. After normalization, the sum of the weights of all candidate reference samples is 1, thus forming a weight set for subsequent eye muscle area estimation.

[0103] It should be noted that the temperature parameter is used to control the sensitivity of weight allocation; when the temperature parameter is small, the weight corresponding to the highly similar candidate reference sample increases further; when the temperature parameter is large, the weight difference between the candidate reference samples decreases. In practical applications, a suitable temperature parameter value can be determined based on historical verification results. In the embodiments of this application, the temperature parameter value ranges from 0.05 to 1; preferably, the temperature parameter value is 0.1.

[0104] Based on the above steps, this step can transform the feature similarity relationship between the candidate reference sample and the live pig to be tested into a quantitative contribution relationship, so that the reference sample that is closer to the live pig to be tested can play a greater role in the subsequent estimation process and improve the reliability of the evaluation results.

[0105] S502. Based on the weights, the reference values ​​of the eye muscle area corresponding to the candidate reference samples are weighted and summed to obtain the estimated value of the eye muscle area.

[0106] Among them, the reference value of eye muscle area is the actual eye muscle area annotation result corresponding to the candidate reference sample; the estimated value of eye muscle area is the predicted eye muscle area of ​​the final output of the live pig to be tested.

[0107] In one possible implementation, the weights corresponding to each candidate reference sample and the reference values ​​of the eye muscle area corresponding to each candidate reference sample are obtained; the reference values ​​of the eye muscle area are weighted and fused according to the corresponding weights, so that the candidate reference samples with larger weights have a greater impact on the final result, and the candidate reference samples with smaller weights have a smaller impact on the final result; after completing the weighted calculation of all candidate reference samples, the fusion result is output as the estimated value of the eye muscle area of ​​the live pig to be tested.

[0108] It should be noted that when there are many candidate reference samples, the candidate reference samples with the highest similarity ranking can be selected first to participate in the weighted calculation; when there are few candidate reference samples, all candidate reference samples can be used to participate in the calculation, thereby ensuring the stability of the estimation results.

[0109] Preferably, the weights corresponding to the candidate reference samples Satisfy the following formula:

[0110] estimated area of ​​eye muscles Satisfy the following formula:

[0111] in, Indicates the first in the candidate reference sample j The weights of the reference values ​​for eye muscle area corresponding to each cosine similarity. and They represent the first and second candidates in the reference sample, respectively. j The and the first k Cosine similarity, Indicates temperature parameter, Indicates the first in the candidate reference sample k Reference values ​​for eye muscle area corresponding to each cosine similarity. Indicates the first in the candidate reference sample k The weights of the eye muscle area reference values ​​corresponding to each cosine similarity.

[0112] Based on the above steps, this step makes full use of the eye muscle area information contained in multiple highly similar reference samples for fusion estimation. Compared with the method of directly using a single reference sample for prediction, it can reduce the error caused by abnormal samples and improve the accuracy and stability of the eye muscle area assessment results.

[0113] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0114] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of this application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and modifications of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and modifications.

Claims

1. A non-contact method for assessing the area of ​​live pig eye muscles, characterized in that, include: Acquire the three-dimensional morphological data and basic data of the live pig to be tested; the basic data includes breed information and age information; The feature extraction backbone network of the feature extraction network is used to extract features from the three-dimensional morphology data to obtain posture-invariant features that are independent of changes in pig posture. The pose-invariant features are mapped using the feature mapping subnetwork of the feature extraction network to obtain area-sensitive features that are sensitive to changes in eye muscle area. Candidate reference samples are determined by matching the basic data of the live pig to be tested and the similarity between the area-sensitive features and the reference area-sensitive features in the pre-set feature library; wherein, the data elements in the pre-set feature library include: breed information, age information, reference value of eye muscle area and reference area-sensitive features; the pre-set feature library is constructed by training samples with labeled reference values ​​of eye muscle area. The estimated value of the eye muscle area of ​​the live pig to be tested is determined based on the reference value of the eye muscle area corresponding to the candidate reference sample.

2. The method according to claim 1, characterized in that, The feature extraction network includes: a feature extraction backbone network and a feature mapping subnetwork; The feature extraction backbone network is used to extract pose-invariant features from the three-dimensional topography data. The feature mapping subnetwork is used to map the pose-invariant features into area-sensitive features.

3. The method according to claim 2, characterized in that, The training process of the feature extraction backbone network includes: Construct a first training set, which contains three-dimensional morphological data of multiple different poses corresponding to the same live pig; Based on the first training set, multiple learnable agent vectors are constructed to form an agent tensor; The 3D topography data from the first training set is input into the feature extraction backbone network to obtain pose-invariant features; Calculate the angle between the pose-invariant features and each learnable agent vector; A first loss function is constructed based on the included angle; Based on the first loss function, the parameters of the feature extraction backbone network are optimized using a gradient iterative algorithm until convergence, thus obtaining the trained feature extraction backbone network.

4. The method according to claim 3, characterized in that, The first loss function Satisfy the following formula: in, This represents the angle between the pose-invariant feature extracted from the current input 3D topography data and the i-th surrogate vector. , These are the pose-invariant features extracted from the current sample. For the surrogate vectors corresponding to other live pigs; This represents the cosine similarity between pose-invariant features and their corresponding proxy vectors, with 0.2 representing the preset angle boundary and 30 representing the scaling factor. n This indicates the category number of the live pig to which the currently input 3D topography data belongs; This represents the total number of live pigs in the first training set, and also corresponds to the number of surrogate vectors in the surrogate tensor; This indicates that the pose-invariant features extracted from the current input 3D topography data are related to the first... n The angle between the proxy vectors; It is an exponential operation function; It is the natural logarithm function.

5. The method according to claim 3, characterized in that, The first loss function increases the cosine similarity between the pose-invariant feature corresponding to the current sample and its corresponding surrogate vector, and decreases the cosine similarity between the pose-invariant feature corresponding to the current sample and other surrogate vectors. This constrains the pose-invariant features corresponding to the same live pig to cluster towards their corresponding surrogate vectors and separates them from the surrogate vectors corresponding to other live pigs, thereby obtaining pose-invariant features that are independent of pose changes.

6. The method according to claim 2, characterized in that, The training process of the feature mapping subnetwork includes: Construct a second training set, which includes multi-pose 3D topographic data with labeled eye muscle area reference values; Three-dimensional morphological data corresponding to multiple different ocular muscle area reference values ​​of the same breed and the same age are selected from the second training set and sorted in ascending order of ocular muscle area reference values ​​to obtain a candidate three-dimensional morphological data sequence. The candidate 3D shape data sequence is input into the trained feature extraction backbone network to extract the corresponding pose-invariant features. The pose-invariant features are input into the feature mapping subnetwork to obtain the corresponding area-sensitive features; A second loss function is constructed based on the feature distance between the area-sensitive features and the difference between the corresponding eye muscle area reference values; Based on the second loss function, the parameters of the feature mapping subnetwork are optimized using a gradient iterative algorithm until convergence, resulting in the trained feature mapping subnetwork.

7. The method according to claim 1, characterized in that, The process of matching candidate reference samples based on the basic data of the live pig to be tested and the similarity between the area-sensitive features and reference area-sensitive features in a pre-set feature library includes: Obtain the breed and age information of the live pig to be tested; Select data elements from the preset feature library that are identical to the variety information and age information; Calculate the cosine similarity between the area-sensitive feature and each reference area-sensitive feature in the data element; Data elements with a cosine similarity lower than a preset similarity threshold are removed, and the remaining data elements are identified as candidate reference samples.

8. The method according to claim 1, characterized in that, The construction process of the pre-set feature library includes: Obtain multi-pose 3D topographic data with labeled eye muscle area reference values; The reference area-sensitive features of the multi-pose 3D topography data are extracted using a feature extraction network. Construct data elements that include variety information, age information, reference values ​​for eye muscle area, and reference area-sensitive features; Calculate the cosine similarity between data elements corresponding to different reference values ​​of ocular muscle area under the same variety and age information; When the proportion of data elements with a cosine similarity greater than a preset similarity threshold exceeds a preset proportion threshold, the corresponding data elements are removed. A pre-built feature library is constructed based on the retained data elements.

9. The method according to claim 1, characterized in that, The step of determining the estimated value of the eye muscle area of ​​the live pig to be tested based on the reference value of the eye muscle area corresponding to the candidate reference sample includes: The cosine similarity of the candidate reference samples is normalized based on the temperature parameter to determine the weight of each candidate reference sample. Based on the weights, the reference values ​​of eye muscle area corresponding to the candidate reference samples are weighted and summed to obtain the estimated value of eye muscle area. Wherein, the weights corresponding to the candidate reference samples Satisfy the following formula: The estimated area of ​​the eye muscles Satisfy the following formula: in, Indicating the first candidate reference sample j The weights of the reference values ​​for eye muscle area corresponding to each cosine similarity. and These respectively represent the first in the candidate reference samples. j The and the first k Cosine similarity, This indicates the temperature parameter. Indicating the first candidate reference sample k Reference values ​​for eye muscle area corresponding to each cosine similarity. Indicating the first candidate reference sample k The weights of the reference values ​​for eye muscle area corresponding to each cosine similarity.

10. A non-contact live pig eye muscle area assessment system, applied to the method described in any one of claims 1-9, characterized in that, The system includes: The data acquisition unit is used to acquire the three-dimensional morphological data of the live pig to be tested and the basic data of the live pig to be tested; the basic data includes breed information and age information; The feature extraction unit is used to extract features from the three-dimensional morphology data through the feature extraction backbone network of the feature extraction network to obtain posture-invariant features that are independent of changes in pig posture; and to map the posture-invariant features using the feature mapping sub-network of the feature extraction network to obtain area-sensitive features that are sensitive to changes in eye muscle area. The candidate reference sample extraction unit is used to match the basic data of the live pig to be tested, the similarity between the area-sensitive features and the reference area-sensitive features in the pre-set feature library, and determine the candidate reference samples; wherein, the data elements in the pre-set feature library include: breed information, age information, eye muscle area reference value, and reference area-sensitive features; the pre-set feature library is constructed through training samples with labeled eye muscle area reference values; The eye muscle area assessment unit is used to determine the estimated value of the eye muscle area of ​​the live pig to be tested based on the eye muscle area reference value corresponding to the candidate reference sample.

Images