Forest monitoring method and system based on image recognition

Abstract

The invention belongs to the technical field of recognition monitoring, and particularly relates to a forest monitoring method and system based on image recognition, and the method comprises the steps: setting a forest monitoring region, and building a monitoring network through the remote sensing of an unmanned plane and the communication of the Internet of Things based on a set monitoring point; performing data processing on the historical image data and the real-time image data to obtain processed historical image data and real-time image data; analyzing the processed historical image data, and constructing a forest monitoring level; and training and extracting the processed real-time image data in an image training and extracting mode, comparing the obtained feature data of the real-time image data with the feature data of the historical image data, and adjusting the forest monitoring frequency according to the comparison result and the constructed forest monitoring level. According to the invention, real-time, efficient and accurate monitoring of the forest is realized, the monitoring frequency can be dynamically adjusted according to the comparison result of the real-time image and the historical image, and the efficiency and safety of forest monitoring are significantly improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of identification and monitoring, and in particular relates to a forest monitoring method and system based on image recognition. Background Art

[0002] Current forest monitoring technologies primarily rely on manual inspections, video surveillance, and some rudimentary methods based on image recognition. While manual inspections are intuitive, they are inefficient and struggle to cover large areas. Video surveillance can enable automated monitoring, but its coverage is limited and its adaptability to complex environments is poor. In recent years, forest monitoring technologies based on image recognition have gradually emerged, improving monitoring efficiency through methods such as video frame selection and splicing. However, these methods still have certain limitations.

[0003] However, existing forest monitoring technologies have numerous shortcomings. Manual inspections rely on manpower, are easily limited by natural conditions such as weather and terrain, and are difficult to achieve continuous monitoring. Video surveillance equipment has limited coverage, making comprehensive forest monitoring impossible and prone to misjudgment in complex environments. Furthermore, while image recognition-based monitoring methods (such as CN114863350A) have improved monitoring efficiency to a certain extent, their image data processing capabilities still need improvement, especially in terms of image resolution, noise processing, and feature extraction, making it difficult to meet the rapid response requirements for emergencies such as forest fires. Summary of the Invention

[0004] The purpose of the present invention is to provide a forest monitoring method and system based on image recognition. By building a monitoring network through drone remote sensing and Internet of Things communication, and combining image enhancement, noise reduction, feature extraction and data clustering technologies, it realizes real-time, efficient and accurate monitoring of forests, and can dynamically adjust the monitoring frequency according to the comparison results of real-time images and historical images, significantly improving the efficiency and safety of forest monitoring.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions: The forest monitoring method based on image recognition includes the following steps: S1. Set a forest monitoring area and set up q monitoring points around the forest monitoring area. Based on the set monitoring points, a monitoring network is built through drone remote sensing and Internet of Things communication. S2. Based on the constructed monitoring network, real-time image data and historical image data are collected in the forest monitoring area in real time, and the historical image data and the real-time image data are processed to obtain processed historical image data and real-time image data; S3. Analyze the processed historical image data and construct a forest monitoring level. The process includes: S3.1. Perform training extraction on the processed historical image data using an image training extraction method to obtain feature data of the historical image data; S3.2. Classify the feature data of the historical image data by data clustering to obtain classified historical image data, and construct a forest monitoring level based on the classified historical image data; S4. Perform training extraction on the processed real-time image data through image training extraction, compare the feature data of the real-time image data with the feature data of the historical image data, and adjust the forest monitoring frequency according to the comparison result and the constructed forest monitoring level.

[0006] Preferably, in S1, building a monitoring network through drone remote sensing and Internet of Things communication includes the following steps: S1.1. Based on the Q monitoring points set up around the forest monitoring area, install a 3D laser imaging device at each monitoring point, and collect image data within the forest monitoring area using the installed 3D laser imaging device; Based on the set forest monitoring area, operate the drone to collect image data within the forest monitoring area; Number the q monitoring points and drones , set the number Number the drone. Number the q monitoring points; S1.2. Set monitoring points The coordinates of is the origin of coordinates, according to Coordinates establish a three-dimensional monitoring coordinate system based on monitoring points Update and confirm coordinates of other monitoring points and drones; S1.3. Save the coordinates of each monitoring point and build a monitoring network based on the connection status of the three-dimensional laser shooting equipment and the drone, wherein each monitoring point is set to be connected to the drone separately.

[0007] Preferably, in said S2, obtaining the processed historical image data and real-time image data comprises the following steps: S2.1. Processing the acquired historical image data by a data processing method to obtain processed historical image data; S2.2. Process the collected real-time image data through a data processing method to obtain processed real-time image data.

[0008] Preferably, in said S2.1, the process of obtaining the processed historical image data is: Enhance the acquired historical image data through image enhancement algorithm; Based on the grayscale value distribution in the historical image data, the pixel probability distribution of the acquired historical image data is calculated: ; Among them, P represents the pixel probability of the acquired historical image data, Represents the number of pixels with grayscale value u, N is the total number of pixels; Construct the image cumulative distribution function based on the calculated pixel probability distribution of the historical image data after acquisition: ; Where L represents the total number of grayscale values ​​in the image, Represents the gray value of the rth gray value after being mapped by the cumulative distribution function; Based on the constructed image cumulative distribution function, the acquired historical image data is enhanced through data mapping; The historical image data after image enhancement is set as the processed historical image data.

[0009] Preferably, in said S2.2, the process of obtaining the processed real-time image data is: The noise reduction algorithm is used to reduce the noise of the collected real-time image data, which can be expressed as: ; in, is a noisy image, is the image after denoising, is the noise on the image; Set the number of image frames to M, and synthesize the real-time image data after noise reduction through the multi-frame averaging algorithm. The synthetic image formula is shown below: ; in, Indicates the coordinate position of the pixel point, is the grayscale mean of the pixel after synthesis, The lth frame image is at position Gray value of For an ideal noise-free image at position The gray value below The lth frame image is at position The grayscale value of the noise under ; The real-time image data after multi-frame averaging is set as the processed real-time image data.

[0010] Preferably, in S3.1, obtaining feature data of historical image data includes the following steps: S3.1.1. Evenly divide the processed historical image data into m image data blocks; S3.1.2. Input the m divided image data blocks into the convolutional neural network. After the convolutional neural network receives the input image data blocks, the convolution layer in the convolutional neural network performs convolution calculations on the input image data blocks based on the set convolution kernel size, convolution step size, and initial weights, extracts the corresponding eigenvalues, and performs multiplication and accumulation. The convolution calculation formula is as follows: ; Where I represents the input image data block, Q represents the weight of the corresponding convolution kernel, b represents the bias value, and F represents the eigenvalue obtained by convolution calculation; S3.1.3. Assume that the convolutional layers and pooling layers in the convolutional neural network are stacked continuously. After the initial extraction of the convolutional layer is completed, the obtained feature values ​​are processed by the activation function and input into the pooling layer; S3.1.4. Optimize the features of the image data blocks using the SSE algorithm during the stacked convolution and pooling process. S3.1.5. Pass the features of the optimized image data block into the fully connected layer of the convolutional neural network; S3.1.6. Expand and combine the feature data through the fully connected layer to obtain a feature array and save it; S3.1.7. Set the feature array obtained from each historical image data as the feature data of the historical image data.

[0011] Preferably, in S3.1.4, optimizing the features of the image data block by using the SSE algorithm includes the following steps: The SSE algorithm performs 1×1 convolution on the features of the image data block to reduce the channel dimension to one dimension, and then passes Function, which normalizes the data and multiplies it with the features of the original image data block to complete the screening of the feature space information of the image data block; The calculation formula of the SSE algorithm is as follows: ; The function is expressed as: ; Among them, the characteristics of the input image data block are , is the output after pooling the input features, is the kth input feature in Y, J is the number of input features, H and W are the height and width of the image data block respectively, 、 Represents the coordinates of the image data block features within the convolutional neural network channel, Indicates the image data block features in The eigenvalue at for Function, y represents The input of the function, e is a natural constant.

[0012] Preferably, in S3.2, constructing the forest monitoring level based on the classified historical image data includes the following steps: Randomly select c feature arrays from the saved feature array as cluster centers, and each group of cluster centers represents a type; Calculate the distance between other saved feature arrays and the cluster center, and add other feature arrays to the nearest cluster center based on the calculated distance until all feature arrays are classified; After the classification is completed, the cluster center of each type is recalculated by the mean algorithm, and iterative clustering is performed based on the recalculated cluster center until the classification of all feature arrays converges, and the classified historical image data is output; Experts score the classified historical image data and construct a forest monitoring level based on the scoring results.

[0013] Preferably, in said S4, adjusting the forest monitoring frequency according to the comparison result and the constructed forest monitoring level comprises the following steps: The feature data of the real-time image data is compared with the feature data of the corresponding classified historical image data by similarity comparison. The similarity comparison formula is as follows: ; in, The similarity between the feature data D of the real-time image data and the feature data E of the corresponding classified historical image data; Set a similarity threshold. When the calculation result is greater than or equal to the set threshold, it means that the feature data of the real-time image data is similar to the feature data of the corresponding classified historical image data. When the calculation result is less than the set threshold, it means that the feature data of the real-time image data and the corresponding classified historical image data are not similar; The scores of the corresponding classified historical image data that are similar to the real-time image data are used as the scores of the real-time image data, and the forest monitoring frequency is adjusted according to the forest monitoring level corresponding to the score.

[0014] The forest monitoring system based on image recognition is used to implement the above method, including: Data acquisition module, used to collect real-time image data and historical image data; A data processing module is used to process the collected real-time image data and historical image data; A data analysis module is used to analyze the processed real-time image data and historical image data; A comparison module is used to compare the analyzed real-time image data with the historical image data; The monitoring adjustment module is used to adjust the monitoring frequency according to the comparison results.

[0015] The beneficial effects of the present invention are: The present invention sets a forest monitoring area and sets q monitoring points around the set forest monitoring area. At the same time, a monitoring network is constructed based on the set monitoring points through drone remote sensing and Internet of Things communication. After the construction is completed, real-time image data and historical image data are collected and processed in real time. After the processing is completed, the processed real-time image data and historical image data are analyzed and feature extracted by image analysis. Finally, the feature data of the real-time image data is compared with the feature data of the historical image data. The forest monitoring frequency is adjusted according to the comparison result, thereby improving the frequency of forest monitoring.

[0016] The present invention solves the problems of single video monitoring range and insufficient monitoring range by using drone remote sensing and Internet of Things communication to build a monitoring network, and improves the security of forest monitoring.

[0017] The present invention processes the collected real-time image data and historical image data separately. First, the collected historical image data is processed by image enhancement, thereby improving the accuracy of historical image data analysis. Then, the collected real-time image data is processed by noise reduction and multi-frame synthesis, thereby improving the real-time performance of forest monitoring and meeting the needs of rapid response to emergencies such as forest fires.

[0018] The present invention trains and extracts processed historical image data through an image training extraction method, and classifies the feature data of the historical image data through a data clustering method after the extraction is completed; after the classification is completed, a forest monitoring level is constructed based on the classified historical image data, thereby improving the rationality of forest monitoring.

[0019] The present invention compares the feature data of real-time image data with the feature data of historical image data through a similarity comparison method, thereby improving the accuracy of real-time image analysis and helping to more accurately detect early signs of forest fires or illegal intrusions, thereby better achieving the purpose of forest monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 Schematic diagram of the method flow of the present invention; Figure 2 Schematic diagram of the system module of the present invention. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the present invention are described clearly and completely below with reference to the accompanying drawings.

[0022] Example 1: Figure 1 As shown, the present invention provides a forest monitoring method based on image recognition, comprising: S1. Set a forest monitoring area and set up q monitoring points around the forest monitoring area. Based on the set monitoring points, a monitoring network is built through drone remote sensing and Internet of Things communication. S2. Based on the constructed monitoring network, real-time image data and historical image data are collected in the forest monitoring area in real time, and the historical image data and the real-time image data are processed to obtain processed historical image data and real-time image data; S3. Analyze the processed historical image data and construct a forest monitoring level. The process includes: S3.1. Perform training extraction on the processed historical image data using an image training extraction method to obtain feature data of the historical image data; S3.2. Classify the feature data of the historical image data by data clustering to obtain classified historical image data, and construct a forest monitoring level based on the classified historical image data; S4. Perform training extraction on the processed real-time image data through image training extraction, compare the feature data of the real-time image data with the feature data of the historical image data, and adjust the forest monitoring frequency according to the comparison result and the constructed forest monitoring level.

[0023] In S1, building a monitoring network using drone remote sensing and IoT communication includes the following steps: S1.1. Based on the Q monitoring points set up around the forest monitoring area, install a 3D laser imaging device at each monitoring point, and collect image data within the forest monitoring area using the installed 3D laser imaging device; Based on the set forest monitoring area, operate the drone to collect image data within the forest monitoring area; Number the q monitoring points and drones , set the number Number the drone. Number the q monitoring points; S1.2. Set monitoring points The coordinates of is the origin of coordinates, according to Coordinates establish a three-dimensional monitoring coordinate system based on monitoring points Update and confirm coordinates of other monitoring points and drones; S1.3. Save the coordinates of each monitoring point and build a monitoring network based on the connection status of the three-dimensional laser shooting equipment and the drone, wherein each monitoring point is set to be connected to the drone separately.

[0024] In S2, obtaining the processed historical image data and real-time image data includes the following steps: S2.1. Process the acquired historical image data through a data processing method to obtain processed historical image data: Enhance the acquired historical image data through image enhancement algorithm; Based on the grayscale value distribution in the historical image data, the pixel probability distribution of the acquired historical image data is calculated: ; Among them, P represents the pixel probability of the acquired historical image data, Represents the number of pixels with grayscale value u, N is the total number of pixels; Construct the image cumulative distribution function based on the calculated pixel probability distribution of the historical image data after acquisition: ; Where L represents the total number of grayscale values ​​in the image, Represents the gray value of the rth gray value after being mapped by the cumulative distribution function; Based on the constructed image cumulative distribution function, the acquired historical image data is enhanced through data mapping; The historical image data after image enhancement is set as the processed historical image data.

[0025] S2.2. Processing the collected real-time image data by a data processing method to obtain processed real-time image data: The noise reduction algorithm is used to reduce the noise of the collected real-time image data, which can be expressed as: ; in, is a noisy image, is the image after denoising, is the noise on the image; Set the number of image frames to M, and synthesize the real-time image data after noise reduction through the multi-frame averaging algorithm. The synthetic image formula is shown below: ; in, Indicates the coordinate position of the pixel point, is the grayscale mean of the pixel after synthesis, The lth frame image is at position Gray value of For an ideal noise-free image at position The gray value below The lth frame image is at position The grayscale value of the noise under ; The real-time image data after multi-frame averaging is set as the processed real-time image data.

[0026] In S3.1, obtaining feature data of historical image data includes the following steps: S3.1.1. Evenly divide the processed historical image data into m image data blocks; S3.1.2. Input the m divided image data blocks into the convolutional neural network. After the convolutional neural network receives the input image data blocks, the convolution layer in the convolutional neural network performs convolution calculations on the input image data blocks based on the set convolution kernel size, convolution step size, and initial weights, extracts the corresponding eigenvalues, and performs multiplication and accumulation. The convolution calculation formula is as follows: ; Where I represents the input image data block, Q represents the weight of the corresponding convolution kernel, b represents the bias value, and F represents the eigenvalue obtained by convolution calculation; S3.1.3. Assume that the convolutional layers and pooling layers in the convolutional neural network are stacked continuously. After the initial extraction of the convolutional layer is completed, the obtained feature values ​​are processed by the activation function and input into the pooling layer; Set the feature value after the convolution layer extraction is completed to , V is the number of eigenvalues, each eigenvalue The corresponding weight is ; The extracted eigenvalues ​​are processed through the activation function to obtain the processed eigenvalues; The activation function processing formula is as follows: ; in, is the corresponding activation function, and y is the processed eigenvalue; S3.1.4. Optimize the features of the image data blocks using the SSE algorithm during the stacked convolution and pooling process: The SSE algorithm performs 1×1 convolution on the features of the image data block to reduce the channel dimension to one dimension, and then passes Function, which normalizes the data and multiplies it with the features of the original image data block to complete the screening of the feature space information of the image data block; The calculation formula of the SSE algorithm is as follows: ; The function is expressed as: ; Among them, the characteristics of the input image data block are , is the output after pooling the input features, is the kth input feature in Y, J is the number of input features, H and W are the height and width of the image data block respectively, 、 Represents the coordinates of the image data block features within the convolutional neural network channel, Indicates the image data block features in The eigenvalue at for Function, y represents The input of the function, e is a natural constant; S3.1.5. Pass the features of the optimized image data block into the fully connected layer of the convolutional neural network; S3.1.6. Expand and combine the feature data through the fully connected layer to obtain a feature array and save it; S3.1.7. Set the feature array obtained from each historical image data as the feature data of the historical image data.

[0027] In S3.2, constructing forest monitoring levels based on classified historical image data includes the following steps: Randomly select c feature arrays from the saved feature array as cluster centers, and each group of cluster centers represents a type; Calculate the distance between other saved feature arrays and the cluster center, and add other feature arrays to the nearest cluster center based on the calculated distance until all feature arrays are classified; After the classification is completed, the cluster center of each type is recalculated by the mean algorithm, and iterative clustering is performed based on the recalculated cluster center until the classification of all feature arrays converges, and the classified historical image data is output; Experts score the classified historical image data and construct a forest monitoring level based on the scoring results.

[0028] For example, the scoring range is set to 0-100, with level 1 being 80-100, marked as excellent, level 2 being 70-80, marked as good, level 3 being 60-70, marked as passing, and level 4 being 0-60, marked as failing; The monitoring frequency of drones at level 1 is set to once every 24 hours, the monitoring frequency of drones at level 2 is set to once every 12 hours, the monitoring frequency of drones at level 3 is set to once every 6 hours, and the monitoring frequency of drones at level 4 is set to once every 2 hours.

[0029] In S4, adjusting the forest monitoring frequency based on the comparison results and the constructed forest monitoring level includes the following steps: The feature data of the real-time image data is compared with the feature data of the corresponding classified historical image data by similarity comparison. The similarity comparison formula is as follows: ; in, The similarity between the feature data D of the real-time image data and the feature data E of the corresponding classified historical image data; Set a similarity threshold. When the calculation result is greater than or equal to the set threshold, it means that the feature data of the real-time image data is similar to the feature data of the corresponding classified historical image data. When the calculation result is less than the set threshold, it means that the feature data of the real-time image data and the corresponding classified historical image data are not similar, and the real-time monitoring and comparison calculation will continue; The scores of the corresponding classified historical image data that are similar to the real-time image data are used as the scores of the real-time image data, and the forest monitoring frequency is adjusted according to the forest monitoring level corresponding to the score.

[0030] Example 2: This example makes the following improvements based on Example 1: First, a light intensity sensor is set up in the monitoring area. Then, in S2.1, before performing image enhancement processing on the historical image data, the light intensity sensor is used to obtain historical light intensity information in the forest monitoring area. Dynamically adjust the contrast stretching parameters in the image enhancement algorithm based on historical light intensity information. When the light intensity is low, increase the contrast stretching intensity to enhance the image brightness and details; when the light intensity is high, reduce the contrast stretching intensity appropriately to avoid overexposure. An improved histogram equalization algorithm is introduced, combining light intensity information to perform segmented equalization on the image's histogram. The image's grayscale value range is divided into multiple intervals, and the equalization parameters of each interval are dynamically adjusted according to the light intensity to better preserve the details of the dark and bright areas in the image. During image enhancement, the image is segmented and processed based on the illumination changes in the local area. A more aggressive enhancement strategy is used for areas with large illumination changes, while a more moderate enhancement strategy is used for areas with small illumination changes, achieving global and local adaptive enhancement.

[0031] For real-time image data, the same image enhancement processing can also be performed based on the real-time light intensity information before noise reduction to improve image quality.

[0032] Example 3: This example makes the following improvements based on Example 1: In S2.2, when performing noise reduction processing on the collected real-time image data, not only a noise reduction algorithm is used, but also an adaptive filtering algorithm is used to process the real-time image data. The specific steps are as follows: Acquire noise intensity information of real-time image data and dynamically adjust the parameters of the noise reduction algorithm according to the noise intensity to meet the image noise reduction requirements under different noise intensities; At the same time, the real-time image data is further processed based on an adaptive filtering algorithm whose filter parameters are adaptively adjusted according to the local features of the image to better preserve image details and remove noise; The real-time image data processed by the noise reduction algorithm and the adaptive filtering algorithm are fused to obtain the final processed real-time image data.

[0033] This improved method can more effectively remove noise from real-time image data while better preserving image detail information, thereby improving the accuracy and reliability of subsequent forest monitoring based on image recognition.

[0034] Example 4: Figure 2 As shown, the forest monitoring system based on image recognition is used to implement the methods in Examples 1 to 3, including: Data acquisition module, used to collect real-time image data and historical image data; A data processing module is used to process the collected real-time image data and historical image data; A data analysis module is used to analyze the processed real-time image data and historical image data; A comparison module is used to compare the analyzed real-time image data with the historical image data; The monitoring adjustment module is used to adjust the monitoring frequency according to the comparison results.

[0035] Example 5: A forest monitoring device based on image recognition includes a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The methods in Examples 1 to 3 are implemented by executing the computer program by the processor.

Claims

1. A forest monitoring method based on image recognition, characterized in that: The following steps are involved: S1. Set a forest monitoring area and set up q monitoring points around the forest monitoring area. Based on the set monitoring points, a monitoring network is built through drone remote sensing and Internet of Things communication. S2. Based on the constructed monitoring network, real-time image data and historical image data are collected in the forest monitoring area in real time, and the historical image data and the real-time image data are processed to obtain processed historical image data and real-time image data; S3. Analyze the processed historical image data and construct a forest monitoring level. The process includes: S3.

1. Perform training extraction on the processed historical image data using an image training extraction method to obtain feature data of the historical image data; S3.

2. Classify the feature data of the historical image data by data clustering to obtain classified historical image data, and construct a forest monitoring level based on the classified historical image data; S4. Perform training extraction on the processed real-time image data through image training extraction, compare the feature data of the real-time image data with the feature data of the historical image data, and adjust the forest monitoring frequency according to the comparison result and the constructed forest monitoring level.

2. The forest monitoring method based on image recognition according to claim 1, characterized in that: In S1, building a monitoring network through drone remote sensing and IoT communication includes the following steps: S1.

1. Based on the Q monitoring points set up around the forest monitoring area, install a 3D laser imaging device at each monitoring point, and collect image data within the forest monitoring area using the installed 3D laser imaging device; Based on the set forest monitoring area, operate the drone to collect image data within the forest monitoring area; Number the q monitoring points and drones , set the number Number the drone. Number the q monitoring points; S1.

2. Set monitoring points The coordinates of is the origin of coordinates, according to Coordinates establish a three-dimensional monitoring coordinate system based on monitoring points Update and confirm coordinates of other monitoring points and drones; S1.

3. Save the coordinates of each monitoring point and build a monitoring network based on the connection status of the three-dimensional laser shooting equipment and the drone, wherein each monitoring point is set to be connected to the drone separately.

3. The forest monitoring method based on image recognition according to claim 1, characterized in that: In the above-mentioned S2, obtaining the processed historical image data and real-time image data includes the following steps: S2.

1. Processing the acquired historical image data by a data processing method to obtain processed historical image data; S2.

2. Process the collected real-time image data through a data processing method to obtain processed real-time image data.

4. The forest monitoring method based on image recognition according to claim 3, characterized in that: In S2.1, the process of obtaining the processed historical image data is as follows: Enhance the acquired historical image data through image enhancement algorithm; Based on the grayscale value distribution in the historical image data, the pixel probability distribution of the acquired historical image data is calculated: ； Among them, P represents the pixel probability of the acquired historical image data, Represents the number of pixels with grayscale value u, N is the total number of pixels; Construct the image cumulative distribution function based on the calculated pixel probability distribution of the historical image data after acquisition: ； Where L represents the total number of grayscale values ​​in the image, Represents the gray value of the rth gray value after being mapped by the cumulative distribution function; Based on the constructed image cumulative distribution function, the acquired historical image data is enhanced through data mapping; The historical image data after image enhancement is set as the processed historical image data.

5. The forest monitoring method based on image recognition according to claim 3, characterized in that: In S2.2, the process of obtaining the processed real-time image data is as follows: The noise reduction algorithm is used to reduce the noise of the collected real-time image data, which can be expressed as: ； in, is a noisy image, is the image after denoising, is the noise on the image; Set the number of image frames to M, and synthesize the real-time image data after noise reduction through the multi-frame averaging algorithm. The synthetic image formula is shown below: ； in, Indicates the coordinate position of the pixel point, is the grayscale mean of the pixel after synthesis, The lth frame image is at position Gray value of For an ideal noise-free image at position The gray value below The lth frame image is at position The grayscale value of the noise under ; The real-time image data after multi-frame averaging is set as the processed real-time image data.

6. The forest monitoring method based on image recognition according to claim 1, characterized in that: In S3.1, obtaining feature data of historical image data includes the following steps: S3.1.

1. Evenly divide the processed historical image data into m image data blocks; S3.1.

2. Input the m divided image data blocks into the convolutional neural network. After the convolutional neural network receives the input image data blocks, the convolution layer in the convolutional neural network performs convolution calculations on the input image data blocks based on the set convolution kernel size, convolution step size, and initial weights, extracts the corresponding eigenvalues, and performs multiplication and accumulation. The convolution calculation formula is as follows: ； Where I represents the input image data block, Q represents the weight of the corresponding convolution kernel, b represents the bias value, and F represents the eigenvalue obtained by convolution calculation; S3.1.

3. Assume that the convolutional layers and pooling layers in the convolutional neural network are stacked continuously. After the initial extraction of the convolutional layer is completed, the obtained feature values ​​are processed by the activation function and input into the pooling layer; S3.1.

4. Optimize the features of the image data blocks using the SSE algorithm during the stacked convolution and pooling process. S3.1.

5. Pass the features of the optimized image data block into the fully connected layer of the convolutional neural network; S3.1.

6. Expand and combine the feature data through the fully connected layer to obtain a feature array and save it; S3.1.

7. Set the feature array obtained from each historical image data as the feature data of the historical image data.

7. The forest monitoring method based on image recognition according to claim 6, characterized in that: In S3.1.4, optimizing the features of the image data block using the SSE algorithm includes the following steps: The SSE algorithm performs 1×1 convolution on the features of the image data block to reduce the channel dimension to one dimension, and then passes Function, which normalizes the data and multiplies it with the features of the original image data block to complete the screening of the feature space information of the image data block; The calculation formula of the SSE algorithm is as follows: ； The function is expressed as: ； Among them, the characteristics of the input image data block are , is the output after pooling the input features, is the kth input feature in Y, J is the number of input features, H and W are the height and width of the image data block respectively, 、 Represents the coordinates of the image data block features within the convolutional neural network channel, Indicates the image data block features in The eigenvalue at for Function, y represents The input of the function, e is a natural constant.

8. The forest monitoring method based on image recognition according to claim 6, characterized in that: In S3.2, constructing the forest monitoring level based on the classified historical image data includes the following steps: Randomly select c feature arrays from the saved feature array as cluster centers, and each group of cluster centers represents a type; Calculate the distance between other saved feature arrays and the cluster center, and add other feature arrays to the nearest cluster center based on the calculated distance until all feature arrays are classified; After the classification is completed, the cluster center of each type is recalculated by the mean algorithm, and iterative clustering is performed based on the recalculated cluster center until the classification of all feature arrays converges, and the classified historical image data is output; Experts score the classified historical image data and construct a forest monitoring level based on the scoring results.

9. The forest monitoring method based on image recognition according to claim 8, characterized in that: In S4, adjusting the forest monitoring frequency according to the comparison result and the established forest monitoring level includes the following steps: The feature data of the real-time image data is compared with the feature data of the corresponding classified historical image data by similarity comparison. The similarity comparison formula is as follows: ； in, The similarity between the feature data D of the real-time image data and the feature data E of the corresponding classified historical image data; Set a similarity threshold. When the calculation result is greater than or equal to the set threshold, it means that the feature data of the real-time image data is similar to the feature data of the corresponding classified historical image data. When the calculation result is less than the set threshold, it means that the feature data of the real-time image data and the corresponding classified historical image data are not similar; The scores of the corresponding classified historical image data that are similar to the real-time image data are used as the scores of the real-time image data, and the forest monitoring frequency is adjusted according to the forest monitoring level corresponding to the score.

10. A forest monitoring system based on image recognition, used to implement the method according to any one of claims 1 to 9, characterized in that: include: Data acquisition module, used to collect real-time image data and historical image data; A data processing module is used to process the collected real-time image data and historical image data; A data analysis module is used to analyze the processed real-time image data and historical image data; A comparison module is used to compare the analyzed real-time image data with the historical image data; The monitoring adjustment module is used to adjust the monitoring frequency according to the comparison results.

Images