Intelligent termite AI image recognition method and system and computer storage medium
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-10
- Publication Date
- 2026-03-27
Smart Images
Figure CN121746775A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application provides an intelligent termite AI image recognition method and system and a computer storage medium, and belongs to the technical field of computer vision. BACKGROUND
[0002] Termites are a serious worldwide pest that causes damage to housing construction, water conservancy dams, and forest crops, etc. with the characteristics of concealment, universality, and severity. Traditional termite monitoring and identification mainly relies on manual exploration. Professional personnel find exposed signs of termite activity, such as mud, mud lines, and split flying holes, through visual observation, knocking and listening, and probe detection. This method is highly dependent on the experience and responsibility of the inspectors, and is not only inefficient, but also difficult to detect termite nests hidden in the structure and damage early, often causing damage when discovered. In recent years, with the development of technology, some auxiliary detection methods have emerged, such as using sound and infrared temperature sensors. However, these methods usually have high false positive rates, are easily disturbed by the environment, or cannot accurately judge the termite species and the extent of the damage. At the same time, although simple image recognition technology has been tried to be applied to termite identification, the existing technology is mostly based on a single visible light image, which can only preliminarily identify exposed termite individuals, and has insufficient recognition ability for termite activity traces similar to the surrounding environment in color and texture (such as fine mud roads and decay traces), and is difficult to achieve precise segmentation and quantitative analysis in a complex background. In addition, the existing methods have single functions, and usually can only provide binary judgments of "yes" or "no", lacking the ability to deeply mine and comprehensively analyze key information such as termite species, damage range, and activity trend, and cannot provide comprehensive and reliable basis for prevention and control decisions. Therefore, there is an urgent need in the field for a termite monitoring solution that can achieve early, accurate, and intelligent identification. SUMMARY
[0003] The application provides an intelligent termite AI image recognition method, system, and computer storage medium, which solves the problem of early and accurate identification and intelligent diagnosis of termites and their concealed activity traces in the prior art. The technical solutions adopted are as follows: An intelligent termite AI image recognition method, comprising the following steps: Obtaining raw image data collected by a multispectral image sensor deployed at a monitoring point, the raw image data containing image information of a visible light band and at least one non-visible light band; Performing registration and fusion processing on the raw image data, and performing scale normalization and optical distortion correction on the fused image to construct standardized multi-channel image data for model input; inputting the standardized multi-channel image data into a termite recognition model trained by multi-task learning, the model performing a target detection task and a pixel-level segmentation task in parallel to simultaneously output a first recognition result containing a termite existence judgment, a preliminary species classification, and a location bounding box, and a second recognition result of an accurate pixel-level mask representing a termite activity trace area; generating an integrated diagnostic report based on the relevance analysis of the first recognition result and the second recognition result, the diagnostic report containing at least a termite infestation risk level assessment and a trace distribution map.
[0004] Further, the acquisition of the original image data includes: synchronously acquiring visible light images and thermal infrared images of the same scene by coordinating the acquisition timing of the visible light camera and the infrared thermal imager; The triggering condition includes dynamic temperature anomaly change detection based on a thermal infrared image sequence, and joint triggering judgment combined with motion area analysis in a visible light image.
[0005] Further, the registration and fusion processing of the original image data includes: aligning visible light and non-visible light band images based on a feature point matching algorithm, and using a pyramid transformation-based image fusion method to fuse information of different wavebands into the same multi-channel image; The optical distortion correction uses a calibration model based on internal sensor parameters to perform geometric correction on the fused image.
[0006] Further, the termite recognition model uses an encoder-decoder architecture, in which the encoder shares weights for multi-band feature extraction, and the decoder branch includes a task decoder for target detection and a task decoder for semantic segmentation; the model is trained end-to-end by jointly optimizing a detection loss function and a segmentation loss function.
[0007] Further, the recognition result further contains one or more of the following information in addition to the judgment information of the existence of termites: species classification information of the recognized termites, approximate quantity statistics in the image, and specific types of the recognized activity traces.
[0008] Further, the generation of prompt information according to the recognition result includes: when the recognition result indicates the existence of termites, automatically generating a structured early warning report that integrates the alarm level, image acquisition position coordinates, discovery time, at least one evidence image containing a recognition box and its corresponding recognition confidence, and sending the early warning report to one or more preset receiving endpoints through a specified communication protocol.
[0009] Further, the method further comprises: storing the original image data related to each identification task, the corresponding identification result and the generated prompt information into a history database after association; and based on the data samples accumulated in the history database, triggering the parameter updating and optimization process of the termite identification model regularly or according to specific conditions.
[0010] In a second aspect, the present application further provides an intelligent termite AI image recognition system for any of the intelligent termite AI image recognition methods, comprising: A multispectral image acquisition module is configured to acquire original image data containing visible light band and at least one non-visible light band information through a multispectral image sensor deployed at a monitoring point. An image preprocessing and fusion module is configured to perform registration, fusion, scale normalization and optical distortion correction on the original image data to construct standardized multi-channel image data. A multi-task identification module is configured to input the standardized multi-channel image data into a termite identification model trained by multi-task learning, and perform target detection and pixel-level segmentation tasks in parallel, and output first and second identification results. A diagnosis report generation module is configured to generate an integrated diagnosis report based on the correlation analysis of the first and second identification results.
[0011] Further, the system further comprises: a data storage module configured to store the original image data related to each identification task, the corresponding identification result and the generated prompt information; and a model optimization module configured to trigger the parameter updating and optimization of the termite identification model in the intelligent identification module based on the historical data accumulated in the data storage module regularly or according to specific conditions.
[0012] In a third aspect, the present application further provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the intelligent termite AI image recognition method of any of the aspects.
[0013] The present application has the following advantages: The intelligent termite AI image recognition method, system and computer storage medium provided by the application have the following advantages. First, the multispectral image sensing technology is introduced, and the information of visible light and infrared bands is comprehensively utilized, so that the data dimension and quality of the image information are significantly improved. The system can penetrate part of the surface cover or utilize the unique heat characteristics of the termite colony for detection, thereby fundamentally overcoming the bottleneck of the sudden decrease in recognition ability of the traditional visible light recognition under poor light, complex background or hidden traces, and achieving earlier and more sensitive discovery of termites and their initial activity traces. Second, the deep learning model architecture based on multi-task learning is innovatively adopted. The model extracts common features through a shared encoder, and then processes tasks in parallel through independent target detection and pixel-level segmentation decoders. This design not only optimizes the utilization rate of computing resources, but more importantly, realizes the synchronous completion of accurate positioning, classification of termite targets and accurate contour extraction of their activity traces. The results of the two can be mutually verified and enhanced, thereby greatly improving the accuracy and reliability of the recognition results and avoiding misjudgment or omission that may be caused by a single task model. Third, through spatial correlation analysis of the target detection result and the pixel-level segmentation mask, an integrated diagnostic report containing risk level evaluation and trace distribution map can be intelligently generated. This makes the output result go beyond the simple "yes / no" judgment and is transformed into deep decision support information containing species information, damage range, activity intensity and even development trend prediction, thereby greatly improving the practical value of the result and providing a direct and scientific basis for accurate formulation of prevention and control measures and efficient allocation of resources. Fourth, the technology closed loop constructed from multi-source data acquisition, intelligent processing to continuous learning optimization ensures that the system has good adaptability and evolution ability, can optimize the model performance through the accumulated field data, adapts to the changes of termite morphology and activity characteristics in different regions and different environments, and thus guarantees the effectiveness and stability of long-term deployment of the system, thereby laying a solid technical foundation for realizing the regional and intelligent termite long-term monitoring and early warning network. BRIEF DESCRIPTION OF DRAWINGS
[0014] Figure 1 The flowchart of the intelligent termite AI image recognition method provided by one of the embodiments of the application is shown in the figure. Figure 2 The system block diagram of the intelligent termite AI image recognition system provided by one of the embodiments of the application is shown in the figure. DETAILED DESCRIPTION
[0015] The preferred embodiments of the application are described below with reference to the accompanying drawings, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the application, and are not used to limit the application.
[0016] Reference Figure 1 The intelligent termite AI image recognition method provided by the embodiment of the application includes the following steps: S1, acquiring raw image data collected by a multispectral image sensor deployed at a monitoring point, the raw image data containing image information of a visible light band and at least one non-visible light band; S2, performing registration and fusion processing on the raw image data, and performing scale normalization and optical distortion correction on the fused image to construct standardized multi-channel image data for model input; S3, inputting the standardized multi-channel image data into an ant identification model trained by multi-task learning, the model performing target detection task and pixel-level segmentation task in parallel to simultaneously output first identification results containing ant existence judgment, preliminary classification of species and location bounding box, and second identification results representing precise pixel-level mask of ant activity trace area; S4, generating an integrated diagnostic report based on the relevance analysis of the first identification results and the second identification results, the diagnostic report containing at least ant infestation risk level assessment and trace distribution map.
[0017] It should be noted that in some embodiments, a multispectral image acquisition device is deployed at a monitoring point with high incidence of termites, such as building foundation, garden trees, etc. The device integrates a visible light camera and a near-infrared sensor, and acquires raw image data according to a preset acquisition strategy: the visible light band covers 400-760nm, the near-infrared band selects 900-1000nm, and two types of band images are synchronously acquired under the same scene; the raw image is first registered and aligned by ORB feature point matching algorithm, then fused into a 3-channel image (R channel for visible light, G channel for near-infrared, B channel for detail enhancement superposition of both) by Laplacian pyramid fusion method; after fusion, scale normalization is performed according to 640x640 pixel size, and optical distortion is corrected by a polynomial distortion correction model based on the intrinsic matrix and distortion coefficient of sensor calibration, to obtain standardized multi-channel image data; the data is input into an ant identification model trained by multi-task learning, the model uses ResNet50 as a shared encoder to extract cross-band features, the decoder branch includes YOLOv8 detection head and U-Net segmentation head, and outputs first identification results (ant existence, species category such as Taiwan white ant / black-winged soil ant, location bounding box) and second identification results (ant trail, pixel-level mask of feeding trace) in parallel; the relevance is verified by calculating the overlap rate (IOU) of the bounding box and the mask, if IOU≥0.7, the recognition is confirmed to be valid, and finally a diagnostic report is generated, which contains "low / medium / high" three-level infestation risk level (determined based on the number of termites and the coverage area of traces) and trace distribution heat map.
[0018] Preferably, the acquisition of raw image data comprises: The visible light camera and the infrared thermal imager are coordinated to acquire visible light images and thermal infrared images in the same scene. The trigger condition includes dynamic temperature abnormal change detection based on a thermal infrared image sequence, and combined motion region analysis in a visible light image for joint trigger judgment.
[0019] In some embodiments, the original image data acquisition is realized by coordinating a visible light CMOS camera (resolution 1920x1080) and a non-cooled infrared thermal imager (temperature measurement range -20-60℃), both of which are calibrated and collected in time by a GPIO synchronous trigger signal, ensuring that the time difference of image collection in the same scene is less than or equal to 10 ms, and the collection frequency is set to 10 fps. The trigger condition adopts a double-condition joint judgment: first, the temperature sequence of the continuous 5 frames of images output by the infrared thermal imager is analyzed, the difference between the temperature of each frame of pixels and the ambient baseline temperature is calculated, and if there is an area with abnormal temperature (exceeding the baseline by ±3℃) of 3 consecutive frames and an area of 50 pixels or more, a preliminary collection instruction is triggered; then, the three-frame difference method is used to extract the motion region in the visible light image at the same time, and if the area of the motion region is 30 pixels or more and the overlap rate with the infrared abnormal region is 0.5 or more, the collection is confirmed to be triggered, and the visible light and infrared thermal image data at this moment are stored synchronously, avoiding false collection caused by a single trigger condition.
[0020] Preferably, the registration and fusion processing of the original image data includes: The visible light and non-visible light band images are aligned based on a feature point matching algorithm, and a pyramid transformation-based image fusion method is used to fuse the information of different wave bands into the same multi-channel image. The optical distortion correction uses a calibration model based on the internal parameters of the sensor to perform geometric correction on the fused image.
[0021] In some embodiments, the image registration and fusion processing steps are as follows: first, the ORB feature point matching algorithm is used to extract the key feature points of the visible light and near-infrared images, the Euclidean distance between the feature points is calculated for rough matching, and the RANSAC algorithm is used to eliminate the false matching points to obtain the homography matrix between the images, realizing the pixel-level alignment of the two wave band images. The feature point matching distance calculation formula is: wherein, and are the coordinates of the two feature points, is the feature point scale factor, The formula, which considers coordinate, scale, and rotation information to improve matching accuracy, uses the feature point rotation angle. Image fusion employs a pyramid transformation-based method, constructing a 5-layer Laplacian pyramid. A weighted average fusion rule is applied to each layer of the pyramid (0.6 for visible light and 0.4 for near-infrared light), and then the fused image is obtained through inverse pyramid transformation. Optical distortion correction uses the Zhang Zhengyou calibration method, which pre-obtains the sensor's intrinsic parameter matrix (focal length, principal point coordinates) and distortion coefficients (radial distortion, tangential distortion). Based on the calibration model, coordinate mapping is performed on each pixel of the fused image to correct geometric deviations caused by lens distortion.
[0022] Preferably, the termite identification model adopts an encoder-decoder architecture, wherein the encoder shares weights for multi-band feature extraction, and the decoder branch includes a task decoder for target detection and a task decoder for semantic segmentation; the model is trained end-to-end by jointly optimizing the detection loss function and the segmentation loss function.
[0023] In some embodiments, the termite identification model adopts an integrated encoder-decoder architecture: the encoder uses a ResNet50 network, removing the original fully connected layers and retaining the first four convolutional blocks to extract deep semantic features from multi-band images, and the encoder weights are shared between detection and segmentation tasks to reduce parameter redundancy; the decoder contains two parallel branches: the detection task decoder is based on the YOLOv8 detection head, which outputs target bounding boxes and classification probabilities through adaptive anchor box generation and multi-scale feature fusion; the segmentation task decoder is based on the U-Net architecture, which fuses features from different levels of the encoder through skip connections to output pixel-level semantic segmentation masks; the model training adopts an end-to-end joint optimization strategy, and the loss function is defined as: in, CIoU loss function (used to optimize the localization accuracy of detected bounding boxes). This is the Dice loss function (used to optimize the pixel-level matching degree of segmentation masks). , The weight coefficients are set based on the fact that the detection task has a higher priority than the segmentation task. During training, data augmentation techniques such as random flipping, brightness adjustment, and Gaussian noise addition are used. The batch size is set to 16, and the initial learning rate is 0.001, which is dynamically adjusted through a cosine annealing strategy.
[0024] Preferably, in addition to the information on whether termites are present, the identification result further includes one or more of the following: species classification information of the identified termites, approximate quantity statistics in the image, and specific types of activity traces identified.
[0025] In some embodiments, the identification result expands multi-dimensional information outside the basic judgment: species classification information is obtained through the feature vector output by the model encoder, input into the fully connected classification layer, which contains 3 output nodes (corresponding to Taiwan white ant, black-winged soil white ant, and other white ants), adopts Softmax activation function to output the probability of each category, and takes the maximum probability as the species determination result; the approximate number of white ants is counted based on the boundary box of the first identification result and the segmentation mask of the second identification result, the mask is analyzed for connected domain, each connected domain corresponds to a white ant individual, and the repeated count is removed in combination with the overlap rate of the boundary box to obtain the statistical number; the specific type of activity trace is determined through the morphological and texture features of the segmentation mask: if the mask is long strip-shaped and the texture is continuous, it is determined as “ant trail”; if the mask is irregular patch-shaped and the gray value is significantly different from the surrounding, it is determined as “tunneling trace”; if the mask is scattered small dot-shaped and distributed in clusters, it is determined as “white ant excrement”, and the above information is integrated into the identification result.
[0026] Preferably, the method further comprises: when the identification result indicates the presence of white ants, automatically generating a structured early warning report integrating the alarm level, image acquisition position coordinates, discovery time, at least one evidence image containing the identification box and the corresponding identification confidence, and sending the early warning report to one or more preset receiving endpoints through a specified communication protocol.
[0027] In some embodiments, when the identification result indicates the presence of white ants (confidence ≥ 0.8), a structured early warning report is automatically generated: the report contains the device unique identification ID, the GPS latitude and longitude coordinates of the image acquisition position, the acquisition timestamp (accurate to seconds), the early warning level (corresponding to the risk level of the diagnostic report), 3 key evidence images (original fusion image, detection image with boundary box, segmentation image with mask) and the identification confidence of each image; the report data is packaged in JSON format, containing a CRC32 data verification field (used to verify the transmission integrity); through a specified communication protocol: in a short-distance scene, the LoRa protocol is used to transmit to the on-site gateway, and in a long-distance scene, the MQTT protocol is used to push to the cloud monitoring platform, while sending push notifications to the operation and maintenance personnel's mobile APP according to the preset configuration, the receiving endpoint supports up to 5 preset addresses, the push timeout time is set to 30s, and if it exceeds, it will be automatically retransmitted twice.
[0028] Preferably, the method further comprises: associating and storing the original image data related to each identification task, the corresponding identification result and the generated prompt information into the historical database; based on the data samples accumulated in the historical database, periodically or according to specific conditions, triggering the parameter updating and optimization process of the white ant identification model.
[0029] In some embodiments, during the method execution, the raw image data, standardized image data, first / second recognition result, and pre-warning report of each identification task are stored in the historical database in association with a unique task ID through the local SD card and cloud server dual storage mode; the database is managed by "year-month" partitioning, and each piece of data contains index tags (monitoring point position, collection time, and recognition result type); the model updating and optimization trigger conditions include periodic triggering (automatic start every 30 days) and conditional triggering (start when the database adds more than 2000 effective samples, or the recognition error rate is greater than or equal to 5% for three consecutive times); the optimization process adopts a transfer learning strategy, uses the existing trained model as the pre-training weight, loads the new samples in the historical database (divided into training set and validation set in a ratio of 7:3), freezes the weights of the first three convolution blocks of the encoder, only fine-tunes the parameters of the top layer of the encoder and the decoder, the optimizer is selected as SGD, the learning rate is set to 0.0001, and after 100 epochs of iteration, if the recognition accuracy of the validation set improves by more than 3%, the model parameters are updated and the original model is replaced, and the update log (update time, number of new samples, and accuracy change value) is recorded.
[0030] Reference Figure 2 The present application also provides an embodiment, an intelligent termite AI image recognition system for any of the intelligent termite AI image recognition methods, comprising: A multispectral image acquisition module for acquiring raw image data containing visible light band and at least one non-visible light band information through a multispectral image sensor deployed at the monitoring point; An image preprocessing and fusion module for registering, fusing, scaling, and optical distortion correction of the raw image data to construct standardized multi-channel image data; A multi-task recognition module for inputting the standardized multi-channel image data into a termite recognition model trained by multi-task learning, performing target detection and pixel-level segmentation tasks in parallel, and outputting first and second recognition results; An intelligent diagnosis report generation module for generating an integrated diagnosis report based on the correlation analysis of the first and second recognition results.
[0031] The intelligent termite AI image recognition system realizes data interaction through industrial Ethernet, and is overall arranged in an outdoor protection box at a termite monitoring point. The multispectral image acquisition module includes a visible light CMOS camera with a resolution of 1920*1080 and a non-cooled infrared thermal imager with a temperature measurement range of -20-60 DEG C. The two are calibrated and collected in synchronization through a GPIO trigger signal, and the image data is transmitted to the preprocessing module through an HDMI interface. The image preprocessing and fusion module uses an NVIDIA Jetson Xavier NX edge computing unit, which is internally provided with an ORB feature point matching algorithm and a Laplacian pyramid fusion algorithm. The dual-band images are first aligned (the feature point matching threshold is set to 0.7), and then a 3-channel image is generated. Subsequently, the optical distortion is corrected by Zhang Zhengyou calibration method, and finally a standardized multi-channel image with a resolution of 640*640 pixels is output. The multi-task recognition module is arranged in the same edge computing unit and carries a ResNet50+YOLOv8+U-Net integrated model trained by multi-task learning. The model weight file is stored in a local SSD hard disk, and the inference is accelerated through a TensorRT framework. The single-frame image recognition time is less than or equal to 200 ms. The first recognition result includes the existence of termites, species classification, and boundary box, and the second recognition result includes the pixel-level mask of the ant path / termite trace. The intelligent diagnosis report generation module is developed based on the Python Flask framework. The recognition effectiveness is verified by calculating the IOU (threshold set to 0.7) of the boundary box and the mask. The termite number, trace coverage area, and invasion risk level are combined to generate a structured diagnosis report including the risk level, trace distribution heat map, and collection location GPS coordinates. The report can be pushed to the monitoring platform through the HTTP protocol.
[0032] Preferably, the system further comprises a data storage module for storing original image data, corresponding recognition results and generated prompt information related to each recognition task; and a model optimization module for updating and optimizing the termite recognition model in the intelligent recognition module based on historical data accumulated in the data storage module, periodically or triggered according to specific conditions.
[0033] The data storage module and the model optimization module are added, and data intercommunication is realized with the original module through TCP / IP protocol. The data storage module adopts a "local + cloud" dual storage architecture: the local storage selects a 1TB industrial-grade SSD hard disk, stores the original image data, standardized image, recognition result and diagnosis report according to the directory structure of "monitoring point ID-capture timestamp", supports local data query and export; the cloud storage synchronizes the local data to the Ali Cloud OSS object storage through the MQTT protocol, manages the data by zones according to years, sets the data retention period to 5 years, and supports multi-terminal access. The model optimization module is integrated in the cloud server, and the trigger conditions include regular triggering (automatic start every 30 days) and conditional triggering (new effective samples ≥2000 or recognition error rate ≥5%); the optimization process adopts a transfer learning strategy, extracts new samples from the cloud database (divided into training set and validation set according to 7:3), freezes the weights of the first three convolution blocks of the ResNet50 encoder, only fine-tunes the top layer and decoder parameters, and trains using the SGD optimizer (initial learning rate 0.0001, 100 epochs); if the validation set recognition accuracy improves by ≥3%, a new model weight file is generated, pushed to the edge computing unit through OTA, and the original model is automatically replaced and the update log (including update time, sample quantity, accuracy change value) is recorded; at the same time, the model optimization module supports remote configuration of training parameters, and the loss function weight coefficient can be adjusted according to the termite species distribution characteristics of the monitoring point.
[0034] In a third aspect, the present application also provides a computer readable storage medium having stored thereon a computer program, which, when executed by a processor, implements the intelligent termite AI image recognition method according to any one of the preceding aspects.
[0035] Storage medium - any of various types of memory devices or storage devices. The term "storage medium" is intended to include an installation medium, e.g., a CD-ROM, floppy disks, or tape device; a computer system memory or random access memory such as DRAM, DDRRAM, SRAM, EDO RAM, Rambus RAM, etc.; a non-volatile memory such as a magnetic medium (e.g., a hard disk or floppy disk), or optical medium (e.g., a CD-ROM); registers or other similar types of memory elements; or a non-tangible medium such as a web server or other network connection (e.g., especially a network accessible storage facility). The memory can also be a combination of different types of storage media. The storage medium can be external to the first computer system where the program is executed, or it can be within the first computer system. The storage medium can be a separate computer system that is connected to the first computer system through a network, such as the Internet. The second computer system can provide program instructions to the first computer system for execution. The term "storage medium" can include two or more memory devices that reside in different locations, e.g., in different computer systems that are connected through a network. The storage medium can store program instructions that are executable by one or more processors (e.g., as a computer program). The embodiments provided by the present application synchronously acquire visible light and non-visible light band image data by deploying a multi-spectrum image acquisition device, realize efficient integration of cross-band information by combining feature point matching and pyramid fusion technology, effectively avoid the recognition limitations of single-band images through optical distortion correction processing, improve the detection capability of termites and activity traces in complex environments (such as dark and blocked scenes), and lay a high-quality data foundation for subsequent identification tasks. Relying on the integrated model trained by multi-task learning, the shared encoder extracts deep cross-band features, and the target detection and pixel-level segmentation tasks are executed in parallel, which can not only accurately output core information such as the existence of termites, species classification, and location bounding box, but also generate accurate masks of activity traces, while realizing the expansion functions such as quantity statistics and trace type determination, greatly improving the comprehensiveness and accuracy of the identification results, and reducing false positives and omissions. Through the correlation analysis, a structured diagnostic report and a hierarchical early warning mechanism are generated, which integrates key information such as risk level, location coordinates, and evidence images, and pushes them to the preset receiving end through multiple protocols, ensuring that the operation and maintenance personnel can respond quickly and accurately, and timely curb the spread of termite invasion and reduce the damage risk of buildings, gardens and other facilities. The "local + cloud" dual storage architecture ensures data security and traceability, and the model iteration optimization mechanism based on historical data continuously improves the adaptation ability of the model to different termite species and complex scenes through transfer learning, prolonging the service life and application value of the system. The system uses edge computing to accelerate inference, and the single-frame identification time is low. In combination with the modular design and flexible deployment scheme, it is suitable for various monitoring point environments, simplifies the installation and maintenance process, and balances real-time and practicality, providing efficient and intelligent technical support for termite control work, significantly improving the scientificity and forward-looking nature of the control work, and reducing the cost of manpower and time.
[0036] Of course, the storage medium provided by the embodiments of the present application contains computer executable instructions, which are not limited to the intelligent termite AI image recognition method as above, but can also perform related operations in the intelligent termite AI image recognition method provided by any embodiment of the present application. Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and variations.
Claims
1. A smart termite AI image recognition method, characterized in that, Includes the following steps: Acquire raw image data collected by a multispectral image sensor deployed at a monitoring point, the raw image data containing image information in the visible light band and at least one non-visible light band; The original image data is registered and fused, and the fused image is scaled and optical distortion corrected to construct standardized multi-channel image data for model input. The standardized multi-channel image data is input into a termite recognition model trained by multi-task learning. The model performs target detection and pixel-level segmentation tasks in parallel to simultaneously output a first recognition result containing termite presence judgment, preliminary species classification and location bounding box, and a second recognition result containing a precise pixel-level mask representing the region of termite activity traces. Based on the correlation analysis between the first identification result and the second identification result, an integrated diagnostic report is generated, which includes at least a termite infestation risk level assessment and a trace distribution map.
2. The intelligent termite AI image recognition method according to claim 1, characterized in that, The acquisition of raw image data includes: By coordinating the acquisition timing of visible light cameras and infrared thermal imagers, visible light images and thermal infrared images of the same scene can be acquired simultaneously. Dynamic temperature anomaly detection based on thermal infrared image sequences, combined with motion region analysis in visible light images for joint trigger judgment.
3. The intelligent termite AI image recognition method according to claim 1, characterized in that, The registration and fusion processing of the original image data includes: The visible and non-visible light band images are aligned based on the feature point matching algorithm, and the information of different bands is fused into the same multi-channel image using the pyramid transformation-based image fusion method. The optical distortion correction employs a calibration model based on sensor internal parameters to perform geometric correction on the fused image.
4. The intelligent termite AI image recognition method according to claim 1, characterized in that, The termite identification model adopts an encoder-decoder architecture, in which the encoder shares weights for multi-band feature extraction, and the decoder branch includes a task decoder for target detection and a task decoder for semantic segmentation; the model is trained end-to-end by jointly optimizing the detection loss function and the segmentation loss function.
5. The intelligent termite AI image recognition method according to claim 4, characterized in that, The first identification result also includes one or more of the following: species classification information of the identified termites, approximate number statistics in the image, and specific types of activity traces identified.
6. The intelligent termite AI image recognition method according to claim 1, characterized in that, The step of generating a prompt message based on the identification result includes: when the identification result indicates the presence of termites, automatically generating a structured early warning report. This report integrates the alarm level, image acquisition location coordinates, discovery time, at least one evidence image containing an identification box and its corresponding identification confidence level, and sending the early warning report to one or more preset receiving endpoints through a specified communication protocol.
7. The intelligent termite AI image recognition method according to claim 1, characterized in that, Also includes: The original image data, corresponding recognition results, and generated prompts related to each recognition task are associated and stored in the historical database; Based on data samples accumulated in the historical database, the parameters of the termite identification model are updated and optimized periodically or according to specific conditions.
8. An intelligent termite AI image recognition system, used in the intelligent termite AI image recognition method according to any one of claims 1-7, characterized in that, include: A multispectral image acquisition module is used to acquire raw image data containing information of the visible light band and at least one non-visible light band through a multispectral image sensor deployed at the monitoring point. The image preprocessing and fusion module is used to register, fuse, scale normalize and correct optical distortion of the original image data to construct standardized multi-channel image data. The multi-task recognition module is used to input the standardized multi-channel image data into the termite recognition model trained by multi-task learning, perform target detection and pixel-level segmentation tasks in parallel, and output the first recognition result and the second recognition result. intelligent The diagnostic report generation module is used to generate an integrated diagnostic report based on the correlation analysis between the first and second identification results.
9. The intelligent termite AI image recognition system according to claim 8, characterized in that, Also includes: The data storage module is used to store the original image data, corresponding recognition results, and generated prompts related to each recognition task; The model optimization module is used to periodically update and optimize the parameters of the termite identification model in the intelligent identification module based on the historical data accumulated in the data storage module or triggered according to specific conditions.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the intelligent termite AI image recognition method as described in any one of claims 1 to 7.