Chicken health state on-line monitoring method integrating lightweight network and feature repair

Abstract

The invention discloses a chicken health state on-line monitoring method fusing a lightweight network and feature restoration, and the method comprises the following steps: S1, carrying out the collection and preprocessing of the image data of a chicken, obtaining a data set, and dividing the data set into a training set and a verification set; s2, constructing a chicken head orientation recognition model of the improved YOLOv8n; s3, image data of a to-be-monitored chicken is collected and preprocessed, and the chicken head orientation is determined based on the optimal recognition model; s4, carrying out the restoration and feature extraction of the cockscomb in the image, and obtaining the features: the color of the cockscomb, the area of the lateral cockscomb and the width of the normal cockscomb; s5, performing multi-feature fusion decision making, and defining a scoring function S; and S6, based on comparison between the scoring function S and a threshold value, health grade division is realized.

Description

Technical Field

[0001] This invention relates to the fields of intelligent livestock and poultry farming, computer vision, deep learning and edge computing technologies, specifically a method for online monitoring of the health status of caged chickens that integrates lightweight networks and feature repair. Background Technology

[0002] Automated disease monitoring technology in large-scale layer hen farming is a key link in achieving high-quality development of modern poultry industry. Currently, farms generally rely on manual experience to identify sick chickens, which suffers from prominent problems such as high labor intensity, low detection efficiency, and high false negative rates. Especially in large-scale farming environments, failure to detect sick chickens in a timely and accurate manner may lead to the spread of diseases and a decline in production performance. Therefore, developing a real-time, non-contact, and automated online monitoring method for chicken health status has significant practical implications and broad application prospects. Summary of the Invention

[0003] To address the problems existing in the background art, this invention proposes a...

[0004] Technical solution:

[0005] An online monitoring method for chicken health status integrating lightweight networks and feature repair, comprising the following steps:

[0006] S1. Image data acquisition and preprocessing of chickens to obtain a dataset, and to divide it into training and validation sets;

[0007] S2. Construct an improved YOLOv8n chicken head orientation recognition model. The YOLOv8n model uses the StarNet lightweight network as the backbone, embeds the SimSPPF module for multi-scale feature fusion, and integrates the GAM global attention mechanism. The model is trained using the dataset obtained in S1 to obtain the optimal recognition model.

[0008] S3. Acquire and preprocess image data of the chickens to be monitored, and determine the chicken's head orientation based on the optimal recognition model;

[0009] S4. Chicken comb restoration and feature extraction in the image, obtaining features: chicken comb color, lateral chicken comb area and frontal chicken comb width;

[0010] S5. Multi-feature fusion decision, defining the scoring function S:

[0011] S=αA'+βW'+γC′

[0012] In the formula, α, β, γ are weights, A' is the normalized lateral comb area, W' is the normalized frontal comb width, and C′ is the normalized comb color index.

[0013] S6. Based on the comparison between the scoring function S and the threshold, the health level is classified.

[0014] Specifically, in S1:

[0015] Data acquisition: Cameras are deployed using a suspended installation structure with a 20° downward angle and a horizontal distance of 1.4m;

[0016] Preprocessing: The original video frames are geometrically calibrated using a radial distortion correction model to eliminate the deformation error of the chicken coop crossbar; effective detection areas are retained by ROI cropping, and keyframes are extracted based on time interval sampling method. The data is cleaned to remove blurry and severely occluded data, thereby constructing the initial dataset.

[0017] Specifically, in S4:

[0018] S4-1, Lab Space Segmentation: Convert the image to Lab space, extract the a* channel and perform OTSU threshold segmentation;

[0019] S4-2, Occlusion Repair: Generate a horizontal mask for the fractured area and input it into the DeepFillv2 network for repair;

[0020] S4-3, Feature Quantization:

[0021] Comb color: Calculate the average value C (a) of the comb area;

[0022] Lateral view of rooster comb area: Calculate the number of connected pixels A in the binary image;

[0023] Width of the rooster comb in the correct position: Calculate the aspect ratio R of the minimum bounding rectangle.

[0024] Preferred:

[0025] Converting from RGB color space to Lab color space requires first converting to XYZ color space. The conversion formula from RGB color space to XYZ color space is as follows:

[0026]

[0027] The formula for converting from XYZ color space to Lab color space is:

[0028]

[0029] In the formula, X n Y n and Z n It is the standard value for reference white point, and D65 white point is commonly used.

[0030] Preferred, X n Y n and Z nTake 95.047, 100.0, and 108.883.

[0031] Preferably, α = 0.4, β = 0.4, and γ = 0.6.

[0032] Preferably, in step S6, three thresholds {T1, T2} are determined, and the health level is classified by comparing the scoring function S with the thresholds.

[0033]

[0034] Level 0 represents healthy, while levels 1-3 correspond to mild, moderate, and severe abnormalities, respectively.

[0035] Preferably, T1 = 0.55, T2 = 0.84.

[0036] Beneficial effects of the present invention

[0037] (1) The chicken head orientation recognition model innovatively adopts a coupled architecture of StarNet lightweight network and GAM attention mechanism, which reduces the number of model parameters and improves inference speed while maintaining recognition accuracy. The cross-scale feature pyramid constructed by the SimSPPF module effectively solves the problem of local feature loss caused by cage occlusion.

[0038] (2) The comb feature restoration algorithm is based on the a* component extraction algorithm established in the Lab color space, combined with the DeepFillv2 image restoration model, to achieve complete segmentation of the comb region. Verification shows that this algorithm significantly enhances robustness to interference factors such as insufficient lighting in the chicken coop, diverse chicken behaviors, and cage occlusion.

[0039] (3) A multi-feature fusion decision model for chicken health assessment was constructed. Based on three key characterization parameters—comb area A, width W, and color index C—normalization was performed to obtain unified scoring indicators A′, W′, and C′. A weighted linear scoring function "S=αA′+βW′+γ"C′ was introduced, where the weight combination was determined based on random forest feature importance analysis and expert scoring weight fusion, and its robustness was verified through ±10% perturbation experiments. From the perspective of the chicken's head, the model used different feature combinations for conditional discrimination and determined the grading threshold using expert consensus labels to achieve the classification of chicken health status. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the system of the present invention.

[0041] Figure 2 This is a technical roadmap for the present invention.

[0042] Figure 3 This is a flowchart of the rooster comb feature extraction algorithm of the present invention.

[0043] Figure 4 The figure shows the experimental results of the multi-feature fusion decision model recognition of the present invention. Detailed Implementation

[0044] The present invention will be further described below with reference to embodiments, but the scope of protection of the present invention is not limited thereto:

[0045] This invention proposes a method for estimating chicken weight based on audio technology, implemented using a chicken weight estimation system. Combined with... Figure 1 The system comprises a data acquisition layer, a data processing layer, and a data display layer. The data acquisition layer collects video data from a camera, preprocesses it, and then transmits it to an edge computing device. The data processing layer performs model detection based on an improved YOLOv8n chicken head orientation recognition model and a DeepFillv2 chicken comb repair model. The data display layer displays the health level of the chickens. Its overall technical approach is as follows: Figure 2 As shown, it includes:

[0046] (1) Data Acquisition and Preprocessing

[0047] The deployment scheme employs a suspended installation structure with a 20° downward angle and a horizontal distance of 1.4m, with the main camera focusing on the second layer of chicken coops. To address the barrel distortion issue caused by wide-angle lenses, a radial distortion correction model is used to geometrically calibrate the original video frames, eliminating deformation errors of the chicken coop crossbars. Effective detection areas are retained through ROI cropping, and keyframes are extracted using a time-interval sampling method (one frame is extracted every 5 seconds). The data is then cleaned to remove blurry and severely occluded data, thus constructing the initial dataset.

[0048] (2) Chicken head orientation recognition model based on improved YOLOv8n

[0049] To achieve accurate classification and recognition of the chicken's head orientation, the YOLOv8n model structure was optimized: a lightweight neural network, StarNet, was used to replace the original backbone network, significantly enhancing feature extraction capabilities; a simplified version of the spatial pyramid pooling module, SimSPPF, was introduced to improve the model's computational efficiency; and the GAM global attention mechanism was integrated into the C2f module to effectively capture key information in multiple dimensions such as channel, spatial width, and height, strengthening feature interactions between different dimensions. These improvements collectively enhanced the model's accuracy and robustness in recognizing the chicken's head orientation in complex scenes.

[0050] (3) Chicken comb repair model based on DeepFillv2

[0051] Based on the chicken head recognition model, a chicken comb repair model based on DeepFillv2 is proposed to remove the occlusion of the comb area by the cage bars and reduce the false negative rate. This algorithm demonstrates stable segmentation performance under multiple interferences, including complex low-light conditions, diverse chicken postures, and cage structure occlusion, providing reliable technical support for health monitoring of caged chickens in large-scale intensive farming scenarios.

[0052] (4) Online monitoring method for chicken health status

[0053] The model structure allows for parameter fine-tuning based on different chicken ages, breeds, or breeding environments, exhibiting strong engineering controllability and scenario adaptability. This method achieves good detection results throughout the entire life cycle of caged chickens.

[0054] The present invention will be further described below with reference to specific embodiments:

[0055] (1) Data Acquisition and Preprocessing: The experiment used Jingfen No. 6 laying hens from breeding chicken houses 1-5 at Shijiazhuang Yukou Poultry Farm as the research subjects, adopting a standardized three-layer stacked cage rearing model (cage size 40cm×40cm×50cm, density 4 chickens / cage). The data acquisition module included six components: camera, network hard disk video recorder, switch, computer, power system, and portable hard drive. Based on the actual production conditions of the farm, the video data acquisition covered the entire life cycle of the chickens, and the acquired data was stored on the hard disk video recorder. The audio files were in WAV format. Data acquisition lasted 24 hours a day, with the chicken house temperature around 28.5℃ and humidity around 45% during the acquisition process.

[0056] (2) Lightweight chicken head orientation recognition:

[0057] By improving the YOLOv8n model, using the StarNet lightweight network as the backbone, embedding the SimSPPF module for multi-scale feature fusion, and integrating the GAM global attention mechanism, the model outputs the orientation labels of the chicken head in the front or side position.

[0058] (3) Combination Figure 3 Chicken comb repair and feature extraction:

[0059] 1) Lab Color Space Segmentation: Convert the image to Lab color space, extract the a* channel, and perform OTSU thresholding segmentation. Converting RGB color space to Lab color space requires first converting to XYZ color space. The conversion formula from RGB to XYZ color space is as follows:

[0060]

[0061] The formula for converting from XYZ color space to Lab color space is:

[0062]

[0063] In the formula, X n Y n and Z n This is the standard value for the reference white point; the commonly used one is D65 white point. X n Y n and Z n The default values ​​are usually 95.047, 100.0, and 108.883.

[0064] 2) Occlusion Repair: Generate a horizontal mask for the fractured area and input it into the DeepFillv2 network for repair (Adam optimizer).

[0065] lr = 0.0001);

[0066] 3) Feature quantization:

[0067] Comb color: Calculate the average value C (a) of the comb area;

[0068] Chicken comb area (lateral view): Calculate the number of pixels A in the connected components of the binary image;

[0069] Chicken comb width (positive position): Calculate the aspect ratio R of the minimum bounding rectangle;

[0070] (4) Multi-feature fusion decision-making:

[0071] 1) Feature normalization:

[0072] The lateral comb area A, the central comb width W, and the color index C are normalized according to thresholds to obtain normalized features A′, W′, C′∈[0,1]. Here, W′ is inversely proportional to the comb width W, indicating that a greater degree of comb inversion indicates poorer health. The width feature W′ is reverse-normalized so that W′ decreases as the width W increases, reflecting the negative impact of comb width on health. The color index C is considered normal only if it falls within the normal range, i.e., C′ is set to 1; when it deviates from this range, C′ needs adjustment, with greater deviation indicating poorer health. The formulas for calculating A′, W′, and C′ are as follows:

[0073]

[0074] D max =max{141-C min C max -155}

[0075] 2) Weighted scoring model:

[0076] Depending on the orientation (frontal or lateral) of the acquired image, different feature combinations and scoring logics are used:

[0077] Posterior image: mainly consider the width and color features of the rooster's comb, and set the weight of the area feature to 0, i.e., S=βW'+γC'.

[0078] Lateral view: mainly consider the area and color features of the rooster comb, and set the weight of the width feature to 0, i.e. S=αA'+γC′.

[0079] The health scoring function S is defined as follows:

[0080] S=αA'+βW'+γC′ (5)

[0081] The weights α, β, and γ are set based on the importance of the features (α = 0.4, β = 0.4, γ = 0.6) to ensure that the contribution of each feature to the final decision is proportional to its discriminative power. (The weights are verified using the feature importance validation method of Random Forest.)

[0082] 3) Health Classification: Based on 300 sets of standardized data with ground-truth labels, three thresholds {T1, T2} are determined, with typical values ​​of T1 = 0.55 and T2 = 0.84. Health levels are classified by comparing the scoring function S with the thresholds.

[0083]

[0084] Level 0 represents healthy, while levels 1-3 correspond to mild, moderate, and severe abnormalities, respectively.

[0085] To verify the efficiency and reliability of the method of this invention, the experimental results of the multi-feature fusion decision model were identified as follows. Figure 4 As can be seen, the model correctly identified 2949 out of a total of 3056 chicken head images, achieving an accuracy rate of 96.5%. This indicates that the model can effectively identify the health status of chickens, and it achieved good results in identifying various health levels, demonstrating good classification performance. Of the 3118 chicken heads appearing in the captured images, 3056 were detected, with a false negative rate controlled at approximately 2%. Due to limitations in the selected chicken head orientation and differences in the distance between the comb and the camera, a small number of samples showed slight misclassification, mainly distributed between the healthy (health_0) and slightly abnormal (health_1) categories. Overall, the proportion of healthy chickens misclassified as sick chickens was approximately 1.2%, and the proportion of sick chickens misclassified as healthy chickens was approximately 1.0%, further validating the model's efficiency and reliability in monitoring chicken health status.

[0086] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A method for online monitoring of chicken health status integrating lightweight networks and feature repair, characterized in that... It includes the following steps: S1. Image data acquisition and preprocessing of chickens to obtain a dataset, and to divide it into training and validation sets; S2. Construct an improved YOLOv8n chicken head orientation recognition model. The YOLOv8n model uses the StarNet lightweight network as the backbone, embeds the SimSPPF module for multi-scale feature fusion, and integrates the GAM global attention mechanism. The model is trained using the dataset obtained in S1 to obtain the optimal recognition model. S3. Acquire and preprocess image data of the chickens to be monitored, and determine the chicken's head orientation based on the optimal recognition model; S4. Chicken comb restoration and feature extraction in the image, obtaining features: chicken comb color, lateral chicken comb area and frontal chicken comb width; S5. Multi-feature fusion decision, defining the scoring function S: S=αA'+βW'+γC ′ In the formula, α, β, γ are weights, A' is the normalized lateral comb area, W' is the normalized frontal comb width, and C... ′ The normalized rooster comb color index; S6. Based on the comparison between the scoring function S and the threshold, the health level is classified.

2. The method according to claim 1, characterized in that... In S1, specifically: Data acquisition: Cameras are deployed using a suspended installation structure with a 20° downward angle and a horizontal distance of 1.4m; Preprocessing: The original video frames are geometrically calibrated using a radial distortion correction model to eliminate deformation errors of the chicken coop crossbar; The effective detection areas are retained by ROI cropping, and keyframes are extracted based on time interval sampling. The data is cleaned to remove blurry and severely occluded data, thereby constructing the initial dataset.

3. The method according to claim 1, characterized in that... In S4, specifically: S4-1, Lab Space Segmentation: Convert the image to Lab space, extract the a* channel and perform OTSU threshold segmentation; S4-2, Occlusion Repair: Generate a horizontal mask for the fractured area and input it into the DeepFillv2 network for repair; S4-3, Feature Quantization: chicken Comb color: Calculate the average a value C for the comb area; Lateral view of rooster comb area: Calculate the number of connected pixels A in the binary image; Width of the rooster comb in the correct position: Calculate the aspect ratio R of the minimum bounding rectangle.

4. The method according to claim 3, characterized in that: Converting from RGB color space to Lab color space requires first converting to XYZ color space. The conversion formula from RGB color space to XYZ color space is as follows: The formula for converting from XYZ color space to Lab color space is: In the formula, X n Y n and Z n It is the standard value for reference white point, and D65 white point is commonly used.

5. The method according to claim 3, characterized in that... X n Y n and Z n Take 95.047, 100.0, and 108.

883.

6. The method according to claim 1, characterized in that... α = 0.4, β = 0.4, γ = 0.

6.

7. The method according to claim 1, characterized in that... In S6, three thresholds {T1,T2} are determined, and the health level is classified by comparing the scoring function S with the thresholds. Level 0 represents healthy, while levels 1-3 correspond to mild, moderate, and severe abnormalities, respectively.

8. The method according to claim 7, characterized in that T1 = 0.55, T2 = 0.84.

Images