Search

CN-121998924-A - Agricultural risk target object intelligent point identification method, device, equipment and storage medium

CN121998924ACN 121998924 ACN121998924 ACN 121998924ACN-121998924-A

Abstract

The application relates to the technical field of artificial intelligence and financial science and technology, and discloses a method, a device, equipment and a medium for identifying intelligent points of agricultural risk targets. The method comprises the steps of running a lightweight artificial intelligent model on a user terminal, preprocessing and extracting features of an obtained object image of an agricultural risk object to generate a multi-layer feature image, carrying out high-frequency detail enhancement, self-adaptive feature fusion and weighted correction on the multi-layer feature image to obtain a fusion feature image, carrying out target detection or density estimation based on the fusion feature image, selecting a detection parameter set from a preset detection parameter library according to an identification scene, carrying out reasonement or post-processing on an identification result according to the selected detection parameter set to obtain an adjusted identification result, carrying out multi-target tracking based on the adjusted identification result, establishing target identity association, combining repeated detection records, calculating the object quantity of the object to generate a final point result, and displaying the final point result and the identification frame on a terminal interface.

Inventors

  • YUAN FANG

Assignees

  • 中国平安财产保险股份有限公司

Dates

Publication Date
20260508
Application Date
20260115

Claims (10)

  1. 1. The intelligent point identification method for the agricultural risk targets based on the front-end artificial intelligence is characterized by comprising the following steps of: Operating a light artificial intelligent model on a user terminal, and preprocessing and extracting features of the obtained object image of the agricultural risk object to generate a multilayer feature map; Carrying out high-frequency detail enhancement, self-adaptive feature fusion and weighting correction based on channel statistics on the multi-layer feature map to obtain a fusion feature map; performing target detection or density estimation based on the fusion feature map to obtain a recognition result containing the target position, the category and the confidence coefficient; selecting a detection parameter set from a preset detection parameter library according to the identification scene, and carrying out reasonement or post-processing on the identification result according to the selected detection parameter set to obtain an adjusted identification result; Performing multi-target tracking based on the adjusted identification result, establishing target identity association, merging repeated detection records, calculating the number of targets, and generating a final point result; and displaying the final point result and the identification frame on a terminal interface.
  2. 2. The method of claim 1, wherein the lightweight artificial intelligence model is built based on MobileNetV, YOLOv5-nano or YOLOv-nano network architecture and is obtained by structural optimization of: Deleting channels lower than a threshold value and adjusting convolution kernel connection of a corresponding layer according to the comparison result of the absolute value sum of the weight of each convolution channel of the network structure and a preset threshold value; converting floating point weight and an activation calculation unit in the lightweight artificial intelligent model into a low-bit integer representation form, and introducing corresponding proportion parameters into the network structure, wherein the proportion parameters are used for unifying numerical scales of different layers, so that the lightweight artificial intelligent model is executed in a fixed-point operation mode; The trained teacher model is used as a supervision source, probability distribution or middle layer characteristics outputted by the teacher model are mapped to be soft targets, and the lightweight artificial intelligent model is trained by minimizing a difference function between the lightweight artificial intelligent model output and the soft targets.
  3. 3. The method of claim 1, wherein the multi-layer feature map is formed from multi-level convolution layer outputs of a model backbone network, and wherein the high frequency detail enhancement, adaptive feature fusion and weighting comprises: enhancing high-frequency components reflecting object edges and contour information in the feature map; Weighting and fusing the feature images from different scales according to the space weight and the channel weight to generate an intermediate fusion feature image; and carrying out global average pooling on the intermediate fusion feature map to obtain channel statistics, and carrying out weighted correction on each channel based on the channel statistics to obtain a final fusion feature map.
  4. 4. The method of claim 1, wherein the detection parameter library stores a plurality of detection parameter sets with a phase tag as an index, each detection parameter set including a confidence threshold, a non-maximum suppression threshold, an anchor frame size set, and an image preprocessing parameter set, wherein the selecting a detection parameter set from a preset detection parameter library according to an identification scene, and re-reasoning or post-processing the identification result according to the selected detection parameter set, to obtain an adjusted identification result, comprises: Determining a current stage based on user input, time information or an image classification result and outputting a stage label, indexing the detection parameter library by the stage label and extracting a corresponding detection parameter set; Applying the detection parameter set to a detection or post-processing process, wherein a confidence threshold is used for screening detection frames, a non-maximum suppression threshold is used for merging overlapping detection frames, an anchor frame size is used for adjusting a candidate frame proportion, and an image preprocessing parameter is used for adjusting resolution and color channel weight of an input image; The result after the detection or post-treatment is the recognition result after the stage self-adaptive adjustment.
  5. 5. The method of claim 1, wherein the multi-target tracking comprises: extracting position information and appearance characteristic vectors of each detection frame from the adjusted identification result, and calculating a target matching score between adjacent video frames according to the similarity between the space displacement of the coordinates of the center point of the detection frame and the appearance characteristic vectors; Establishing an inter-frame target corresponding relation according to the matching score to form a target track; The detection records belonging to the same target track are combined, and only a single target object is counted in counting statistics so as to inhibit repeated detection.
  6. 6. The method according to any one of claims 1-5, wherein after displaying the final point result and the identification frame on the terminal interface, further comprising: uploading encrypted object identification summary information to a background system, wherein the identification summary information comprises an identification result, count data, a time stamp, geographical position information, a device identifier and a summary image subjected to downsampling; And the identification abstract information is encrypted by an encryption algorithm before uploading, and corresponding original image data and intermediate identification cache data are deleted from the local after the uploading is completed.
  7. 7. The method of claim 6, wherein the method temporarily stores the identification digest information before encryption at the terminal when in a network-less or weak network environment; when detecting that network connection is available, performing SM4 encryption on the identification summary information and uploading the identification summary information to a background system; And automatically clearing the locally temporarily stored identification abstract information and related intermediate data after the uploading is completed.
  8. 8. An agricultural risk target object intelligent point identification system, characterized in that the device comprises: The image acquisition module is used for acquiring the object image of the agricultural risk object and providing image data for the AI reasoning module; The AI reasoning module is used for running the lightweight artificial intelligent model on the user terminal, preprocessing and extracting the obtained object image of the agricultural risk object to generate a multilayer feature map; the characteristic enhancement module is used for carrying out high-frequency detail enhancement, self-adaptive characteristic fusion and weighting correction based on channel statistics on the multi-layer characteristic map so as to obtain a fusion characteristic map; The parameter self-adaptive module is used for executing target detection or density estimation based on the fusion feature map, obtaining a recognition result comprising the position, the category and the confidence coefficient of a target object, selecting a detection parameter set from a preset detection parameter library according to a recognition scene, and carrying out re-reasoning or post-processing on the recognition result according to the selected detection parameter set to obtain an adjusted recognition result; The tracking counting module is used for executing multi-target tracking based on the adjusted identification result, establishing target identity association, merging repeated detection records, calculating the number of target objects and generating a final point result; And the display module is used for displaying the final point result and the identification frame on a user terminal interface.
  9. 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 the processor implements the steps of the agricultural risk object intelligent point identification method of any one of claims 1 to 7 when the computer program is executed.
  10. 10. A computer readable storage medium storing a computer program, wherein the computer program when executed by a processor implements the steps of the agricultural risk object intelligent point identification method of any one of claims 1 to 7.

Description

Agricultural risk target object intelligent point identification method, device, equipment and storage medium Technical Field The application relates to the technical fields of financial science and technology and artificial intelligence, in particular to a method, a device, equipment and a storage medium for identifying intelligent points of agricultural risk targets. Background In the agricultural insurance business process, counting (counting) of the number of agricultural insurance targets (such as tree fruits, livestock in a farm, artificial forest seedlings and the like) is a core basic link of policy acceptance verification, risk level assessment and claim settlement and loss assessment verification, and the accuracy of the result directly determines the reliability of agricultural insurance business data so as to influence rights and interests of insurance companies and insurance applicant. On one hand, the traditional manual checking mode is limited by manual subjective judgment errors, visual fatigue and scene interference (such as fruit shielding by branches and leaves of crops and dense livestock activities), the point error rate of small targets (such as flowers of fruit trees in the flowering period and dense fattening pigs in the housing) or dense-distribution targets (such as flowers of the fruit trees in the colony period and dense fattening pigs in the housing) is higher, the basic requirement of agricultural insurance business on point accuracy cannot be met, and the finally output point result still has obvious deviation and cannot be used as reliable data basis of agricultural insurance business. In summary, in the existing agricultural risk standard object point number technology, the defect that the high-precision standard object point number is the most core cannot be realized, which directly leads to frequent settlement and verification, risk assessment distortion and settlement and damage disputes in agricultural insurance service, and severely restricts standardization and efficient development of agricultural insurance service. Disclosure of Invention The application provides a method, a device, equipment and a storage medium for identifying intelligent points of agricultural risk targets, which are used for solving the technical problem that the high-precision target point number cannot be realized. An agricultural risk target intelligent point identification method based on front-end artificial intelligence comprises the following steps: Operating a light artificial intelligent model on a user terminal, and preprocessing and extracting features of the obtained object image of the agricultural risk object to generate a multilayer feature map; Carrying out high-frequency detail enhancement, self-adaptive feature fusion and weighting correction based on channel statistics on the multi-layer feature map to obtain a fusion feature map; performing target detection or density estimation based on the fusion feature map to obtain a recognition result containing the target position, the category and the confidence coefficient; selecting a detection parameter set from a preset detection parameter library according to the identification scene, and carrying out reasonement or post-processing on the identification result according to the selected detection parameter set to obtain an adjusted identification result; Performing multi-target tracking based on the adjusted identification result, establishing target identity association, merging repeated detection records, calculating the number of targets, and generating a final point result; and displaying the final point result and the identification frame on a terminal interface. An agricultural risk object intelligent point identification system, the device comprising: The image acquisition module is used for acquiring the object image of the agricultural risk object and providing image data for the AI reasoning module; the AI reasoning module is used for running the lightweight artificial intelligent model on the user terminal, preprocessing and extracting the obtained object image of the agricultural risk object to generate a multilayer feature map; the characteristic enhancement module is used for carrying out high-frequency detail enhancement, self-adaptive characteristic fusion and weighting correction based on channel statistics on the multi-layer characteristic map so as to obtain a fusion characteristic map; The parameter self-adaptive module is used for executing target detection or density estimation based on the fusion feature map, obtaining a recognition result comprising the position, the category and the confidence coefficient of a target object, selecting a detection parameter set from a preset detection parameter library according to a recognition scene, and carrying out re-reasoning or post-processing on the recognition result according to the selected detection parameter set to obtain an adjusted recognition result; The trac