CN-121982705-A - Ginger stem leaf plant diseases and insect pests identification optimization method based on image identification

CN121982705ACN 121982705 ACN121982705 ACN 121982705ACN-121982705-A

Abstract

The invention discloses an image recognition-based ginger stem and leaf disease and pest recognition optimization method, in particular relates to the technical field of image data processing, and aims to solve the problem that the long-term recognition accuracy is reduced due to the fact that the existing fixed model cannot adapt to slow drifting of field characteristics; the method quantifies the characteristic drift degree by monitoring the topological structure stability of the model characteristic space, judges whether the model needs to be updated according to the characteristic drift degree, positions a core characteristic region causing model confusion based on historical misjudgment data when the model needs to be updated, directionally collects a new sample in a corresponding field region, optimizes model parameters, achieves prospective assessment and accurate updating of the model cognitive state, and guarantees long-term stability and high efficiency of the identification system in a dynamic environment.

Inventors

CHEN CHAO
ZHU JIABAO
REN XUEXIANG
JIANG BENLI
LU XIANYONG

Assignees

安徽省农业科学院经济作物研究所

Dates

Publication Date: 20260505
Application Date: 20260408

Claims (10)

1. The ginger stem and leaf pest identification optimization method based on image identification is characterized by comprising the following steps of: S1, acquiring a ginger stem leaf digital image in a current period, and identifying the ginger stem leaf digital image by using a deployed classification model to obtain a current identification result and a corresponding current feature set; s2, acquiring a plurality of historical feature sets generated by the classification model in a plurality of continuous historical periods, and respectively carrying out topological structure analysis on the plurality of historical feature sets to obtain respective historical topological structure description information; s3, comparing the current feature set with the historical topological structure description information, and calculating to obtain the topological difference degree representing the deviation degree of the current feature set structure; S4, judging whether the classification model needs to be updated according to the topology difference degree; s5, when updating is needed, identifying a core feature fuzzy area which causes confusion of the classification model based on misjudgment data in the current identification result and the historical identification result, and directionally acquiring a new training sample image from an actual field area corresponding to the core feature fuzzy area; and S6, extracting features from the new training sample image, and performing parameter optimization on the classification model to finish updating.
2. The optimization method for identifying plant diseases and insect pests of ginger stem and leaf based on image identification as set forth in claim 1, wherein S1 includes: acquiring a plurality of ginger stem leaf digital images covering different illumination conditions in a current period; Respectively inputting a plurality of ginger stem leaf digital images into deployed classification models, outputting the disease and pest category of each ginger stem leaf digital image by the classification models as a current recognition result, and synchronously extracting the activation values of preset layers in the classification models to form image feature vectors of the corresponding images; the set of image feature vectors of all images belonging to the current period is determined as the current feature set.
3. The optimization method for identifying plant diseases and insect pests of ginger stem and leaf based on image identification as set forth in claim 1, wherein S2 includes: Acquiring a plurality of history feature sets stored in a plurality of continuous history periods by a classification model, wherein each history feature set comprises history feature vectors of all images in the corresponding history period; For each history feature set, calculating the distance between any two history feature vectors in the history feature set, and constructing a graph structure for representing the connection relation between the history feature vectors according to a distance threshold; And analyzing the quantity and scale distribution of connected components of the graph structure corresponding to each historical characteristic set, and generating historical topological structure description information for describing the topological morphology of the characteristic space of each historical period.
4. The optimization method for identifying plant diseases and insect pests of ginger stem and leaf based on image identification according to claim 1, wherein S3 comprises: based on the distance between any two current feature vectors in the current feature set, and according to the distance threshold value which is the same as the description information of the constructed historical topological structure, constructing a current graph structure corresponding to the current feature set; Analyzing the quantity and the scale distribution of connected components of the current graph structure to obtain the description information of the current topological structure; matching and matching the current topological structure description information with each historical topological structure description information, and calculating the distribution distance between the description information on the multidimensional feature; And carrying out aggregation calculation on the distribution distances corresponding to all the history periods to obtain the topological difference degree representing the deviation degree of the current feature set structure.
5. The optimization method for identifying plant diseases and insect pests of ginger stem and leaf based on image recognition according to claim 4, wherein the distribution distance is obtained by calculating the statistical distance between the current topological structure description information and each historical topological structure description information on the scale distribution of connected components.
6. The optimization method for identifying plant diseases and insect pests of ginger stem and leaf based on image identification as set forth in claim 1, wherein S4 includes: acquiring a current topological difference degree and a historical topological difference degree sequence corresponding to a plurality of continuous historical periods; analyzing the statistical distribution characteristics of the historical topological difference sequence to determine a difference fluctuation range; And comparing the current topological difference degree with the difference degree fluctuation range, and judging that the classification model needs to be updated when the current topological difference degree continuously exceeds the difference degree fluctuation range.
7. The optimization method for identifying plant diseases and insect pests of ginger stem and leaf based on image identification according to claim 1, wherein S5 comprises: Extracting misjudgment feature vectors corresponding to samples misjudged by the classification model in the current recognition result and the historical recognition result; in the feature space, performing density clustering on the misjudgment feature vectors, and identifying a region in which the misjudgment feature vectors are densely distributed as a core feature fuzzy region; determining a geographic range to be sampled in an actual field area according to the position information of the original image corresponding to the misjudgment feature vector in the core feature fuzzy area; And acquiring a new ginger stem and leaf digital image in a determined geographic range as a new training sample image.
8. The optimization method for identifying ginger stem and leaf diseases and insect pests based on image identification according to claim 7, wherein density clustering adopts a density-based noise application spatial clustering method, and a core feature fuzzy region is identified by setting a neighborhood radius and a minimum sample number parameter.
9. The optimization method for identifying plant diseases and insect pests of ginger stem and leaf based on image identification as set forth in claim 1, wherein S6 includes: extracting image feature vectors from the new training sample images to form a new sample feature set; combining the new sample feature set with part of sample feature vectors selected from the historical feature set to form a model optimization training set; performing parameter iterative optimization on the classification model based on the model optimization training set; And replacing the deployed classification model by using the classification model after parameter optimization to finish updating.
10. The method for identifying and optimizing plant diseases and insect pests of ginger stem and leaf based on image identification according to claim 9, wherein constraint terms are added in a loss function in the process of parameter iterative optimization to limit the variation amplitude of the output result of the classification model after parameter optimization on the historical feature set relative to the deployed classification model.

Description

Ginger stem leaf plant diseases and insect pests identification optimization method based on image identification Technical Field The invention relates to the technical field of image data processing, in particular to a ginger stem leaf plant diseases and insect pests identification optimization method based on image identification. Background In the field of ginger planting, timely and accurately identifying stem and leaf diseases and insect pests is a precondition for effective prevention and control. In the prior art, an image recognition method can be adopted for auxiliary recognition, the quality is improved by collecting digital images of ginger stems and leaves, preprocessing the images, extracting characteristic information related to diseases and insect pests in the images, discriminating by utilizing a classification model which is trained in advance based on sample data, and finally outputting a classification recognition result of the diseases and insect pests. However, once the classification model for discriminating in the method is trained, the internal parameters and decision rules thereof are kept fixed, and in the actual production environment, pathogenic bacteria population or external conditions of the ginger plant diseases and insect pests may change, so that the symptom features presented on the image thereof may be subject to insignificant dynamic evolution, and the fixed classification model is difficult to adapt to such slow feature drift, so that the accuracy of identifying the actual plant diseases and insect pests in the field by the model gradually decreases with the passage of time, and the agricultural management requirement of long-term and stable monitoring of the plant diseases and insect pests cannot be met. Disclosure of Invention In order to overcome the defects in the prior art, the invention provides an image recognition-based ginger stem and leaf pest and disease damage recognition optimization method for solving the problems in the prior art. In order to achieve the above purpose, the present invention provides the following technical solutions: A ginger stem leaf plant diseases and insect pests identification optimization method based on image identification comprises the following steps: S1, acquiring a ginger stem leaf digital image in a current period, and identifying the ginger stem leaf digital image by using a deployed classification model to obtain a current identification result and a corresponding current feature set; s2, acquiring a plurality of historical feature sets generated by the classification model in a plurality of continuous historical periods, and respectively carrying out topological structure analysis on the plurality of historical feature sets to obtain respective historical topological structure description information; s3, comparing the current feature set with the historical topological structure description information, and calculating to obtain the topological difference degree representing the deviation degree of the current feature set structure; S4, judging whether the classification model needs to be updated according to the topology difference degree; s5, when updating is needed, identifying a core feature fuzzy area which causes confusion of the classification model based on misjudgment data in the current identification result and the historical identification result, and directionally acquiring a new training sample image from an actual field area corresponding to the core feature fuzzy area; and S6, extracting features from the new training sample image, and performing parameter optimization on the classification model to finish updating. Further, S1 includes: acquiring a plurality of ginger stem leaf digital images covering different illumination conditions in a current period; Respectively inputting a plurality of ginger stem leaf digital images into deployed classification models, outputting the disease and pest category of each ginger stem leaf digital image by the classification models as a current recognition result, and synchronously extracting the activation values of preset layers in the classification models to form image feature vectors of the corresponding images; the set of image feature vectors of all images belonging to the current period is determined as the current feature set. Further, S2 includes: Acquiring a plurality of history feature sets stored in a plurality of continuous history periods by a classification model, wherein each history feature set comprises history feature vectors of all images in the corresponding history period; For each history feature set, calculating the distance between any two history feature vectors in the history feature set, and constructing a graph structure for representing the connection relation between the history feature vectors according to a distance threshold; And analyzing the quantity and scale distribution of connected components of the graph structure corresponding