A method and system for assessing the body size and weight of laying hens based on image recognition

Abstract

This invention relates to the field of intelligent livestock and poultry farming technology, and discloses a method and system for assessing the body size and weight of laying hens based on image recognition. The method includes: acquiring video data of laying hen behavior and environmental parameter data; processing the video data to identify laying hen behavior patterns, while simultaneously analyzing the correlation between environmental parameter data and behavior to establish a state representation; constructing a strategy network to achieve stress state detection and automatic selection of the optimal assessment time, and extracting body shape features under stable postures; using the body shape features at the selected time, combined with the identified behavioral state information, to construct a weight assessment model; learning the feature distribution of the normal assessment process of the weight assessment model, and detecting abnormal assessment results through dual indicators of reconstruction error and probability deviation. This invention achieves high-precision and high-reliability assessment of laying hen body size and weight by integrating behavioral analysis, environmental perception, and quality control.

Description

Technical Field

[0001] This invention relates to the field of intelligent livestock and poultry farming technology, and more specifically, to a method and system for assessing the body size and weight of laying hens based on image recognition. Background Technology

[0002] With the large-scale development of the poultry industry, accurate monitoring of laying hen weight is of great significance for assessing their health status, optimizing feed formulation, and predicting production performance. Traditional methods for obtaining laying hen weight mainly rely on manual weighing, but this method is labor-intensive and inefficient. More importantly, the weighing process can cause stress to the laying hens, affecting their normal physiological state and egg production performance.

[0003] To address these issues, researchers began exploring non-contact weight assessment methods based on computer vision technology. Existing technologies mainly include single-image-based body shape feature extraction and machine learning-based weight prediction models. These methods establish a mapping model from image features to weight by analyzing the correlation between the physical characteristics of laying hens (such as body length, height, and chest circumference) and weight.

[0004] However, existing image recognition weight assessment technologies still have the following technical shortcomings:

[0005] Existing methods primarily rely on static image analysis, failing to adequately consider the impact of hen posture changes on body shape feature extraction. In actual farming environments, hen postures change frequently, including standing, crouching, foraging, and drinking. The appearance of the same hen varies significantly under different postures, leading to inconsistent accuracy of static image-based assessment methods. Furthermore, current technologies lack comprehensive consideration of environmental factors. Environmental parameters such as temperature, humidity, and light intensity significantly affect hen behavior, inducing stress behaviors (such as rapid running, wing flapping, and gathering). These stress behaviors interfere with the accurate extraction of body shape features, but existing methods have not established a correlation between environmental parameters and hen behavior. The correlation analysis mechanism of chicken behavior status; existing assessment methods lack an intelligent selection mechanism for the optimal assessment time. Due to the dynamic and complex nature of laying hen behavior, weight assessment at any time may encounter unfavorable assessment conditions (such as laying hens being in a state of stress, unstable posture, etc.). Existing technologies have failed to achieve automatic optimization of assessment timing, affecting the consistency and reliability of assessment results. Existing technologies lack quality control and anomaly detection mechanisms for assessment results. In complex breeding environments, due to factors such as changes in light, shading, and equipment failure, the assessment system may produce abnormal results, but existing methods cannot automatically identify and handle these abnormal situations, resulting in insufficient overall reliability of the assessment system.

[0006] Therefore, there is a need for an intelligent weight assessment technology that can comprehensively consider the behavioral patterns of laying hens, the influence of environmental factors, the timing of assessment, and the quality control of results, in order to improve the accuracy, stability, and practicality of non-contact weight assessment of laying hens. Summary of the Invention

[0007] This invention provides a method and system for assessing the body size and weight of laying hens based on image recognition, solving the technical problems of traditional laying hen weight measurement requiring manual grasping, which leads to stress response and unstable assessment accuracy.

[0008] This invention provides a method for assessing the body size and weight of laying hens based on image recognition, comprising:

[0009] Video data of laying hen behavior and environmental parameter data were acquired using a multimodal data acquisition system;

[0010] A three-dimensional convolutional neural network is used to process video data to identify the behavior patterns of laying hens. At the same time, the correlation between environmental parameter data and behavior is analyzed to establish a state representation.

[0011] Based on the state representation output, a policy network is constructed using a reinforcement learning algorithm to realize stress state detection and automatic selection of the best evaluation time, and to extract body shape features under stable posture.

[0012] By utilizing the body shape features selected at the time of the policy network and combining them with the identified behavioral state information, a weight assessment model is constructed to achieve the mapping from image features to weight.

[0013] By using a variational autoencoder to learn the feature distribution of the normal assessment process of a weight assessment model, abnormal assessment results are detected by reconstructing error and probability bias, thus achieving self-verification of the assessment process.

[0014] Furthermore, the collection of video data on laying hen behavior includes:

[0015] Multiple high-definition cameras were deployed to monitor the activity area of ​​the laying hens from different angles;

[0016] The acquired video sequences are preprocessed, including spatial resolution normalization, temporal length normalization, noise filtering, and illumination compensation.

[0017] The preprocessed video sequence is time-aligned with environmental parameter data to form multimodal data pairs.

[0018] Furthermore, the three-dimensional convolutional neural network includes:

[0019] The input layer is used to receive the standardized video sequence;

[0020] Multiple 3D convolutional blocks, each block containing two 3D convolutional layers and one 3D max pooling layer;

[0021] The temporal attention module is used to focus on keyframes of behavioral changes;

[0022] Global average pooling layers and fully connected layers are used to generate behavioral feature vectors.

[0023] Furthermore, the establishment state representation includes:

[0024] Concatenate the behavioral feature vector and the environmental feature vector;

[0025] The interaction between behavioral features and environmental features is learned through feature fusion networks.

[0026] Generate a comprehensive feature vector representing the current state of the laying hen.

[0027] Furthermore, the reinforcement learning algorithm models the evaluation timing selection problem as a Markov decision process, wherein:

[0028] The state space is the set of all state representations;

[0029] The action space is a binary action space;

[0030] The reward function is defined based on the consistency and stability of the evaluation results.

[0031] Furthermore, the weight assessment model employs a multi-branch fusion network structure, including:

[0032] The body shape feature branch is used to process body shape features;

[0033] The behavioral feature branch is used to process behavioral features.

[0034] The environmental features branch is used to process environmental features.

[0035] The fusion layer is used to weightedly fuse the output features of the three branches through an attention mechanism;

[0036] The regression layer is used to map the fused features to weight values.

[0037] Furthermore, the variational autoencoder includes:

[0038] An encoder maps input features to distribution parameters in the latent space;

[0039] The decoder reconstructs vectors from the latent distribution into vectors in the original feature space;

[0040] The sampling process uses reparameterization techniques to sample latent variables from a normal distribution.

[0041] Furthermore, the anomaly assessment results include:

[0042] Calculate the reconstruction error, which represents the difference between the reconstructed data and the original data.

[0043] Calculate the probability bias, which represents the probability density of the data in the learned distribution;

[0044] The anomaly score is defined as a weighted sum of reconstruction error and probability bias;

[0045] The validity of the evaluation results is determined by comparing the abnormal scores with preset thresholds.

[0046] Furthermore, during the process of detecting abnormal evaluation results, when an abnormal evaluation result is detected, the following operations are performed:

[0047] The detected abnormal evaluation results are marked as suspicious results;

[0048] Trigger the reassessment process and select a new assessment time;

[0049] Record the characteristics of abnormal patterns for subsequent model updates;

[0050] An alert is generated when multiple consecutive assessments show abnormalities.

[0051] This invention provides an image recognition-based system for assessing the body size and weight of laying hens, used to execute the aforementioned image recognition-based method for assessing the body size and weight of laying hens, comprising:

[0052] The multimodal data acquisition module is used to acquire video data of laying hen behavior and environmental parameter data;

[0053] The behavioral-environment co-analysis module is used to identify the behavioral patterns of laying hens and analyze the correlation between environmental parameters and behavior;

[0054] The optimal assessment timing selection module is used to realize stress state detection and automatic selection of the optimal assessment timing;

[0055] The weight assessment module is used to map image features to weight.

[0056] The quality control and anomaly detection module is used to enable self-verification of the evaluation process.

[0057] The beneficial effects of this invention are: by selecting the best evaluation time through behavioral analysis, the evaluation error caused by posture changes is reduced, and the consistency of multiple evaluations of the same laying hen is improved;

[0058] The behavioral state correction mechanism further improves the assessment accuracy, reduces the bias of assessment results under different behavioral states, and makes the assessment results more stable and reliable. Multimodal feature fusion provides more comprehensive information input, improves the average accuracy of weight assessment, and is an improvement compared to methods that only use image features.

[0059] Based on the analysis of environmental factors, the accuracy of the evaluation under different light conditions is reduced. Traditional methods have large fluctuations in accuracy when the light changes significantly, while the present invention reduces the fluctuations and improves the robustness to adapt to different light conditions, enabling the present invention to operate stably in various aquaculture environments.

[0060] Environmental parameter sensing and stress state detection enable this invention to automatically avoid unsuitable assessment times, improve the reliability of assessment under extreme environmental conditions, extend the stable operation time compared to traditional methods, and reduce maintenance requirements.

[0061] By using generative models for anomaly detection, this invention can automatically identify and filter low-quality evaluation results, thereby improving the reliability and credibility of the evaluation results and ensuring their high quality.

[0062] The quality control mechanism enables the present invention to automatically adjust the evaluation strategy and trigger a re-evaluation process when an anomaly is detected. The evaluation success rate is higher than that of traditional methods. It can also automatically update the model according to the characteristics of the anomaly pattern, so that the performance can be continuously optimized over time.

[0063] This invention eliminates the need for manual weighing, improving assessment efficiency, reducing labor intensity and costs. The non-contact assessment method reduces stress in laying hens, improves egg production performance, and provides early warning of abnormal weight changes, offering timely intervention opportunities for breeding management. It also reduces mortality during the breeding cycle and improves feed conversion rate. Attached Figure Description

[0064] Figure 1 This is a flowchart of an image recognition-based method for assessing the body size and weight of laying hens in this invention;

[0065] Figure 2 This is a bar chart comparing the accuracy of the method of this invention with that of traditional evaluation methods under different conditions;

[0066] Figure 3 It is a line graph comparing the evaluation errors of the traditional method and the method of the present invention under different posture conditions;

[0067] Figure 4 This is a scatter plot showing the relationship between the confidence index and the actual accuracy in the method of this invention.

[0068] Figure 5This is a radar chart comparing the performance of the system of this invention and a conventional system under different lighting conditions;

[0069] Figure 6 This is a bar chart showing the detection rate and false alarm rate of variational autoencoder anomaly detection under different types of anomalies.

[0070] Figure 7 This is a pie chart showing the proportion of benefits generated after the application of the technology of this invention in a breeding farm. Detailed Implementation

[0071] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0072] At least one embodiment of the present invention discloses a method for assessing the body size and weight of laying hens based on image recognition, such as... Figure 1 As shown, it includes:

[0073] Step 1: Acquire video data of laying hen behavior and environmental parameter data through a multimodal data acquisition system;

[0074] This step utilizes a multimodal data acquisition system to obtain video data of laying hen behavior and environmental parameter data, forming a complete information input. Specifically, it includes the following sub-steps:

[0075] Step 1.1, short-time video sequence acquisition;

[0076] A short-time video sequence acquisition system is deployed, comprising multiple high-definition cameras that monitor the activity area of ​​the laying hens from different angles. Each camera acquires video at a rate of 25 frames per second, with each acquisition lasting 5 to 10 seconds, forming a short-time video sequence containing 125-250 frames. The acquisition system automatically initiates the acquisition process based on preset time intervals (e.g., once per hour) or specific triggering conditions (e.g., sudden changes in environmental parameters).

[0077] The acquired video sequences underwent preprocessing, including:

[0078] Spatial resolution standardization: Adjust the video captured by different cameras to a uniform resolution of 1280×720;

[0079] Time length standardization: Video sequences of different lengths are uniformly trimmed or padded to 8 seconds (200 frames).

[0080] Noise filtering: Apply a Gaussian filter to remove random noise from the video;

[0081] Lighting compensation: Adaptive lighting compensation is performed based on the overall brightness of the scene to reduce the impact of lighting changes.

[0082] The preprocessed video sequence is stored in a standard format (such as MP4) and additional metadata such as timestamps and camera IDs is added for easy subsequent processing.

[0083] Step 1.2, Environmental parameter collection;

[0084] A network of environmental parameter sensors, including temperature sensors, humidity sensors, light intensity sensors, and air quality sensors, is deployed synchronously to collect various parameters of the aquaculture environment in real time. The sensor acquisition frequency is set to once per minute to capture subtle changes in the environment.

[0085] The environmental parameter data has undergone preliminary processing, including:

[0086] Outlier detection and correction: Identify and handle abnormal readings caused by sensor malfunctions or interference;

[0087] Data smoothing: Applying methods such as moving averages to smooth parameter values ​​that fluctuate significantly;

[0088] Data standardization: Standardizing parameter values ​​of different dimensions to the same range (e.g., 0-1).

[0089] Feature extraction: Calculate derived features such as the rate of change and fluctuation range of environmental parameters.

[0090] The processed environmental parameter data is stored in a structured format (such as CSV) and associated with the video sequence of the corresponding time period to form multimodal data pairs.

[0091] Step 2: Use a 3D convolutional neural network to process the video data, identify the behavior patterns of laying hens, and analyze the correlation between environmental parameter data and behavior to establish a state representation.

[0092] This step utilizes a 3D convolutional neural network to process the video sequence, identify typical behavioral patterns of laying hens, and analyze the correlation between environmental parameters and behavior to establish a comprehensive state representation. Specifically, it includes the following sub-steps:

[0093] Step 2.1, recognizing the behavior of laying hens;

[0094] A three-dimensional convolutional neural network was applied to analyze preprocessed short-time video sequences to identify typical behavioral patterns of laying hens. The three-dimensional convolutional neural network used consisted of the following components:

[0095] Input layer: Receives the normalized video sequence, with a shape of [200, 720, 1280, 3] (number of frames × height × width × number of channels).

[0096] The 3D convolutional layer group contains 5 3D convolutional blocks, each block containing two 3D convolutional layers and one 3D max pooling layer, progressively extracting spatiotemporal features.

[0097] First convolutional block: 32 3×3×3 convolutional kernels, stride (1, 1, 1), number of output channels 32;

[0098] Second convolutional block: 64 3×3×3 convolutional kernels, stride (1, 1, 1), number of output channels 64;

[0099] Third convolutional block: 128 3×3×3 convolutional kernels, stride (1, 1, 1), number of output channels 128;

[0100] Fourth convolutional block: 256 3×3×3 convolutional kernels, stride (1, 1, 1), number of output channels 256;

[0101] Fifth convolutional block: 512 3×3×3 convolutional kernels, stride (1, 1, 1), number of output channels 512;

[0102] Each convolutional layer is followed by a batch normalization layer and a ReLU activation function;

[0103] The pooling layer kernel size after each convolutional block is 2×2×2, with a stride of (2, 2, 2).

[0104] Temporal Attention Module: Introduces a temporal-dimensional attention mechanism to focus on keyframes of behavioral changes.

[0105] Self-attention computation is applied to the time dimension to generate time weights;

[0106] Weights are multiplied by feature maps to enhance the feature representation of keyframes;

[0107] Global average pooling layer: performs average pooling on the feature map in both spatial and temporal dimensions to generate a fixed-length feature vector;

[0108] Fully connected layer: Contains two fully connected layers. The first layer has an output dimension of 256, and the second layer has an output dimension of 128.

[0109] Output layer: Outputs the probability distribution of multiple behavior categories through the softmax function.

[0110] The application of 3D convolutional neural networks in the task of recognizing the behavior of laying hens has the following advantages:

[0111] It can capture features in both spatial and temporal dimensions simultaneously, making it suitable for handling dynamic behavior patterns;

[0112] Three-dimensional convolutional kernels can effectively extract local spatiotemporal features and identify subtle behavioral differences;

[0113] The time-attention mechanism can automatically focus on key moments of behavioral change, improving recognition accuracy;

[0114] Deep network structures can learn complex hierarchical feature representations and adapt to diverse behavioral patterns.

[0115] In its implementation, this 3D convolutional neural network was optimized for specific behaviors of laying hens:

[0116] By using a smaller 3×3×3 convolution kernel, the number of parameters is reduced while maintaining good feature extraction capabilities;

[0117] By increasing the number of channels layer by layer, the network's ability to express complex patterns is improved;

[0118] Incorporate a time-attention mechanism to pay special attention to critical moments in the behavioral transitions of laying hens;

[0119] Batch normalization and residual connections are used to improve training stability and convergence speed.

[0120] The trained network can identify the current behavioral state of laying hens from input videos and output behavioral feature vectors. This vector encodes the spatiotemporal pattern information of the behavior of laying hens.

[0121] Step 2.2, Environment-Behavior Correlation Analysis;

[0122] The environmental parameter data and behavior recognition results are time-aligned to analyze the correlation between environmental factors and laying hen behavior patterns. Specific steps include:

[0123] Time alignment: interpolating environmental parameter data to the same time point as the video frame;

[0124] Correlation analysis: Calculate the statistical correlation between each environmental parameter and the behavioral category.

[0125] Point-to-binary correlation coefficient: Calculates the correlation between continuous environmental parameters (such as temperature and humidity) and binary behavioral states (such as whether or not one is in a state of stress);

[0126] Cramer's V coefficient: Calculates the strength of the association between discrete environmental parameters (such as light level) and multi-class behavioral states;

[0127] Mutual information analysis: Calculates the nonlinear correlation between environmental parameters and behavioral categories;

[0128] Threshold determination: Based on the results of correlation analysis, determine the normal range and abnormal threshold of each environmental parameter;

[0129] Environmental feature extraction: Extracting environmental feature vectors from raw environmental parameters and their derived features. :

[0130] Original environmental parameters:

[0131] Temperature (°C);

[0132] Relative humidity (%)

[0133] Illumination intensity (lux);

[0134] Ammonia concentration (ppm);

[0135] Carbon dioxide concentration (ppm);

[0136] Air velocity (m / s);

[0137] Derived features:

[0138] Temperature and Humidity Index (THI) = 0.8 × Temperature + Relative Humidity × (Temperature - 14.4) / 100 + 46.4;

[0139] Short-term rate of change of environmental parameters (the magnitude of change within 10 minutes).

[0140] The volatility (standard deviation) of environmental parameters;

[0141] Periodic characteristics of environmental parameters (extracted via Fourier transform);

[0142] Multi-parameter combined index (such as the combined influence index of temperature and humidity);

[0143] Principal component analysis (PCA) was used to reduce the dimensionality of the above features, retaining the principal components necessary to explain 95% of the variance, thus forming the final environmental feature vector. .

[0144] Step 2.3, State representation construction;

[0145] behavioral feature vector and environmental feature vectors By splicing together the data, a comprehensive state representation can be constructed:

[0146] ;

[0147] in Indicates time The state representation vector; Represents a behavioral feature vector; Represents the environmental feature vector; This represents a vector concatenation operation;

[0148] This status indicates that the information on laying hen behavior and environmental conditions has been integrated, providing a complete decision-making basis for selecting the optimal timing for subsequent assessment.

[0149] State representation Further processing is performed using a feature fusion network, which contains multiple fully connected layers and normalization layers to learn the interaction between behavioral and environmental features, generating a more expressive state representation. The final state representation has a dimension of 128 and contains rich joint behavioral-environment information.

[0150] Step 3: Based on the state representation output, a policy network is constructed using a reinforcement learning algorithm to realize stress state detection and automatic selection of the best evaluation time, and to extract body shape features under stable posture.

[0151] This step, based on reinforcement learning algorithms, constructs a policy network to achieve stress state detection and automatic selection of the optimal evaluation time, extracting body shape features under stable postures. Specifically, it includes the following sub-steps:

[0152] Step 3.1, evaluate the timing selection problem modeling;

[0153] The problem of timing the evaluation is modeled as a Markov Decision Process (MDP), where:

[0154] State space: All possible state representations constructed in step 2 A set;

[0155] Action space: binary action space , where 0 means "no evaluation" and 1 means "evaluation";

[0156] Reward function: Based on the definition of consistency and stability of the evaluation results, according to the embodiments of this application, when the "evaluation" action is selected and the state stability is higher than a threshold... Positive rewards are given when the "evaluate" action is selected but the state stability is below the threshold. When a negative reward is given, a neutral reward is given when the "do not evaluate" action is selected;

[0157] State transition: determined by environmental dynamics and changes in laying hen behavior.

[0158] This modeling approach transforms the timing of assessment into a reinforcement learning problem, with the goal of learning an optimal policy to assess weight at the appropriate time (such as when the laying hen is in a stable posture and environmental conditions are favorable).

[0159] Step 3.2, Policy Network Construction and Training;

[0160] Construct a policy network to map states to actions. This network consists of the following components:

[0161] Input layer: Receives 128-dimensional state representation ;

[0162] Feature extraction layer: consists of 3 fully connected layers, each followed by batch normalization and ReLU activation function;

[0163] Attention layer: applies a self-attention mechanism to focus on key elements in the state vector;

[0164] Output layer: Outputs the probability distribution of actions through the softmax function.

[0165] Alternatively, in the implementation of the self-attention mechanism, a multi-head attention structure can be adopted, using four attention heads to process the state vector in parallel, and then concatenating the results and fusing them through a linear projection layer. This multi-head structure can focus on different aspects of the state features from different angles, improving the richness of the feature representation.

[0166] The policy network is trained using a policy gradient algorithm, specifically Proximal Policy Optimization (PPO). This algorithm improves training stability by pruning the objective function to limit the policy update step size. It's important to note that the core idea of ​​PPO is to limit the gap between the new and old policies, preventing excessively large policy updates from causing training instability. In its implementation, this is achieved by calculating the probability ratio between the new and old policies and limiting it within a certain range.

[0167] In some implementations, in addition to the PPO algorithm, TrustRegionPolicyOptimization (TRPO) or AdvantageActor-Critic (A2C) algorithms can also be used for training. TRPO ensures training stability by constraining the KL divergence of each policy update, while A2C improves data efficiency through parallel environment interaction. The most suitable algorithm can be selected for implementation based on the specific application scenario.

[0168] The training data comes from interactive experiences between simulated environments and real-world aquaculture scenarios. The training process includes:

[0169] Pre-training phase: Supervised learning is performed using expert-annotated data at the "optimal evaluation time";

[0170] Reinforcement learning phase: Continuously optimize the strategy through interaction with the environment;

[0171] Fine-tuning phase: Adaptive adjustments are made using data from specific farms.

[0172] Step 3.3, Stress state detection;

[0173] Based on a pre-trained policy network, automatic detection of stress states is achieved. When laying hens are under stress, their behavioral patterns typically exhibit abnormal behaviors such as rapid movement and frequent wing flapping, which are reflected in the state representation. middle.

[0174] By learning pattern recognition capabilities, policy networks can assign lower probabilities of "evaluation" actions to these stress states, thereby automatically avoiding unsuitable evaluation opportunities.

[0175] ;

[0176] in Represents the policy function; Indicates the action of "evaluation"; Represents the conditional probability operator; This represents the state at time t;

[0177] The system also has explicit stress state detection rules, such as:

[0178] The probability of the "stress" category in the behavioral characteristics exceeds the threshold;

[0179] Environmental parameters changed drastically;

[0180] The behavior category changes frequently within a short period of time.

[0181] When a stress state is detected, the system will automatically delay the assessment operation until the state returns to normal.

[0182] Step 3.4, execute the action at the optimal time;

[0183] When the policy network outputs the "evaluate" action ( When this occurs, the system performs the following operations:

[0184] Extract the body shape features of laying hens from the current video frame;

[0185] Record environmental parameters and behavioral status information;

[0186] Mark this moment as the optimal time for evaluation;

[0187] This triggers the reasoning process of the weight assessment model.

[0188] In this way, the system can automatically select the moment when the laying hen is in a stable posture and the environmental conditions are suitable for weight assessment, thereby improving the accuracy and consistency of the assessment.

[0189] Step 4: Using the policy network to select body shape features at the appropriate time, and combining them with the identified behavioral state information, a weight assessment model is constructed to realize the mapping from image features to weight.

[0190] This step utilizes body shape characteristics at a selected time point, combined with behavioral state information, to construct a weight assessment model that considers behavioral state, achieving an accurate mapping from image features to weight. Specifically, it includes the following sub-steps:

[0191] Step 4.1, Body shape feature extraction;

[0192] Extracting the body shape characteristics of laying hens from video frames at the optimal assessment time. The body shape feature extraction process includes:

[0193] Object detection: Use the YOLOv5 model to detect the location of laying hens in video frames and generate bounding boxes;

[0194] Instance segmentation: MaskR-CNN is used to accurately segment the detected laying hens and obtain their outlines;

[0195] Morphological feature calculation: A series of morphological features are calculated based on the segmentation results. The specific calculation method is as follows:

[0196] Body length: The maximum length of the hen's outline is obtained by using the minimum bounding rectangle method. The cv2.minAreaRect() function is used to calculate the minimum bounding rectangle, and its longer side is taken as the body length.

[0197] Body height: Based on the length of the short side of the minimum bounding rectangle, combined with the projection of the outline in the vertical direction, the maximum height perpendicular to the body length direction is calculated using the numpy.max() function;

[0198] Body width: Obtain the contour point set using cv2.findContours() at different positions of the contour (front, middle, and rear sections), and calculate the horizontal distance to obtain the width of each part;

[0199] Area: The pixel area enclosed by the outline can be directly calculated using the cv2.contourArea() function;

[0200] Perimeter: The pixel perimeter of the outline is calculated using the cv2.arcLength() function;

[0201] Shape factor: Calculates shape features such as roundness (4π × area / perimeter²) and rectangularity (area / area of ​​minimum circumscribed rectangle) based on area and perimeter;

[0202] Pose angle: Use the cv2.fitEllipse() function to fit the ellipse and obtain the angle between the ellipse's principal axis and the horizontal direction;

[0203] Deep feature extraction: A pre-trained ResNet-50 model is used to extract deep features from the laying hen images. The specific steps are as follows:

[0204] Input data preparation: The laying hen image regions obtained by MaskR-CNN segmentation are cropped and scaled to 224×224 pixels, and then standardized (mean subtracted and standard deviation divided).

[0205] Feature extraction: Input the preprocessed image into the pre-trained ResNet-50 model (pre-trained on the ImageNet dataset) and extract the 2048-dimensional feature vector output by the last convolutional layer;

[0206] Feature dimensionality reduction: Principal component analysis (PCA) was used to reduce the 2048-dimensional eigenvectors to 256 dimensions, retaining approximately 95% of the variance information.

[0207] Body shape feature vector The generation process is as follows:

[0208] Morphological feature vector construction: The 12 morphological features calculated above, including body length, body height, body width (front, middle, and rear sections), area, perimeter, roundness, rectangularity, and posture angle, are used to form a basic feature vector.

[0209] Feature standardization: Min-Max standardization is performed on morphological features to unify the numerical range to the [0,1] interval;

[0210] Feature fusion: The standardized 12-dimensional morphological feature vector is concatenated with the 256-dimensional depth feature vector to form a 268-dimensional body shape feature vector. ;

[0211] Feature weighting: Based on the experimental results on the validation set, morphological features and deep features are assigned weights of 0.4 and 0.6 respectively to balance the contributions of the two types of features.

[0212] The final generated body shape feature vector It contains a wealth of information characterizing the body size of laying hens, providing a basis for weight assessment.

[0213] Step 4.2, Feature fusion and weight assessment model construction;

[0214] Body shape characteristics Behavioral characteristics and environmental characteristics A weight assessment model is constructed by fusion. This model employs a multi-branch fusion network structure:

[0215] Body shape characteristic branch: processing It contains three fully connected layers to generate intermediate representations. ;

[0216] Behavioral feature branch: processing It contains two fully connected layers to generate intermediate representations. ;

[0217] Environmental characteristics branch: processing It contains two fully connected layers to generate intermediate representations. ;

[0218] Fusion layer: Combines the output features of the three branches By splicing and weighting the data through an attention mechanism, a generation is achieved. ;

[0219] Regression layer: Features are fused through a fully connected layer. Mapped to weight value .

[0220] In some implementations, the fusion layer can employ a cross-attention mechanism, enabling mutual attention and influence between feature branches. For example, body shape features can be used as query vectors, while behavioral and environmental features serve as key-value pairs, achieving interactive fusion between features through attention computation. This approach can better capture the correlations between features of different modalities.

[0221] According to embodiments of this application, the entire model can be viewed as a multi-input single-output function, receiving three different features as inputs and outputting the predicted weight value of laying hens. Furthermore, it should be understood that the three feature branches first process their respective features to generate intermediate representations, which are then fused through an attention mechanism and finally mapped to weight values ​​via a regression layer.

[0222] Alternatively, in the implementation of the regression layer, a quantitative regression approach can be adopted, which transforms the weight prediction problem into a classification problem and discretizes the weight values. For example, the weight range can be divided into multiple intervals, and a classification network can be used to predict the probability that the weight of the laying hen falls into each interval. Then, a weighted average is calculated as the final prediction result. This approach can better handle the multimodal characteristics of weight distribution and improve prediction accuracy.

[0223] Step 4.3, Behavioral State Correction;

[0224] To further improve the accuracy of the assessment, a behavioral state correction mechanism is introduced. Under different behavioral states, the relationship between the body shape characteristics and weight of laying hens may vary slightly. For example, the abdomen may be slightly swollen when foraging, while the body may be more curled up when resting.

[0225] Behavioral state modification is achieved in the following ways:

[0226] Dedicated evaluation parameters are trained for each typical behavioral state;

[0227] Select the corresponding parameter set based on the behavior recognition results;

[0228] Apply behavior-specific calibration coefficients to adjust the prediction results.

[0229] The corrected weight prediction can be expressed as:

[0230] ;

[0231] in This represents the corrected predicted weight value; This represents the original predicted weight value output by the model; This represents the calibration coefficient based on the behavioral state.

[0232] Step 4.4, Model Training and Optimization;

[0233] The weight assessment model was trained using real-world labeled data, including images of laying hens, behavioral videos, environmental parameters, and actual weight values. Mean Squared Error (MSE) was used as the loss function during training.

[0234] ;

[0235] in This represents the mean squared error loss function; The total number of samples; This represents the summation operator; Indicates the first The model predicts body weight for each sample; Indicates the first The true actual weight of each sample; This represents the squaring operation.

[0236] To improve the model's generalization ability, the following strategies are adopted:

[0237] Data augmentation: Applying techniques such as random cropping, rotation, and scaling to increase the diversity of training data;

[0238] Regularization: Use L2 regularization and Dropout to prevent overfitting;

[0239] Batch normalization: accelerates training and improves stability;

[0240] Learning rate scheduling: A cosine annealing learning rate strategy is adopted;

[0241] Cross-validation: Use K-fold cross-validation to evaluate model performance.

[0242] Step 5: Use variational autoencoders to learn the feature distribution of the normal assessment process of the weight assessment model, and detect abnormal assessment results by using reconstruction error and probability bias as dual indicators to achieve self-verification of the assessment process.

[0243] This step utilizes a variational autoencoder (VAE) to learn the feature distribution of a normal evaluation process. It then detects abnormal evaluation results using both reconstruction error and probability bias as indicators, thus achieving self-verification of the evaluation process. Specifically, it includes the following sub-steps:

[0244] Step 5.1, Construction of variational autoencoder;

[0245] A variational autoencoder model is constructed to learn the data distribution of the normal evaluation process. The variational autoencoder consists of two main parts: an encoder and a decoder.

[0246] Encoder: The distribution parameters that map input features to the latent space;

[0247] Input layer: Receives evaluation process features, including body shape features, environmental parameters, behavioral features, and predicted weight, with 256 input dimensions;

[0248] Hidden layers: Contain 3 fully connected layers;

[0249] First layer: 256 neurons, using the LeakyReLU activation function (slope of 0.2);

[0250] Second layer: 128 neurons, using the LeakyReLU activation function (slope of 0.2);

[0251] The third layer consists of 64 neurons, using the LeakyReLU activation function (slope of 0.2).

[0252] Output layer: Two parallel fully connected layers;

[0253] Mean layer: Outputs the mean of the latent variables, with a dimension of 32;

[0254] Variance layer: Outputs the log-variance of the latent variables, with a dimension of 32.

[0255] Decoder: Reconstructs vectors from the sampled latent distribution into vectors in the original feature space;

[0256] Input layer: Receives latent vectors sampled from a normal distribution, with a dimension of 32;

[0257] Hidden layers: Contain 3 fully connected layers;

[0258] First layer: 64 neurons, using the LeakyReLU activation function (slope of 0.2);

[0259] Second layer: 128 neurons, using the LeakyReLU activation function (slope of 0.2);

[0260] The third layer consists of 256 neurons, using the LeakyReLU activation function (slope of 0.2).

[0261] Output layer: Fully connected layer, outputs reconstructed features with the same dimensions as the input features.

[0262] The sampling process of variational autoencoders is achieved using reparameterization techniques, which involves sampling noise from a standard normal distribution and then transforming it using the mean and standard deviation to obtain latent variable samples.

[0263] The application of variational autoencoders in quality control for laying hen weight assessment has the following advantages:

[0264] It can learn to evaluate the distribution characteristics of data, rather than simply mapping them point by point;

[0265] Probabilistic modeling allows for the quantification of the uncertainty of assessment results;

[0266] Regularization of the latent space enables the model to have better generalization ability;

[0267] It can detect anomalies by using both reconstruction error and probability bias as indicators.

[0268] In specific application scenarios, the variational autoencoder has been specifically optimized for evaluating quality control:

[0269] The input features contain multimodal information, improving the comprehensiveness of anomaly detection;

[0270] By employing a smaller latent space dimension (32-dimensional), the model is forced to learn a compact feature representation;

[0271] Use LeakyReLU instead of traditional ReLU to improve the gradient vanishing problem;

[0272] The reconstruction term and KL divergence term in the loss function are balanced by weighting coefficients to adapt to specific task requirements.

[0273] Step 5.2, Normal Mode Learning;

[0274] The variational autoencoder is trained using a large amount of data from the normal evaluation process, allowing it to learn the data distribution characteristics of normal evaluation. The training data includes:

[0275] Body type ;

[0276] Environmental parameters ;

[0277] Behavioral characteristics ;

[0278] Predicted weight ;

[0279] Evaluate consistency metrics (such as the standard deviation of multiple evaluations).

[0280] The following strategies were used during the training process:

[0281] Training is performed using only the evaluation results that have been verified as correct.

[0282] Apply data augmentation techniques to expand the training samples;

[0283] An early stopping strategy is employed to prevent overfitting.

[0284] Use batch normalization to improve training stability.

[0285] Once trained, the model is able to capture the data distribution characteristics of a normal evaluation process, providing a foundation for anomaly detection.

[0286] Step 5.3, Anomaly Assessment and Detection;

[0287] An anomaly detector is constructed based on a trained variational autoencoder. Anomaly assessment and detection are achieved by calculating the following two metrics:

[0288] Reconstruction error The difference between the reconstructed data and the original data is measured by the squared Euclidean distance between the original features and the reconstructed features:

[0289] ;

[0290] in This represents the reconstruction error of input x; Represents the square of the Euclidean distance; Represents the original input features; Indicates reconstruction features;

[0291] probability bias : Measures the probability density of data in the learned distribution, i.e., evaluates the negative log-likelihood of the sample.

[0292] According to embodiments of this application, anomaly scoring The result is calculated by weighting these two indicators:

[0293] ;

[0294] in Indicates input Abnormal scores; Indicates input Reconstruction error; Weighting coefficients representing probability bias; Indicates input The probability bias.

[0295] The anomaly detection process is as follows: extract the feature vector of the current evaluation process; input the variational autoencoder to obtain the reconstruction results and latent variable distribution parameters; calculate the anomaly score; compare the anomaly score with a preset threshold, and if the anomaly score is higher than the threshold, it is determined to be an anomaly evaluation.

[0296] Optionally, the anomaly detection threshold can be determined adaptively rather than as a fixed value. For example, it can be set to the mean plus a certain number of standard deviations based on the statistical distribution of historical evaluation data, or the threshold can be determined using quantile methods. This adaptive threshold can better adapt to the characteristics of different breeding environments and laying hen breeds.

[0297] Step 5.4, Evaluation result correction and feedback;

[0298] When an abnormal evaluation result is detected, the system performs the following operations:

[0299] Mark the assessment result as suspicious;

[0300] Trigger the reassessment process and select a new assessment time;

[0301] Record the characteristics of abnormal patterns for subsequent model updates;

[0302] If multiple consecutive assessments show abnormalities, an alarm will be generated to indicate a possible system malfunction or abnormality in the laying hens.

[0303] For evaluation results that pass quality control, the system calculates a confidence index:

[0304] ;

[0305] in This represents the confidence level of input x; This represents the ratio of the anomaly score to the threshold. This represents the abnormal score for input x; Indicates the threshold for abnormal scoring;

[0306] The final output format of the evaluation results is as follows:

[0307] ;

[0308] in This is the corrected predicted weight value. It's the confidence level. It is the evaluation timestamp.

[0309] A system for assessing the body size and weight of laying hens based on image recognition, used to execute the aforementioned method for assessing the body size and weight of laying hens based on image recognition, includes:

[0310] The multimodal data acquisition module is used to acquire video data of laying hen behavior and environmental parameter data;

[0311] The behavioral-environment co-analysis module is used to identify the behavioral patterns of laying hens and analyze the correlation between environmental parameters and behavior;

[0312] The optimal assessment timing selection module is used to realize stress state detection and automatic selection of the optimal assessment timing;

[0313] The weight assessment module is used to map image features to weight.

[0314] The quality control and anomaly detection module is used to enable self-verification of the evaluation process.

[0315] Here, the present invention provides an implementation example:

[0316] The test scenario was a commercial chicken farm with 10,000 laying hens, using a cage-raising model, divided into 5 chicken houses. Each chicken house was equipped with 8 high-definition cameras and a complete network of environmental parameter sensors for real-time acquisition of video data and environmental parameters. The system was deployed on the farm's edge server and connected to each sensor device via a local area network.

[0317] The test subjects were white Leghorn laying hens from the same batch, aged between 180 and 270 days, at their peak egg production. Before the test, 200 laying hens were randomly selected and manually weighed to obtain their actual weight data as the evaluation benchmark. The average weight of these laying hens was 1.83 kg, with a standard deviation of 0.21 kg.

[0318] The following describes the specific implementation process of the system in actual operation:

[0319] Multimodal data acquisition: The system automatically triggers the data acquisition process once per hour. In a typical acquisition process, eight cameras work simultaneously, each recording 8 seconds of video at a rate of 25 frames per second, generating approximately 1.6GB of raw video data. Simultaneously, environmental sensors collect parameters such as current temperature (26.5℃), humidity (62%), light intensity (320 lux), and air quality index. All data is preprocessed and then transmitted to an edge server for further processing.

[0320] Behavior-Environment Joint Analysis: Taking a video clip captured by camera No. 3 in chicken house No. 1 as an example, the system uses a 3D convolutional neural network to identify 12 laying hens in the video and analyze their behavioral states. Eight hens were in a "foraging" state, three were in a "resting" state, and one was in a "walking" state. Environmental parameter analysis showed that the current conditions were within the normal range with no abnormal fluctuations. The system generates a state representation for each laying hen, consisting of a concatenation of a behavioral feature vector (128 dimensions) and an environmental feature vector (32 dimensions), used for subsequent evaluation timing selection.

[0321] Optimal evaluation timing: After receiving the state representation, the policy network calculates the evaluation action probability for each of the 12 laying hens. The results show that the evaluation action probabilities for 8 hens are below the threshold of 0.35 (indicating they are not suitable for evaluation), mainly because they are in motion or have unstable postures. The evaluation action probabilities for the remaining 4 hens are 0.82, 0.79, 0.76, and 0.58, respectively, all exceeding the system-set threshold of 0.5, and therefore are marked as suitable targets for evaluation. For these targets, the system extracts body shape features from the current video frame for weight assessment.

[0322] Weight Assessment Model: Taking a laying hen with an assessment action probability of 0.82 as an example, the system first obtains its precise contour through object detection and instance segmentation, calculating a body length of 32.7 cm, a body height of 18.5 cm, a body width (chest) of 12.3 cm, and a contour area of ​​495.8 square centimeters. These morphological features are combined with depth features to form a body shape feature vector. Then, the system inputs body shape features, behavioral features, and environmental features into a multi-branch fusion network. After feature fusion and regression calculation, a preliminary weight prediction of 1.78 kg is obtained. Considering that the laying hen is in a "resting" state, the system applies a behavioral correction coefficient of 1.02, and the final predicted weight is 1.82 kg.

[0323] Quality control and anomaly detection: The variational autoencoder receives the feature vector from this evaluation process, calculates the reconstruction error as 0.035, the probability bias as 1.28, and the comprehensive anomaly score as 0.31, which is lower than the system's set threshold of 0.5. Therefore, it is judged as a normal evaluation result. The system calculates the confidence level of this evaluation as 0.38 and records the final result as {1.82kg, 0.38, 2023-04-15 10:23:45}.

[0324] During a complete evaluation cycle across the entire farm, the system evaluated a total of 3,245 laying hens, averaging approximately 200 hens per hour, far exceeding the efficiency of manual weighing. The evaluation results showed that the average weight of the laying hens was 1.81 kg, with the weight distribution conforming to a normal distribution, and the deviation from manual sampling results was less than 3%.

[0325] Data collection and analysis over one month of system operation verified the effectiveness of the two core technologies presented in this solution:

[0326] Accuracy verification of the assessment: 100 laying hens were selected for a 7-day follow-up assessment, with manual weighing used as a control. The results showed that:

[0327] The assessment error of the same laying hen under different postures was reduced from ±150g in the traditional method to ±47g in this method;

[0328] The accuracy rate of the assessment (the proportion of predicted weight that differs from actual weight within ±50g) has increased from 78% using traditional methods to 96%.

[0329] The consistency of multiple assessments (the coefficient of variation of the results of multiple assessments of the same laying hen) decreased from 0.082 to 0.021;

[0330] The difference in assessment accuracy under different lighting conditions (morning, noon, and evening) decreased from 28% to 6.5%.

[0331] Application Benefit Verification: In aquaculture management practice, the benefits brought by this system include:

[0332] The assessment efficiency is improved by about 9.7 times, and 200 laying hens can be assessed per hour, while manual methods can only process about 20 hens per hour;

[0333] The detection rate of abnormal weight reached 91.5%, and potential health problems were detected on average 45 hours in advance;

[0334] The mortality rate during the breeding cycle decreased from 3.8% to 3.2%, a reduction of 15.8%.

[0335] Stress response decreased, and egg production increased by approximately 6.3%.

[0336] Feed formulation adjustments based on body weight distribution improved feed conversion rate by 6.8%.

[0337] These verification results fully demonstrate the advantages of this technical solution in improving assessment accuracy, enhancing environmental adaptability, achieving self-quality control, and bringing application benefits, providing strong technical support for modern egg-laying hen farming.

[0338] like Figures 2 to 7The figures shown are: a bar chart comparing the accuracy of the method of the present invention and the traditional evaluation method under different conditions; a line graph comparing the evaluation error of the traditional method and the method of the present invention under different posture conditions (unit: g); a scatter plot showing the relationship between the evaluation confidence index and the actual accuracy in the method of the present invention; a radar chart showing the performance evaluation of the system of the present invention and the traditional system under different lighting conditions; a bar chart showing the detection rate and false alarm rate of the variational autoencoder anomaly detection under different types of anomalies; and a pie chart showing the proportion of benefits after the application of the technology of the present invention in a breeding farm.

[0339] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for assessing the body size and weight of laying hens based on image recognition, characterized in that, include: Video data of laying hen behavior and environmental parameter data were acquired using a multimodal data acquisition system; A three-dimensional convolutional neural network is used to process video data to identify the behavior patterns of laying hens. At the same time, the correlation between environmental parameter data and behavior is analyzed to establish a state representation. The three-dimensional convolutional neural network includes: The input layer is used to receive the standardized video sequence; Multiple 3D convolutional blocks, each block containing two 3D convolutional layers and one 3D max pooling layer; The temporal attention module is used to focus on keyframes of behavioral changes; Global average pooling layers and fully connected layers are used to generate behavioral feature vectors; The establishment status representation includes: Concatenate the behavioral feature vector and the environmental feature vector; The interaction between behavioral features and environmental features is learned through feature fusion networks. Generate a comprehensive feature vector representing the current state of the laying hen; Based on the state representation output, a policy network is constructed using a reinforcement learning algorithm to realize stress state detection and automatic selection of the best evaluation time, and to extract body shape features under stable posture. By utilizing the body shape features selected at the time of the policy network and combining them with the identified behavioral state information, a weight assessment model is constructed to achieve the mapping from image features to weight. By using a variational autoencoder to learn the feature distribution of the normal assessment process of a weight assessment model, abnormal assessment results are detected by reconstructing error and probability bias, thus achieving self-verification of the assessment process.

2. The method for assessing the body size and weight of laying hens based on image recognition according to claim 1, characterized in that, The collection of video data on laying hen behavior includes: Multiple high-definition cameras were deployed to monitor the activity area of ​​the laying hens from different angles; The acquired video sequences are preprocessed, including spatial resolution normalization, temporal length normalization, noise filtering, and illumination compensation. The preprocessed video sequence is time-aligned with environmental parameter data to form multimodal data pairs.

3. The method for assessing the body size and weight of laying hens based on image recognition according to claim 1, characterized in that, The reinforcement learning algorithm models the evaluation timing selection problem as a Markov decision process, where: The state space is the set of all state representations; The action space is a binary action space; The reward function is defined based on the consistency and stability of the evaluation results.

4. The method for assessing the body size and weight of laying hens based on image recognition according to claim 1, characterized in that, The weight assessment model employs a multi-branch fusion network structure, including: The body shape feature branch is used to process body shape features; The behavioral feature branch is used to process behavioral features. The environmental features branch is used to process environmental features. The fusion layer is used to weightedly fuse the output features of the three branches through an attention mechanism; The regression layer is used to map the fused features to weight values.

5. The method for assessing the body size and weight of laying hens based on image recognition according to claim 1, characterized in that, The variational autoencoder includes: An encoder maps input features to distribution parameters in the latent space; The decoder reconstructs vectors from the latent distribution into vectors in the original feature space; The sampling process uses reparameterization techniques to sample latent variables from a normal distribution.

6. The method for assessing the body size and weight of laying hens based on image recognition according to claim 1, characterized in that, The results of the anomaly assessment include: Calculate the reconstruction error, which represents the difference between the reconstructed data and the original data. Calculate the probability bias, which represents the probability density of the data in the learned distribution; The anomaly score is defined as a weighted sum of reconstruction error and probability bias; The validity of the evaluation results is determined by comparing the abnormal scores with preset thresholds.

7. The method for assessing the body size and weight of laying hens based on image recognition according to claim 1, characterized in that, During the process of detecting abnormal evaluation results, when an abnormal evaluation result is detected, the following operations are performed: The detected abnormal evaluation results are marked as suspicious results; Trigger the reassessment process and select a new assessment time; Record the characteristics of abnormal patterns for subsequent model updates; An alert is generated when multiple consecutive assessments show abnormalities.

8. A system for assessing the body size and weight of laying hens based on image recognition, characterized in that, A method for assessing the body size and weight of laying hens based on image recognition as described in any one of claims 1-7 includes: The multimodal data acquisition module is used to acquire video data of laying hen behavior and environmental parameter data; The behavioral-environment co-analysis module is used to identify the behavioral patterns of laying hens and analyze the correlation between environmental parameters and behavior; The optimal assessment timing selection module is used to realize stress state detection and automatic selection of the optimal assessment timing; The weight assessment module is used to map image features to weight. The quality control and anomaly detection module is used to enable self-verification of the evaluation process.

Images