Infrared monitoring method for footpad temperature of caged meat duck
By using unmanned inspection equipment with visual and far-infrared acquisition units in three-dimensional caged ducks, combined with semantic segmentation and target detection models, the problem of accuracy in monitoring the body temperature of caged ducks has been solved, realizing non-contact and accurate foot temperature measurement, and ensuring the reliability and accuracy of health assessment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING AGRI MECHANIZATION INST MIN OF AGRI
- Filing Date
- 2026-02-27
- Publication Date
- 2026-07-21
AI Technical Summary
In three-dimensional cage-raised meat ducks, traditional body temperature monitoring methods are inefficient, have a strong lag, and cannot accurately reflect the true physiological state of the meat ducks. Thermal imaging technology is severely affected by cage netting and feathers, and cannot accurately link temperature abnormalities to specific individuals.
Using unmanned inspection equipment equipped with vision and far-infrared acquisition units, combined with semantic segmentation and target detection models, the equipment identifies and removes cage netting obstructions, completes the foot contour through generative adversarial networks, achieves non-contact and accurate measurement of foot temperature, and assesses the health status of ducks by combining time-synchronized triggering mechanisms.
It enables non-contact, accurate body temperature monitoring in intensive cage environments, avoiding animal stress, providing reliable measurement results, accurately pinpointing the temperature zone of the paws, reducing stress response and the risk of disease transmission, and providing reliable health assessment data.
Smart Images
Figure CN121740240B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of infrared monitoring technology, and more specifically, to an infrared monitoring method for the foot temperature of caged ducks. Background Technology
[0002] With the intensive development of the breeding industry, three-dimensional cage farming has become the mainstream model for meat duck farming because it can greatly improve space utilization and breeding density. However, this model also brings huge challenges to individual health monitoring. Meat ducks are poultry, and body temperature is one of the most sensitive and direct indicators of their health status. In traditional farming, veterinarians or managers rely on experience to observe, which is inefficient, has a strong lag, and is prone to missing early symptoms. Contact body temperature measurement (such as rectal thermometer) is completely impractical in large-scale farming and will cause serious animal stress.
[0003] Although thermal imaging technology can be used for non-contact temperature measurement, its direct application to caged ducks presents several challenges. First, the dense metal cage mesh severely obscures the duck's body, making it difficult for the thermal imager to obtain a complete thermal image of the body surface. Second, the thick feathers covering the main parts of the duck's body result in a surface temperature far lower than the core body temperature, which is greatly affected by environmental ventilation. This leads to large fluctuations and poor correlation in the measured values, making it impossible to accurately reflect the true physiological state. Furthermore, thermal imaging images cannot clearly distinguish closely adjacent individuals, making it impossible to accurately link temperature anomalies to specific individuals in specific cage locations. Therefore, we propose an infrared monitoring method for the foot temperature of caged ducks. Summary of the Invention
[0004] The purpose of this invention is to provide an infrared monitoring method for the foot temperature of caged ducks, in order to solve the problems mentioned in the background art.
[0005] To achieve the above objectives, the present invention provides a method for infrared monitoring of foot temperature in caged ducks, comprising the following steps:
[0006] Step 1: Use unmanned inspection equipment equipped with a visual acquisition unit and a far-infrared acquisition unit to move around the cage-raising area and collect visible light images and infrared thermal images of the meat ducks.
[0007] Step 2: Use a semantic segmentation algorithm to process the visible light image, identify and segment the metal cage pixel region, remove the metal cage pixel region from the image to obtain the de-occluded image, and preliminarily locate the effective area of the duck's foot in the de-occluded image based on the target detection model.
[0008] As a further improvement to this technical solution, the identification and segmentation of the metal mesh pixel region includes the following steps:
[0009] Multiple sets of visible light images of caged ducks with labeled cage netting pixels were extracted from a pre-stored database of standard duck foot contours. A semantic segmentation algorithm with an encoder-decoder structure was used to perceive and segment the cage netting pixels, generating de-occluded images, where:
[0010] The encoder is responsible for extracting the mesh features of the image, including repeating mesh patterns, linear edge features, and texture information related to the metal material;
[0011] The decoder is responsible for reconstructing the pixel-level position and shape of the cage net in the original image.
[0012] As a further improvement to this technical solution, the preliminary localization of the effective area of the duck's foot in the de-occluded image based on the target detection model includes the following steps:
[0013] The bounding boxes of the feet of ducks in the standard contour database of duck feet were labeled with age attribute labels to form a localization dataset.
[0014] The YOLO v8 model was used as the basic framework to train the localization dataset, and the weights of the classification loss and localization loss were adjusted using the bounding box regression loss function to generate the object detection model.
[0015] The de-occluded image is used to extract features and input into the target detection model to identify the foot position, contour, and size parameters of ducks of different ages, generate the effective foot area, and complete the initial foot localization.
[0016] As a further improvement to this technical solution, the standard contour database of duck feet also includes:
[0017] Average foot size, contour features, foot temperature, and growth and development data for three age groups: 3-7 days old, 8-29 days old, and 30-37 days old.
[0018] The contour feature parameters include aspect ratio, contour curvature, and key node distribution features.
[0019] Step 3: Morphological analysis of the effective area of the foot. Based on the aspect ratio, contour curvature, and key node distribution characteristics of the feet of ducks of different ages, incomplete feet of ducks that are obscured by the cage are identified. By combining the standard contour of the feet of the corresponding age with the generative adversarial network algorithm, the contour information of the incomplete feet is completed and the complete foot contour is output.
[0020] As a further improvement to this technical solution, the morphological analysis of the effective area of the foot, based on the aspect ratio, contour curvature, and key node distribution characteristics of the feet of ducks of different ages, identifies incomplete duck feet obscured by the cage, including the following steps:
[0021] Receive the image patch corresponding to the bounding box of the effective area of the foot, convert it into a binary image, and extract its outermost pixel set to obtain the actual observed contour of the current foot.
[0022] Based on the standard profile database of duck feet, the age corresponding to the foot size parameters is predicted. Standard data matching the predicted age of the current foot is called and calculated. The actual observed profile is compared with the standard data, including:
[0023] Aspect Ratio Analysis: Calculate the aspect ratio of the minimum bounding rectangle of the observed profile. If the difference between the ratio and the standard data is higher than the critical threshold, an incomplete signal is output, indicating that the profile is lost due to occlusion.
[0024] Contour curvature analysis: Calculate the curvature of each point on the contour line. If a smooth curve or break is detected at the position of a high curvature point, an incomplete signal is output.
[0025] Key node distribution feature verification: Detect key anatomical nodes on the contour, compare the number and relative position of the detected nodes with the standard topology in the standard data, and output an incomplete signal if they do not match.
[0026] As a further improvement to this technical solution, the generative adversarial network algorithm includes a generator and a discriminator. The generator generates a predicted contour map of the occluded part based on the standard contour features of the foot corresponding to the age of the child through a convolutional neural network. The discriminator judges the similarity between the predicted contour and the standard contour. When the similarity is ≥90%, the generator outputs the completed foot contour.
[0027] Step 4: Using a time-synchronized triggering mechanism, acquire far-infrared thermal imaging images of the foot positioning area, extract feature values representing foot temperature within the area, and assess the health status of the individual duck based on the feature values and the age information of the duck corresponding to the foot.
[0028] As a further improvement to this technical solution, a time-synchronized triggering mechanism is adopted to acquire far-infrared thermal imaging images of the foot positioning area, including the following steps:
[0029] The spatial geometric transformation relationship between the computer vision acquisition unit and the far-infrared acquisition unit is calculated, and the transformation matrix is pre-calibrated.
[0030] The pixel coordinates of the complete foot outline are identified and mapped onto the synchronously acquired infrared thermal image as the region of interest.
[0031] Based on the conversion of infrared thermal images into temperature matrices, the temperature matrix of the region of interest is extracted as a feature value representing the temperature of the sole of the foot.
[0032] As a further improvement to this technical solution, a regional average temperature algorithm is used to extract the feature value representing the foot temperature within the region. The average value of the temperature matrix of the region of interest is calculated as the final foot temperature data, eliminating the temperature interference of edge pixels.
[0033] As a further improvement to this technical solution, based on the characteristic values and the age information of the duck corresponding to the foot, the health status of the individual duck is assessed, including temperature change trend analysis, specifically:
[0034] If the temperature of the sole of the foot is within the sub-health threshold range for three consecutive monitoring tests and shows an upward or downward trend, it is judged as a sub-healthy state.
[0035] If the foot temperature is monitored for three consecutive times and is within the abnormal threshold range, it is determined to be an abnormal state, triggering an audible and visual alarm and a cloud notification.
[0036] The sub-health threshold range is ±0.3℃ above the normal range, and the abnormal threshold range is ±0.5℃ above the normal range.
[0037] The normal range is as follows: 38.5-41.5℃ for ducklings aged 3-7 days, 38.3-41.3℃ for medium-sized ducks aged 8-29 days, and 38.0-41.0℃ for adult ducks aged 30-37 days.
[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0039] This infrared monitoring method for foot temperature in caged ducks achieves truly non-contact and accurate body temperature monitoring of individual ducks in a densely caged environment. It bypasses the two long-standing fundamental technical obstacles of "feather insulation" and "cage nets obscuring the body," solving the problem of stress-free and accurate body temperature monitoring in caged poultry. It avoids capture stress, and the measurement of foot temperature has a high correlation with core body temperature, making the results more reliable. Furthermore, the entire process is non-contact, requiring no human contact with the ducks, thus avoiding animal stress caused by contact temperature measurement. It balances monitoring accuracy and breeding welfare, reducing the risk of growth inhibition or disease transmission caused by stress reactions.
[0040] Simultaneously, it outputs a complete foot outline, avoiding temperature sampling errors caused by occlusion. The complete outline can accurately locate the effective temperature measurement area of the foot, eliminating occlusion interference and ensuring that the extracted temperature feature values closely match the actual foot temperature. This provides a reliable data foundation for core body temperature inference and enables precise binding of temperature data to specific cage locations and individual ducks. It avoids the problem of being unable to distinguish adjacent individuals due to incomplete outlines, providing a clear target association basis for subsequent establishment of individual health records and anomaly tracing. Attached Figure Description
[0041] Figure 1 This is a flowchart illustrating the overall structure of the present invention;
[0042] Figure 2 This is a flowchart illustrating the process of identifying incomplete duck feet obscured by the cage, according to the present invention.
[0043] Figure 3 This is a schematic diagram illustrating the principle of the occlusion removal image of the present invention. Detailed Implementation
[0044] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0045] Example 1
[0046] Please see Figure 1 As shown in the figure, this embodiment provides a method for infrared monitoring of foot temperature in caged ducks, including the following steps:
[0047] Step 1: Using unmanned inspection equipment equipped with a visual acquisition unit and a far-infrared acquisition unit, the equipment moves within the caged duck area to collect visible light images and infrared thermal images of the ducks. The visual acquisition unit can be a visible light camera, and the far-infrared acquisition unit can be an infrared thermal imager. The unmanned inspection equipment can be a tracked robot that moves along a preset path within the caged duck house. Through a hardware synchronization triggering mechanism, it is ensured that the visible light images and infrared thermal images collected from the same scene at any given time are strictly aligned in time and space. The visible light images are used for high-precision target recognition and segmentation, and the infrared thermal images are used for temperature information extraction.
[0048] Step 2: Use a semantic segmentation algorithm to process the visible light image, identify and segment the metal cage pixel region, remove the metal cage pixel region from the image to obtain the de-occluded image, and preliminarily locate the effective area of the duck's foot in the de-occluded image based on the target detection model.
[0049] Furthermore, such as Figure 3 As shown, the process of identifying and segmenting the pixel region of the metal mesh cage includes the following steps:
[0050] By maintaining a camera distance of 1.5-2.5 meters from the cage net and using a tripod to avoid motion blur, at least 5000 original images were captured, covering all dimensions of combined scenes, specifically including: different lighting (natural light / supplementary light / backlight), different time of day, different seasons, different cage net densities, thicknesses, newness and age (including rust), shooting angles (front view, oblique view, overhead view), feed, feathers, water stains attached to the cage net, and occlusion caused by different postures of the foreground ducks. Professionals used annotation tools to outline each metal cage net wire in the images at the pixel level, labeling it as the foreground ("cage net"), and labeling all other areas (ducks, background, etc.) as the background, forming a dataset stored in the standard outline database of duck feet;
[0051] Multiple sets of visible light images of caged ducks with labeled cage netting pixels were extracted from a pre-stored database of standard duck foot contours. A semantic segmentation algorithm with an encoder-decoder structure was used to perceive and segment the cage netting pixels, generating de-occluded images, where:
[0052] The encoder (such as ResNet) is responsible for extracting the cage features of the image, including repeating grid patterns, linear edge features, and texture information related to metal material. Specifically, with ResNet and other networks as the backbone, the input image is processed through a series of convolutional layers. The shallow network captures low-level features, such as the edges, corners, short lines, and bright spots produced by metal reflections of the cage wire. As the network deepens, the middle network combines short lines to identify intermediate features, such as regular intersections and line segments in specific directions, which helps to robustly perceive the cage even if it is occluded, contaminated, or reflected in some areas.
[0053] The decoder is responsible for restoring the pixel-level position and shape of the wire mesh in the original image. Specifically, it gradually expands the feature map size through upsampling operations. The key technique lies in skip connections: the upsampling results during the decoding process are fused with the shallow feature map at the same scale as the encoder. The shallow feature map retains rich details and accurate edge position information, which helps the decoder accurately draw the wire mesh outline on the correct pixels when restoring the size. Before the final output, the edges of the segmented wire mesh are fine-tuned using the shallowest original edge features to ensure that the segmentation boundary is aligned with the sub-pixel edge of the wire, avoiding jagged or blurry boundaries.
[0054] Then, the preliminary localization of the effective area of the duck's foot in the unoccluded image based on the target detection model includes the following steps:
[0055] First, semantic segmentation algorithms with encoder-decoder structures are used to perceive and segment cage net pixels in the visible light images already stored in the standard contour database of duck feet, generating unoccluded images. Professionals then use rectangular boxes to accurately annotate the bounding boxes of the duck feet in the unoccluded images of the standard contour database of duck feet, and attach age attribute labels (e.g., feet of ducklings aged 1-10 days, feet of growing ducks aged 11-30 days), forming a localization dataset. The initial dataset in the standard contour database of duck feet contains feet from ducklings to slaughter at various ages, with huge differences in size and shape, covering feet in different postures such as standing, walking, hanging, and resting. It includes images of feet blending with the background of residual feed, water stains, and feces, as well as cases where feet are partially occluded by adjacent ducks or the crossbars at the bottom of the cage, ensuring broad coverage.
[0056] The YOLO v8 model was used as the basic framework to train the localization dataset. A bounding box regression loss function was employed to adjust the weights of the classification and localization losses. The YOLO v8 model used was a nano model (balancing speed and accuracy), with an initial learning rate of 0.01 and 500 iterations. The SGD optimizer was used. The GAN generator consisted of 4 convolutional layers, and the discriminator consisted of 3 convolutional layers. The similarity threshold of 90% was set based on a large number of contour pairing samples for validation. Specifically:
[0057] Bounding box regression loss (localization loss): CIoU Loss or EIoU Loss is used instead of traditional IoU Loss. This type of loss not only considers the overlap area between the predicted box and the ground truth box, but also the consistency of the center point distance and aspect ratio. It can more precisely guide the model to learn how to generate bounding boxes that closely fit the actual contour of the foot, which is the core of achieving accurate localization.
[0058] Classification loss: We use a weighted cross-entropy loss function to address the potential imbalance in the number of foot samples at different ages, ensuring that the model has good recognition capabilities for all age categories.
[0059] Loss weight adjustment: In the total loss function of the model, the weight ratio of localization loss to classification loss is dynamically or manually adjusted. Given that the accurate localization (coordinates, size) of the foot is crucial for subsequent contour completion and infrared temperature measurement in this task, the weight of localization loss is usually increased appropriately to guide the model to prioritize the accuracy of the bounding box position, and then the accuracy of age classification.
[0060] Initializing with YOLOv8 weights pre-trained on a large public dataset (such as COCO) can accelerate model convergence and improve performance. The prepared dataset is then input into the model for training. The network weights are continuously optimized using the backpropagation algorithm to minimize the loss function defined above. The model's performance is evaluated on an independent validation set. Key metrics include: mean accuracy (AP), recognition accuracy for different age categories, and IoU value for bounding box localization. The model is then ensured to meet preset accuracy and recall requirements to generate an object detection model.
[0061] The de-occluded image is used as input for feature extraction into a target detection model. The model's backbone network (such as CSPDarknet) performs multi-level feature extraction on the input image, constructing a feature pyramid. Shallow features contain rich detailed information (such as foot edges and textures), suitable for detecting small targets; deep features contain high-level semantic information, understanding the overall concept of "foot". The Feature Pyramid Network (FPN) and Path Aggregation Network (PAN) structures fuse and enhance features from different levels, and make predictions on feature maps at three different scales. This ensures that the foot can be effectively detected regardless of its size in the image, identifying the foot position, contour, and size parameters of ducks of different ages, generating an effective foot region, and completing the initial foot localization, including:
[0062] Location: The center point of the bounding box, expressed in pixel coordinates;
[0063] Contour: The rectangular region circumscribed by the bounding box of the foot, which is the base region for subsequent more detailed pixel-level contour analysis (such as morphological analysis and GAN completion);
[0064] Size parameters: The width and height of the bounding box (in pixels), combined with known camera calibration parameters, can indirectly estimate the actual physical size of the foot, providing auxiliary information for age verification.
[0065] It is worth noting that the aforementioned standard outline database of duck feet also includes:
[0066] Average foot size, contour features, foot temperature, and growth and development data were collected for three age groups: 3-7 days old (2-5cm), 8-29 days old (5-8cm), and 30-37 days old (8-12cm).
[0067] The contour feature parameters include aspect ratio, contour curvature, and key node distribution features.
[0068] Step 3: Morphological analysis of the effective area of the foot. Based on the aspect ratio, contour curvature, and key node distribution characteristics of the feet of ducks of different ages, incomplete feet of ducks that are obscured by the cage are identified. By combining the standard contour of the feet of the corresponding age with the generative adversarial network algorithm, the contour information of the incomplete feet is completed and the complete foot contour is output.
[0069] Because many feet are obscured by the cage's structural members, their visible outlines are incomplete. Directly using these incomplete outlines for infrared thermography would lead to incorrect temperature sampling areas, affecting the accuracy of health assessments. Therefore, if... Figure 2 As shown, the morphological analysis of the effective area of the foot, based on the aspect ratio, contour curvature, and key node distribution characteristics of the feet of ducks of different ages, identifies incomplete feet of ducks obscured by the cage, including the following steps:
[0070] The image patch corresponding to the bounding box (ROI region) of the effective area of the foot is received, converted into a binary image (e.g., the foot pixels are separated from the background by thresholding), and its outermost pixel set is extracted to obtain the actual observed contour of the current foot.
[0071] Based on the standard profile database of duck feet, the age corresponding to the foot size parameters is predicted. Standard data matching the predicted age of the current foot is called and calculated. The actual observed profile is compared with the standard data, including:
[0072] Aspect Ratio Analysis: Calculate the aspect ratio of the smallest bounding rectangle of the observed profile. If the difference between the ratio and the standard data (e.g., 3-7 day old feet are short and round, with a small aspect ratio; 30-37 day old feet are long and spread out, with a large aspect ratio) is higher than the critical threshold, then an incomplete signal is output, indicating that the profile is lost due to occlusion, such as the toes or webs being missing.
[0073] Contour curvature analysis: Calculate the curvature of each point on the contour line. A complete foot contour should include specific high curvature points (such as sharp toes) and low curvature segments (such as smooth webbed edges). If a smooth curve or break is detected at the location of a high curvature point (based on the distribution characteristics of key nodes), it indicates that the area is occluded and an incomplete signal is output.
[0074] Key node distribution feature verification: Detect key anatomical nodes on the contour (usually through pre-trained shape models or corner detection algorithms), such as toe tips and webbed connection points. Compare the number and relative positions of the detected nodes with the standard topology in the standard data. If they do not match, for example, the standard should have 4 toes, but only 2 are detected, then an incomplete signal is output.
[0075] Furthermore, the generative adversarial network algorithm includes a generator and a discriminator. The generator generates a predicted contour map of the occluded area based on the standard foot contour features corresponding to the child's age using a convolutional neural network. The discriminator determines the similarity between the predicted contour and the standard contour. When the similarity is ≥90%, the generator outputs the completed foot contour. Specifically:
[0076] When an incomplete signal is received, a conditional generative adversarial network is used. The generator (G) is input to a binary mask of the incomplete contour and an age label (as a condition), while the discriminator (D) is input to the real complete contour or the contour output by the generator. The age label is also used for judgment.
[0077] Then, a large number of complete, age-labeled foot contour masks need to be collected. Simulating the occlusion of cage wire rods, a portion is randomly erased from the complete mask to artificially create incomplete-complete contour pairing data pairs, which are used to train the network to learn the mapping relationship from incomplete to complete.
[0078] During the completion process, the incomplete foot contour mask and its age information are input into the pre-trained generator (G). The encoder part inside the generator first understands the existing shape of the input incomplete contour. The conditional information (age) guides the network to recall the size proportion and shape prior of the foot at the corresponding age from the "memory" of the standard database. Based on this, the decoder part imagines and draws the reasonable shape of the occluded part and generates a predicted complete contour mask. The discriminator forces the generator not only to complete the shape, but also to make the completed contour look realistic, that is, to conform to the general growth law and morphological characteristics of the foot of the duck at that age, and to avoid producing strange or unreasonable shapes.
[0079] Note: The completed contour output by the generator is smoothed to eliminate any possible minor jagged edges. The completed contour is then quickly compared with a standard contour database to verify that its aspect ratio, area, etc., have returned to the normal range. This step serves as a quality check to ensure that the completed result conforms to anatomical common sense.
[0080] Step 4: Using a time-synchronized triggering mechanism, acquire far-infrared thermal imaging images of the foot positioning area, extract feature values representing foot temperature within the area, and assess the health status of the individual duck based on the feature values and the age information of the duck corresponding to the foot.
[0081] The process involves using a time-synchronized triggering mechanism to acquire far-infrared thermal imaging images of the foot positioning area, including the following steps:
[0082] When the inspection equipment moves to the predetermined monitoring position, the control system sends a synchronization pulse signal, simultaneously triggering both devices (the visual acquisition unit and the far-infrared acquisition unit) to perform exposure acquisition. This ensures that the visual acquisition unit and the far-infrared acquisition unit capture the same physical scene at the same microsecond level. Even a slight delay of milliseconds can cause image misalignment, as the duck's feet may move. Hardware synchronization fundamentally eliminates measurement errors caused by time asynchrony. Before deployment, the visual acquisition unit and the far-infrared acquisition unit, which are rigidly installed together, need to be jointly calibrated using a special calibration board that can be recognized in both wavelength bands (e.g., a checkerboard pattern with heating points). Multiple sets of images are captured simultaneously, and the spatial geometric transformation relationship between the visual acquisition unit and the far-infrared acquisition unit (including rotation, translation, scaling, and lens distortion parameters) is calculated. The transformation matrix is pre-calibrated.
[0083] The pixel coordinates of the complete foot outline are identified and mapped onto the synchronously acquired infrared thermal image as the region of interest.
[0084] Based on the conversion of infrared thermal images into temperature matrices (each pixel corresponds to a temperature value), the temperature matrix of the region of interest is extracted as a feature value representing the temperature of the sole of the foot.
[0085] Then, feature values representing foot temperature are extracted within the region. The average temperature algorithm is used to calculate the average temperature matrix of the region of interest as the final foot temperature data. Temperature interference from edge pixels (5 pixels outside the outline) is eliminated, and environmental compensation correction is performed on the extracted foot temperature to reduce the impact of environmental heat radiation on the absolute accuracy of the measurement. Specifically, environmental compensation can be combined with the real-time ambient temperature and ventilation speed of the duck house. A linear regression model is used to construct the compensation formula: Corrected temperature = Measured temperature + 0.02 × (Ambient reference temperature - Real-time ambient temperature) - 0.01 × Ventilation speed, where the ambient reference temperature is set to 25℃ (suitable growth temperature for meat ducks).
[0086] Furthermore, based on the aforementioned characteristic values and the age information of the duck corresponding to the foot, the health status of the individual duck is assessed, including temperature change trend analysis, specifically:
[0087] If the temperature of the sole of the foot is within the sub-health threshold range for three consecutive monitoring tests and shows an upward or downward trend, it is judged as a sub-healthy state.
[0088] If the foot temperature is within the abnormal threshold range for three consecutive monitoring tests, or if it exceeds the abnormal threshold range once and fails to return to the normal range for the next two tests, it is determined to be an abnormal state, triggering an audible and visual alarm and a cloud notification.
[0089] The sub-health threshold range is ±0.3℃ above the normal range, and the abnormal threshold range is ±0.5℃ above the normal range.
[0090] The normal range is as follows: 38.5-41.5℃ for ducklings aged 3-7 days, 38.3-41.3℃ for medium-sized ducks aged 8-29 days, and 38.0-41.0℃ for adult ducks aged 30-37 days.
[0091] Therefore, all assessment results, raw data, and triggered actions are fully recorded to form individual health records. The accumulated "early warning - final clinical diagnosis" result data can be used to periodically verify and optimize temperature thresholds and trend judgment rules, forming a closed-loop optimization. Among them, the "three consecutive confirmations" mechanism effectively filters out occasional temperature fluctuations caused by brief environmental interference (such as instantaneous strong winds, brief exposure to cold water), measurement noise, or instantaneous animal activity. A single abnormal reading will not immediately trigger the highest level alarm, but will be given an observation period. It will only be triggered when the abnormality is confirmed to be persistent. This makes the alarm information more reliable, avoids false alarms, and enhances the management's trust in the system.
[0092] Finally, the age information, foot temperature data, health status assessment results, and corresponding image data of each individual duck are stored in the cloud server to establish an individual health record, supporting the traceability of historical data and the iterative optimization of the monitoring model.
[0093] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.
Claims
1. A method for infrared monitoring of foot temperature in caged meat ducks, characterized in that, Includes the following steps: Unmanned inspection equipment equipped with visual acquisition units and far-infrared acquisition units is used to move and collect visible light images and infrared thermal images of meat ducks in the cage-raising area. A semantic segmentation algorithm was used to process the visible light image, identify and segment the pixel region of the metal cage mesh, and remove the pixel region of the metal cage mesh from the image to obtain the de-occluded image. Based on the target detection model, the effective area of the foot of the meat duck was initially located in the de-occluded image. Morphological analysis of the effective area of the foot: Based on the aspect ratio, contour curvature, and key node distribution characteristics of the feet of ducks of different ages, incomplete feet of ducks that are obscured by the cage are identified. By combining the standard contour of the feet of the corresponding age with the generative adversarial network algorithm, the contour information of the incomplete feet is completed and the complete foot contour is output. Using a time-synchronized triggering mechanism, far-infrared thermal imaging images of the foot positioning area are acquired, and feature values representing foot temperature are extracted within the area. Based on the feature values and the age information of the duck corresponding to the foot, the health status of the individual duck is assessed. The morphological analysis of the effective area of the foot, based on the aspect ratio, contour curvature, and key node distribution characteristics of the feet of ducks of different ages, identifies incomplete feet of ducks that are obscured by the cage, including the following steps: Receive the image patch corresponding to the bounding box of the effective area of the foot, convert it into a binary image, and extract its outermost pixel set to obtain the actual observed contour of the current foot. Based on the standard profile database of duck feet, the age corresponding to the foot size parameters is predicted. Standard data matching the predicted age of the current foot is called and calculated. The actual observed profile is compared with the standard data, including: Aspect Ratio Analysis: Calculate the aspect ratio of the minimum bounding rectangle of the observed profile. If the difference between the ratio and the standard data is higher than the critical threshold, an incomplete signal is output, indicating that the profile is lost due to occlusion. Contour curvature analysis: Calculate the curvature of each point on the contour line. If a smooth curve or break is detected at the position of a high curvature point, an incomplete signal is output. Key node distribution feature verification: Detect key anatomical nodes on the contour, compare the number and relative position of the detected nodes with the standard topology in the standard data, and output an incomplete signal if they do not match. The generative adversarial network algorithm includes a generator and a discriminator. The generator generates a predicted contour map of the occluded part based on the standard contour features of the foot corresponding to the age of the child through a convolutional neural network. The discriminator judges the similarity between the predicted contour and the standard contour. When the similarity is ≥90%, the generator outputs the completed foot contour.
2. The infrared monitoring method for foot temperature of caged meat ducks according to claim 1, characterized in that: The process of identifying and segmenting the pixel region of the metal mesh includes the following steps: Multiple sets of visible light images of caged ducks with labeled cage netting pixels were extracted from a pre-stored database of standard duck foot contours. A semantic segmentation algorithm with an encoder-decoder structure was used to perceive and segment the cage netting pixels, generating de-occluded images, where: The encoder is responsible for extracting the mesh features of the image, including repeating mesh patterns, linear edge features, and texture information related to the metal material; The decoder is responsible for reconstructing the pixel-level position and shape of the cage net in the original image.
3. The infrared monitoring method for foot temperature of caged meat ducks according to claim 2, characterized in that: The method for initially locating the effective area of the duck's foot in the unoccluded image based on the object detection model includes the following steps: The bounding boxes of the feet of ducks in the standard contour database of duck feet were labeled with age attribute labels to form a localization dataset. The YOLO v8 model was used as the basic framework to train the localization dataset, and the weights of the classification loss and localization loss were adjusted using the bounding box regression loss function to generate the object detection model. After feature extraction from the de-occluded image, it is input into the target detection model to identify the foot position, contour, and size parameters of ducks of different ages, generate the effective foot area, and complete the initial foot localization.
4. The infrared monitoring method for foot temperature of caged meat ducks according to claim 3, characterized in that: The standard outline database of duck feet also includes: Average foot size, contour features, foot temperature, and growth and development data for three age groups: 3-7 days old, 8-29 days old, and 30-37 days old. The contour feature parameters include aspect ratio, contour curvature, and key node distribution features.
5. The infrared monitoring method for foot temperature of caged meat ducks according to claim 4, characterized in that: Using a time-synchronized triggering mechanism, far-infrared thermal imaging images of the foot positioning area are acquired, including the following steps: The spatial geometric transformation relationship between the computer vision acquisition unit and the far-infrared acquisition unit is calculated, and the transformation matrix is pre-calibrated. The pixel coordinates of the complete foot outline are identified and mapped onto the synchronously acquired infrared thermal image as the region of interest. Based on the conversion of infrared thermal images into temperature matrices, the temperature matrix of the region of interest is extracted as a feature value representing the temperature of the sole of the foot.
6. The infrared monitoring method for foot temperature of caged meat ducks according to claim 5, characterized in that: The feature value representing the foot temperature is extracted within the region using a region average temperature algorithm. The average value of the temperature matrix of the region of interest is calculated as the final foot temperature data, eliminating temperature interference from edge pixels.
7. The infrared monitoring method for foot temperature of caged meat ducks according to claim 6, characterized in that: Based on the aforementioned characteristic values and the age information of the duck corresponding to the foot, the health status of the individual duck is assessed, including temperature change trend analysis, specifically: If the temperature of the sole of the foot is within the sub-health threshold range for three consecutive monitoring tests and shows an upward or downward trend, it is judged as a sub-healthy state. If the foot temperature is monitored for three consecutive times and is within the abnormal threshold range, it is determined to be an abnormal state, triggering an audible and visual alarm and a cloud notification. The sub-health threshold range is ±0.3℃ above the normal range, and the abnormal threshold range is ±0.5℃ above the normal range. The normal range is as follows: 38.5-41.5℃ for ducklings aged 3-7 days, 38.3-41.3℃ for medium-sized ducks aged 8-29 days, and 38.0-41.0℃ for adult ducks aged 30-37 days.