Deep neural network based group density monitoring system for chouioia clementi

Abstract

The present application relates to the technical field of image recognition, in particular to a group density monitoring system for the pine moth based on a deep neural network, which comprises an image acquisition module, a density label generation module, a feature compression path module, a region identification module and a insect group density determination module.In the present application, the accuracy of insect target extraction is improved by identifying the needle structure region and screening the texture color difference image block, the precision of grade determination is enhanced by generating a label atlas in combination with the relationship between the insect arrangement density and the area, the attention of the network to key features is improved by realizing path compression through the feature map channel response contrast activation threshold, the spatial precision of insect group distribution analysis is strengthened by associating the response point positioning with the insect body boundary, the consistency of determination is optimized by constructing a closed-loop feedback mechanism through dynamic correction of the response result and the original label, and the recognition precision is overall improved, the artificial intervention is reduced, and the adaptive ability and the automation level of density monitoring are enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image recognition, and in particular to a cypress moth population density monitoring system based on a deep neural network. BACKGROUND

[0002] The technical field of image recognition relates to an artificial intelligence method for recognizing, classifying, and detecting target objects in images by analyzing and processing image information, and the core matters include image acquisition, feature extraction, target detection and classification, and the commonly used methods include deep learning model training and inference process of convolutional neural networks. Image recognition technology is widely used in face recognition, traffic monitoring, biometric recognition, agricultural pest and disease detection, etc. Among them, traditional cypress moth population density monitoring refers to estimating the number of insect bodies and evaluating the density based on manual visual inspection or image acquisition through traditional image processing methods such as edge detection, template matching, image threshold segmentation, etc. This kind of method relies on manual intervention or uses feature engineering to extract the morphological features of insect bodies for quantity determination, which has a high error rate and a low level of automation. A system for constructing and implementing a convolutional neural network model for image recognition tasks based on a deep neural network is constructed to complete the insect population density determination. The system generally uses image acquisition equipment to obtain forest area pest image data and trains a recognition model through the construction of a convolutional neural network to realize insect population estimation.

[0003] The prior art relies on manual visual judgment or traditional image processing methods to perform insect body quantity estimation in cypress moth image recognition. The methods mainly include edge detection, template matching, threshold segmentation, etc. These methods are easily affected by environmental interference and inconsistency of insect body morphology when dealing with forest area images with rich image features and complex backgrounds, resulting in large errors in the feature extraction process. Since these methods often use fixed templates and rules for judgment, they cannot adapt to the dynamic changes of insect body density over time, season, and regional differences, and have insufficient generalization ability when dealing with diverse image scenarios. For example, when the insect bodies in the image are partially occluded, locally overlapped, or the background color is close to the target insect body, traditional segmentation and edge detection are prone to misjudgment and missed detection, which seriously affects the accurate determination of insect population density. In addition, the traditional processing flow often lacks linkage analysis of the relationship between the target quantity or distribution, and cannot judge the insect population aggregation state from the macro distribution structure. Only relying on individual recognition for density judgment leads to a serious deviation between the label generation and the actual insect situation. Highly dependent on manual adjustment of parameters and judgment results, the system has low operation efficiency and weak automation capability, and is difficult to perform large-scale and high-frequency forest area monitoring tasks, which causes lag in actual pest warning and prevention and control deployment. SUMMARY

[0004] The application aims to solve the problems in the prior art and provides a cypress pine moth population density monitoring system based on a deep neural network.

[0005] To achieve the above-mentioned purpose, the application adopts the following technical scheme, the cypress pine moth population density monitoring system based on a deep neural network comprises:

[0006] The image acquisition module acquires cypress forest monitoring equipment image frames, identifies typical needle leaf structure regions, extracts image blocks with abnormal texture and color features, screens images according to contour continuity and region frequency, and generates a suspected cypress pine moth region set;

[0007] The density label generation module analyzes the matching relationship between insect texture density and region area based on the suspected cypress pine moth region set, judges the aggregation trend, and generates a region insect population density label atlas in combination with the target quantity and image feature comparison results;

[0008] The feature compression path module calls the region label in the region insect population density label atlas, inputs the image into a deep neural network, extracts a feature map channel response, compares the number and amplitude of the channel responses, screens active channels, records path numbers, and generates a compressed path index set;

[0009] The region identification module extracts a compressed path feature map according to the compressed path index set, locates a high response point, and associates it with insect boundary coordinates, analyzes the distribution concentration degree, and generates an insect response distribution map;

[0010] The insect population density determination module calls the response dense region in the response distribution map, compares it with the original grade of the label atlas, identifies the deviation and updates the label after correction, and generates an insect population density monitoring result set.

[0011] As a further scheme of the application, the suspected cypress pine moth region set comprises typical needle leaf structure regions, texture difference image blocks, and color feature abnormal image blocks, the region insect population density label atlas comprises insect texture arrangement density labels, image region area grade labels, and identifiable target quantity density labels, the compressed path index set comprises active channel routing numbers, channel response amplitude threshold records, and feature map channel reservation identifiers, and the insect response distribution map comprises insect high response position coordinates, insect aggregation region morphology boundaries, and insect distribution concentration degree grades, and the insect population density monitoring result set comprises insect population density grade results, label deviation correction results, and insect population spatial distribution change results.

[0012] As a further scheme of the application, the image acquisition module comprises:

[0013] The image frame acquisition submodule acquires a sequence of image frames taken by the cypress forest monitoring device in a continuous time period, numbers and stores the image frames in a difference time period, records the spatial resolution and shooting time label of each image frame, integrates the image frames after sorting according to the time label, and obtains an image frame sequence set;

[0014] The needle leaf structure recognition submodule extracts the RGB color channel data and texture gradient information in each image based on the image frame sequence set, performs color clustering and texture direction clustering calculation on the pixel blocks, judges whether the typical needle leaf structure discrimination rule is met, aggregates the spatial continuous regions and removes isolated spots, and acquires a needle leaf region contour map;

[0015] The difference region screening submodule calls the needle leaf region contour map, performs color difference ratio calculation and texture variance detection on the region, obtains a suspected difference block saliency value in combination with the contour closure index and the region appearance frequency, compares the suspected difference block saliency value with a set region saliency threshold value, extracts the image blocks higher than the region saliency threshold value and marks the spatial position, and obtains a suspected cypress caterpillar region set.

[0016] As a further scheme of the present application, the density label generation module comprises:

[0017] The texture density calculation submodule acquires the texture direction distribution of the insect body in the region image block and the area information of the image block based on the suspected cypress caterpillar region set, counts the number of texture arrangement units of the insect body, performs data processing according to the texture direction information and the corresponding area in the image block, extracts the texture structure arrangement density characteristics of the insect body in the region, integrates the texture density correlation information of all image blocks in the image sequence, and obtains an insect body texture density value set;

[0018] The insect body aggregation judgment submodule calls the insect body texture density value set, extracts the insect body number per unit area index according to the region image area and the number of identifiable insect bodies, and performs aggregation trend analysis on the difference region in combination with the insect body overlap degree and the variation characteristics of the number of insect bodies in the image block in the sequence to obtain an insect body aggregation degree value. The result is compared with a set reference value to generate an aggregation classification label;

[0019] The density map construction submodule calls the aggregation classification label, performs color partition mapping on the region according to the color coding rule corresponding to the difference aggregation level, extracts the coordinate information of the image block and establishes a label layer matrix, fuses the layer and the original image contour, adds the region insect body identification number, and generates a region insect swarm density label map.

[0020] As a further scheme of the present application, the specific calculation formula for the aggregation trend analysis on the difference region is:

[0021] ;

[0022] The operation obtains the insect aggregation degree value, compares the result with the set reference value, and generates an aggregation classification label;

[0023] wherein, represents the insect aggregation degree value, represents the number of insects per unit area of the image block, represents the average number of insects per unit area of the entire region, represents the number of insects per unit area of the image block, represents the number of insects per unit area of the image block, represents the standard deviation of the number of insects in the image block in the image sequence, represents the gradient amplitude of the change in the texture direction of the insects in the image block, represents the gradient amplitude of the change in the texture direction of the insects in the image block, represents the sum of all image blocks in the region.

[0024] As a further scheme of the present application, the feature compression path module comprises:

[0025] The label identification extraction submodule extracts the label number information corresponding to the image block based on the label layer divided in the region insect population density label atlas, identifies the image blocks with repeated numbers in the same region and uniformly identifies them, selects the image block at the center position of the region to establish a number mapping relationship, obtains the corresponding relationship between the label number and the image block number, and generates a region label mapping table;

[0026] The response channel screening submodule calls the region label mapping table, inputs the mapped image block into the neural network structure, extracts the channel response in the layer feature map, detects the response position number and response amplitude in each channel, calculates the response intensity value of the channel, compares the result with the channel activation threshold, screens the channel numbers that meet the conditions, and performs number arrangement to generate an activated channel number set;

[0027] The compression path index generation submodule extracts the path identifier and spatial position label of each channel in the network according to the activated channel number set, establishes a mapping structure of the channel number and the path number, integrates and sorts the channel path information in all image blocks, and generates a compression path index set.

[0028] As a further scheme of the present application, the region identification module comprises:

[0029] The feature map extraction submodule extracts the feature map output content of the corresponding channel according to the path number in the compression path index set, classifies and integrates the layer information under the path according to the image block number, uniformly reorganizes the spatial position index, obtains the structured feature map data, and generates a compression path feature map set;

[0030] The high-response positioning submodule extracts response position and response amplitude information in the channel image based on the compressed path feature map set, identifies response points meeting threshold conditions, locates coordinates in the image space, constructs a complete spatial index structure, aggregates coordinate information in all image blocks, and generates a high-response pixel coordinate set;

[0031] The shape matching analysis submodule extracts boundary layer data of the insect aggregation area according to the high-response pixel coordinate set, matches the spatial relationship between each response point and the corresponding boundary area, counts the number of response points in the boundary area, calculates the response point distribution density in the area, and generates an insect response distribution map.

[0032] As a further scheme of the present application, the insect population density determination module comprises:

[0033] The response area comparison submodule extracts the spatial position and dense area boundary of the response point based on the insect response distribution map, counts the number of response points in the area, combines the original grade mark of the corresponding position in the area insect population density label atlas, completes the spatial pairing operation of the response area and the label grade, obtains the grade comparison result under the position consistent area, and generates a label grade deviation distribution value;

[0034] The label deviation calculation submodule identifies the position range of the deviation according to the label grade deviation distribution value, extracts the number and spatial distribution density of the response points in the corresponding area, judges the difference between the original grade mark and the actual response, calculates the grade adjustment index of the area, and merges the offset, to generate a grade adjustment coefficient matrix;

[0035] The density grade updating submodule updates the grade mark of the area according to the grade adjustment coefficient matrix, calls the original label grade label information, replaces the grade field in all deviation areas, updates the insect population grade mark of the area, records the updated grade value of each position, integrates all adjusted label contents, and generates an insect population density monitoring result set.

[0036] As a further scheme of the present application, the image frame refers to a continuous image sequence collected by a monitoring camera device within a fixed time interval, and each frame is a three-channel RGB image composed of pixels;

[0037] The texture and color features refer to the repetition and color distribution of pixels in the image in space arrangement, and the texture can obtain directionality and contrast indicators through a gray level co-occurrence matrix, and the color features are obtained through histogram or mean value calculation of the RGB three channels;

[0038] The contour continuity refers to the closure degree of the target edge in the image in space, which can be determined by the contour closure degree after the target contour is extracted by an edge detection algorithm.

[0039] As a further scheme of the present application, the image feature comparison result refers to a fitting degree comparison result of the insect body image and the features of the labeled target in the training sample library, and the fitting degree matching is performed in combination with the color histogram, the texture direction distribution and the shape features;

[0040] The response dense area refers to an image area formed by the pixel points with a response value exceeding a set activation threshold value in the network feature map.

[0041] Compared with the prior art, the present application has the following advantages and positive effects:

[0042] In the present application, the needle-leaf structure area is recognized, and the texture color difference image block is screened, so as to improve the accuracy of insect body target extraction, and the label atlas is generated in combination with the relationship between the insect body arrangement density and the area, so as to enhance the accuracy of grade determination, the response comparison activation threshold value of the feature map channel is used to realize path compression, so as to improve the attention of the network to the key features, the response point positioning is associated with the insect body boundary, so as to strengthen the spatial accuracy of the swarm distribution analysis, the response result is compared with the original label to realize dynamic correction, a closed-loop feedback mechanism is constructed, the determination consistency is optimized, and the overall recognition accuracy is improved, the artificial intervention is reduced, and the adaptive ability and the automation level of the density monitoring are enhanced. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 The system flowchart of the present application is shown in the figure;

[0044] Figure 2 The acquisition flowchart of the image acquisition module in the present application is shown in the figure;

[0045] Figure 3 The acquisition flowchart of the density label generation module in the present application is shown in the figure;

[0046] Figure 4 The acquisition flowchart of the feature compression path module in the present application is shown in the figure;

[0047] Figure 5 The acquisition flowchart of the area recognition module in the present application is shown in the figure;

[0048] Figure 6 The acquisition flowchart of the swarm density determination module in the present application is shown in the figure. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below in combination with the figures and examples. It should be understood that the specific examples described herein are only used to explain the present application, and are not used to limit the present application.

[0050] In the description of the present application, it should be understood that the terms "length", "width", "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In addition, in the description of the present application, "a plurality of" means two or more, unless otherwise explicitly and specifically limited.

[0051] Please refer to Figure 1 The deep neural network-based cypri-pine moth population density monitoring system comprises:

[0052] The image acquisition module acquires image frame data shot by the cypri-pine forest monitoring device, identifies needle leaf structure regions meeting color clustering characteristics and texture directionality distribution rules, extracts image blocks different from texture and color characteristics, records contour continuity and region occurrence frequency as screening basis, and generates a suspected cypri-pine moth region set;

[0053] A typical needle leaf structure region refers to a region having the following image characteristics: the Euclidean distance between the color value of the main color cluster after K-means clustering and the needle leaf green reference color (R=85G=100B=75) is less than 10; the texture directionality is evaluated by the direction energy index of the gray level co-occurrence matrix, the energy value in the main direction is more than 0.75, and the texture variance is less than 4.0, indicating that the region has a stable directional arrangement structure;

[0054] The texture directionality distribution rule is constructed based on the gray level co-occurrence matrix, four directions of 0°, 45°, 90° and 135° are selected, and the direction energy and contrast index of the image block are calculated. When the energy value of a single direction exceeds a set threshold (such as 0.75) and is accompanied by low variance (such as <4.0), it is considered that the image block has a significant directional feature, and can be determined as a candidate region of typical needle leaf structure;

[0055] The image frame data refers to a continuous image sequence collected by the monitoring camera device within a fixed time interval, each frame is a three-channel RGB image composed of pixels, and is used to record the visual information of the target region of the cypri-pine forest;

[0056] Texture and color characteristics refer to the repetition and color distribution of pixels in the spatial arrangement of the image. Texture can obtain directionality and contrast index through the gray level co-occurrence matrix, and color characteristics can be obtained through the histogram or mean value calculation of the RGB three channels;

[0057] Contour continuity refers to the degree of closure of the target edge in the image in space, which can be determined by the contour closure degree after the target contour is extracted by the edge detection algorithm, and is used to assist in judging the structural integrity of the target region in the image;

[0058] The density label generation module analyzes the matching relationship between the arrangement density of the texture structure of the insect body and the area of the image region based on the suspected C. chui region set, judges the aggregation trend of the insect body in the region, and generates a region insect population density label atlas in combination with the identifiable target quantity and the image feature comparison result;

[0059] Arrangement density refers to the spatial distribution density of the texture structure of the insect body per unit area, which is determined according to the proportional relationship between the number of target textures in the image region and the area, and is used to reflect the aggregation trend of the insect population;

[0060] The image feature comparison result refers to the fitting degree comparison result of the insect body image and the features of the labeled target in the training sample library, which is performed in combination with the color histogram, texture direction distribution and shape feature fitting degree matching;

[0061] The feature compression path module calls the label identification of the region in the region insect population density label atlas, inputs the image block corresponding to the label into the deep neural network structure, extracts the response of the channel in the feature map, compares the response number and amplitude information of the active position in the channel with the set channel activation threshold, retains the channel whose activation level reaches the channel activation threshold, records the activation channel number to construct a channel index set, and generates a compressed path index set;

[0062] The response condition refers to the activation feedback state of the channel in the neural network to the input image region, which is commonly indicated by the average activation value, the activation region size or the non-zero proportion index, and is used to determine whether the channel is effective to the current input;

[0063] The channel activation threshold refers to the minimum activation level boundary preset for screening the feature channel, which can be set by statistically analyzing the average response level of the channel in the training sample;

[0064] The region recognition module extracts the feature map result output by the compressed path according to the path number marked by the compressed path index set, locates the pixel points whose response amplitude exceeds the preset response threshold, and associates the coordinates with the morphological boundary in the insect aggregation region, analyzes the distribution concentration degree, and generates an insect response distribution map;

[0065] The morphological boundary refers to the contour boundary of the insect body formed by the image segmentation or edge detection algorithm in the image, which represents the boundary area between the target and the background, and is commonly used as the basis for coordinate matching and region positioning;

[0066] The insect population density determination module calls the response point dense area in the insect body response distribution map, compares it with the original grade label in the area insect population density label map, identifies whether the label is deviated, adjusts the original grade label according to the difference, updates the result, and generates the insect population density monitoring result set;

[0067] The response point dense area refers to the image area formed by the pixel points with response values exceeding the set activation threshold value in space in the network feature map, which is used to estimate the distribution degree of the insect body in the target image area.

[0068] The suspected cecropia moth area set includes the typical needle leaf structure area, the texture difference image block, and the color feature abnormal image block. The area insect population density label map includes the insect body texture arrangement density label, the image area area grade label, and the identifiable target number density label. The compression path index set includes the active path routing number, the channel response amplitude threshold record, and the feature map channel reservation mark. The insect body response distribution map includes the insect body high response position coordinate, the insect body aggregation area shape boundary, and the insect body distribution concentration degree grade. The insect population density monitoring result set includes the insect population density grade result, the label deviation correction result, and the insect population spatial distribution change result.

[0069] Please refer to Figure 2 , the image acquisition module includes:

[0070] The image frame acquisition submodule acquires the image frame sequence photographed by the cypress forest area monitoring device in a continuous period, numbers and stores the image frames in different time periods, records the spatial resolution and photographing time label of each image frame, integrates the image frames after sorting according to the time label, and obtains the image frame sequence set;

[0071] The image frame sequence of the continuous period of the Shubai Forest Area monitoring device is obtained by using a fixed high-definition camera device with automatic time synchronization and high-frequency sampling function. The device is continuously sampled from 9:00 to 15:00 every day from June 10, 2023 to July 10, 2023. The device model is XMG-2000, the shooting interval is every 10 seconds, the storage format is JPEG, each image is attached with time stamp and position information, the coordinate information is recorded by the GPS module of the device with an accuracy of ±0.5 meters, the naming rule of the collected images is "LQ_YYYYMMDD_HHMMSS.jpg", for example, 20230612_103210 corresponds to the image frame at 10:32:10 on June 12, after the image collection is completed, all the images are sorted by time label, in the actual application, 10 days of sunny weather data are selected, a total of 21600 images are collected, the spatial resolution of each image is recorded as 4000x3000 pixels, the clear frames are screened out, the dark, blurred and jitter frames are removed, for example, the removed image is LQ_20230612_104510.jpg, the remaining images are constructed into an image frame sequence set and segmented according to the time interval to construct a subsequence for subsequent structure recognition and time series analysis, the total number of effective image frames is 16320 frames, which constitutes the image frame sequence set.

[0072] The needle structure recognition submodule is based on the image frame sequence set, extracts the RGB color channel data and texture gradient information in each image, performs color clustering and texture direction clustering calculation on the pixel blocks, judges whether the typical needle structure discrimination rule is met, aggregates the spatial continuous area and removes the isolated blobs, and obtains the needle area contour map;

[0073] Based on the image frame sequence set, the color matrix and texture feature distribution of all pixel blocks in the image are extracted, each image is divided into image blocks with a size of 40x40 pixels, and each image contains about 7500 image blocks. For each image block, the RGB channel three-dimensional value is extracted to construct a color difference vector and calculate the Euclidean distance with the standard color of the typical needle structure. The standard color is set as R=85G=100B=75. In the example, the image block number of the image block LQ_20230612_103210_frame001_block0057 is R=90G=110B=80. The color difference ratio is calculated as follows:

[0074] ;

[0075] The texture features of the image blocks in the gray scale image are counted again, the energy, contrast and variance are extracted through the gray level co-occurrence matrix, the variance is selected as the representative index, and the variance value of the example block is In the region contour identification, if the image block color difference ratio is between 1015, the texture variance is between 3.06.0, and the center distance between blocks is less than 50 pixels, it is marked as a candidate needle structure region, contour tracking is performed using OpenCV, and isolated spots with an area less than 400 pixels² are removed, and then the closure degree is extracted as the ratio of the contour area to the convex hull area. The example image block area is 460 pixels², and the convex hull area is 530 pixels², and the closure degree is calculated as follows:

[0076] ;

[0077] The frequency of the image block appearing continuously is recorded, and the example image block appears 5 times in the continuous subsequence. Further, the image block region set that meets the above conditions is saved in a numbered form, which is prepared for subsequent screening and scoring.

[0078] The difference region screening submodule calls the needle region contour map, performs color difference ratio calculation and texture variance detection on the region, and combines the contour closure degree index and the specific calculation formula of the region appearance frequency:

[0079] ;

[0080] The operation obtains the suspected difference block saliency value and compares it with the set region saliency threshold value, extracts the image block higher than the region saliency threshold value and marks the spatial position, and obtains the suspected C. sinensis moth region set;

[0081] wherein, represents the suspected difference block saliency value, represents the color difference ratio of the i-th image block, represents the mean value of all image block color difference ratios, represents the texture variance value of the i-th image block, represents the appearance frequency of the i-th image block in the image frame sequence, represents the contour closure degree value of the i-th image block, and n represents the total number of image blocks;

[0082] The needle region contour map is called, and the parameters of all image block numbers inside it are extracted and scored. The extracted parameters include image block color difference ratio , regional average color difference , texture variance , image block appearance frequency in the sequence , and contour closure degree . It is assumed that the contour region contains 5 image blocks, and the parameters are shown in the following table:

[0083] Table 1 Image block saliency score calculation parameter table

[0084] ;

[0085] Based on the above data, the saliency score is calculated by substituting the formula:

[0086] ;

[0087] Image block 1: ;

[0088] Image block 2: ;

[0089] Image block 3: ;

[0090] Image block 4: ;

[0091] Image block 5: ;

[0092] The total numerator is:

[0093] ;

[0094] The denominator is:

[0095] ;

[0096] The saliency score is:

[0097] ;

[0098] The results show that the image block set presents high difference in color difference deviation, texture complexity and continuous appearance frequency, and the saliency score exceeds the preset screening threshold 12.0, which can be judged as a suspected Sabina pine moth area, and the image block number of the set is marked and output as an image coordinate range, which is used for subsequent regional aggregation processing and recognition model training to obtain a suspected Sabina pine moth area set. The usefulness of the formula is that by introducing color difference deviation, texture complexity root weight, frequency weighting and other calculation methods, the suspiciousness of the image area is scored and judged by integrating multiple parameters, realizing the integrated evaluation ability of difference characteristics;

[0099] The operation logic of the formula is reflected in the multi-dimensional fusion of the color, texture, time sequence and structural features of the image block. The difference between the color difference ratio and the overall average value is used to reflect the deviation degree of the image block in color, representing the difference in color between the image block and the typical background area. The larger the value, the more abnormal the color of the area. The texture variance is processed by square root and participates in the calculation. The square root operation is used to reduce the influence of the fluctuation of the variance value on the score, avoid the dominance of the extreme texture complexity in the significance calculation, and make the score more stable. At the same time, it reflects the influence of the texture structure on spatial heterogeneity. The sum of the two represents the comprehensive difference of the single image block in color and texture dimensions. Then the combined value is multiplied by the frequency of the image block in the image sequence, in order to strengthen the weight assignment of the time continuity area, so that the abnormal area with high continuity gets a higher score. After summing up the weighted values of all image blocks, the product of the total number of image blocks and the average contour closure is used as the denominator for normalization. The purpose is to balance the score as a whole through the integrity of the structure, limit the influence of the incomplete structure area on the judgment result, and then ensure that the score logic takes into account the color difference significance, texture complexity, time sequence stability and structural continuity at the same time.

[0100] The suspected difference block significance value is a score index for comprehensively measuring the abnormal degree of a region in the image in multiple feature dimensions, mainly used to measure whether the image block significantly deviates from the typical structural features of coniferous forest background. The higher the value, the stronger the difference of the region in color distribution, texture morphology, time continuity, etc. The value is calculated by integrating the color difference deviation, texture variance and the frequency of repeated appearance in the image sequence, and then normalized by combining the contour closure. Therefore, it can fully reflect the abnormal degree of the image block in color space, texture structure and time sequence characteristics. The score value is an important basis for judging whether the image region belongs to the suspected cypriatte moth damage region, and is the core evaluation quantity in the image difference detection process.

[0101] Please refer to Figure 3 , the density label generation module comprises:

[0102] The texture density calculation sub-module obtains the texture direction distribution of the insect body in the region image block and the area information of the image block based on the suspected cypriatte moth region set, counts the number of texture arrangement units of the insect body, processes the data according to the texture direction information and the corresponding area in the image block, extracts the texture structure arrangement density features of the insect body in the region, integrates the texture density correlation information of all image blocks in the image sequence, and obtains the texture density value set of the insect body.

[0103] Image blocks were divided according to image frames, and texture direction channels were constructed for each region. The block size in each frame was set to 40×40 pixels. The gray-level co-occurrence matrix was extracted for each image block, and the continuous length of the main texture direction chain was calculated. The area of ​​the corresponding image block was collected as a reference unit for ratio calculation. The texture density index was constructed by statistically analyzing the gray-level texture direction intensity and the total area of ​​the image block. The image frame range was selected from the measured data in June 2023. Five image blocks numbered A1 to A5 were extracted, and their number of insects, image block area, number of insect overlaps, texture direction gradient value, and standard deviation of the number of image blocks in the sequence were recorded. Based on this, a structured index system was established, and the parameters are shown in Table 2.

[0104] Table 2. Insect aggregation parameters

[0105] ;

[0106] As shown in Table 2, the texture distribution values ​​of the corresponding image blocks were collected and statistically standardized by looping within the image frame. The area was uniformly 800 pixels². The structural feature extraction parameters of each block were archived according to the image processing module to establish a set of insect texture density values.

[0107] The insect aggregation judgment submodule calls the insect texture density value set, extracts the insect quantity index per unit area based on the area of ​​the region image and the number of identifiable insects, and combines the insect overlap degree and the variation characteristics of the number of insects in the image patch in the sequence to perform aggregation trend analysis on the difference region. The specific calculation formula is as follows:

[0108] ;

[0109] The aggregation degree value of the insects is obtained by calculation, and the result is compared with the set baseline value to generate an aggregation classification label;

[0110] in, Indicates the degree of aggregation of the insects. Indicates the first The number of insects per unit area of ​​each image patch. This represents the average number of insects per unit area across the entire region. Indicates the first Number of insect bodies overlapping per unit area in an image patch This represents the standard deviation of the number of worms in a given image block within an image sequence. This represents the gradient magnitude of the change in the direction of the insect texture within the image patch. Indicates all within the region Summation operation is performed on each image patch;

[0111] Call the worm texture density value set, according to the area of the image and the number of identifiable worms, the image block data in table 2 is processed item by item, first calculate the number of worms per unit area ratio , each image block is =0.03, =0.0375, =0.035, =0.04375, =0.0275, the total number of worms in the area is 139, the area is 4000 pixel², then the average number of worms per unit area is , calculate the deviation of each image block , respectively =0.00475, =0.00275, =0.00025, =0.009, =0.00725, respectively get the number of worms per unit area , each image block is =0.075, =0.09375, =0.085, =0.1125, =0.0625, add them together and multiply by the number of standard deviations , respectively as follows: = , = , = , = , = , the sum of the above numerators is 1.4827, and the sum of the texture direction change amplitude of all image blocks is 0.32, 0.28, 0.35, 0.40, 0.25, and the total is 1.6, which is substituted into the following formula:

[0112] ;

[0113] According to the above calculation process, the worm aggregation value is 0.9267, which is compared with the reference interval setting value, the aggregation value is between 0.7 and 1.1, and the worm aggregation value is generated;

[0114] The operation logic of the formula embodies a multi-dimensional comprehensive measurement mechanism from the insect body distribution density, spatial aggregation phenomenon to the image texture change. Firstly, the absolute value of the difference between the number of insect bodies per unit area in each image block and the average insect body density in the region is processed to construct a local number offset index, which reflects the fluctuation degree of local insect body density. Then, the offset value is added to the number of insect bodies per unit area in the image block to unify the number difference and the spatial aggregation phenomenon into a synthetic index representing the "local aggregation abnormal intensity". Then, the synthetic index is multiplied by the standard deviation of the number of insect bodies in the image block in the image sequence to incorporate the temporal fluctuation into the evaluation, highlighting the scoring weight of the area with repeated occurrence or significant change. In the numerator part, the weighted aggregation of the above-mentioned index in each image block is completed, while in the denominator part, the sum of the texture direction change amplitude of the image block is calculated to reflect the structural complexity of the entire region as the normalization coefficient of the numerator, and through division, a unified scoring dimension is constructed to output the aggregation degree score representing the aggregation degree of the insect bodies in the region.

[0115] The insect body aggregation degree value is a comprehensive index for measuring whether the insect bodies in the image region have an aggregation trend in terms of spatial distribution, number change and texture structure characteristics. Its specific meaning lies in reflecting whether the number of insect bodies in the region is relatively concentrated, whether there is an overlapping accumulation phenomenon per unit area, and the fluctuation of such distribution characteristics in the image sequence. By integrating the number difference per unit area, the insect body overlap density and the change amplitude of the number in time sequence, while considering the adjustment effect of the change intensity of the image block texture direction on the regional complexity, a quantitative value is outputted which can be used for aggregation level classification and regional hierarchical labeling. The larger the value, the more the insect bodies tend to concentrate in the local region, and the stronger the aggregation.

[0116] The density map construction submodule calls the aggregation hierarchical label, maps the region according to the color coding rule corresponding to the difference aggregation level, extracts the coordinate information of the image block and establishes the label layer matrix, fuses the layer with the original image contour, adds the regional insect body identification number, and generates the regional insect swarm density label map.

[0117] The aggregation level label is called to divide the value into three aggregation levels according to the set level interval, and the value below 0.7 is the low aggregation level, which is set as a blue mark, the interval 0.7 to 1.1 is the middle aggregation level, which is marked as yellow, and the value greater than 1.1 is the high aggregation level, which is marked as red, the aggregation value is 0.9267, the corresponding level is middle, the image blocks are arranged according to the number, the coordinates of each image block are mapped to the two-dimensional label layer structure, and the corresponding RGB color coding is given according to the aggregation level, the layer information is superimposed and rendered with the original image boundary contour, the image block identification number and color mark are output together, the aggregation density coding graph is generated on the image block aggregation density color layer according to the coordinate order, the layer image structure is uniformly numbered and the label image is exported, and the regional insect density label atlas is obtained.

[0118] Please refer to Figure 4 The feature compression path module comprises:

[0119] The label identification extraction submodule extracts the label number information corresponding to the image block based on the label layer divided in the regional insect density label atlas, identifies the image blocks with repeated numbers in the same region and uniformly identifies them, selects the image block at the center position in the region to establish the number mapping relationship, obtains the corresponding relationship between the label number and the image block number, and generates a regional label mapping table.

[0120] The coordinate index field of each image block in the layer matrix is read, and the number field and the label field are compared, the multiple image blocks under the same label number are identified through the image block number, it is judged whether the image blocks belong to the same region and the number attribution relationship is established, for each image block set corresponding to the label number, the image block with the smallest number sequence or the closest geometric center position to the center of the region is selected as the representative block identification, the number of the representative image block is taken as the unique identification of the region, and the index mapping between the label number and the image block number is established, the triple corresponding relationship between the number sequence, the label attribution and the representative mark of all image blocks is established, the structure is recorded and the label mapping field is sorted and output, and the regional label mapping table is generated.

[0121] The response channel screening submodule calls the regional label mapping table, inputs the mapping image block into the neural network structure, extracts the channel response in the layer feature map, detects the response position number and response amplitude in each channel, calculates the response intensity value of the channel, compares the result with the channel activation threshold, screens the channel numbers meeting the conditions, and performs number arrangement to generate an activated channel number set.

[0122] The image block corresponding to the image block number recorded in the region label mapping table is input into the neural network structure for forward propagation to obtain the response state of each channel in the feature map layer, extract the response position number and corresponding response amplitude data in each channel, and count the response average value of all channels in each image block as a response feature index. The response feature index is compared with the set channel activation threshold, which is set to 0.61. The value is determined to be the most suitable for channel screening according to the matching curve between the different amplitude response regions and the annotation accuracy in the sample image. It is judged whether the response intensity of each channel reaches the set threshold. If the condition is met, the channel number is recorded as the active channel number. The number of channels meeting the condition in each image block is counted and a screening result index is established. The screening data is shown in Table 3:

[0123] Table 3 Image block channel response statistics table

[0124] ;

[0125] As shown in Table 3, the average response amplitude of image block B1 is 0.62, and the channel screening number is 5. The average response amplitude of B3 is 0.66, and the channel screening number is 6, both of which meet the activation threshold condition. The average value of B4 is 0.53, and the channel screening number is only 3. The channel screening result is recorded by combining the response amplitude and the screening number information, and the active channel number set is generated.

[0126] The compression path index generation submodule extracts the path identifier and spatial position label of each channel in the network according to the active channel number set, establishes a mapping structure of channel number and path number, integrates and sorts the channel path information in all image blocks, and generates a compression path index set;

[0127] According to the image block number and the screening channel number recorded in the active channel number set, the path information corresponding to each active channel is extracted from the neural network structure, the path structure identifier mapped by the channel number is read, and the path index field is established in combination with the convolution layer number. The image block number and the path number corresponding to the active channel under it are paired to generate the mapping relationship between the image block and the path structure. The path number results of all image blocks are summarized and sorted according to the image block number to construct a path structure mapping set, output the number list, and generate a compression path index set.

[0128] Please refer to Figure 5 , the region recognition module comprises:

[0129] The feature map extraction submodule extracts the feature map output content of the corresponding channel according to the path number in the compression path index set, classifies and integrates the layer information under the path according to the image block number, uniformly reorganizes the spatial position index, obtains the structured feature map data, and generates a compression path feature map set.

[0130] According to the path number in the compressed path index set, the feature map output content of the corresponding channel is extracted, the image block number and the path number are paired, the channel number index of each image block in the path structure is read, the matched layer output result is obtained from the feature cache structure, the extracted multiple channel images are merged by path number, the image block position is reorganized through the space position index field, and the layer content corresponding to all path numbers is summarized in the image block unit. In the example of image block number B1, the number of path numbers is 5 and the number of channels is 8. The image reorganization operation is performed on the channel data in order according to the path order, and the mapping structure between the image block number and the channel layer is constructed accordingly. After the feature channel aggregation of all image blocks is completed, the integrated layer data is uniformly output, and the compressed path feature map set is generated.

[0131] The high-response positioning sub-module extracts the response position and response amplitude information in the channel image based on the compressed path feature map set, identifies the response points that meet the threshold condition, locates the coordinates in the image space, constructs a complete space index structure, and summarizes the coordinate information in all image blocks to generate a high-response pixel coordinate set.

[0132] Based on the compressed path feature map set, the pixel response data of the channel layer in each image block is scanned point by point, all pixel response positions and their corresponding response amplitude information are extracted, the number of response positions in the channel image is counted, the response distribution structure is obtained, the maximum response value and the average response value are selected from the response data, the overall response level of the image block is calculated, and the average response threshold is further set as the screening basis to judge whether each response position meets the threshold condition and is marked as an active response point. In the example of image block number B2, a total of 14 response positions are identified, the maximum response value is 0.68, the average response threshold is set to 0.60, and 8 response points that meet the condition are selected. In image block number B5, 10 response points are selected, and in image block B3, 6 response points are selected. All response pixel positions that meet the condition are re-collected according to the image block number and a space index structure is established to generate a high-response pixel coordinate set. The statistical data in this process is shown in Table 4:

[0133] Table 4 Image block feature response extraction data table

[0134] ;

[0135] As shown in Table 4, 9 response points are selected in image block B1 and 7 response points are selected in image block B4. The number of response points varies with the number of channels and the maximum response value, showing a difference distribution, which provides a basis for spatial positioning for regional matching.

[0136] The morphological matching analysis submodule extracts the boundary layer data of the insect aggregation area according to the high-response pixel coordinate set, matches the spatial relationship of each response point and the corresponding boundary area, counts the number of response points in the boundary area, calculates the response point distribution density in the area, and generates an insect response distribution map;

[0137] The position index of the response pixel in all image blocks is subjected to a coordinate matching operation, the spatial boundary box data in each boundary area is read, it is judged whether the response point is located in the boundary area, the belonging area is determined through the spatial inclusion relationship, in the response point coordinate set of the image block numbered B1, it is identified that 5 coordinate points fall into the insect boundary inner area, the number of response points falling into the boundary area in the image block B5 is 8, and the number of response points falling into the boundary area in the image block B3 is 3, after counting the number of response points contained in each boundary area, the spatial density value is calculated based on the area parameter and the number of response points, the response concentration calculation operation of the insect area is completed, and an insect response distribution map is generated.

[0138] Please refer to Figure 6 The insect swarm density determination module comprises:

[0139] The response area comparison submodule extracts the spatial position and dense area boundary of the response point based on the insect response distribution map, counts the number of response points in the area, combines the original grade mark of the corresponding position in the area insect swarm density label atlas, completes the spatial pairing operation of the response area and the label grade, obtains the grade comparison result of the position consistent area, and generates a label grade deviation distribution value;

[0140] Based on the spatial information in the worm response distribution map, the specific coordinate values and the boundary range of the area where all the response points are located are extracted, each response point is classified according to the layer number and coordinate index, the total number of response points in each area is calculated according to the image block principle, and the boundary profile of the response dense area is mapped in combination with the pixel density structure of the layer. On this basis, the label data layer spatially overlapping with the response area in the region worm density label atlas is obtained, the original grade mark field of each spatial cell in the label atlas is extracted, the spatial position index correspondence relationship between the response area and the label grade field is established, and the corresponding matching group between the actual performance of the response point density and the grade annotation field is formed. The response area and the spatial unit with the same number in the label atlas are compared, the number of response points in each spatial position under the response area is read during the comparison, and is paired with the label grade field value at the position. The grade difference value statistical operation is performed on the record items with the same spatial position number but different grade annotations, whether there are inconsistent grade record items is judged, for the cells with grade difference value, the grade deviation point is marked, for example, in the image block with region number R08, the number of response points is 38, the corresponding label grade is 2, and the system expects the grade to be 3 according to the number of response points. Therefore, the position is recorded as a deviation point. The difference value information of all identified deviation positions is recorded and marked to establish a complete label grade deviation record structure and generate a label grade deviation distribution value.

[0141] The label deviation calculation submodule identifies the position range of the deviation according to the label grade deviation distribution value, extracts the number and spatial distribution density of response points in the corresponding area, judges the difference between the original grade mark and the actual response, calculates the grade adjustment index of the area, and merges the offset to generate a grade adjustment coefficient matrix.

[0142] According to the label grade deviation distribution value, the spatial position number of all existing deviations is extracted, the response point density data under each number is read, the numerical difference between the original grade value of each position and the expected grade value derived from the point density of the current response area is calculated, the absolute amplitude of the difference value is calculated, and the direction attribute of the grade promotion or decline is judged. Then, the response point number, point density average of the deviation position point and the area information of the image block where the point is located are called to weight the difference value, establish a grade deviation intensity value field to reflect the necessary degree of grade correction of each position, and record the number, original grade, expected grade, difference amplitude, response point number and area information of all deviation points in a table structure. Part of the data record is shown in Table 5:

[0143] Table 5 Grade deviation area index statistical table

[0144] ;

[0145] As shown in Table 5, the area R03 and R08 appear underestimation of the level, the difference is 1, the number of response points is greater than 35, the area R14 appears overestimation of the level, the difference is -1, and the number of points is significantly less. The statistical information of all the deviation areas is collected to construct the level adjustment coefficient matrix, and the level adjustment coefficient matrix is generated.

[0146] The density level updating submodule updates the level field of all deviation areas according to the level adjustment coefficient matrix, replaces the original label level marking information, records the updated level value of each position, integrates all the adjusted label content, and generates the insect population density monitoring result set.

[0147] According to the level adjustment coefficient matrix, the spatial position number and the adjustment amplitude value recorded therein are called one by one, the level field corresponding to the position in the original label level atlas is found, and the original level value field is read for judgment processing. If the adjustment amplitude is positive, the level is raised, and if it is negative, the level is lowered. During the operation process, the level adjustment value must be controlled not to exceed the maximum range of the system defined level, i.e. all level fields must be limited between 1 and 5 levels, otherwise the boundary value will be replaced. In actual application, when the original level of area R08 is 2 and the adjustment coefficient is +1, the level field of this position is updated to 3. The updated value is recorded in the update log and marked as adjusted. For all adjusted positions, the output operation is performed to build the updated label level field record table, and the label level atlas data layer structure after updating is output, the level field correction processing of the label data set is completed, and the insect population density monitoring result set is generated.

[0148] The above is only a preferred embodiment of the present application, and does not limit the form of the present application. Any skilled person in the art can modify or change the above disclosed technical content to equivalent embodiments applied to other fields, but any simple modification, equivalent change and modification made according to the technical essence of the present application to the above embodiments without departing from the technical solution content of the present application still belongs to the protection scope of the present application.

Claims

1. A deep neural network-based cinnabaria lativittata population density monitoring system, characterized in that, The system comprises: The image acquisition module acquires image frames of the cypress forest monitoring device, identifies needle leaf structure regions meeting color clustering characteristics and texture directionality distribution rules, extracts image blocks with abnormal texture and color characteristics, screens images according to contour continuity and region frequency, and generates a suspected cypress moth region set; The density label generation module analyzes the matching relationship between insect texture density and region area based on the suspected cypress moth region set, judges the aggregation trend, and generates a region insect population density label atlas in combination with the target quantity and image feature comparison result; The feature compression path module calls the region label in the region insect population density label atlas, inputs the image block corresponding to the label into a deep neural network structure, extracts a feature map channel response, compares the number and amplitude of the channel responses, screens active channels, records the active channel numbers to construct a channel index set, and generates a compressed path index set; The region identification module extracts a compressed path feature map according to the compressed path index set, locates pixel points with a response amplitude exceeding a preset response threshold, associates the pixel points with insect boundary coordinates, analyzes the concentration degree of the distribution, and generates an insect response distribution map; The insect population density determination module calls the response dense region in the response distribution map, compares it with the original level of the label atlas, identifies the deviation, updates the label after correction, and generates an insect population density monitoring result set.

2. The deep neural network-based thrips population density monitoring system of claim 1, wherein, The suspected cypress moth region set includes typical needle leaf structure regions, texture difference image blocks, and color feature abnormal image blocks. The region insect population density label atlas includes insect texture arrangement density labels, image region area level labels, and identifiable target quantity density labels. The compressed path index set includes active channel routing numbers, channel response amplitude threshold records, and feature map channel retention identifiers. The insect response distribution map includes insect high-response position coordinates, insect aggregation region morphology boundaries, and insect distribution concentration degree levels. The insect population density monitoring result set includes insect population density level results, label deviation correction results, and insect population spatial distribution change results. 3.The deep neural network-based thrips population density monitoring system of claim 1, wherein The image acquisition module comprises: The image frame acquisition submodule acquires a sequence of image frames photographed by the cypress forest monitoring device in consecutive time periods, numbers and stores the image frames in different time periods, records the spatial resolution and photographing time label of each image frame, integrates the image frames in order of the time label, and obtains a sequence of image frame sets; The needle leaf structure identification submodule extracts RGB color channel data and texture gradient information in each image based on the sequence of image frame sets, performs color clustering and texture directionality clustering calculation on pixel blocks, judges whether the typical needle leaf structure discrimination rule is met, aggregates spatially continuous regions and removes isolated blobs, and obtains a needle leaf region contour map; The difference region screening submodule calls the needle leaf region contour map, performs color difference ratio calculation and texture variance detection on the regions, obtains a suspected difference block saliency value in combination with the contour closure index and region occurrence frequency, compares the suspected difference block saliency value with a set region saliency threshold value, extracts image blocks higher than the region saliency threshold value and marks the spatial positions, and obtains a suspected cypress moth region set.

4. The deep neural network-based forest tent caterpillar population density monitoring system of claim 3, wherein, The density label generation module comprises: The texture density calculation sub-module obtains the texture direction distribution of the insect body in the region image block and the area information of the image block based on the suspected C. chouyoia region set, counts the number of texture arrangement units of the insect body, processes data according to the texture direction information and the corresponding area in the image block, extracts the texture structure arrangement density features of the insect body in the region, integrates the texture density correlation information of all image blocks in the image sequence, and obtains a set of insect body texture density values; The insect body aggregation judgment sub-module calls the set of insect body texture density values, extracts the number of insect bodies per unit area according to the area of the region image and the number of identifiable insect bodies, and combines the insect body overlap degree and the change feature of the number of insect bodies in the image block in the sequence to analyze the aggregation trend of the difference region and obtain an insect body aggregation degree value. The result is compared with the set reference value to generate an aggregation classification label; The density atlas construction sub-module calls the aggregation classification label, maps the region according to the color coding rule corresponding to the difference aggregation level, extracts the coordinate information of the image block and establishes a label layer matrix, fuses the layer with the original image contour, adds the region insect body identification number, and generates a region insect swarm density label atlas.

5. The deep neural network-based forest tent caterpillar population density monitoring system of claim 4, wherein, The specific calculation formula for analyzing the aggregation trend of the difference region is: ； The operation obtains the insect body aggregation degree value, the result is compared with the set reference value, and the aggregation classification label is generated; in, Indicates the degree of aggregation of the insects. Indicates the first The number of insects per unit area of ​​each image patch. This represents the average number of insects per unit area across the entire region. Indicates the first Number of insect bodies overlapping per unit area in an image patch This represents the standard deviation of the number of worms in a given image block within an image sequence. This represents the gradient magnitude of the change in the direction of the insect texture within the image patch. Indicates all within the region Summation operation is performed on each image patch.

6. The deep neural network-based forest tent caterpillar population density monitoring system of claim 4, wherein, The feature compression path module includes: The label identification extraction sub-module extracts the label number information corresponding to the image block based on the label layer divided in the region insect swarm density label atlas, identifies the image blocks with repeated numbers in the same region and uniformly identifies them, selects the image block at the center position in the region to establish a number mapping relationship, obtains the corresponding relationship between the label number and the image block number, and generates a region label mapping table; The response channel screening sub-module calls the region label mapping table, inputs the mapping image block into the neural network structure, extracts the channel response situation in the layer feature map, detects the response position number and response amplitude in each channel, calculates the response intensity value of the channel, compares the result with the channel activation threshold value, screens the channel numbers meeting the condition, and performs number arrangement to generate an activated channel number set; The compression path index generation sub-module extracts the path identifier and spatial position index of each channel in the network according to the activated channel number set, establishes a mapping structure of the channel number and the path number, integrates the channel path information in all image blocks and sorts and archives them, and generates a compression path index set.

7. The deep neural network-based forest tent caterpillar population density monitoring system of claim 6, wherein, The region identification module includes: The feature map extraction sub-module extracts the feature map output content of the corresponding channel according to the path number in the compression path index set, classifies and integrates the layer information under the path according to the image block number, uniformly reorganizes the spatial position index, obtains the structured feature map data, and generates a compression path feature map set; The high-response positioning submodule extracts response position and response amplitude information in the channel image based on the compressed path feature map set, identifies response points that meet a response threshold condition, locates coordinates in the image space, constructs a complete spatial index structure, aggregates coordinate information in all image blocks, and generates a high-response pixel coordinate set; The shape matching analysis submodule extracts boundary layer data of a worm body aggregation area based on the high-response pixel coordinate set, matches the spatial relationship between each response point and the corresponding boundary area, counts the number of response points in the boundary area, calculates the response point distribution density in the area, and generates a worm body response distribution map.

8. The deep neural network-based forest tent caterpillar population density monitoring system of claim 7, wherein, The worm population density determination module includes: The response area comparison submodule extracts the spatial position and dense area boundary of the response point based on the worm body response distribution map, counts the number of response points in the area, combines the original grade mark at the corresponding position in the area worm population density label atlas, completes the spatial pairing operation of the response area and the label grade, obtains the grade comparison result in the position consistent area, and generates a label grade deviation distribution value; The label deviation calculation submodule identifies the position range of the deviation based on the label grade deviation distribution value, extracts the number and spatial distribution density of response points in the corresponding area, judges the difference between the original grade mark and the actual response situation, calculates the grade adjustment index of the area, and merges the offset, to generate a grade adjustment coefficient matrix; The density grade updating submodule updates the grade mark of the area based on the grade adjustment coefficient matrix, calls the original label grade label information, replaces the grade field in all deviation areas, records the updated grade value at each position, integrates all adjusted label contents, and generates a worm population density monitoring result set.

9. The deep neural network-based forest tent caterpillar population density monitoring system of claim 1, wherein, The image frame refers to a sequence of continuous images collected by a monitoring camera device at a fixed time interval, and each frame is a three-channel RGB image composed of pixels. The texture and color features refer to the repetition and color distribution of pixels in the image in space, and the texture can obtain directionality and contrast indicators through a gray level co-occurrence matrix, and the color features can be obtained through an RGB three-channel histogram or mean value calculation. The contour continuity refers to the closure degree of the target edge in the image in space, which can be determined by the contour closure degree after the target contour is extracted by an edge detection algorithm.

10. The deep neural network-based forest tent caterpillar population density monitoring system of claim 1, wherein, The image feature comparison result refers to the fitting degree comparison result of the worm image and the labeled target in the training sample library, which combines color histograms, texture direction distribution, and shape features to perform fitting degree matching. The response dense area refers to an image area formed by pixels with response values exceeding a set activation threshold in space in the network feature map.

Images