Agricultural insurance subject matter intelligent point identification method, device and equipment and storage medium

Abstract

The invention relates to the technical field of artificial intelligence and financial science and technology, can be applied to an insurance system, and discloses an agricultural insurance subject matter intelligent point identification method and device, equipment and a medium. The method comprises the steps of operating a lightweight artificial intelligence model on a user terminal, and performing preprocessing and feature extraction on an obtained agricultural insurance subject matter image to generate a multi-layer feature map; performing high-frequency detail enhancement, adaptive feature fusion and weighted correction on the multi-layer feature map to obtain a fused feature map; performing target detection or density estimation based on the fused feature map; selecting a detection parameter group from a preset detection parameter library according to the recognition scene, and performing re-reasoning or post-processing on the recognition result according to the selected detection parameter group to obtain an adjusted recognition result; executing multi-target tracking based on the adjusted recognition result, establishing target identity association, combining repeated detection records, calculating the number of subject matters, and generating a final point result; and displaying the final counting result and the identification frame on a terminal interface.

Description

Technical Field

[0001] This application relates to the fields of financial technology and artificial intelligence technology, and in particular to a method, device, equipment and storage medium for intelligent point identification of agricultural insurance objects. Background Technology

[0002] In the agricultural insurance business process, the counting of agricultural insurance objects (such as fruit trees, livestock, and seedlings in plantations) is a core foundational step in policy underwriting verification, risk level assessment, and claims settlement calculation. The accuracy of the results directly determines the reliability of agricultural insurance business data, which in turn affects the rights and interests of insurance companies and policyholders.

[0003] Currently, the counting of agricultural insurance targets mainly relies on two types of technical solutions, but both have key drawbacks that prevent the counting of targets with high accuracy. On the one hand, traditional manual counting methods are limited by subjective judgment errors, visual fatigue, and scene interference (such as crop branches and leaves obscuring fruit, or dense livestock activity). The counting error rate is relatively high for small targets (such as young fruit trees or young livestock) or densely distributed targets (such as flowering fruit trees or densely fattened pigs in pens). In some complex scenarios, the error rate is even higher. This cannot meet the basic requirements of agricultural insurance for counting accuracy, and the final output counting results still have significant deviations and cannot be used as a reliable data basis for agricultural insurance business.

[0004] In summary, the most critical flaw in existing agricultural insurance target counting technology is its inability to achieve high-precision target counting. This directly leads to doubts about policy verification, distorted risk assessment, and frequent disputes over claims settlement in agricultural insurance business, severely restricting the standardization and efficiency of agricultural insurance operations. Summary of the Invention

[0005] This application provides a method, apparatus, device, and storage medium for intelligent point identification of agricultural insurance targets, in order to solve the technical problem of being unable to achieve high-precision point identification of targets.

[0006] A method for intelligent point identification of agricultural insurance targets based on front-end artificial intelligence includes: A lightweight artificial intelligence model is run on the user terminal to preprocess and extract features from the acquired images of agricultural insurance targets to generate multi-layer feature maps. The multi-layer feature map is subjected to high-frequency detail enhancement, adaptive feature fusion, and weighted correction based on channel statistics to obtain a fused feature map; Based on the fused feature map, target detection or density estimation is performed to obtain recognition results including the target's location, category, and confidence level; Based on the recognition scenario, a set of detection parameters is selected from a preset detection parameter library, and the recognition result is re-inferred or post-processed according to the selected set of detection parameters to obtain an adjusted recognition result. Based on the adjusted identification results, multi-target tracking is performed, target identity associations are established, duplicate detection records are merged, the number of targets is calculated, and the final point count result is generated. The final score result and the recognition box are displayed on the terminal interface.

[0007] A smart point identification system for agricultural insurance targets, the device comprising: The image acquisition module is used to acquire images of agricultural insurance targets and provide image data to the AI ​​inference module; The AI ​​inference module is used to run a lightweight artificial intelligence model on the user terminal to preprocess and extract features from the acquired agricultural insurance target images to generate multi-layer feature maps. The feature enhancement module is used to perform high-frequency detail enhancement, adaptive feature fusion, and weighted correction based on channel statistics on the multi-layer feature map to obtain a fused feature map. The parameter adaptive module is used to perform target detection or density estimation based on the fused feature map to obtain the recognition result including the target's location, category and confidence level. It selects a set of detection parameters from a preset detection parameter library according to the recognition scenario, and re-infers or post-processes the recognition result based on the selected set of detection parameters to obtain an adjusted recognition result. The tracking and counting module is used to perform multi-target tracking based on the adjusted identification results, establish target identity associations, merge duplicate detection records and calculate the number of targets, and generate the final count result. The display module is used to display the final point result and the recognition box on the user terminal interface.

[0008] In one of the solutions provided in this application, high-frequency detail enhancement, adaptive feature fusion, and weighted correction based on channel statistics are used. In addition, scene-specific detection parameter groups (such as the difference between parameters in the young fruit stage and the mature stage) and multi-target tracking are used to solve the agricultural insurance-specific problems of target characteristic differences at different growth stages and repeated counting of video streams. This overcomes the problem of insufficient adaptability of traditional general detection models in agricultural insurance scenarios. The above processing improves the recognition rate and reduces the point count error, thus achieving high-precision point count targets. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1This is a schematic diagram of an application environment for an intelligent point recognition method for agricultural insurance targets based on front-end artificial intelligence, according to one embodiment of this application. Figure 2 This is a flowchart illustrating a method for intelligent point identification of agricultural insurance targets based on front-end artificial intelligence in one embodiment of this application; Figure 3 This is a schematic diagram of the structure of an intelligent point recognition system for agricultural insurance targets based on front-end artificial intelligence in one embodiment of this application; Figure 4 This is a schematic diagram of the structure of a computer device according to one embodiment of this application. Detailed Implementation

[0011] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] like Figure 1 As shown in the embodiments of this application, a method for intelligent point counting and identification of agricultural insurance targets based on front-end artificial intelligence is provided. In this method, a lightweight artificial intelligence model is run on a user terminal to preprocess and extract features from the acquired images of agricultural insurance targets to generate multi-layer feature maps. The multi-layer feature maps are then subjected to high-frequency detail enhancement, adaptive feature fusion, and weighted correction based on channel statistics to obtain a fused feature map. Based on the fused feature map, target detection or density estimation is performed to obtain a recognition result including the target's location, category, and confidence level. A set of detection parameters is selected from a preset detection parameter library according to the recognition scenario, and the recognition result is re-inferred or post-processed based on the selected detection parameter set to obtain an adjusted recognition result. Based on the adjusted recognition result, multi-target tracking is performed to establish target identity associations, merge duplicate detection records, and calculate the number of targets to generate a final point count result. The final point count result and the recognition box are displayed on the terminal interface, primarily for achieving high-precision point counting of agricultural insurance targets.

[0013] It should be noted that the user terminal can be a mobile phone or other front-end device, and there are no specific restrictions.

[0014] like Figure 2 In one embodiment, a method for intelligent point identification of agricultural insurance targets based on front-end artificial intelligence is provided, comprising the following steps: S101. Run a lightweight artificial intelligence model on the user terminal to preprocess and extract features from the acquired agricultural insurance target images to generate multi-layer feature maps. S102. Perform high-frequency detail enhancement, adaptive feature fusion, and weighted correction based on channel statistics on the multi-layer feature map to obtain a fused feature map; S103. Perform target detection or density estimation based on the fused feature map to obtain an identification result including the target's location, category, and confidence level; S104. Select a set of detection parameters from the preset detection parameter library according to the recognition scenario, and re-infer or post-process the recognition result based on the selected set of detection parameters to obtain the adjusted recognition result. S105. Based on the adjusted identification results, perform multi-target tracking, establish target identity association, merge duplicate detection records and calculate the number of targets, and generate the final point count result; S106. Display the final point count result and the recognition box on the terminal interface.

[0015] The user terminal refers to a mobile device (such as a smartphone or tablet) equipped with a camera and front-end computing capabilities, used to run a lightweight artificial intelligence model and acquire images of agricultural insurance targets, supporting real-time preview and offline inference. The lightweight artificial intelligence model in this application is a deep learning model optimized specifically for agricultural insurance point counting tasks. For example, it can be modified based on lightweight networks such as MobileNetV3, YOLOv5-nano, or YOLOv8-nano, possessing characteristics of small parameter count and low computational cost, and can be directly embedded into front-end code for execution. It should be noted that the modified network is described in subsequent embodiments and will not be elaborated here.

[0016] Multi-layer feature maps are feature representations output by multiple convolutional layers of the backbone network of a lightweight artificial intelligence model (such as MobileNetV3-Small). Different layers correspond to different scales of information: shallow feature maps retain high-frequency details such as target edges and contours, while deep feature maps contain global information such as target category and semantics.

[0017] High-frequency detail enhancement refers to strengthening high-frequency information such as target edges and contours in feature maps to adapt to the recognition of small targets (such as fruits and young animals); adaptive feature fusion refers to fusing feature maps of different scales in a dynamic learning manner rather than fixed fusion, prioritizing the retention of key scale information; channel statistics weighted correction refers to performing global average pooling on the fused feature map to obtain channel statistics, and adjusting the weight of each channel based on the statistics to enhance the distinction between the target and the background.

[0018] The preset detection parameter library is a set of parameters indexed by the scene / growth stage, storing multiple sets of detection parameters (such as confidence threshold, NMS threshold, anchor box size, and image preprocessing parameters). Multi-target tracking refers to establishing the target's motion trajectory by associating the position and appearance features of targets in adjacent frames, merging duplicate detection records of the same target, and avoiding duplicate counting.

[0019] To facilitate understanding, this example uses a specific agricultural insurance scenario for counting. This embodiment focuses on counting apples during the young fruit stage in an apple orchard: farmers / insurance surveyors use a smartphone (user terminal) to run an agricultural insurance app and count the fruit on apple trees in the young fruit stage. At this stage, the apples are small (e.g., 2-3 cm in diameter), green in color similar to the leaves, and densely packed, easily obscuring the fruit. Traditional manual counting has a high error rate. The method provided in this embodiment utilizes a lightweight front-end model to achieve accurate identification of small targets and count without duplication, thus solving the problem of difficult counting during the young fruit stage.

[0020] In this example, the user terminal runs a lightweight AI model. Preprocessing and feature extraction generate multi-layer feature maps. After the agricultural insurance app on the smartphone is launched, it loads the lightweight AI model with embedded front-end code, allowing it to run without relying on the network. The user follows the app's guidance to align the camera with the apple tree canopy, and the camera captures real-time video frames (e.g., 30 frames per second). The model performs preprocessing on the input frames: increasing the image resolution from the default 640x640 to 1280x1280 to preserve the details of the young fruit; applying a transformation matrix that suppresses the green channel and enhances the red channel to weaken leaf interference and highlight the green young fruit; the preprocessed image is input into the model's backbone network, which outputs three sets of multi-layer feature maps at different scales through multiple convolutional layers (e.g., conv2, conv4, conv6 layers). The conv2 layer outputs a shallow feature map (preserving the edges of the young fruit), and the conv6 layer outputs a deep feature map (containing semantic information of the young fruit).

[0021] For the shallow feature map (conv2 layer), depthwise separable convolution is used to reduce computation, and an attention module is used to enhance high-frequency components such as the edges and contours of young fruits, suppress leaf texture interference, and generate a detail-enhanced feature map. The detail-enhanced feature map and the deep feature maps (conv4 and conv6 layers) are input into a multi-scale adaptive fusion module: dynamic weights are calculated for the three sets of feature maps and weighted summation is performed to generate an intermediate fused feature map. Global average pooling is performed on the intermediate fused feature map to obtain the statistics (mean and variance) of each channel. The channel dependencies are learned through two fully connected layers, and the learned global context weights are weighted point by point to the feature map to finally obtain the fused feature map (highlighting the young fruit region and suppressing background noise). Based on the fused feature map, the detection head performs target detection, locates the detection box (location information) of each young fruit in the image, classifies it as an apple young fruit (category information), and outputs the confidence score of each young fruit (such as 0.35, 0.42, etc.); generates the initial recognition result: containing 12 detection boxes, corresponding to 12 candidate young fruits, with confidence scores distributed between 0.28 and 0.55 (some low-confidence candidate boxes are blurry young fruits).

[0022] Based on the user-inputted scene tags for the young apple stage, the app indexes a preset detection parameter library and extracts a set of detection parameters specific to the young apple stage: Confidence threshold: 0.3 (lower than 0.7 for the mature stage, to avoid missing blurry young fruits); NMS threshold: 0.3 (lower than 0.45 for the flowering stage, to separate densely overlapping young fruit detection boxes); Anchor frame size: 32x32, 48x48 (adapting to the small size of young fruits, different from the 64x64 anchor frame for the mature stage); The above parameter set is applied to the initial recognition results: Confidence level filtering: retain candidate boxes with a confidence level ≥ 0.3, and remove two fuzzy candidate boxes with confidence levels of 0.28 and 0.29; NMS filtering: Perform NMS (threshold 0.3) on the remaining 10 candidate boxes and merge two detection boxes with an overlap of more than 30% (corresponding to the same immature fruit). Morphological filtering: Remove one slender detection frame with an aspect ratio greater than 3 (determined as a false leaf detection); The final adjusted recognition result was: 8 valid detection frames, corresponding to 8 young fruits, with no missed detections or false detections.

[0023] For video streams (non-static images) captured by a slowly moving camera, multi-target tracking is performed. From the adjusted recognition results, the position information (center point coordinates, width and height) and appearance feature vector (mean of pixels within the detection box, texture features) of each young fruit detection box are extracted. Between adjacent frames (such as frame t and frame t+1), the target matching score is calculated by IOU spatial matching and appearance feature similarity (IOU > 0.5 and appearance similarity > 0.8 are judged as the same target). Combined with Kalman filtering, the target motion trajectory is predicted. The target trajectory is established for the detection boxes of 5 consecutive frames. It is found that the detection box A in frame t and the detection box A1 in frame t+1 have a matching score of 0.92, and are judged as the same young fruit. Their detection records are merged (only counted once during counting). All target trajectories are traversed, and the number of independent trajectories is counted as 8, generating the final point count result: a total of 8 young apples. According to the rendering logic, the following is overlaid on the camera preview screen in the UI interface of the agricultural insurance app: Recognition box: The positions of 8 young fruits are marked with green rectangles, and the category of young apples is marked inside the box; Counting result: The current count is displayed in real time in the upper right corner of the interface: 8 fruits, and the text is highlighted for easy viewing by users; As can be seen, this embodiment addresses the pain point of difficulty in counting small targets during the young fruit stage by: enhancing high-frequency details, adaptive feature fusion, and weighted correction based on channel statistics; the model runs locally on the mobile device (without network dependence), and the counting results are updated in real time with the video stream, solving the problems of difficulty in counting targets in remote orchards without network access and the low efficiency of traditional manual methods; through scenario-based detection parameter groups (differences between young fruit stage and mature fruit stage parameters) and multi-target tracking, it solves the problems of differences in target characteristics at different growth stages and repeated counting in the video stream specific to agricultural insurance, overcoming the problem of insufficient adaptability of traditional general target detection models in agricultural insurance scenarios. Through the above processing, the recognition rate is improved and the counting error is reduced, achieving high-precision target counting.

[0024] In one embodiment, the lightweight AI model is built based on MobileNetV3, YOLOv5-nano, or YOLOv8-nano network architectures, and is obtained through the following structural optimizations: S01. Based on the comparison result of the absolute value of the weights of each convolutional channel in the network structure with a preset threshold, delete channels that are lower than the threshold and adjust the convolutional kernel connections of the corresponding layers. S02. Convert the floating-point weights and activation calculation units in the lightweight artificial intelligence model into low-bit integer representations, and introduce corresponding scaling parameters into the network structure. The scaling parameters are used to unify the numerical scale between different layers, so that the lightweight artificial intelligence model can be executed in a fixed-point operation mode. S03. Using the trained teacher model as the supervision source, and the probability distribution of the teacher model output or the intermediate layer feature mapping as the soft target, the lightweight artificial intelligence model is trained by minimizing the difference function between the output of the lightweight artificial intelligence model and the soft target.

[0025] In this embodiment, specifically, the lightweight AI model infrastructure refers to an agricultural insurance target detection model based on MobileNetV3, YOLOv5-nano, or YOLOv8-nano networks. It is defined as a deep learning model with minimal computational and memory usage, optimized for the capabilities of front-end terminals (such as smartphones). The initial architecture includes a backbone feature extraction network (such as MobileNetV3-Small) and a target detection head (such as the YOLOv8-nano detection head). The modified lightweight AI model can select optimization techniques for channels based on the importance of the channel weights in each convolutional layer. By deleting redundant channels with low weight contribution, the number of model parameters and computational load are reduced, while preserving detection accuracy as much as possible. Optimization techniques convert 32-bit floating-point weights and activation values ​​in the model to low-bit (such as 8-bit or 16-bit) integers to optimize computational resource efficiency, and introduce scaling parameters (such as scaling factors and offsets) to unify the numerical scale between layers, ensuring that the accuracy loss of fixed-point operations is controllable. Furthermore, the lightweight AI model, acting as a student model, is supervised by a high-precision teacher model. The training of the lightweight student model aims to bring the front-end model close to or reach the level of the back-end AI. This is achieved by minimizing the output difference (soft target loss) between the student and teacher models, while preserving the detection capabilities of the teacher model. It should be understood that the teacher model refers to the high-precision back-end model used for distillation. This can be a non-lightweight agricultural insurance target detection model (such as YOLOv8-large, MobileNetV3-Large+ complex detection head), possessing high recognition accuracy but with a large computational load. It is only used for supervision during the offline training phase of the lightweight AI model in this embodiment.

[0026] This embodiment uses the scenario of counting fattening pigs in a pig farm as the application object to illustrate the implementation process. It should be understood that the agricultural insurance point counting app is run on a regular smartphone (user terminal) to perform real-time point counting of fattening pigs densely packed in pig pens and in various postures (lying down, eating, walking). Although the initial model built based on YOLOv8-nano is relatively lightweight, it still suffers from slow inference speed and high memory consumption when running on a mobile device. This embodiment modifies the model to a front-end usable version with fast inference, low memory consumption, and high accuracy to meet the real-time point counting requirements.

[0027] In this embodiment, YOLOv8-nano is selected as the base network, with CSPDarknet-nano as the backbone network and YOLOv8 lightweight detection head as the detection head. The fattening pig labeled dataset (containing images of different pen lighting, pig postures, and densities) mentioned in the optimization for agricultural insurance scenarios is used to pre-train the basic YOLOv8-nano model, so that the model initially has the ability to detect fattening pigs. The pre-trained YOLOv8-nano model is iterated through all convolutional layers (such as conv1-conv5 layers in the backbone and conv6-conv8 layers in the neck). The sum of the absolute values ​​of the weights of each convolutional channel in each layer is calculated. The larger this value is, the higher the contribution of the channel to feature extraction. Based on the resource efficiency optimization requirements, a preset pruning threshold (such as 0.1, i.e., retaining channels with a sum of absolute weights ≥ 0.1) is set to perform channel filtering on each convolutional layer. For example, the backbone's conv3 layer originally contained 64 channels. The sum of the absolute weights of 12 channels was calculated to be < 0.1, which was determined to be redundant channels and directly deleted. The convolutional kernel connections of adjacent layers are adjusted synchronously. For example, after deleting 12 channels from the conv3 layer, the number of input channels in the subsequent conv4 layer is adjusted from 64 to 52 to ensure the continuity of the network structure. The pruned model is fine-tuned using a fattening pig dataset to recover the accuracy loss caused by pruning. Finally, the number of model parameters and memory usage will be significantly reduced after pruning, thus improving the inference speed. The floating-point weights (32-bit float) and activation units (32-bit float) of the pruned model are converted into 8-bit integer (int8) representations. The specific processing is as follows: For each layer, a scaling factor S_w (S_w = (maximum weight - minimum weight) / (2^8 - 1)) and an offset Z_w (Z_w = -round(minimum weight / S_w)) are calculated for the weights. The floating-point weights are then converted to int8 using the formula w_int8 = round(w_float / S_w) + Z_w. During model inference, a scaling factor S_a and an offset Z_a are calculated in real-time for the activation values ​​of each layer, also converted to int8 for calculation.

[0028] By embedding scaling parameters (i.e., S_w, Z_w, S_a, Z_a) between different layers of the model, the int8 values ​​of different layers are scaled uniformly during computation. For example, before the output activation value (int8) of the conv4 layer is passed to the conv5 layer, it is converted into a floating-point equivalent value by the inverse conversion of S_w and Z_w of the conv5 layer, and then it is calculated with the weights of the conv5 layer to avoid the precision loss caused by the misalignment of values ​​between layers.

[0029] In this embodiment, to meet the requirement of the front-end model approaching the level of the back-end AI, YOLOv8-large is selected as the teacher model. This model has high parameter count, computational cost, and accuracy, but it cannot be run on the front end and is only used for offline distillation. A hybrid loss function is constructed, with the output of the teacher model as the soft objective and the output of the student model (quantized YOLOv8-nano) as the hard objective. The KL divergence (measuring the difference between the two) between the class probability distribution of the student model output and the probability distribution of the teacher model output is calculated, with a weight of 0.7. The YOLO loss between the student model output and the ground truth annotations (fattening pig detection boxes, categories) is calculated, with a weight of 0.3. The total loss = 0.7 × KL divergence loss + 0.3 × YOLO loss. The quantized student model is trained using the fattening pig dataset through distillation. By minimizing the total loss, the student model learns the feature extraction capabilities of the teacher model, thereby improving accuracy and inference speed.

[0030] As can be seen, this embodiment addresses the core contradiction between lightweight front-end models and high accuracy. It reduces the number of parameters through channel pruning, improves inference speed through fixed-point quantization, and enhances accuracy through knowledge distillation, achieving a balance between lightweight deployment and high-precision detection. This aligns with the requirement for front-end models to approach the level of back-end AI. The optimized model improves inference speed on ordinary smartphones, resolving the initial model's slow speed and high resource consumption, meeting the deployment requirements of the agricultural insurance app on mobile devices. The optimized model requires no high-end hardware support and can run on the smartphones of ordinary farmers / surveyors, significantly reducing the implementation cost of agricultural insurance point counting technology and reflecting the design goals of high efficiency, convenience, and cost reduction.

[0031] In one embodiment, the multi-layer feature map is formed by the output of multi-level convolutional layers of the model backbone network; in step S102, that is, the high-frequency detail enhancement, adaptive feature fusion and weighting include: S1021. Enhance the high-frequency components in the feature map that reflect the edge and contour information of the target object; S1022. Weighted fusion of feature maps from different scales according to spatial weight and channel weight to generate intermediate fused feature maps; S1023. Perform global average pooling on the intermediate fused feature map to obtain channel statistics, and perform weighted correction on each channel based on the channel statistics to obtain the final fused feature map.

[0032] In this embodiment, the backbone network refers to the core network structure responsible for feature extraction in a lightweight artificial intelligence model. The backbone network progressively transforms the input image into feature representations of different levels of abstraction through multiple convolutional layers, forming the basis for generating multi-layer feature maps. Multi-layer feature maps are sets of features sequentially output by the multi-level convolutional layers of the backbone network. Different layers of feature maps carry different information: shallow feature maps (outputs of convolutional layers closer to the input layer) retain high-frequency details such as the edges and contours of agricultural insurance targets, while deep feature maps (outputs of convolutional layers closer to the output layer) contain global information such as target category and semantics. For feature maps of different scales, a weighted sum of spatial and channel weights learned dynamically is used to prioritize retaining feature information crucial to the current scene. Channel statistics refer to the statistical results (such as mean and variance) of each channel obtained by performing global average pooling on the fused feature maps, reflecting the contribution of each channel to target recognition. Channel weighted correction refers to adjusting the weights of each feature channel based on the channel statistics, enhancing channels useful for target recognition, weakening redundant channels, and improving the target discrimination of the fused feature maps.

[0033] In this embodiment, the lightweight AI model loaded on the agricultural insurance app on a smartphone allows surveyors to capture real-time images of young forests using their phone's camera. After the images are input into the backbone network, features are extracted step by step through multiple convolutional layers. The second convolutional layer (conv2, shallow layer) outputs a feature map of size 320x320, which preserves high-frequency details such as the edges of the young tree trunk and small branches and leaves, but contains a lot of background noise from weeds. The fourth convolutional layer (conv4, middle layer): outputs a feature map with a size of 160x160, with a moderate level of abstraction, which can distinguish the overall outline of the sapling from part of the background. The 6th convolutional layer (conv6, deep layer): outputs a feature map with a size of 80x80. It has the highest level of abstraction and contains semantic category information such as saplings, weeds, and soil, but loses some small object details. The three sets of feature maps output from layers conv2, conv4, and conv6 are collected to form a multi-layer feature map, which provides input for subsequent enhancement and fusion.

[0034] Enhancement is performed on the shallow feature map (conv2 layer, 320x320) in the multi-layer feature map according to the high-frequency detail injection path: depthwise separable convolution (3x3 kernel) is performed on the conv2 feature map to reduce the computational cost while preserving high-frequency details such as the edges of the sapling trunk and the branching of the leaves, and filtering out some low-frequency background noise; an SE attention module is introduced to calculate the importance weight of each channel of the conv2 feature map. Channels carrying the edge information of the sapling (such as channels 12 and 18) are given high weights (0.8-0.9), and channels carrying the background of weeds (such as channels 5 and 23) are given low weights (0.1-0.2).

[0035] Adaptive fusion is performed on the shallow feature map (conv2 enhanced version) enhanced with high-frequency details, the mid-level feature map (conv4), and the deep feature map (conv6). For each spatial location and channel of the three feature maps, spatial weights and channel weights are calculated separately. In dense sapling areas (such as the center of the image), spatial weights are biased towards the shallow feature map (weight α=0.6) to prioritize the preservation of edge details. In background areas (such as weedy areas in the corner of the image), spatial weights are biased towards the deep feature map (weight γ=0.7) to prioritize the use of semantic information to distinguish the background. Channels that carry the semantics of saplings (such as the 8th channel of conv6 layer) are given high weights (0.8), and redundant channels are given low weights (0.2). The intermediate fused feature map is generated by performing a weighted summation using the formula F_mid=α×conv2 enhanced version+β×conv4+γ×conv6 (α, β, γ are dynamically learned weight tensors, and α+β+γ=1). For the intermediate fused feature map (F_mid), a weighted correction rule based on channel statistics is applied. Global average pooling is performed on F_mid to compress the 160x160 two-dimensional feature map into a 1x1 channel vector (length consistent with the number of channels in the feature map, such as 64 dimensions). This vector represents the channel statistics; the higher the statistic value, the greater the contribution of the corresponding channel to target recognition. The channel statistics are then input into two fully connected layers (the first layer is 64 to 32 dimensions with ReLU activation; the second layer is 32 to 64 dimensions with Sigmoid activation) to learn the correction weights for each channel (values ​​ranging from 0 to 1). For example: Channel carrying the outline of the sapling: Correction weight = 0.9 (enhanced); Channel carrying weed background: Correction weight = 0.2 (weakening); The learned correction weights are multiplied point-by-point with each channel of the intermediate fused feature map (F_mid) to obtain the final fused feature map (F_final). In this feature map, the sapling features are significantly enhanced, and background noise is further suppressed.

[0036] The final fused feature map (F_final, 160x160) is input into the model's target detection head (such as the YOLOv8-nano detection head). This feature map can then be used to accurately locate the saplings and output the recognition results. Compared to the solution without enhanced fusion, the false negative rate for saplings is significantly reduced, fully meeting the requirements for the number of saplings in plantations.

[0037] As can be seen, this embodiment effectively addresses the pain point of low feature discrimination between small targets / complex backgrounds in agricultural insurance. Through high-frequency detail enhancement (complementing details), adaptive fusion (optimizing scale), and channel correction (strengthening targets), it improves the discrimination between saplings and the background, reduces the false negative rate, and meets the requirements for high-precision point counting. Whether it's small targets like saplings and young fruits, or medium to large targets like livestock, the multi-layer feature enhancement fusion logic can improve recognition accuracy by strengthening key features and optimizing the fusion effect, providing a high-quality feature foundation for subsequent target detection and point counting. In one embodiment, the detection parameter library stores multiple sets of detection parameter groups indexed by stage labels. Each set of detection parameter groups includes a confidence threshold, a non-maximum suppression threshold, a set of anchor box sizes, and a set of image preprocessing parameters. In step S104, namely, selecting a set of detection parameters from a preset detection parameter library according to the recognition scenario, and re-inferring or post-processing the recognition result based on the selected set of detection parameters to obtain an adjusted recognition result, the following steps are included: S1041. Determine the current stage based on user input, time information, or image classification results, and output a stage label; index the detection parameter library with the stage label and extract the corresponding detection parameter group; S1042. The detection parameter group is applied to the detection or post-processing process, wherein the confidence threshold is used to filter detection boxes, the non-maximum suppression threshold is used to merge overlapping detection boxes, the anchor box size is used to adjust the candidate box ratio, and the image preprocessing parameters are used to adjust the resolution and color channel weights of the input image. S1043. The result after the detection or post-processing is the identification result after stage adaptive adjustment.

[0038] The detection parameter library is a structured set of parameters uniquely indexed by stage labels. It pre-stores multiple sets of detection parameters adapted to different agricultural insurance identification scenarios (such as different crop growth stages or different livestock density scenarios), forming the basis for stage-based adaptation. Stage labels identify the characteristic scenario in which the current agricultural insurance target is located (e.g., apple flowering stage, wheat heading stage in crop growth; dense pig pen in livestock scenario). These labels can be determined through user input, time information, and image classification, serving as the key basis for indexing the detection parameter library.

[0039] The detection parameter group is a collection of parameters in the detection parameter library that are bound to a single stage label. It contains four core types of parameters: Confidence threshold: The confidence threshold for filtering the detection boxes output by the model. In scenes with ambiguous targets (such as the young fruit stage), the threshold should be lowered to avoid missed detections. Non-maximum suppression (NMS) threshold: The criterion for merging overlapping detection boxes. In scenes with dense targets (such as during flowering season), the threshold should be lowered to avoid mistakenly deleting real targets. Anchor frame size set: Candidate box parameters that fit the typical size of the target. Small target scenes (such as young fruit, flowers) need to be configured with small-scale anchor frames; Image preprocessing parameter set: Parameters to optimize the quality of the input image, including resolution adjustment (preserving details of small targets) and color enhancement (distinguishing between targets and background). Stage adaptive adjustment refers to the logic of determining stage labels, calling corresponding parameters, and applying parameter optimization for detection / post-processing, so that the model outputs recognition results that are adapted to the current scene, thus solving the problem of insufficient accuracy of traditional fixed parameters that are adapted to all scenes.

[0040] For example, this embodiment uses the scenario of counting apple blossoms during the flowering period as an example: Insurance surveyors use a smartphone to run an agricultural insurance app to count the blossoms on apple trees in bloom. At this time, apple blossoms have the characteristics of being small targets (e.g., 2-3cm in diameter), densely distributed (e.g., 10-20 blossoms per branch), and having high contrast with green leaves but easily overlapping. If fixed detection parameters (e.g., resolution 640x640, NMS threshold 0.45, confidence threshold 0.5) are used, the false negative rate will be high due to overlapping blossoms being mistakenly deleted by NMS and loss of details of small blossoms. By using the stage-adaptive parameter adjustment method of this embodiment, the problem of low accuracy in counting dense small targets is solved by calling the apple blossom period-specific detection parameter group.

[0041] In this embodiment, a preset detection parameter library needs to be constructed first. Following the logic of the crop growth cycle-visual feature mapping library, a detection parameter library indexed by stage labels is constructed. For example, the core structure is as follows: Determine the current stage and output the stage label: Initial user input: The surveyor opens the Apple Points function in the Agricultural Insurance App, selects the apple-flowering period in the pop-up window of the growth stage on the interface, and manually selects the apple-flowering period based on on-site observation; Image classification model validation: The app automatically launches a lightweight image classification model to perform coarse classification on the apple tree canopy images captured in real time by the camera. The model identifies visual features of dense white small targets against a green leaf background and outputs a classification result of apple-flowering stage, which is consistent with the user input, thus confirming that the current stage is labeled as apple-flowering stage. The app combines phenological data of apple tree variety (Red Fuji), planting location (Yantai, Shandong), and current date (mid-April) to verify that mid-April is the flowering period for Yantai Red Fuji apples, further confirming that the stage label is correct and preventing users from making incorrect selections.

[0042] The detection parameter library is indexed by stage labels, and the corresponding detection parameter groups are extracted.

[0043] According to the index parameter library rules, the app extracts a unique set of detection parameters from the preset detection parameter library, using apple-flowering period as the index, as follows: Image preprocessing parameters: resolution 1280x1280, red channel weight increased by 20%, green channel weight decreased by 10%; Model inference parameters: confidence threshold 0.3; Anchor frame size set: 32x32, 48x48 (adapts to small flower sizes, avoiding the omission of small flowers by large anchor frames); Post-processing parameters: NMS threshold 0.3 (adapts to scenes with densely overlapping flowers to prevent deletion of the real target due to high overlap).

[0044] The detection parameter set is applied to the detection / post-processing to obtain the adjusted recognition result.

[0045] The extracted apple-flowering period parameter set was embedded into the entire process of image preprocessing → model inference → postprocessing, and the difference in effect between fixed parameters and adaptive parameters was compared: 1. Image preprocessing stage: Applying image preprocessing parameters Fixed parameters (no adaptive): Using 640x640 resolution, no color channel adjustment, some small flower details are compressed and lost; Adaptive parameters (apple-flowering period): The image resolution is increased to 1280x1280, retaining small flower details with a diameter of <3cm; at the same time, the red channel is enhanced and the green channel is suppressed, which improves the contrast between white flowers and green leaves and reduces background interference.

[0046] 2. Model inference stage: Applying confidence threshold and anchor frame size Fixed parameters: Using a 0.5 confidence threshold and 64x64 anchor boxes, a large number of blurry small flowers are filtered out due to confidence < 0.5, and small flowers cannot be detected due to anchor box mismatch; Adaptive parameters: Using a 0.3 confidence threshold (reducing missed detections), and loading small anchor boxes of 32x32 and 48x48 sizes, the model can identify more low-confidence but real small flowers.

[0047] 3. Post-processing: Applying NMS thresholding Fixed parameters: Using a threshold of 0.45 NMS, densely overlapping flower detection boxes (overlap > 30%) were misclassified as the same target, resulting in only 35 valid detection boxes being retained, leading to a high false negative rate. Adaptive parameters: Using a threshold of 0.3 NMS, only detection boxes with overlap > 30% were merged (to avoid false deletion), resulting in 78 valid detection boxes being retained, thus reducing the false negative rate.

[0048] Output adjusted recognition results After applying the above parameters, the results of apple blossom recognition were obtained: 78 valid detection boxes, each corresponding to an apple blossom, including the blossom location (coordinates), category (apple blossom), and confidence level (0.3-0.6). There were no obvious missed detections or false detections, which fully met the point requirements of dense small targets during the flowering period.

[0049] As can be seen, this embodiment addresses the core pain point of adapting fixed parameters to all scenarios. Traditionally, people dislike using a fixed set of parameters to address the entire crop growth cycle / different scenarios for livestock, leading to missed detections of small targets and false deletions of dense targets. This embodiment, through the association of stage labels and parameter libraries, reduces the missed detection rate of flowering flowers, adapting to the need for dynamic changes in the characteristics of the insured object in agricultural insurance scenarios, aligning with the core innovation of optimization for agricultural insurance scenarios. Users do not need to manually adjust complex parameters (such as NMS thresholds and anchor frame sizes); they only need to select a stage label (or the system automatically determines it) to call the optimal parameters, making it particularly suitable for non-professional farmers / surveyors. Through stage-adaptive parameter settings, the model can cover multiple scenarios such as apple flowering, young fruit, and ripening stages, without needing to train a separate model for each scenario, reducing development costs and possessing scalability (the model can adapt to new scenarios, requiring only an update to the parameter library).

[0050] In one embodiment, step S105, the multi-target tracking includes: S1051. Extract the position information and appearance feature vector of each detection box from the adjusted recognition result; S1052. Between adjacent video frames, calculate the target matching score based on the similarity between the spatial displacement of the detection box center point coordinates and the appearance feature vector. S1053. Based on the matching score, establish the inter-frame target correspondence to form the target trajectory; S1054. Merge detection records belonging to the same target trajectory and count them as a single target in the counting statistics to suppress duplicate detection.

[0051] Location information refers to the spatial coordinate data of each object detection box in the adjusted recognition results, including the center point coordinates (x, y) and width and height (w, h) of the detection box. It is the core basis for judging the spatial displacement of the target. Appearance feature vector: visual feature data used to distinguish different targets. Target matching score: the metric for determining the association between targets between adjacent video frames. It can be calculated by weighting spatial displacement matching (IOU, Intersection over Union) and appearance feature similarity (such as cosine similarity). The higher the score, the greater the probability that the two detection boxes belong to the same target (for example, the threshold is usually set to 0.5, and a score ≥ 0.5 is considered a match).

[0052] The target trajectory is established by inter-frame target matching, which is a sequence of the same target's positions in consecutive video frames, ensuring continuous tracking even if the target is briefly occluded. All detection records of the same target trajectory are merged into one, and only counted once during counting.

[0053] To make it easier to understand, we will use a specific application scenario as an example.

[0054] Based on the completed and adjusted recognition results (e.g., 12 fattening pig detection boxes are output in frame t), execute: Location information extraction: For each detection box, read its spatial coordinate data: center point coordinates (x_t,i,y_t,i) (i=1 to 12, representing the i-th detection box in frame t), width and height (w_t,i,h_t,i), and store them as a location vector P_t,i=(x_t,i,y_t,i,w_t,i,h_t,i); Appearance feature vector extraction: For the fattening pig image within each detection box, extract the ear tag position (if present) and key points of the back contour (such as the coordinates of 5 feature points), and convert them into a 128-dimensional vector; Supplementing pixel features: Calculate the pixel mean and texture variance of the RGB three channels within the detection box, and generate a 64-dimensional vector; The features are merged into an appearance feature vector F_t,i = (ear tag / outline 128-dimensional vector + pixel 64-dimensional vector), totaling 192 dimensions, to ensure that the features of different fattening pigs can be distinguished (such as different ear tag numbers and differences in coat color will result in different F_t,i).

[0055] Take two consecutive frames (frame t and frame t+1) and calculate the matching score as follows: Preprocessing for frame t+1: Similar processing is used to obtain the position vector P_{t+1,j} (j=1 to 13, the number of detection boxes varies slightly due to the activity of the pig herd) and the appearance feature vector F_{t+1,j} of frame t+1; Spatial displacement matching (IOU calculation): For P_t,i in frame t and P_{t+1,j} in frame t+1, calculate the intersection-union ratio (IOU) of the two detection boxes = (intersection area of ​​the two boxes) / (union area of ​​the two boxes). The larger the IOU, the smaller the spatial displacement of the target, and the higher the probability that they belong to the same target (e.g., IOU=0.6 means that the two boxes mostly overlap and the spatial matching degree is high). Appearance feature similarity calculation: For F_t,i and F_{t+1,j}, calculate the cosine similarity Sim = (F_t,i・F_{t+1,j}) / (||F_t,i||×||F_{t+1,j}||). The closer Sim is to 1, the more similar the appearance features are, and the higher the probability that they belong to the same target (e.g., Sim = 0.92 means that the two targets have the same ear tags and fur color). Weighted matching score calculation: The score is calculated by weighting spatial matching (0.4 for spatial matching and 0.6 for appearance matching) and the score is calculated as Score = 0.4 × IOU + 0.6 × Sim. For example, for a pair (P_t,1 and P_{t+1,2}), the IOU is 0.5 and Sim is 0.9, so the score is 0.4 × 0.5 + 0.6 × 0.9 = 0.74 (≥ 0.5 threshold, considered a potential match).

[0056] For each target (P_t,i) in frame t, use Kalman filtering to predict its expected position P'{t+1,i} in frame t+1 (based on the target's motion speed and direction estimation), narrowing the matching search range (only searching for P_{t+1,j} within 50 pixels around P'{t+1,i}), thus reducing the computational load; Input the matching scores of all (P_t,i,P_{t+1,j}) into the Hungarian algorithm to solve for the maximum matching score set. That is, each P_t,i matches only 1 P_{t+1,j} (score ≥ 0.5), unmatched P_{t+1,j} are determined as newly entered targets, and unmatched P_t,i are determined as targets leaving the screen; For a successfully matched target: add P_{t+1,j} to the trajectory corresponding to P_t,i (e.g., for a target with trajectory ID=1, add the position P_{t+1,2} at frame t+1). For the new target: Assign a new trajectory ID (e.g., ID=13) and establish a new trajectory; For targets that leave the frame: retain the trajectory record (still included in the total count); Example results: From frame t to t+10, a total of 12 valid trajectories were established (corresponding to 12 fattening pigs), with no new or departing targets.

[0057] Trajectory Traversal and Deduplication: Traverse all valid trajectories (12 in total), with each trajectory corresponding to one fattening pig. Even if a pig is detected in 5 consecutive frames (e.g., from left to right in the shot), all its detection records belong to the same trajectory (ID=5). The number of valid trajectories was counted (12), and the final count result was 12 fattening pigs, completely eliminating the possibility of inflated counts caused by multiple detections of the same pig; if a pig was briefly obscured by other pigs (not detected from frame t+3 to t+5), when it was detected again in frame t+6, it still matched the original trajectory (ID=8) to avoid misjudging it as a new target.

[0058] As can be seen, this embodiment effectively addresses the pain point of repeated counting of dynamic targets in agricultural insurance scenarios. By establishing trajectories and deduplication, the counting error is reduced, further improving the accuracy of high-precision counting. Furthermore, deduplication can be achieved for both fattening pigs and poultry through location + appearance dual-feature tracking, eliminating the need to develop separate tracking logic for different targets.

[0059] In one embodiment, after step S106, that is, after displaying the final point result and the recognition box on the terminal interface, the method further includes: S107. Upload the encrypted target identification summary information to the backend system; the identification summary information includes identification results, count data, timestamp, geographical location information, device identifier, and downsampled summary image; the identification summary information is encrypted using an encryption algorithm before uploading, and the corresponding original image data and intermediate identification cache data are deleted from the local system after uploading.

[0060] In this embodiment, the identification summary information refers to a data set that reflects the core information of the counting results without uploading the original image. It includes the identification results (category and quantity of the target), counting data (final count), timestamp (counting completion time), geographical location information (GPS coordinates), and device identifier (unique ID of the terminal). Its core function is to reduce the risk of privacy leakage while ensuring the validity of the data.

[0061] For example, the encryption algorithm can be a national standard encryption algorithm, used to encrypt the identification digest information. It is a symmetric encryption algorithm with high security and efficiency, and is suitable for the lightweight encryption needs of mobile devices, ensuring that the uploaded data is not stolen or tampered with during transmission.

[0062] Raw image data refers to the original images / video frames of agricultural insurance targets captured by the user terminal's camera. This data is cached only in the device's RAM and must be automatically erased after recognition is complete. It is not uploaded to the backend system and is a core aspect of privacy protection. Intermediate recognition cache data refers to temporary data generated during model inference, including multi-layer feature maps, initial detection boxes, and unadjusted recognition results. This data must be deleted synchronously with the raw images to avoid the risk of residual data leakage.

[0063] This embodiment addresses the core pain point of sensitive data leakage in agricultural insurance. Traditional solutions, which involve uploading raw images, easily leak policyholder privacy (such as pigpen facilities and individual pig characteristics). This embodiment, by uploading only summary information, encryption, and local data deletion, ensures that the raw images are not uploaded to the cloud, the transmitted data is encrypted, and there are no local residues, thus meeting the protection requirements for sensitive data. Furthermore, the summary information includes core business data such as point results, GPS, and timestamps, ensuring that the backend can verify the authenticity of the points, achieving both data usability and privacy security. Compared to uploading the original 1280x1280 image (approximately 2MB per image), the total size of the downsampled 320x320 summary image (approximately 100KB per image) + JSON data (approximately 1KB) is only 101KB, significantly reducing transmission bandwidth consumption.

[0064] In one embodiment, when in a network-free or weak network environment, the method temporarily stores the identification digest information before encryption locally on the terminal; When a network connection is detected to be available, the identification summary information is encrypted using SM4 and uploaded to the backend system; after the upload is completed, the locally stored identification summary information and related intermediate data are automatically cleared.

[0065] A network-free / weak network environment refers to a scenario where the user terminal (such as a smartphone) cannot establish an effective data transmission link. This includes high-frequency agricultural insurance scenarios such as remote farms and mountainous forests. Specifically, it manifests as the disappearance of 4G / 5G signals, WiFi connection failure, or network speed that is too low (unable to complete data upload). In this case, it is impossible to transmit data to the backend system in real time.

[0066] Local temporary storage refers to storing the unencrypted identification digest information in a secure and persistent storage area on the user terminal (non-volatile storage, such as a built-in encrypted partition on the phone). This temporary data is cached in RAM to ensure that the temporary data is not lost after power failure (such as when the phone is turned off). Network connectivity availability means that the terminal detects a stable 4G / 5G signal or a normal WiFi connection (able to access the Internet), meeting the bandwidth and stability requirements for encrypted data upload.

[0067] This solution addresses the core pain point of agricultural insurance where point calculation results cannot be uploaded in offline or weak network environments. Through secure local storage and automatic uploading with a network connection, it ensures that point calculation results are not lost offline and can be traced back when a network connection is available, preventing interruptions to agricultural insurance point calculation services due to network issues. Temporary data is stored in an encrypted partition on the terminal and protected by lightweight AES to prevent local theft. Uploading uses SM4 national cryptographic encryption, and all temporary and intermediate data is immediately cleared after uploading, mitigating offline storage risks and meeting regulatory requirements for sensitive data protection. Users no longer need to manually record point calculation results (such as paper records); the app automatically completes the entire process of temporary storage, network detection, uploading, and clearing, reducing manual operations by surveyors. It also avoids secondary counting issues due to data loss, reducing time and labor costs for both insurance companies and farmers.

[0068] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0069] In one embodiment, an intelligent point identification system for agricultural insurance targets is provided, which corresponds one-to-one with the intelligent point identification method for agricultural insurance targets described in the above embodiments. For example... Figure 3 As shown, the intelligent point identification system for agricultural insurance targets includes: Image acquisition module 301 is used to acquire images of agricultural insurance targets and provide image data to the AI ​​inference module; AI inference module 302 is used to run a lightweight artificial intelligence model on the user terminal to preprocess and extract features from the acquired agricultural insurance target images to generate multi-layer feature maps. The feature enhancement module 303 is used to perform high-frequency detail enhancement, adaptive feature fusion, and weighted correction based on channel statistics on the multi-layer feature map to obtain a fused feature map. The parameter adaptive module 304 is used to perform target detection or density estimation based on the fused feature map to obtain a recognition result including the target's location, category and confidence level. It selects a set of detection parameters from a preset detection parameter library according to the recognition scenario, and re-infers or post-processes the recognition result based on the selected set of detection parameters to obtain an adjusted recognition result. The tracking and counting module 305 is used to perform multi-target tracking based on the adjusted identification results, establish target identity associations, merge duplicate detection records and calculate the number of targets, and generate the final count result; Display module 306 is used to display the final point result and the recognition box on the user terminal interface.

[0070] For specific limitations regarding the intelligent point-counting and identification system for agricultural insurance assets, please refer to the limitations on the intelligent point-counting and identification method for agricultural insurance assets mentioned above, which will not be repeated here. Each module in the aforementioned intelligent point-counting and identification system for agricultural insurance assets can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0071] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown. The computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. When executed by the processor, the computer program implements the functions of an intelligent point-counting system for agricultural insurance assets, or implements the steps of an intelligent point-counting method for agricultural insurance assets.

[0072] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the function of an intelligent point identification system for agricultural insurance targets, or executes the steps of an intelligent point identification method for agricultural insurance targets. To avoid repetition, these steps will not be described here.

[0073] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When executed by a processor, the computer program implements the function of an intelligent point identification system for agricultural insurance targets, or when executed, implements the steps of an intelligent point identification method for agricultural insurance targets. To avoid repetition, these steps will not be repeated here.

[0074] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0075] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0076] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for intelligent point identification of agricultural insurance targets based on front-end artificial intelligence, characterized in that, include: A lightweight artificial intelligence model is run on the user terminal to preprocess and extract features from the acquired images of agricultural insurance targets to generate multi-layer feature maps. The multi-layer feature map is subjected to high-frequency detail enhancement, adaptive feature fusion, and weighted correction based on channel statistics to obtain a fused feature map; Based on the fused feature map, target detection or density estimation is performed to obtain recognition results including the target's location, category, and confidence level; Based on the recognition scenario, a set of detection parameters is selected from a preset detection parameter library, and the recognition result is re-inferred or post-processed according to the selected set of detection parameters to obtain an adjusted recognition result. Based on the adjusted identification results, multi-target tracking is performed, target identity associations are established, duplicate detection records are merged, the number of targets is calculated, and the final point count result is generated. The final score result and the recognition box are displayed on the terminal interface.

2. The method according to claim 1, characterized in that, The lightweight AI model is built on MobileNetV3, YOLOv5-nano, or YOLOv8-nano network architectures and is obtained through the following structural optimizations: Based on the comparison between the absolute value of the weights of each convolutional channel in the network structure and a preset threshold, channels below the threshold are deleted and the convolutional kernel connections of the corresponding layers are adjusted. The floating-point weights and activation units in the lightweight artificial intelligence model are converted into low-bit integer representations, and corresponding scaling parameters are introduced into the network structure. These scaling parameters are used to unify the numerical scale between different layers, enabling the lightweight artificial intelligence model to execute in a fixed-point arithmetic manner. Using a trained teacher model as the source of supervision, and the probability distribution of the teacher model output or the intermediate layer feature map as the soft objective, the lightweight AI model is trained by minimizing the difference function between the output of the lightweight AI model and the soft objective.

3. The method according to claim 1, characterized in that, The multi-layer feature map is formed by the output of multi-level convolutional layers of the model's backbone network; the high-frequency detail enhancement, adaptive feature fusion, and weighting include: Enhance the high-frequency components in the feature map that reflect the edge and contour information of the target object; Feature maps from different scales are weighted and fused according to spatial weights and channel weights to generate an intermediate fused feature map. Global average pooling is performed on the intermediate fused feature map to obtain channel statistics, and each channel is weighted and corrected based on the channel statistics to obtain the final fused feature map.

4. The method according to claim 1, characterized in that, The detection parameter library stores multiple sets of detection parameter groups indexed by stage labels. Each detection parameter group includes a confidence threshold, a non-maximum suppression threshold, an anchor box size set, and an image preprocessing parameter set. The step of selecting a detection parameter group from the preset detection parameter library based on the recognition scenario, and then re-inferring or post-processing the recognition result based on the selected detection parameter group to obtain an adjusted recognition result includes: The current stage is determined based on user input, time information, or image classification results, and a stage label is output; the detection parameter library is indexed using the stage label, and the corresponding detection parameter group is extracted; The detection parameter set is applied to the detection or post-processing process, wherein the confidence threshold is used to filter detection boxes, the non-maximum suppression threshold is used to merge overlapping detection boxes, the anchor box size is used to adjust the candidate box ratio, and the image preprocessing parameters are used to adjust the resolution and color channel weights of the input image. The result after the detection or post-processing is the identification result after stage adaptive adjustment.

5. The method according to claim 1, characterized in that, The multi-target tracking includes: The position information and appearance feature vector of each detection box are extracted from the adjusted recognition results; between adjacent video frames, the target matching score is calculated based on the similarity between the spatial displacement of the center point coordinates of the detection box and the appearance feature vector. Based on the matching scores, inter-frame target correspondences are established to form target trajectories; Detection records belonging to the same target trajectory are merged and counted as a single target in the counting statistics to suppress duplicate detections.

6. The method according to any one of claims 1-5, characterized in that, After displaying the final point count and recognition box on the terminal interface, the method further includes: The encrypted target identification summary information is uploaded to the backend system; the identification summary information includes the identification result, count data, timestamp, geographical location information, device identifier, and downsampled summary image; The identification summary information is encrypted using an encryption algorithm before uploading, and the corresponding original image data and intermediate identification cache data are deleted from the local machine after uploading.

7. The method according to claim 6, characterized in that, When in a network-free or weak network environment, the method temporarily stores the identification digest information before encryption locally on the terminal; When a network connection is detected to be available, the identification digest information is encrypted using SM4 and uploaded to the backend system; After uploading, the locally stored recognition summary information and related intermediate data will be automatically cleared.

8. A smart point identification system for agricultural insurance targets, characterized in that, The device includes: The image acquisition module is used to acquire images of agricultural insurance targets and provide image data to the AI ​​inference module; The AI ​​inference module is used to run a lightweight artificial intelligence model on the user terminal to preprocess and extract features from the acquired agricultural insurance target images to generate multi-layer feature maps. The feature enhancement module is used to perform high-frequency detail enhancement, adaptive feature fusion, and weighted correction based on channel statistics on the multi-layer feature map to obtain a fused feature map. The parameter adaptive module is used to perform target detection or density estimation based on the fused feature map to obtain the recognition result including the target's location, category and confidence level. It selects a set of detection parameters from a preset detection parameter library according to the recognition scenario, and re-infers or post-processes the recognition result based on the selected set of detection parameters to obtain an adjusted recognition result. The tracking and counting module is used to perform multi-target tracking based on the adjusted identification results, establish target identity associations, merge duplicate detection records and calculate the number of targets, and generate the final count result. The display module is used to display the final point result and the recognition box on the user terminal interface.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent point identification method for agricultural insurance targets as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent point identification method for agricultural insurance targets as described in any one of claims 1 to 7.

Images