CN-122023997-A - Watermelon fusarium wilt identification method fusing leaf image characteristics and storage medium

CN122023997ACN 122023997 ACN122023997 ACN 122023997ACN-122023997-A

Abstract

The invention discloses a watermelon fusarium wilt identification method and a storage medium fusing leaf image characteristics, and relates to the technical field of intelligent identification of plant diseases, comprising the steps of collecting leaf original digital images and constructing a multi-level analysis structure; the method comprises the steps of carrying out multichannel spectral separation on images to extract spectral components in different physiological states, generating a composite characteristic map fusing apparent and deep physiological information by utilizing a characteristic growth model, carrying out regional deconstruction and characteristic matching on the map according to a multi-hierarchy structure, calculating pathogen conformity indexes of all regions, fusing to generate an integral infection probability map, carrying out spatial clustering on the map to identify a lesion core region, constructing a topological network for disease development based on a spatial morphological relation of the map, and combining network structure parameters and the probability map to output an identification conclusion. The method can improve the detection sensitivity of early symptoms and can realize analysis of disease spatial distribution mode and development dynamics.

Inventors

QIAN WENHAO
CHEN LIPING
YAO WEI

Assignees

湖州市农业科学研究院(湖州市农业科技发展中心)

Dates

Publication Date: 20260512
Application Date: 20260410

Claims (10)

1. The identification method of watermelon fusarium wilt by fusing the image features of the leaves is characterized by comprising the following steps: Collecting an original digital image of a watermelon leaf to be detected, and constructing a multi-layer analysis structure aiming at the image of the watermelon leaf to be detected; performing multi-channel spectral separation on the original digital image, extracting image components reflecting different physiological states, and introducing the separated image components into a feature growth model to generate a composite feature map about the surface states of the blade; performing regional deconstructment on the composite characteristic map according to the multi-layer analysis structure to obtain a plurality of regional characteristic data blocks, and performing layer-by-layer matching and comparison on the regional characteristic data blocks and a pre-stored watermelon fusarium wilt pathogen characteristic library; According to the matching and comparison results, calculating the pathogen coincidence index of each regional characteristic data block; fusing pathogen coincidence indexes of all the subarea characteristic data blocks to generate an integral infection probability map of the watermelon leaves to be detected; Performing spatial cluster analysis on the integral infection probability map, identifying a potential lesion core area, and constructing a topological network describing disease development according to a morphological evolution relationship of the lesion core area; Based on the structural parameters of the topological network and the integral infection probability map, outputting a recognition conclusion of watermelon fusarium wilt.
2. The method for identifying watermelon fusarium wilt by fusing leaf image features according to claim 1, wherein the steps of collecting an original digital image of a watermelon leaf to be detected and constructing a multi-layer analysis structure for the image of the watermelon leaf to be detected comprise: under a standard illumination environment, acquiring a high-resolution color image covering the whole visible surface of the watermelon leaf to be detected by using an image sensor; in the high-resolution color image, automatically identifying the blade outline and segmenting the blade outline from a background area; establishing an initial analysis grid covering the whole blade area by taking the blade profile as a basic boundary; According to the distribution information of the main pulse and the side pulse of the blade, carrying out self-adaptive subdivision on the initial analysis grid to form a plurality of sub-region grid levels with gradually reduced scales; The multi-layer analysis structure is a space analysis framework formed by the initial analysis grid and sub-region grids of all layers, and each grid unit is associated with the layer where the grid unit is located and space position coordinates.
3. The method for identifying watermelon fusarium wilt by fusing leaf image features according to claim 2, wherein the performing multi-channel spectral separation on the original digital image to extract image components reflecting different physiological states specifically comprises: analyzing the intensity value of each pixel point in the high-resolution color image under a red channel, a green channel and a blue channel; combining the red channel intensity value, the green channel intensity value and the blue channel intensity value of each pixel point into a three-dimensional spectrum vector; Classifying and judging the three-dimensional spectrum vector of each pixel point according to the preset spectrum vector reference range of the physiological states of the various blades; Extracting the pixel points which are judged to belong to the same physiological state from the high-resolution color image, and reserving the original spatial position relation of the pixel points to form an independent image layer; Repeating the distinguishing and extracting process to finally obtain a plurality of image components corresponding to the healthy leaf green tissue, the green-losing tissue, the necrotic tissue and the water stain-shaped tissue respectively.
4. The method for identifying watermelon fusarium wilt by fusing image features of leaves according to claim 3, wherein the step of introducing the separated image components into a feature growth model to generate a composite feature map about the surface state of the leaves specifically comprises the following steps: Calculating, for each image component, its statistical distribution of pixels within each grid cell of the multi-layer analysis structure; the pixel statistical distribution comprises pixel density, average spectrum intensity and standard deviation of spectrum intensity; splicing the pixel statistical distributions from different image components in the same grid unit to form a multi-source feature vector of the grid unit; According to the hierarchical relation of the multi-layer analysis structure, summarizing and fusing the multi-source feature vectors of the lower-layer child grid cells into the upper-layer parent grid cells to which the multi-layer child grid cells belong; Traversing all grid cells of all levels to generate a data set which is completely corresponding to the multi-level analysis structure and contains multi-source feature vectors in each cell, wherein the data set is the composite feature map.
5. The method for identifying watermelon fusarium wilt with fused leaf image features according to claim 4, wherein the performing region deconstructing on the composite feature map according to the multi-layer analysis structure to obtain a plurality of partitioned feature data blocks specifically comprises: Analyzing multi-source feature vectors of each grid cell starting from the highest-level grid of the composite feature map; based on similarity measurement of multi-source feature vectors among grid cells, merging adjacent grid cells with similar features to form a primary feature block; In the next hierarchy, performing the same similarity merging operation on all sub-grid cells belonging to the same primary feature block to form a finer secondary feature block; the process is iterated from top to bottom level to level until reaching the bottommost grid of the multi-level analysis structure; the stable feature blocks eventually formed at each level, along with the multi-source feature vectors for all the grid cells they contain, are packed into a single partitioned feature data block.
6. The method for identifying watermelon fusarium wilt with fused leaf image features according to claim 5, wherein the step-by-step matching and comparing the partitioned feature data block with a pre-stored watermelon fusarium wilt pathogen feature library specifically comprises: the watermelon fusarium wilt pathogen feature library stores standard feature vector sets of different infection stages and different disease parts; Extracting central trend characteristics of multi-source characteristic vectors of all grid cells contained in each partition characteristic data block as representative characteristic vectors of the partition characteristic data block; Performing similarity calculation on the representative feature vector of the partition feature data block and each vector in the standard feature vector set; Recording all standard feature vectors with the similarity exceeding a preset threshold value with the representative feature vectors of the partition feature data blocks, and marking infection stage and disease part information corresponding to all standard feature vectors; And the matching and comparing result is a similarity record corresponding to each partition characteristic data block and an associated infection stage and morbidity part information set.
7. The method for identifying watermelon fusarium wilt by fusing leaf image features according to claim 6, wherein the calculating the pathogen conformity index of each regional feature data block according to the matching and comparing results specifically comprises: For a partition characteristic data block, counting the total number of standard characteristic vectors which are recorded in a matching and comparison result and exceed a preset threshold; analyzing an infection stage corresponding to the standard feature vector recorded in the matching and comparison result, and calculating the concentration of the infection stage on a pathological development time sequence; analyzing the disease parts corresponding to the standard feature vectors recorded in the matching and comparison results, and calculating the coincidence degree of the disease parts and the actual positions of the current partition feature data blocks on the blades; synthesizing the total number, the concentration degree on the pathological development time sequence and the fitness of the actual position, and generating a quantized pathogen coincidence index value through weighted calculation; And the pathogen conformity index value is used for representing the conformity degree of the leaf area corresponding to the partition characteristic data block and the watermelon fusarium wilt typical characteristic.
8. The method for identifying watermelon fusarium wilt by fusing leaf image features according to claim 7, wherein the generating the integral infection probability map of the watermelon leaf to be detected by fusing pathogen conformity indexes of all the partitioned feature data blocks specifically comprises: acquiring a blade area range covered by each partition characteristic data block and a pathogen conformity index value obtained by calculation; The space position of the blade is taken as a coordinate, and the pathogen coincidence degree index value of each subarea characteristic data block is assigned to all pixel points in the coverage range of the subarea characteristic data block; For the pixel points covered by the plurality of regional characteristic data blocks, adopting the average value of a plurality of corresponding pathogen coincidence index values as the assignment of the pixel points; carrying out spatial interpolation smoothing treatment on assignment of all pixel points in the blade area to form a probability distribution curved surface which covers the whole blade and has continuously changing numerical values; The integral infection probability map is the probability distribution curved surface, wherein the numerical value of each position represents the probability that the pixel belongs to the disease spot area.
9. The method for identifying watermelon fusarium wilt by fusing leaf image features according to claim 1, wherein the spatial clustering analysis is performed on the overall infection probability map to identify a potential lesion core area, and a topology network describing disease development is constructed according to a morphological evolution relationship of the lesion core area, and specifically comprises the following steps: Setting a probability threshold value for the overall infection probability map, and marking pixel points with probability values higher than the threshold value as candidate disease spots; Based on the spatial proximity of the candidate disease spots, using a density clustering algorithm to aggregate candidate disease spots that are close to each other into a plurality of independent disease spot clusters; Calculating the geometric center of each disease spot cluster, and taking the geometric center as a circle center, and expanding outwards until all pixel points in the cluster are covered to form an initial boundary of a disease spot core area; Analyzing the area, the shape factor and the relative position between the disease spot core areas and the main veins of the blades, and screening out areas which accord with typical spatial distribution characteristics of watermelon fusarium wilt as final potential disease spot core areas; Extracting key points on the boundary contour of each potential lesion core region, and constructing a region growing path according to the contour evolution trend; Establishing directed connecting edges between the nodes by taking each potential lesion core area as a node and taking the space distance between the areas, the probability distribution continuity and the extending direction of the area growth path as the basis; and constructing a topological network representing the spatial association and the evolution direction of the lesion area through the nodes and the directional connecting edges.
10. A storage medium having stored thereon a computer program which, when executed by a processor, implements a watermelon fusarium wilt identification method of merging leaf image features according to any one of claims 1 to 9.

Description

Watermelon fusarium wilt identification method fusing leaf image characteristics and storage medium Technical Field The invention belongs to the technical field of intelligent identification of plant diseases, and particularly relates to a watermelon fusarium wilt identification method and a storage medium fused with leaf image features. Background Currently, a plant disease identification method based on computer vision mainly depends on RGB images in the visible light band. These conventional techniques make disease determinations by extracting color, texture, or shape features of the leaves and using a classifier. However, the visible light image mainly reflects apparent morphology, is not sensitive enough to the change of physiological states in plants, such as chlorophyll degradation, water stress and other early lesions, and has limited disease recognition capability in a latent stage or an initial stage, so that early warning is difficult to realize. In addition, the feature extraction process of the conventional method is often based on full-image or fixed grid division, and the locality and the spatial dynamic characteristics of occurrence and development of diseases on leaf tissues are not fully considered. Most of the existing disease identification schemes consider the problem as an overall classification task, and output results are simple judgment of health or infection, or the suspected area is segmented by relying on manually set thresholds. Such methods lack the ability to automatically discover and pinpoint the core area of the lesion, and are not effective in distinguishing similar lesions caused by different factors. They fail to model and analyze spatial correlations and evolution trends between lesions, only provide static diagnosis, fail to evaluate the severity and stage of development of the disease, and limit the guiding value of targeted interventions in precision agriculture. There is a need for an identification technique that can reveal the physiological state of the leaf from multiple dimensions and can dynamically resolve the spatial spread pattern of disease. Disclosure of Invention The present invention aims to solve at least one of the technical problems existing in the prior art; therefore, the invention provides a method for identifying watermelon fusarium wilt by fusing leaf image features, which comprises the following steps: Collecting an original digital image of a watermelon leaf to be detected, and constructing a multi-layer analysis structure aiming at the image of the watermelon leaf to be detected; performing multi-channel spectral separation on the original digital image, extracting image components reflecting different physiological states, and introducing the separated image components into a feature growth model to generate a composite feature map about the surface states of the blade; performing regional deconstructment on the composite characteristic map according to the multi-layer analysis structure to obtain a plurality of regional characteristic data blocks, and performing layer-by-layer matching and comparison on the regional characteristic data blocks and a pre-stored watermelon fusarium wilt pathogen characteristic library; According to the matching and comparison results, calculating the pathogen coincidence index of each regional characteristic data block; fusing pathogen coincidence indexes of all the subarea characteristic data blocks to generate an integral infection probability map of the watermelon leaves to be detected; Performing spatial cluster analysis on the integral infection probability map, identifying a potential lesion core area, and constructing a topological network describing disease development according to a morphological evolution relationship of the lesion core area; Based on the structural parameters of the topological network and the integral infection probability map, outputting a recognition conclusion of watermelon fusarium wilt. Further, the method for acquiring the original digital image of the watermelon leaf to be detected and constructing a multi-layer analysis structure aiming at the image of the watermelon leaf to be detected specifically comprises the following steps: under a standard illumination environment, acquiring a high-resolution color image covering the whole visible surface of the watermelon leaf to be detected by using an image sensor; in the high-resolution color image, automatically identifying the blade outline and segmenting the blade outline from a background area; establishing an initial analysis grid covering the whole blade area by taking the blade profile as a basic boundary; According to the distribution information of the main pulse and the side pulse of the blade, carrying out self-adaptive subdivision on the initial analysis grid to form a plurality of sub-region grid levels with gradually reduced scales; The multi-layer analysis structure is a space analysis framework formed by the initia