An intelligent insect situation monitoring method and system based on self-distillation and large model generation

Abstract

The present application relates to a kind of intelligent pest monitoring method and system based on self-distillation and large model generation, by constructing the image-pest label image pair with multiple pest labels, and data enhancement is carried out;Subsequently, the teacher network guides the student network to generate prediction results, and the parameter optimization and hardware embedding of student network are completed by combining the way of multi-level distillation.The intelligent pest monitoring system adopts edge computing module in image processing link, so that the overall system is robust to light, background change, finally reduces the missed detection rate, overcomes the technical defects of sample scarcity and difficult to realize full-chain intelligent.

Description

Technical Field

[0001] This invention relates to the field of computer modeling and system technology, and more specifically, to an intelligent insect monitoring method and system based on self-distillation and large model generation. Background Technology

[0002] Currently, pest monitoring in smart agriculture mainly relies on two technological approaches: one is traditional manual inspection and trap counting, which is inefficient and lacks real-time performance; the other is automated monitoring systems based on machine vision, typically consisting of insect-attracting lamps, industrial cameras, image processing units, and remote servers. The latter's typical workflow is as follows: the insect-attracting lamp attracts pests, the camera captures images, the images are uploaded to a cloud server via 4G / 5G networks, and large-scale deep learning models (such as the YOLO series and ResNet) deployed in the cloud are used for identification. Finally, the results are returned to the user or a report is generated.

[0003] In recent years, with the development of edge computing, some devices have attempted to deploy lightweight convolutional neural networks (such as MobileNet and YOLOv3-tiny) on embedded platforms (such as Raspberry Pi and Jetson Nano) to perform inference directly on the device, thereby reducing reliance on the network. At the same time, some studies have begun to introduce Transformer and Large Language Model (LLM) to perform semantic analysis on detection results, generating more intuitive insect infestation reports.

[0004] However, existing lightweight models, limited by the number of parameters, suffer from a false negative rate of over 35% for minute pests (such as fruit flies and aphids) in complex field environments. While traditional pruning and quantization methods can compress model size, they lead to a significant decrease in accuracy (mAP can drop by up to 20%). To improve generalization ability, simply increasing the number of real samples faces the dilemma of high costs for pest image annotation and scarcity of long-tail category samples. Furthermore, existing systems either rely entirely on the cloud, resulting in limited network coverage and early warning delays of over 48 hours, or only output detection boxes at the edge, lacking semantic understanding of pest trends, environmental correlations, and control recommendations. Users still need to manually interpret the data, failing to form a complete closed loop from "seeing insects" to "understanding pest infestations." End-to-end cloud collaboration is inefficient, making it difficult to achieve full-chain intelligence under low bandwidth conditions. Summary of the Invention

[0005] The technical problem this invention aims to solve is how to overcome the technical shortcomings of existing insect infestation monitoring methods, such as high false negative rates, scarce samples, and difficulty in achieving full-chain intelligence. To overcome these shortcomings, this invention provides an intelligent insect infestation monitoring method and system based on self-distillation and large-scale model generation.

[0006] This invention provides a method for constructing an edge computing module for insect infestation monitoring, comprising the following steps: S1: Construct image-insect-labeled image pairs with multiple insect labels, and perform data augmentation on these image pairs to obtain a multi-view dataset; S2: Construct a teacher network and optimize its parameters on the multi-view dataset using supervised learning so that it can extract pest information from images, including pest category, quantity, confidence level, and location. S3: Construct a student network and optimize its parameters using a multi-level distillation method guided by the teacher network, so that it has the ability to extract the insect information from the image, just like the teacher network after parameter optimization. S4: Embed the optimized student network into hardware to obtain an edge computing module for extracting the insect information from the image.

[0007] The edge computing module construction method for insect pest monitoring disclosed in this invention first constructs image-insect-label image pairs with multiple insect labels and performs data augmentation to ensure reasonable views, insect counts, and image quality. Then, a teacher network guides a student network to generate prediction results, and multi-level distillation is used to optimize the parameters of the student network. The student network is a lightweight model, easily embedded in hardware. Through multi-level distillation, the computational load can be reduced while significantly improving measurement accuracy. Because the images used in the parameter optimization process have multiple insect labels, possessing multi-view and cross-view properties, the student network model finally embedded in the hardware reduces the false negative rate and significantly decreases the false positive rate for small pests, thereby improving the detection rate of small pests. Through this series of image construction and parameter optimization processes, the finally constructed edge computing module can successfully complete core detection and analysis, requiring only the upload of a small amount of key data, reducing network dependence and operating costs, thus overcoming the technical shortcomings of existing methods, such as sample scarcity and difficulty in achieving full-chain intelligence.

[0008] In one possible implementation, step S1 includes the following steps: S11: Collect multiple insect images and label them with insect information to obtain image-insect tag image pairs; S12: Use a masked controlled diffusion model to perform data augmentation on each sample of the image-insect label image pair to obtain the multi-view dataset.

[0009] Data augmentation using the Mask Controlled Diffusion Model (MSMC-PestGen) can achieve the technical effect of generating multiple synthetic views with "consistent foreground and varying backgrounds" (such as different lighting, crop backgrounds, and weather conditions) for each real sample, ultimately forming a multi-view training set. These views can share the same foreground mask, so their bounding boxes can be directly derived from the mask without additional manual annotation, thereby further improving the efficiency of teacher network model parameter optimization and making the model robust to changes in lighting and background, adaptable to various scenarios such as farmland, orchards, and tea gardens.

[0010] In one possible implementation, step S3 includes the following steps: S31: Construct a deep feature distillation loss function, a channel relation matrix distillation loss function, and a detection head output distillation loss function, with the output of the parameter-optimized teacher network and the output of the student network as variables; S32: Summate the deep feature distillation loss function, the channel relation matrix distillation loss function, and the detector head output distillation loss function to obtain cross-view... Figure 1 Consistent distillation loss function; S33: The cross-view Figure 1 The consistency distillation loss function is linearly combined with the same-view distillation loss function and the target detection network loss function to obtain the total loss function; S34: Find the optimal network parameters of the student network that make the total loss function take the best value, and use the optimal network parameters as the network parameters of the student network to complete the parameter optimization of the student network.

[0011] By using the network parameters of the lightweight student model as the target to be optimized, the system simultaneously receives real images and multi-view synthesized images during training and performs multi-level distillation to make the output of the student network approximate the output of the teacher network. This makes the student model robust to changes in lighting and background, thereby improving environmental adaptability. Furthermore, the edge computing module does not need to input a large number of images, thus overcoming the technical defects of existing methods, such as the scarcity of samples and the difficulty in achieving full-chain intelligence.

[0012] The second technical solution of the present invention is to provide an intelligent insect monitoring system based on self-distillation and large model generation, the system comprising: Collection chamber; A trapping module, located at the top of the collection chamber, is used to attract insects into the collection chamber. The sensor module is used to collect environmental information including light intensity, temperature, humidity, and rainfall. An image acquisition module, installed inside the acquisition chamber, is used to acquire images of insects; An edge computing module is electrically connected to the image acquisition module and is the edge computing module obtained by the edge computing module construction method for insect monitoring described in the invention, used to extract the insect information from the insect image; The terminal device is electrically connected to both the sensor module and the edge computing module, and is embedded with a large language model to generate an insect infestation analysis report based on the environmental information and the situation information. This report covers an overview of the insect infestation, analysis of influencing factors, assessment of the insect infestation level, short-term development trends, and control recommendations.

[0013] The system disclosed in this invention comprises a collection chamber, a trapping module, a sensor module, an image acquisition module, an edge computing module, and a terminal device. The trapping module attracts insects into the collection chamber, and the image acquisition module acquires images of the insects. Simultaneously, the sensor module collects environmental information including light, temperature, humidity, and rainfall. Then, in conjunction with the edge computing module and a large language model, it generates an insect infestation analysis report covering an overview of the infestation, analysis of influencing factors, assessment of infestation levels, short-term development trends, and control recommendations. This allows the system to not only output detection frames but also automatically generate detailed reports containing infestation trends, influencing factors, and control recommendations, lowering the barrier to interpretation for users, improving decision-making efficiency, and achieving end-to-end intelligent analysis. Furthermore, the control recommendations generated by this system can guide precise pesticide application; experimental results show that it can reduce pesticide usage by approximately 35%. At the same time, the system employs an edge computing module in the image processing stage, making it robust to changes in light and background, reducing the false negative rate while overcoming the technical limitations of sample scarcity and the difficulty in achieving full-chain intelligence.

[0014] In one possible implementation, the trapping module includes: Insect barrier; An insect-attracting lamp, installed on the insect-blocking plate, is used to generate insect-attracting light; A fan is used to create an airflow that introduces the insects into the collection chamber.

[0015] This trapping module, which integrates insect-attracting lamps, impact-type baffles, and fans, combines "light-attracting-impact" and "wind-assisted" technologies to effectively improve the capture efficiency and collection effect of pests.

[0016] In one possible implementation, the sensor module includes a raindrop sensor, a photosensor, a light sensor, a temperature sensor, a humidity sensor, and a light curtain sensor. This enables parallel acquisition of information on light, temperature, humidity, rain, snow, and light curtain, giving the entire system the technical effects of environmental perception and trigger control.

[0017] In one possible implementation, the system further includes an industrial control computer, which is electrically connected to the trapping module, the image acquisition module, and the sensor module, and is used to control the operation of the trapping module, the image acquisition module, and the sensor module according to the environmental information. In one possible implementation, the industrial computer is configured to perform the following steps: A1: Reads the current environment information; A2: Determine whether the trigger command requirements of the rain sensor have been met. If so, then shut down the host computer of the industrial control computer; If not, proceed to the next step; A3: Determine whether the trigger command requirements of the photosensor have been met. If so, proceed to the next step; If not, then shut down the host computer of the industrial control computer; A4: Activate the trapping module; A5: Determine whether the trigger command requirements of the light curtain sensor have been met. If so, proceed to the next step; If not, then return to step A1; A6: Call the image acquisition module to acquire insect images.

[0018] The industrial control computer is configured to read sensor data, i.e., environmental information, and control the operation of the trapping module, image acquisition module, and sensor module based on the reading and comparison results. This allows for data acquisition according to environmental conditions, achieving the beneficial effect of automatic control.

[0019] The third technical solution of the present invention is to provide an intelligent insect infestation monitoring method based on self-distillation and large model generation, applicable to the intelligent insect infestation monitoring system based on self-distillation and large model generation described in the invention, comprising the following steps: T1: The insects are attracted into the collection chamber by the trapping module, and images of the insects are acquired by the image acquisition module. T2: Collects environmental information including light intensity, temperature, humidity, and rainfall through sensor modules; T3: Extract information from the insect images using the edge computing module; T4: Generate an insect infestation analysis report based on the environmental information and the situation information using the terminal device. The report includes an overview of the insect infestation, analysis of influencing factors, assessment of the insect infestation level, short-term development trend, and control recommendations.

[0020] The intelligent insect pest monitoring method disclosed in this invention first attracts insects into the collection chamber through a trapping module and acquires insect images through an image acquisition module; simultaneously, an environmental information including light, temperature, humidity, and rainfall is collected through a sensor module. Then, with the cooperation of an edge computing module and a large language model, an insect pest analysis report is generated, covering an overview of the pest situation, analysis of influencing factors, pest severity assessment, short-term development trends, and control recommendations. This lowers the barrier to interpretation for users, improves decision-making efficiency, and achieves the technical effect of end-to-end intelligent analysis. Furthermore, the control recommendations generated by this method can guide precise pesticide application; experimental results show that it can reduce pesticide usage by approximately 35%. At the same time, the system employs an edge computing module in the image processing stage, making it robust to changes in light and background, reducing the false negative rate while overcoming the technical shortcomings of sample scarcity and difficulty in achieving full-chain intelligence. Attached Figure Description

[0021] Figure 1 This is a flowchart of an edge computing module construction method for insect monitoring disclosed in Embodiment 1 of this application; Figure 2 This is a schematic diagram of teacher-student network parameter optimization disclosed in Embodiment 1 of this application; Figure 3 This is a schematic diagram of the structure of an intelligent insect monitoring system based on self-distillation and large model generation disclosed in Embodiment 2 of this application; Figure 4 This is a schematic diagram of the overall architecture of an intelligent insect monitoring system based on self-distillation and large model generation disclosed in Embodiment 2 of this application; Figure 5 This is the industrial control computer operation flowchart disclosed in Embodiment 2 of this application; Figure 6 This is a flowchart of the insect infestation analysis report generation process disclosed in Embodiment 2 of this application; Figure 7 This is a flowchart of the method disclosed in Embodiment 2 of this application. Detailed Implementation

[0022] First, those skilled in the art should understand that these embodiments are merely used to explain the technical principles of the embodiments of this application and are not intended to limit the scope of protection of the embodiments of this application. Those skilled in the art can make adjustments as needed to adapt to specific application scenarios.

[0023] In the description of the embodiments of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "electrical connection" and "establishing an electrical connection relationship" should be interpreted broadly, that is, it should be understood that two or more parties have an electrical relationship, which can be a connection through a wire, a radio connection, or a combination of both; it can be a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in the embodiments of this application based on the specific circumstances.

[0024] The present application will be further described in detail below using two embodiments, in conjunction with the accompanying drawings and specific embodiments.

[0025] Example 1: See Figure 1 and Figure 2 This application discloses a method for constructing an edge computing module for insect infestation monitoring. Figure 1 Here is a flowchart of the construction method, which includes the following steps: S1: Data preparation and augmentation.

[0026] A multi-view dataset is obtained by constructing image-insect-label image pairs with multiple insect labels and performing data augmentation on these image pairs. Specifically, step S1 includes the following steps: S11: Image Collection. Collect multiple insect images and label them with insect information to obtain image-insect label image pairs. This embodiment uses a public dataset formed by merging datasets IP102 and Pest24. Based on this dataset, the Pest112 dataset, containing 112 pest category labels and over 44,000 images, is constructed.

[0027] S12: Data Augmentation. Data augmentation is performed on each sample of the image-insect label image pair using a mask-controlled diffusion model to obtain the multi-view dataset. Multiple synthetic views with consistent foreground and varying backgrounds (e.g., different lighting, crop backgrounds, weather conditions) are generated for each sample in the dataset using the mask-controlled diffusion model (MSMC-PestGen), forming the multi-view training set. These views share the same foreground mask, so their bounding boxes can be directly derived from the mask without additional manual annotation.

[0028] S2: Teacher model training.

[0029] A teacher network was constructed, and its parameters were optimized using supervised learning on a multi-view dataset. This optimized network was then used to extract pest information from images, including pest category, quantity, confidence level, and location. See also... Figure 2In this embodiment, the heavyweight eighth version of the object detection network (i.e., YOLOv8-s / m / l / x) is used as the teacher network. With this large-capacity model as the teacher network, fully supervised training is carried out on the "real + synthetic" multi-view dataset to obtain a high-precision teacher network, that is, a teacher network with optimized parameters.

[0030] S3: Student model distillation.

[0031] A student network was constructed, and its parameters were optimized using a multi-stage distillation method guided by the teacher network, enabling it to extract insect information from images, similar to the optimized teacher network. See also... Figure 2 In this embodiment, a lightweight version 8 object detection network (i.e., YOLOv8n) is used as the student network. The network parameters of this lightweight student network are used as the target to be optimized. During the training process, real images and multi-view synthesized images are received simultaneously, and multi-level distillation is performed.

[0032] The parameter optimization method of the student network in this embodiment is actually a cross-view cross-multilevel self-distillation (CMLD) framework diagram. Figure 2 A detailed description of the teacher-student network structure, multi-view input, MRAB alignment module, deep feature distillation, channel relation matrix distillation, output distillation, and cross-view... Figure 1 The process of consistency monitoring. See also Figure 2 Specifically, step S3 includes the following steps: S31: Construct deep feature distillation loss function, channel relation matrix distillation loss function, and detector head output distillation loss function, using the outputs of the optimized teacher network and student network as variables. This step includes the following steps: S311: Construction of the Deep Feature Distillation Loss Function. An Attention Alignment Block (MRAB) is introduced after the deep feature layers (e.g., P6, P8) of the teacher and student models. The MRAB consists of a Spatial Attention Module (SAM) and a Channel Attention Module (CAM) connected in series. Residual connections preserve the original features, ensuring consistent channel statistics and spatial focusing capabilities before the features enter the distillation loss calculation. The mean squared error loss between features is calculated, and foreground mask weighting is introduced to enhance the migration of the main pest region. The deep feature distillation loss function is defined as follows: , in, F S , F T These are the deep characteristics of student networks and teacher networks, respectively. S Representing the student network, T Representing the teacher network, M For the foreground mask, ⊙ indicates element-wise multiplication. This is the loss function for deep feature distillation.

[0033] S312: Channel Relationship Matrix Distillation (CRM) Loss Function Construction. The cosine similarity matrix between feature channels is calculated to approximate the channel relationships of the student network to the teacher network, thus preserving higher-order semantic associations. The loss function is defined as follows: , in, R(F) The cosine similarity matrix between channels. The channel relationship matrix is ​​the distillation loss function.

[0034] S313: Construction of the loss function for the detector head output distillation. The class probability distribution (logits) is aligned using KL divergence, and the bounding box regression distribution (DFL distribution) is aligned using L1. The loss function is defined as follows: , in, p For class probability distribution, q The DFL distribution for bounding box regression. The output distillation loss function is set for the detection head.

[0035] S32: Cross-view Figure 1 Construction of the consistency distillation loss function. The deep feature distillation loss function, the channel relation matrix distillation loss function, and the detector head output distillation loss function are summed to obtain the cross-view result. Figure 1 Coherence distillation loss function. For two different synthetic views x of the same pest. (i) x (j) This forces the student model to maintain consistency between its output features on one view and the teacher model's output features on another view (cross-supervision), thereby learning a context-invariant representation of the pest structure. The loss function is defined as: , in, and For view metrics.

[0036] S33: Will cross vision Figure 1 The consistency distillation loss function, the same-view distillation loss function, and the YOLOv8 object detection network loss function are linearly combined to obtain the total loss function. The total loss function is the weighted sum of the original detection loss of the YOLOv8 object detection network and the distillation losses mentioned above: , in, L same For the distillation loss in the same view (i=j case), the distillation weight λ d The value was optimized to 0.7 through experiments.

[0037] S34: Solve the optimal network parameters of the student network that make the total loss function take the best value by using the optimization model solution algorithm, and use the optimal network parameters as the network parameters of the student network to complete the parameter optimization of the student network.

[0038] S4: Lightweight model deployment.

[0039] The optimized student network is embedded in hardware to obtain an edge computing module for extracting insect information from images. In this embodiment, the distilled and optimized YOLOv8n model is converted from PyTorch format to ONNX format, and then converted to a format suitable for Raspberry Pi ARM architecture (.param + .bin) using the NCNN framework. The NCNN inference latency can be reduced to 82.7ms / frame (640×640 input), which is about 6 times faster than PyTorch native inference, while maintaining the same detection accuracy (mAP@50-95 reaches 53.0%).

[0040] Example 2: See Figure 3 and Figure 4 This embodiment further discloses an intelligent insect pest monitoring system based on self-distillation and large model generation. A schematic diagram of the structure of this intelligent insect pest monitoring system is shown below. Figure 3 As shown, the intelligent insect monitoring system includes a collection chamber, a trapping module, a sensor module, an image acquisition module, an industrial control computer, an edge computing module, and terminal devices. The edge computing module is electrically connected to the image acquisition module, the terminal devices are electrically connected to both the sensor module and the edge computing module, and the industrial control computer is electrically connected to the trapping module, the image acquisition module, and the sensor module. Furthermore, the edge computing module is the one obtained by the edge computing module construction method for insect monitoring described in Embodiment 1, used to extract insect information from insect images. This embodiment uses an energy module composed of a solar panel and a battery to power the system, enabling 24-hour unattended operation. Figure 4 This is a schematic diagram of the overall system architecture, showing the composition and data flow of the hardware layer, edge computing layer, large model generation layer, and host computer platform.

[0041] In this intelligent insect monitoring system, the trapping module is located at the top of the collection chamber to attract insects into the chamber. The trapping module includes an insect-blocking plate, an insect-attracting lamp, and a fan. The insect-attracting lamp is located on the insect-blocking plate to generate insect-attracting light; the fan is located in the insect descent channel between the insect-blocking plate and the collection chamber to create an airflow that guides the insects into the collection chamber.

[0042] See Figure 4In this intelligent insect monitoring system, the sensor module is used to collect environmental information including light, temperature, humidity, and rainfall. The sensor module used in this embodiment includes a raindrop sensor, a photosensor, a light sensor, a temperature sensor, a humidity sensor, and a light curtain sensor.

[0043] See Figure 4 In this intelligent insect monitoring system, an image acquisition module is installed inside the acquisition chamber to acquire images of insects. In this embodiment, an industrial-grade camera is used as the image acquisition module to achieve high-resolution imaging when pests fall into the acquisition chamber.

[0044] See Figure 5 To improve the environmental adaptability, data acquisition accuracy, and intelligence level of this intelligent insect monitoring system, this embodiment proposes an industrial control computer scheme to control the operation of the trapping module, image acquisition module, and sensor module based on environmental information; such as Figure 5 As shown, in this embodiment, the industrial control computer is configured to perform the following steps: A1: Read the current environmental information; A2: Determine whether the trigger command requirements of the rain sensor are met based on the environmental information. If yes, shut down the host of the industrial control computer; otherwise, proceed to the next step; A3: Determine whether the trigger command requirements of the photosensitive sensor are met based on the environmental information. If yes, proceed to the next step; otherwise, shut down the host of the industrial control computer; A4: Start the trapping module, i.e., start the insect-attracting lamp and the fan; A5: Determine whether the trigger command requirements of the light curtain sensor are met based on the environmental information. If yes, proceed to the next step; otherwise, return to step A1; A6: Call the image acquisition module to acquire insect images.

[0045] See Figure 6 In this intelligent pest monitoring system, terminal devices are embedded with a large language model to generate pest analysis reports based on environmental and pest information. These reports cover pest overview, influencing factor analysis, pest severity assessment, short-term development trends, and control recommendations. Figure 6 As shown, the large language model in this embodiment is the DeepSeek-R1-1.5B model.

[0046] See Figure 7The following further discloses the usage method of the intelligent insect monitoring system based on self-distillation and large model generation in this embodiment, which includes the following steps: T1: When the light intensity is below the threshold, insects are attracted into the collection chamber through the trapping module, and insect images are acquired through the image acquisition module; T2: Environmental information including light intensity, temperature, humidity, and rainfall is collected through the sensor module; T3: Insect information is extracted from the insect images through the edge computing module; T4: An insect analysis report is generated by the terminal device based on the environmental information and the insect information, covering the insect situation overview, influencing factor analysis, insect situation level assessment, short-term development trend, and control recommendations.

[0047] Experiments show that the intelligent insect monitoring system disclosed in this embodiment has the following beneficial effects: First, the detection accuracy has been greatly improved.

[0048] Through CMLD distillation, YOLOv8n, as the student network, achieved an mAP of 53.0% and an F1 score of 69.3% on the Pest112 dataset, with performance approaching that of YOLOv8x (53.5%), while the computational cost was only 1 / 30 of that of YOLOv8x, which is significantly better than traditional lightweight methods.

[0049] Second, the detection rate of minute pests has increased.

[0050] Thanks to the multi-view synthesis and cross-view distillation of the student network, the system reduces the false alarm rate of tiny pests such as fruit flies and aphids to below 5%, and significantly reduces the false alarm rate.

[0051] Third, it has end-to-end intelligent analysis capabilities.

[0052] By embedding a large language model, this system not only outputs detection boxes but also automatically generates detailed reports containing pest trends, influencing factors, and control recommendations, lowering the barrier to interpretation for users and improving decision-making efficiency.

[0053] Fourth, it has strong environmental adaptability.

[0054] The system's hardware meets IP66 protection standards and can operate continuously at temperatures ranging from -20℃ to 60℃ and humidity levels below 95%. The embedded student network and large language model are robust to changes in lighting and background, making them suitable for various scenarios such as farmland, orchards, and tea gardens.

[0055] Fifth, reduce network dependence and operating costs.

[0056] The edge computing module of this system can complete core detection and analysis, requiring only the upload of a small amount of key data, making it suitable for areas with weak network coverage. The prevention and control suggestions generated by the system can guide precise pesticide application, and experimental results show that it can reduce pesticide usage by about 35%.

[0057] In the description of the embodiments of this application, it should be noted that the terms "inner" and "outer" and other terms indicating direction or positional relationship are based on the direction or positional relationship shown in the drawings. This is only for the convenience of description and does not indicate or imply that the device or component must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this application.

[0058] In the description of this application, the references to terms such as "an embodiment," "some embodiments," "in this embodiment," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0059] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for constructing an edge computing module for insect infestation monitoring, characterized in that, Includes the following steps: S1: Construct image-insect-labeled image pairs with multiple insect labels, and perform data augmentation on these image pairs to obtain a multi-view dataset; S2: Construct a teacher network and optimize its parameters on the multi-view dataset using supervised learning so that it can extract pest information from images, including pest category, quantity, confidence level, and location. S3: Construct a student network and optimize its parameters using a multi-level distillation method guided by the teacher network, so that it has the ability to extract the insect information from the image, just like the teacher network after parameter optimization. S4: Embed the optimized student network into hardware to obtain an edge computing module for extracting the insect information from the image.

2. The method for constructing an edge computing module for insect infestation monitoring according to claim 1, characterized in that, Step S1 includes the following steps: S11: Collect multiple insect images and label them with insect information to obtain image-insect tag image pairs; S12: Use a masked controlled diffusion model to perform data augmentation on each sample of the image-insect label image pair to obtain the multi-view dataset.

3. The method for constructing an edge computing module for insect monitoring according to claim 2, characterized in that, Step S3 includes the following steps: S31: Construct a deep feature distillation loss function, a channel relation matrix distillation loss function, and a detection head output distillation loss function, with the output of the parameter-optimized teacher network and the output of the student network as variables; S32: Summate the deep feature distillation loss function, the channel relation matrix distillation loss function, and the detection head output distillation loss function to obtain the cross-view consistency distillation loss function; S33: The cross-view consistency distillation loss function is linearly combined with the same-view distillation loss function and the target detection network loss function to obtain the total loss function; S34: Find the optimal network parameters of the student network that make the total loss function take the best value, and use the optimal network parameters as the network parameters of the student network to complete the parameter optimization of the student network.

4. An intelligent insect infestation monitoring system based on self-distillation and large-scale model generation, characterized in that, include: Collection chamber; A trapping module, located at the top of the collection chamber, is used to attract insects into the collection chamber. The sensor module is used to collect environmental information including light intensity, temperature, humidity, and rainfall. An image acquisition module, installed inside the acquisition chamber, is used to acquire images of insects; An edge computing module is electrically connected to the image acquisition module and is the edge computing module obtained by the edge computing module construction method for insect monitoring as described in any one of claims 1-3, used to extract the insect information from the insect image; The terminal device is electrically connected to both the sensor module and the edge computing module, and is embedded with a large language model to generate an insect infestation analysis report based on the environmental information and the situation information. This report covers an overview of the insect infestation, analysis of influencing factors, assessment of the insect infestation level, short-term development trends, and control recommendations.

5. The intelligent insect monitoring system based on self-distillation and large model generation according to claim 4, characterized in that, The trapping module includes: Insect barrier; An insect-attracting lamp, installed on the insect-blocking plate, is used to generate insect-attracting light; A fan is used to create an airflow that introduces the insects into the collection chamber.

6. The intelligent insect monitoring system based on self-distillation and large model generation according to claim 5, characterized in that, The sensor module includes a raindrop sensor, a photosensor, a light sensor, a temperature sensor, a humidity sensor, and a light curtain sensor.

7. The intelligent insect monitoring system based on self-distillation and large model generation according to any one of claims 4-6, characterized in that, It also includes an industrial control computer, which is electrically connected to the trapping module, the image acquisition module, and the sensor module, and is used to control the operation of the trapping module, the image acquisition module, and the sensor module according to the environmental information; The intelligent insect monitoring system based on self-distillation and large model generation according to claim 7 is characterized in that the industrial control computer is configured to perform the following steps: A1: Reads the current environment information; A2: Determine whether the trigger command requirements of the rain sensor have been met. If so, then shut down the host computer of the industrial control computer; If not, proceed to the next step; A3: Determine whether the trigger command requirements of the photosensor have been met. If so, proceed to the next step; If not, then shut down the host computer of the industrial control computer; A4: Activate the trapping module; A5: Determine whether the trigger command requirements of the light curtain sensor have been met. If so, proceed to the next step; If not, then return to step A1; A6: Call the image acquisition module to acquire insect images.

8. A smart insect infestation monitoring method based on self-distillation and large model generation, characterized in that, The intelligent insect monitoring system based on self-distillation and large model generation as described in any one of claims 4-8 includes the following steps: T1: The insects are attracted into the collection chamber by the trapping module, and images of the insects are acquired by the image acquisition module. T2: Collects environmental information including light intensity, temperature, humidity, and rainfall through sensor modules; T3: Extract information from the insect images using the edge computing module; T4: Generate an insect infestation analysis report based on the environmental information and the situation information using the terminal device. The report includes an overview of the insect infestation, analysis of influencing factors, assessment of the insect infestation level, short-term development trend, and control recommendations.

Images