CN-122024282-A - Pest feature extraction method, system and equipment based on domain knowledge guided large model

CN122024282ACN 122024282 ACN122024282 ACN 122024282ACN-122024282-A

Abstract

A pest feature extraction method, system and equipment based on domain knowledge guiding large model comprise the steps of constructing a standardized pest knowledge base fusing Wikipedia, professional books and papers, adopting a two-stage strategy, wherein a high-distinction structured feature phrase is extracted from the knowledge base by utilizing a large language model in the first stage, individual core features are extracted through a multi-mode large model by combining pest images, artificial labeling frames and feature phrase prompts in the second stage, cross-mode cosine similarity alignment and confidence-distinction-matching degree multidimensional scoring mechanisms are introduced, and high-quality features are screened.

Inventors

WANG RONGFANG
GUO JIAXUAN
Song Liangdong
LI JIAHAO
JIAO CHANGZHE

Assignees

西安电子科技大学

Dates

Publication Date: 20260512
Application Date: 20260129

Claims (10)

1. The pest characteristic extraction method based on the domain knowledge guiding large model is characterized by comprising the following steps of: S1, acquiring pest basic optical image data through multiple channels, and performing quality control on an image by adopting a definition quantitative screening formula to construct an image data set covering various scenes and forms; S2, constructing a multidimensional hierarchical expertise collecting system, screening authoritative pest knowledge, integrating pest biological information in Wikipedia, authoritative professional books and historical academic treatises, and establishing a standardized knowledge database after agricultural expert calibration; S3, extracting a first-stage characteristic phrase, namely extracting a structural core characteristic phrase according to the life cycle stage of pests by using a large language model based on the standardized knowledge database established in the step S2, calculating the characteristic distinguishing degree to screen high-identification-degree characteristics, and constructing a characteristic phrase database; s4, performing second-stage single-target feature extraction, namely inputting pest basic optical image data meeting definition requirements, a manually marked identification frame and feature phrases corresponding to life cycle stages, extracting individual core features of pests through a multi-mode large model, and calculating feature confidence to quantify extraction reliability; s5, matching and matching cross-modal features, namely measuring the matching degree of the image feature vector and the text feature vector based on a cosine similarity model, eliminating repeated or invalid information, and generating a preliminary feature description result; and S6, feature quality evaluation optimization, namely screening the preliminary result through a multi-dimensional comprehensive scoring function integrating confidence, discrimination and matching degree, retaining high-value features and outputting final pest feature description.
2. The pest feature extraction method according to claim 1, wherein in step S1, the multi-channel acquisition of pest base optical image data includes integrating a public image captured by a web crawler, a field image provided by an agricultural co-worker, and a high quality illustration extracted from agricultural field color printed books and academic papers, based on an IP102 dataset.
3. The pest feature extraction method according to claim 1, wherein in step S2, the standardized knowledge database covers a plurality of types of corn and wheat pests including fall armyworms, cotton bollworms, corn borers, leaf beetles, aphids, spider mites, and gall midges, and forms a unified description of field verification and information for a dispute information organization agricultural expert team in terms of body type, body color, body surface structure, and appendage features at the full life cycle stage of eggs, larvae, pupae, and adults.
4. The pest feature extraction method according to claim 1, wherein in step S3, the large language model inputs a preprocessed expertise text, outputs a noun phrase containing a subject in the form of "feature subject+feature attribute", and extracts a plurality of feature phrases per life cycle stage of each type of pest.
5. The pest feature extraction method according to claim 1, wherein in step S4, the multi-modal large model is driven to generate at least 5 core features covering the shape, color, and texture dimensions by inputting the basic optical image data of the pest, the artificially marked marking box, and the feature phrases of the corresponding stages acquired in the first stage as prompt information.
6. The pest feature extraction method according to claim 1, wherein in step S5, the image feature vector is extracted through a deep convolutional neural network, the text feature vector is extracted through a pre-training model, and the matching degree of the image feature vector and the text feature vector is measured by adopting cosine similarity, so as to determine cross-modal consistency.
7. The pest characteristics extraction method according to claim 1, wherein in step S6, the multi-dimensional integrated scoring function is defined as: Wherein, the For the feature quality score (value range 0-1), For the confidence level of the feature, For the purpose of distinguishing the characteristics, To cross-modality match similarity, the weight coefficient is set to (Satisfy the following ) Pressing down Descending order of arrangement before reservation High value features The quality of the features is ensured.
8.A pest feature extraction system based on domain knowledge guided large models, implementing the method of any one of claims 1 to 7, comprising: An image acquisition and screening module for performing step S1 of claim 1; a expertise construction module for performing step S2 of claim 1; a first stage feature phrase extraction module for performing step S3 of claim 1; A second stage single object feature extraction module for performing step S4 of claim 1; A cross-modality matching alignment module for performing step S5 of claim 1; A feature quality assessment optimization module for performing step S6 of claim 1.
9. The pest feature extraction system of claim 8, wherein the system is applied to intelligent recognition of corn and wheat pests and is extended to pest feature extraction of rice and cotton crops, and is deployed on a PC platform with a high-performance GPU and calls a large language model and a multi-mode large model through an API.
10. A pest characteristic extraction device for guiding a large model based on domain knowledge, comprising: A memory for storing a computer program, data and a model; A processor for implementing the method of any one of claims 1 to 7 when executing the computer program, and the pest feature extraction method of any one of claims 1 to 7 based on the domain knowledge guided large model.

Description

Pest feature extraction method, system and equipment based on domain knowledge guided large model Technical Field The invention belongs to the technical field of intelligent identification of agricultural pests, and particularly relates to a pest feature extraction method, system and equipment based on a domain knowledge guiding large model. Background Agriculture is used as basic industry for guaranteeing global grain safety, and the production efficiency and the output quality of the agriculture are directly related to stable social and economic development. The wide spread and invasion of agricultural pests are core bottlenecks for restricting the improvement of agricultural productivity, the annual global crop yield loss caused by the pests is up to 20% -30% according to the national Food and Agricultural Organization (FAO) report, and the accurate and efficient pest identification is a key premise for constructing a modern monitoring and early warning system and realizing accurate prevention and control. With the rapid development of intelligent agriculture, computer Vision (CV), deep learning and large model technology have become core supports for breaking through the limitations of traditional manual identification and improving the intelligent level of pest identification. Under the promotion of computer vision and deep learning technology, the field of agricultural pest identification has been shifted from the traditional image processing method to an intelligent algorithm model based on data driving. Early methods based on Convolutional Neural Network (CNN), such as improved AlexNet, resNet, obtain preliminary results in pest image classification tasks by virtue of strong local feature extraction capability, but the models have the problems of insufficient global information capture and weak adaptability to complex backgrounds, and are difficult to cope with actual scene challenges such as pest camouflage, morphological variation and the like. In recent years, the architectures such as Vision Transformer (ViT), swin Transformer and the like realize global feature modeling through a multi-head self-attention mechanism, and the architectures such as Mamba and the like are based on a new paradigm of a State Space Model (SSMs), so that the capturing capacity of long-distance dependency is further improved, and a new technical path is provided for fine-grained pest feature extraction. Meanwhile, the rise of the large model provides powerful power for cross-modal information fusion, and the high-performance large models such as DEEPSEEK R and the like can integrate multi-source data characteristics by virtue of the excellent generalization capability of the large models, so that pest identification is promoted to develop from a single-mode fusion direction to a multi-mode fusion direction. While the technology is continuously iterated, the current agricultural pest identification and feature extraction still face a plurality of data bottlenecks, namely firstly, the existing model depends on a single public data set (such as IP 102), scene coverage is limited and a quality screening mechanism is lacked, so that the model is insufficient in actual scenes such as complex background and illumination change in fields, meanwhile, data are not deeply combined with agricultural expertise, and accurate feature characterization is difficult to support, secondly, a feature extraction layer, most methods still stay in single-mode visual feature extraction or only simply splice cross-mode information, and a systematic fusion mechanism guided by expertise is lacked, so that pest core biological features cannot be accurately captured, feature distinction is low, and pest identification difficulties with high similarity between classes are difficult to deal with. Therefore, integration of multisource data and authoritative expertise is needed, a set of accurate, efficient and robust agricultural pest feature extraction system is constructed by means of a large model and multi-mode fusion technology, the limitation of the prior art in data quality, feature accuracy and engineering practicability is broken through, and powerful data support is provided for core applications such as pest identification, monitoring and early warning in intelligent agriculture. 2) Prior art solution The existing pest identification technology mainly relies on single-mode optical image analysis or traditional manual feature extraction (such as SIFT and HOG) and deep learning feature extraction methods [ Peng Gongxing, xu Huiming and Liu Hua ], based on a lightweight crop pest identification model [ J ] of improved ShuffleNet V, agricultural engineering journal, 2022,38 (11): 161-170 ], [ Chen Jiong and Liu Jianhua ], improved YOLOv5 is used for rice pest detection [ J ], [ information technology and informatization, 2023 (7): 165-171 ]. However, three major core challenges exist in actual agricultural scenes, namely, farmland environments h