Multi-species living small target health state detection method fusing machine vision and multi-sensor technology

Abstract

The invention relates to the technical field of living body small target detection, in particular to a multi-species living body small target health state detection method fusing machine vision and a multi-sensor technology. The method comprises the following steps: determining a physiological reference and an abnormality judgment standard; constructing a multi-source data acquisition system to acquire sensor data such as images and infrared temperature and pressure in a typical scene; after the data is preprocessed and marked, an improved convolutional neural network model is constructed, a multi-scale attention module is introduced to enhance the capturing capability of tiny features such as skin texture posture details, and a small target detection head is newly added to improve the detection precision of a small target; after being trained, the model is deployed to an upper computer and is integrated with a sensor system to realize real-time collaborative analysis of vision and sensing data; and the integrated platform outputs multi-species health state evaluation and abnormal early warning to realize efficient and accurate health monitoring of the small target living body. According to the invention, automatic and accurate identification and early warning of abnormal health states of felines and insects are realized.

Description

Technical Field

[0001] This invention relates to the field of live small target detection technology, specifically to a method for detecting the health status of multi-species live small targets that integrates machine vision and multi-sensor technology. Background Technology

[0002] Currently, in the field of live animal health status detection, especially for small targets of multiple species such as felines and insects, traditional methods mainly rely on manual observation or single sensor technology, which have limitations such as low efficiency, strong subjectivity, and susceptibility to environmental interference. With the development of machine vision and sensor technology, some studies have attempted to use image recognition or sensor data for analysis, but most of them are limited to a single technical route and lack an effective mechanism for multi-source information fusion.

[0003] However, in complex scenarios, such as minor lesions on animal surfaces or individual abnormalities in insect populations, existing solutions often struggle to achieve accurate and stable detection. Furthermore, systematic detection standards and adaptation methods have not yet been established to address the physiological differences among different species and at different growth stages. Therefore, there is an urgent need for a health status detection method that can integrate multiple technologies, adapt to various scenarios, and possess strong generalization capabilities to improve automation and detection accuracy. Summary of the Invention

[0004] The purpose of this invention is to address the problems existing in the background technology by proposing a method for detecting the health status of multi-species live small targets that integrates machine vision and multi-sensor technology.

[0005] The technical solution of this invention: a method for detecting the health status of small, multi-species live targets by integrating machine vision and multi-sensor technology, comprising the following specific implementation steps: S1. Conduct basic research on the health mechanisms of multiple species and establish physiological benchmarks and criteria for judging health abnormalities in target species; S2. Construct a multi-source data acquisition system to collect image data of multiple species of live small targets, as well as infrared temperature, pressure, and humidity sensor data in typical scenarios. S3. Preprocess the acquired image data and sensor data; S4. Create a labeled dataset based on the preprocessed data and physiological benchmarks and health abnormality judgment criteria; S5. Construct an improved convolutional neural network model, introduce a multi-scale attention module, and add a small target detection head; S6. Train and encapsulate the improved convolutional neural network model using labeled datasets; S7. Deploy the packaged model to the host computer and integrate it with the multi-source data acquisition system to achieve real-time collaborative analysis of visual data and sensor data, so as to output health status assessment and abnormal warning results.

[0006] Preferably, conducting basic research on the health mechanisms of multiple species specifically includes: For felines, the study divides them into kitten, adult, and senior stages, and defines the baseline ranges for weight and body temperature for each age group. For insects, identify the typical appearance and physiological characteristics of bees and medicinal beetles at different developmental stages; The system systematically reviews common skin diseases in felines and clarifies their corresponding skin and fur appearance characteristics; Summarize the common types of health abnormalities in insects and identify their corresponding physical manifestations.

[0007] Preferably, constructing a multi-source data acquisition system specifically includes: For felines, domestic and outdoor scenes were selected as collection locations, and data were collected during the daily routines of domestic cats and the activity times of stray cats. For insects, breeding sheds and outdoor breeding areas were selected as collection sites, and their active periods were chosen for collection. Using a high-definition camera, image data is collected from various light sources, including natural light, indoor lighting, and supplementary lighting, using shooting angles such as overhead, side, oblique, or macro shots. Infrared temperature sensors are used to collect body surface temperature data, pressure sensors are used to collect pressure feedback data to calculate weight, and temperature and humidity sensors are used to monitor ambient temperature and humidity data.

[0008] Preferably, the preprocessing of the acquired image data and sensor data specifically includes: Denoising, hair or occlusion segmentation, and image enhancement processing are performed on images of felines. Cluster target segmentation and micro-feature magnification processing are performed on insect images; Outlier values ​​are removed from the raw data collected by infrared temperature, pressure, and temperature and humidity sensors. Perform data standardization on the sensor data after removing outliers; The standardized sensor data and image data are precisely aligned over time.

[0009] Preferably, the creation of a labeled dataset specifically includes: Based on the physiological benchmarks and health abnormality judgment criteria, the preprocessed feline images are labeled with health or abnormality categories, and the abnormal areas are marked with their locations. Based on the physiological benchmarks and health abnormality judgment criteria, the preprocessed insect images are labeled with health and disease categories; The labeled image dataset is divided into training set, validation set and test set according to a predetermined ratio.

[0010] Preferably, constructing an improved convolutional neural network model specifically includes: The feature maps output by specific modules in the backbone network are used as the input to the multi-scale attention module; Within the multi-scale attention module, shallow high-resolution detail information of skin texture and local erythema is extracted through the shallow detail feature branch, and deep semantic information of overall body posture and behavioral posture is extracted through the deep semantic feature branch. Global average pooling is performed on the shallow detail feature map and the deep semantic feature map respectively to obtain feature vectors. The feature vectors are then concatenated and sequentially processed by a fully connected layer for dimensionality reduction, ReLU activation function, dimensionality recovery by a fully connected layer, and Sigmoid activation function to obtain the attention weight vector. The attention weight vector is weighted and added to the shallow detail feature map and the deep semantic feature map respectively to obtain the fused feature map as the output of the multi-scale attention module; A new general-purpose small target detection head is added to the convolutional neural network to work in conjunction with the original detection head, in order to adapt to the small target detection needs of different species.

[0011] Preferably, the multi-scale attention module also includes a parallel processing mechanism for global information branches and local feature branches; The global information branch performs horizontal and vertical average pooling to capture global features, and then sequentially undergoes 1×1 convolution dimensionality reduction, Sigmoid activation function to generate preliminary weights, group normalization, and Softmax operation to obtain channel attention distribution. Finally, it is refined into global feature weights through matrix multiplication fusion and average pooling. The local feature branch uses a 3×3 convolution kernel to capture key details of the body surface, and then processes them sequentially through the Softmax function, matrix multiplication to fuse attention weights, and average pooling to obtain local anomaly feature weights. The results of the global information branch and the local feature branch are added together, multiplied element-wise with the original input feature map, and then the final feature map is output through the Sigmoid activation function.

[0012] Preferably, training and encapsulating the improved convolutional neural network model using labeled datasets specifically includes: The model is trained using the training and validation sets, and the learning rate is dynamically adjusted using a cosine annealing strategy. During training, model performance is monitored and parameters are adjusted using the validation set. Once the model's recognition accuracy on the independent test set reaches the predetermined standard, the optimal model is saved as a file in a specific format to complete the model encapsulation.

[0013] Preferably, the packaged model is deployed to a host computer and integrated with a multi-source data acquisition system, specifically including: The packaged model file is deployed to the host computer system, enabling it to perform real-time image recognition and analysis. Integrating pressure sensors, infrared temperature sensors, and temperature and humidity sensors, a real-time collaborative system of visual recognition and sensing monitoring is formed. When detecting felines, pressure sensors and infrared temperature sensors are activated. When detecting insects, activate the temperature and humidity sensor; The health abnormality thresholds for each sensor are set in strict accordance with the findings of basic research on the health mechanisms of multiple species.

[0014] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects: This invention designs a method for detecting the health status of small, multi-species live targets by integrating machine vision and multi-sensor technologies. First, by fusing machine vision and data from multiple types of sensors, a collaborative detection system combining non-contact and contact methods is constructed, enabling comprehensive health status perception of live targets of multiple species, including felines and insects. This overcomes the bottleneck of single-vision technology being susceptible to occlusion, lighting, and small target size in complex scenes, significantly improving the reliability and robustness of health status assessment. Second, addressing the challenge of small target detection, an innovative multi-scale attention module is introduced, along with a dedicated small target detection head, enabling the deep learning model to simultaneously capture the target's global semantic information and local subtle features. This not only enhances the ability to extract and identify key features such as local skin lesions in animals and abnormalities on the body surface of insects, but also effectively improves the model's generalization ability and detection accuracy across different body sizes and species. Furthermore, this method achieves an automated closed loop from data acquisition and model analysis to decision output, enabling the fusion analysis of visual recognition results with sensor-measured body temperature, weight, and environmental parameters, and intelligent judgment based on pre-researched physiological mechanism models. This allows for early detection and accurate warning of health abnormalities, greatly reducing reliance on manual labor and improving monitoring efficiency. Finally, this solution possesses good versatility and scalability, providing valuable reference for the health status monitoring of other small living targets, and has broad application prospects in fields such as smart farming, pet health management, and biodiversity research. Attached Figure Description

[0015] Figure 1This is a flowchart of a method for detecting the health status of small, multi-species live targets that integrates machine vision and multi-sensor technology, as proposed in this invention. Figure 2 This is a structural diagram of the EMA module in the embodiment; Figure 3 This is a diagram of the improved network model structure in the embodiment; Figure 4 The schematic diagram of the LM35 circuit in the embodiment is shown below; Figure 5 The schematic diagram of the HX711 circuit in the embodiment is shown below; Figure 6 The diagram shows the STM32F103C8 microcontroller used in this embodiment. Figure 7 This is a 3D modeling diagram from the embodiment; Figure 8 Assembly view in the embodiment Figure 1 ; Figure 9 Assembly view in the embodiment Figure 2 ; Figure 10 Assembly view in the embodiment Figure 3 ; Figure 11 Assembly view in the embodiment Figure 4 ; Figure 12 This is a detailed list diagram for the embodiment; Reference numerals: 1: Right-angle buffer protection component; 2: Mounting bracket; 3: Ball screw; 4: Connector; 5: Photoelectric displacement sensor; 6: Acrylic sheet; 7: Angle iron; 8: Aluminum profile 1; 9: Directional wheel; 10: Glass 1; 11: Caster connecting component; 12: Glass 2; 13: Aluminum profile 2; 14: Hinge; 15: Glass 3; 16: Depth camera; 17: Infrared temperature sensor; 18: Stepper motor; 19: Camera guide rail connector; 20: Non-contact infrared temperature camera; 21: Aluminum profile 3; 22: Column head with shoulder bolt; 23: Pressure sensor Measurement plate; 24: Pressure sensor; 25: Glass 4; input: Input feature map; h: Height of feature map; c: Number of channels in feature map; w: Width of feature map; g: Number of groups (divides channels into g groups); c / / g: Number of channels in each group (integer division); g*batchsize: Batch size multiplied by number of groups; X: Feature map segmentation or transformation operation; AvgPool: Average pooling layer; Concat+1×1: Concatenation followed by a 1×1 convolution; Sigmoid: Sigmoid activation function; Group Norm: Group Normalization; Softmax: Softmax activation function; Matmul: Matrix multiplication; Softmax: Softmax activation function; Output: Final output of the module; backbone: Backbone network; neck: Neck network; head: Detection head; SPPF: Spatial pyramid pooling fast module; maxpool: Max pooling layer; ConvBNSiLU: Convolution + batch normalization + SiLU activation function; C3: C3 bottleneck structure module; Conv: Convolution module; Concat: Concatenation operation; Detect: Small detection head; Upsample: Upsampling; Bottleneck: Bottleneck structure; Conv2d: 2D convolution; BatchNorm2d: 2D batch normalization; Bbox.Loss: Bounding box loss; Cls.Loss: Classification loss. Detailed Implementation

[0016] Example 1: This invention proposes a method for detecting the health status of small, multi-species live targets by integrating machine vision and multi-sensor technologies, such as... Figure 1 As shown, the specific implementation steps include the following: Step S1: Basic research on the health mechanisms of multiple species; Step S2: Multi-scenario, multi-source data collection; Step S3: Data preprocessing, including enhancing relevant animal images and removing outliers from the data; Step S4: Create a dataset, label animal images, and divide them into training / validation / test sets; Step S5: Construct and improve the convolutional neural network model, based on deep learning algorithms, by introducing relevant attention modules and adding a small target detection head; Step S6: Model training and packaging. Set hyperparameters to train the model, save the optimal model, and package it. Step S7: Deploy multiple modules, deploy the model to the host computer, and build a pressure and infrared temperature sensor monitoring system.

[0017] In an optional embodiment, step S1 involves conducting basic research on the health mechanisms of the target species (taking felines and insects as examples) to lay a theoretical foundation for subsequent data annotation, feature extraction, and sensor threshold setting. Specifically, this includes: S11. Physiological Characteristics of Felines: Felines are clearly divided into three age groups: kittens (1-12 months), adults (1-8 years), and senior cats (8 years and older). Their physical appearance and skin condition are analyzed at each stage, and the baseline ranges for weight and body temperature are precisely defined, as shown in Table 1. Table 1. Physiological characteristics of felines at different age stages ; S12. Insect Physiological Characteristics Research: For bees and medicinal beetles, clarify their typical appearance and physiological characteristics at different developmental stages (such as larval and adult stages), including but not limited to body size, body length, and body surface characteristics. The physiological characteristic benchmarks of bees and medicinal beetles are shown in Table 2. Table 2 Physiological characteristics of insects at different age stages ; S13. Research on the Mechanisms of Abnormal Health in Felines: This study systematically reviews four common types of skin diseases in felines (fungal, parasitic, allergic, and bacterial), and clarifies the typical skin / fur appearance characteristics of each disease. The diagnostic criteria for feline skin diseases are shown in Table 3 below. Table 3. Criteria for Diagnosing Common Skin Diseases in Felines ; S14. Research on the Mechanisms of Insect Health Abnormalities: Summarize common types of health abnormalities in insects (bees, medicinal beetles) (such as Varroa mite infestation, sacbrood disease, fungal infection, and environmental maladaptation), and clarify their typical external manifestations. The judgment criteria are shown in Table 4. Table 4. Criteria for Judging Common Health Abnormalities in Insects .

[0018] In an optional embodiment, step S2, based on the research results of step S1, establishes a multi-source data acquisition system to systematically collect images and sensor data of multiple species of live small targets under different typical scenarios, in order to construct a comprehensive raw database, specifically including: S21. Determine the data collection scenario and time period: For felines, we identified domestic settings (living room, cat bed, balcony) and outdoor settings (abandoned yards, alley corners) as data collection locations, and selected the daily activity times of domestic cats (7:00-8:00 AM and 7:00-8:00 PM) and the activity times of stray cats (5:00-6:00 AM and 9:00-10:00 PM) for data collection. For insects, the collection sites were determined to be breeding sheds (with different temperature and humidity control zones and different breeding density zones) and outdoor breeding areas (around bee hives and wild breeding areas for medicinal insects), and the collection was carried out during their active periods (such as the daytime feeding time of silkworms and the daytime honey-collecting time of bees). S22. Image Data Acquisition: Using a high-definition camera, data is acquired under various light source conditions (natural light, indoor lighting, supplementary lighting, etc.) and at targeted shooting angles. For felines, use overhead shots to show the entire skin, side shots to focus on skin details, and angled shots in the wild; For insects, macro photography and overhead shots from different distances are used to showcase both group and individual characteristics; Ensure that the images collected cover all combinations of different species, living conditions, ages, and health statuses; S23, Sensor Data Acquisition: Simultaneous acquisition of data from multiple sensors: Use an infrared temperature sensor to collect body surface temperature; Pressure sensors are used to collect pressure feedback when a living body or container is placed, which is then used to calculate body weight. Temperature and humidity sensors are used to monitor the temperature and humidity of the rearing environment.

[0019] In an optional embodiment, step S3 performs targeted preprocessing operations on the raw data collected in step S2 to improve data quality and ensure the effectiveness of subsequent model training and the consistency of data fusion, specifically including: S31, Image Data Processing: For images of animals such as felines, the focus is on denoising, hair / occlusion segmentation, and image enhancement. For insect images, the focus is on cluster target segmentation (to avoid interference from individual overlap) and magnification of subtle features (such as changes in the pubescence on the insect's body surface). S32, Sensor Data Processing: Outlier values ​​are removed from the raw data collected by infrared temperature, pressure, temperature and humidity sensors. Perform data standardization on the data after removing outliers; The time series of standardized sensor data and image data are precisely aligned.

[0020] In an optional embodiment, step S4 prepares the preprocessed data into a standardized dataset that can be used for model training, specifically including: S41. Data annotation: Based on the research results of step S1, the preprocessed image data is annotated; For images of felines, label their health / abnormality categories and mark the locations of abnormal areas; For insect images, label their health / disease category; S42. Dataset partitioning: Divide the labeled image dataset into training set, validation set and test set in a ratio of 8:1:1.

[0021] In an optional embodiment, step S5 uses the YOLO series neural network as the base network and performs targeted optimizations and improvements. The core of this optimization is the introduction of a customized multi-scale attention (EMA) module and the addition of a small target detection head to enhance the model's ability to extract and detect key features of small targets across multiple species. Specifically, this includes: S51. Input Feature Map Processing: The feature map output by the C3 module in Backbone is used as the input of the EMA module. Assume that the size of the input feature map is C×H×W, where C represents the number of channels, and H and W represent the height and width of the feature map, respectively. S52. Multi-branch feature extraction, specifically: Shallow detail feature branch: A convolutional layer with a 3×3 small convolutional kernel is used to extract shallow high-resolution detail information such as skin texture and local erythema on the animal surface, and generate shallow fine feature map; Deep semantic feature branch: Using a 5×5 large convolutional kernel or a convolutional layer combined with downsampling operations (such as max pooling), deep semantic information such as the overall body shape and behavioral posture of felines is extracted to obtain a deep semantic feature map; S53, Attention weight calculation, specifically: For shallow detail feature maps F detail and deep semantic feature map F semantic Perform global average pooling to compress the feature map into a C×1×1 vector, resulting in the shallow detail feature vector V. detail and deep semantic feature vector V semantic V detail and V semanticAfter concatenation, a dimensionality reduction operation is performed through a fully connected layer (such as reducing the dimension from 2C to C / r, where r is the dimensionality reduction coefficient, which can be adjusted according to the model's computational resources and performance requirements, and is generally taken as 2-4), to obtain the intermediate vector M; The intermediate vector M is processed by the ReLU activation function, and then the dimension is restored to C through another fully connected layer to obtain the attention weight vector A. The Sigmoid activation function is applied to A to make the weight values ​​range between 0 and 1. S54. Feature Fusion Output: The attention weight vector A is then fused with the shallow detail feature map F. detail and deep semantic feature map F semantic Perform a weighting operation, then add the two weighted feature maps together to obtain the fused feature map F. fusion This feature map contains both subtle feature information of the target object and integrates overall semantic information. As the output of the EMA module, it is fed into subsequent network layers for processing. EMA module structure as follows Figure 2 As shown, this module enhances the capture of multi-scale features of felines through a dual-branch parallel processing mechanism: First, the global information branch: first, average pooling in the horizontal and vertical directions is performed to capture global features, then dimensionality reduction and feature fusion are achieved through 1×1 convolution, then preliminary weights are generated through Sigmoid, and channel attention distribution is obtained by combining group normalization and Softmax operation. Subsequently, feature weights at the global level are extracted through matrix multiplication fusion and average pooling. Secondly, a local feature branch is added: a local feature extraction branch is added, which uses a 3×3 convolution kernel to accurately capture key details of the body surface of different species. For animals, it can capture changes in skin texture (such as roughness, scab texture) and local disease morphology (such as punctate dandruff, small red spots). For insects, it can capture subtle deformities (such as missing wings) and abnormal body color (such as white spots on the body surface of silkworms and color spots on the abdomen of bees). Subsequently, the feature is classified by the Softmax function, and attention weights are fused by matrix multiplication (strengthening the feature weights of animal disease areas and insect abnormal body shapes, and weakening interference from family background, shed clutter, etc.). Then, the redundant information is compressed by the average pooling operation, and finally the feature weights focused on local anomalies are obtained. The result of this branch is added to the result of the global feature branch, and after element-wise multiplication with the original input feature map, the corresponding dimension feature map is output by the Sigmoid activation function. This effectively enhances the model's efficiency in capturing and utilizing differences in the body surface of animals of different sizes (such as kittens and adult cats), subtle skin diseases of animals (such as early ringworm, slight hair loss), and subtle abnormal features of insects (such as poor larval development). S55. Introduce a small target detection head, specifically: To address the limitations of the original model's three detection heads in detecting small targets across multiple species (such as difficulty in accurately identifying small patches of ringworm in animals, minor body anomalies in insects, or inability to adapt to small targets such as kittens and insect larvae), a new general-purpose small target detection head has been added. This head adapts to the detection needs of different species through adaptive parameter adjustment—enhancing the identification of small skin lesions in animals and optimizing the detection of individual anomalies in cluster environments for insects. The improved four detection heads work collaboratively, enabling simultaneous detection tasks across multi-scale feature maps (from low-resolution global features to high-resolution local features). This achieves deep fusion of multi-scale information—covering the overall health status of larger targets like adult cats with the large-scale detection head, while capturing subtle animal symptoms, juvenile characteristics, and individual insect anomalies with the newly added small-target detection head. By rationally coordinating the weight allocation of features across different scales and species, the accuracy of detecting health anomalies in small targets across multiple species (animals and insects) is significantly improved. For details of the improved network structure model, please refer to [link to network structure model]. Figure 3 .

[0022] In an optional embodiment, step S6 trains and encapsulates the optimized model from step S5, specifically including: S61. Model Training: Train the model using the training and validation sets defined in step S4, and dynamically adjust the learning rate using a cosine annealing strategy, as shown in the following formula: ; in, This represents the learning rate at the t-th training epoch. This represents the minimum value of the learning rate; This represents the maximum value of the learning rate; Indicates the current training round; This represents the total number of training rounds in a complete cosine annealing cycle; S62. Model Packaging: During training, the model performance is monitored in real time using the validation set, and parameters are adjusted to avoid overfitting. Once the model's recognition accuracy on the independent test set reaches the predetermined standard, the optimal model is saved as a .pt file to complete the model packaging for subsequent deployment.

[0023] In an optional embodiment, step S7 integrates and deploys the packaged model and hardware system, specifically including: S71. Model Deployment: Deploy the optimal model (.pt file) packaged in step S6 to the host computer system to enable it to perform real-time image recognition and analysis. S72. Build a sensor system, integrating corresponding sensor modules according to the specific detection needs of different species, to form a real-time collaborative system of visual recognition and species-adaptive sensing monitoring: When testing felines, a pressure sensor (used to help fix the posture and measure relevant animal weight data) and an infrared temperature sensor (used to collect body surface temperature) are used. When detecting insects, use a temperature and humidity sensor (to monitor the breeding environment). The selection and operating principles of various sensors (e.g., see the LM35 infrared temperature sensor circuit) Figure 4 For the HX711 pressure sensor circuit, please refer to [link / reference]. Figure 5 For STM32F103C8 series microcontrollers, please refer to [link / reference]. Figure 6 The setting of health abnormality thresholds for different species (see Tables 5 and 6 for details) must strictly follow the mechanistic research results of step S1. Table 5 Infrared Temperature Sensor Thresholds ; Table 6 Pressure Sensor Thresholds .

[0024] Example 2: The present invention proposes a multi-species live small target health status detection device that integrates machine vision and multi-sensor technology. This device is used to execute the multi-species live small target health status detection method that integrates machine vision and multi-sensor technology proposed in Example 1, comprising: Transmission and drive system: The depth camera movement on the left, right and rear sides adopts a combination of ball screw and linear guide to achieve high-precision linear motion. The vertical (up and down) travel is 300mm, and the positioning accuracy is ±0.1mm, ensuring accurate camera height adjustment. The drive motor is a Leadshine DM542 stepper motor, which is matched with a synchronous pulley (transmission ratio 1:1). The moving speed is adjustable from 0.1-8mm / s, which can meet the stability requirements of static shooting and can also flexibly adjust the speed according to the cat's cooperation. Main Structure and Materials: The platform frame is constructed using industrial-grade European standard 4040 aluminum alloy profiles, which are lightweight and have strong load-bearing capacity; the pressure detection plate frame is made of 304 stainless steel with a 2mm thick anti-slip silicone layer on the surface, ensuring the cat's standing safety and preventing displacement of beehives and other boxes, which would affect the accuracy of weight measurement; key transmission components (such as lead screw nuts and guide rail sliders) are made of 45# steel with heat treatment to reduce deformation after long-term use and improve structural stability; Sensor and camera configuration: Through-beam photoelectric positioning sensors are installed at the extreme positions of the lead screws on the left, right, and rear sides for camera movement origin calibration and travel protection to prevent damage from overtravel; the depth camera is a D435i, fixed with an adjustable bracket to ensure that the optical axis is perpendicular to the cat detection surface, with a depth of field coverage of 200-3000mm; the infrared temperature sensor (LM35) is integrated and fixed on the rear side to capture the temperature of cats entering the platform at any time, without the need for additional detection points; Its workflow is as follows: After the platform starts, it first performs a self-check. Each lead screw motor 18 drives the camera 16 back to its initial position, and the pressure detection plate 23 is initialized to zero. When a target is on the pressure plate, the pressure sensor 24 triggers a signal, and the system collects weight data. Then, based on the target's body shape, the system automatically adjusts the camera height to the set threshold. The left, right, and rear motors drive the camera to move sequentially. Each time the camera moves to a preset detection point (e.g., 3 points on the left side, 3 points on the right side, and 2 points on the back), the photoelectric sensor 5 triggers the camera to take pictures, collecting a total of 20 images (including depth data), which are then synchronously transmitted to the host computer. The host computer uses algorithms to fuse and analyze the multi-directional images to identify relevant symptoms and generates a health report by combining body temperature and weight data. It should be noted that in terms of 3D modeling and verification of the platform, the overall structural design was completed using Solidworks software (such as...). Figure 7 (See the 3D model diagram shown). Figure 8-11 This demonstrates the positional relationships of the various components (motor, ball screw, depth camera, pressure plate, etc.). Figure 12 The material (such as aluminum alloy profile model, 45# steel component specifications) and model information of each component are marked in the details column; Furthermore, to achieve efficient feedback, visualization, and early warning of health data for felines (domestic pet cats and stray cats) and insects (bees and medicinal beetles), it is necessary to build an intelligent execution and interaction system based on the FastAPI+React architecture. This system can integrate health detection results from multiple species (feline skin diseases, individual bee abnormalities, beetle development status) and sensor data (feline weight / body temperature, bee hive temperature / weight, beetle rearing box temperature, humidity, and weight), providing real-time feedback to the platform interface and triggering targeted care reminders. The specific construction steps are as follows: (1) The backend uses the FastAPI framework to achieve front-end and back-end separation, and creates RESTful APIs in FastAPI; (2) Define the HTTP request processing logic in the API, extract the animal health test results and corresponding weight and body temperature data from the database, and send these data back to the front end; (3) Use the React framework to build a front-end platform, with WebStorm as the integrated development environment. In React, use the fetch API to send HTTP requests to obtain back-end data. (4) In the React component, display the returned animal health test results, corresponding weight, body temperature data and possible skin disease categories on the platform page.

[0025] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A method for detecting the health status of small, multi-species live targets by integrating machine vision and multi-sensor technology, characterized in that, The specific implementation steps include the following: S1. Conduct basic research on the health mechanisms of multiple species and establish physiological benchmarks and criteria for judging health abnormalities in target species; S2. Construct a multi-source data acquisition system to collect image data of multiple species of live small targets, as well as infrared temperature, pressure, and humidity sensor data in typical scenarios. S3. Preprocess the acquired image data and sensor data; S4. Create a labeled dataset based on the preprocessed data and physiological benchmarks and health abnormality judgment criteria; S5. Construct an improved convolutional neural network model, introduce a multi-scale attention module, and add a small target detection head; S6. Train and encapsulate the improved convolutional neural network model using labeled datasets; S7. Deploy the packaged model to the host computer and integrate it with the multi-source data acquisition system to achieve real-time collaborative analysis of visual data and sensor data, so as to output health status assessment and abnormal warning results.

2. The method for detecting the health status of multi-species live small targets by integrating machine vision and multi-sensor technology according to claim 1, characterized in that, Conducting basic research on the mechanisms of health in multiple species specifically includes: For felines, the study divides them into kitten, adult, and senior stages, and defines the baseline ranges for weight and body temperature for each age group. For insects, identify the typical appearance and physiological characteristics of bees and medicinal beetles at different developmental stages; The system systematically reviews common skin diseases in felines and clarifies their corresponding skin and fur appearance characteristics; Summarize the common types of health abnormalities in insects and identify their corresponding physical manifestations.

3. The method for detecting the health status of multi-species live small targets by integrating machine vision and multi-sensor technology according to claim 2, characterized in that, The construction of a multi-source data acquisition system specifically includes: For felines, domestic and outdoor scenes were selected as collection locations, and data were collected during the daily routines of domestic cats and the activity times of stray cats. For insects, breeding sheds and outdoor breeding areas were selected as collection sites, and their active periods were chosen for collection. Using a high-definition camera, image data is collected from various light sources, including natural light, indoor lighting, and supplementary lighting, using shooting angles such as overhead, side, oblique, or macro shots. Infrared temperature sensors are used to collect body surface temperature data, pressure sensors are used to collect pressure feedback data to calculate weight, and temperature and humidity sensors are used to monitor ambient temperature and humidity data.

4. The method for detecting the health status of multi-species live small targets by integrating machine vision and multi-sensor technology according to claim 3, characterized in that, Preprocessing of the acquired image and sensor data specifically includes: Denoising, hair or occlusion segmentation, and image enhancement processing are performed on images of felines. Cluster target segmentation and micro-feature magnification processing are performed on insect images; Outlier values ​​are removed from the raw data collected by infrared temperature, pressure, and temperature and humidity sensors. Perform data standardization on the sensor data after removing outliers; The standardized sensor data and image data are precisely aligned over time.

5. The method for detecting the health status of multi-species live small targets by integrating machine vision and multi-sensor technology according to claim 4, characterized in that, Creating a labeled dataset specifically includes: Based on the physiological benchmarks and health abnormality judgment criteria, the preprocessed feline images are labeled with health or abnormality categories, and the abnormal areas are marked with their locations. Based on the physiological benchmarks and health abnormality judgment criteria, the preprocessed insect images are labeled with health and disease categories; The labeled image dataset is divided into training set, validation set and test set according to a predetermined ratio.

6. The method for detecting the health status of multi-species live small targets by integrating machine vision and multi-sensor technology according to claim 5, characterized in that, Building an improved convolutional neural network model specifically includes: The feature maps output by specific modules in the backbone network are used as the input to the multi-scale attention module; Within the multi-scale attention module, shallow high-resolution detail information of skin texture and local erythema is extracted through the shallow detail feature branch, and deep semantic information of overall body posture and behavioral posture is extracted through the deep semantic feature branch. Global average pooling is performed on the shallow detail feature map and the deep semantic feature map respectively to obtain feature vectors. The feature vectors are then concatenated and sequentially processed by a fully connected layer for dimensionality reduction, ReLU activation function, dimensionality recovery by a fully connected layer, and Sigmoid activation function to obtain the attention weight vector. The attention weight vector is weighted and added to the shallow detail feature map and the deep semantic feature map respectively to obtain the fused feature map as the output of the multi-scale attention module; A new general-purpose small target detection head is added to the convolutional neural network to work in conjunction with the original detection head, in order to adapt to the small target detection needs of different species.

7. The method for detecting the health status of multi-species live small targets by integrating machine vision and multi-sensor technology according to claim 6, characterized in that, The multi-scale attention module also includes a parallel processing mechanism for global information branches and local feature branches; The global information branch performs horizontal and vertical average pooling to capture global features, and then sequentially undergoes 1×1 convolution dimensionality reduction, Sigmoid activation function to generate preliminary weights, group normalization, and Softmax operation to obtain channel attention distribution. Finally, it is refined into global feature weights through matrix multiplication fusion and average pooling. The local feature branch uses a 3×3 convolution kernel to capture key details of the body surface, and then processes them sequentially through the Softmax function, matrix multiplication to fuse attention weights, and average pooling to obtain local anomaly feature weights. The results of the global information branch and the local feature branch are added together, multiplied element-wise with the original input feature map, and then the final feature map is output through the Sigmoid activation function.

8. The method for detecting the health status of multi-species live small targets by integrating machine vision and multi-sensor technology according to claim 7, characterized in that, The training and encapsulation of the improved convolutional neural network model using labeled datasets specifically includes: The model is trained using the training and validation sets, and the learning rate is dynamically adjusted using a cosine annealing strategy. During training, model performance is monitored and parameters are adjusted using the validation set. Once the model's recognition accuracy on the independent test set reaches the predetermined standard, the optimal model is saved as a file in a specific format to complete the model encapsulation.

9. The method for detecting the health status of multi-species live small targets by integrating machine vision and multi-sensor technology according to claim 8, characterized in that, The packaged model is deployed to the host computer and integrated with the multi-source data acquisition system, specifically including: The packaged model file is deployed to the host computer system, enabling it to perform real-time image recognition and analysis. Integrating pressure sensors, infrared temperature sensors, and temperature and humidity sensors, a real-time collaborative system of visual recognition and sensing monitoring is formed. When detecting felines, pressure sensors and infrared temperature sensors are activated. When detecting insects, activate the temperature and humidity sensor; The health abnormality thresholds for each sensor are set in strict accordance with the findings of basic research on the health mechanisms of multiple species.

Images