Image recognition-based method for screening of superior individual trees

Abstract

The application discloses a forest tree excellent single plant screening method based on image recognition, comprising the following steps: step one: collecting whole plant images and local close-up images of the same forest tree single plant in the breeding stage and establishing a pairing relationship; step two: performing regional normalization to obtain whole plant regions and organ regions; step three: mapping to an improved AlexNet shared convolution main body to generate a basic phenotype feature map and establish a judgment path set; step four: separating and extracting to form ornamental representation, growth representation and stress resistance representation; step five: forming double-view corresponding representation; step six: performing Gromov-Wasserstein optimal transport structure matching to form double-view coupled representation; step seven: jointly judging to generate single plant breeding value representation; and step eight: calculating three types of probabilities and determining a screening result. The application improves the accuracy and structure matching capability of forest tree excellent single plant screening.

Description

Technical Field

[0001] This invention relates to the field of forest tree breeding technology, and in particular to a method for screening superior individual trees based on image recognition. Background Technology

[0002] With the increasing demand for breeding new varieties of ornamental greening trees, image recognition and screening technology for superior individual trees has received widespread attention. Existing breeding screening methods mainly rely on manual visual inspection, experience-based judgment, and simple trait comparison for identifying superior individual trees. However, these methods generally suffer from the following problems in practical applications:

[0003] The collected whole-plant images and close-up images are easily affected by changes in lighting, foliage occlusion, and background interference, resulting in unstable expression of key phenotypic information such as plant shape, flower color, leaf color, and lesions, making it difficult to support reliable screening of a large number of candidate plants. The morphological information of the whole plant and the detailed information of organs are scattered in different images. Existing processing methods usually only analyze a single view independently, making it difficult to establish the correspondence between the whole plant view and the local organ view. This easily leads to a disconnect between the judgment of overall growth and the judgment of local ornamental traits. For complex phenotypic data with coexisting ornamental features, growth features, and stress resistance features, traditional convolutional recognition methods often lack an effective separation mechanism, causing different judgment elements to be mixed in the feature space, affecting the accuracy of the probability assessment of superior plants and the stability of the screening results.

[0004] Therefore, how to provide a method for screening superior individual trees based on image recognition is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an image recognition-based method for screening superior individual trees. This invention combines an improved AlexNet shared convolutional backbone, decision path construction, and Gromov-Wasserstein optimal transmission structure matching to jointly identify and screen whole-tree images and local close-up images of individual trees during the breeding stage. It has the advantages of high screening efficiency, high decision accuracy, and strong breeding adaptability.

[0006] The image recognition-based method for screening superior individual trees according to an embodiment of the present invention includes the following steps:

[0007] Step 1: Collect whole-tree images and close-up images of individual trees in the same forest during the breeding stage, establish pairing relationships according to individual tree identifiers, and form a candidate individual tree image set;

[0008] Step 2: Perform region normalization on the candidate single-plant image set to determine the whole-plant region and organ region;

[0009] Step 3: Map the whole plant region and organ region to the improved AlexNet shared convolutional backbone to generate basic phenotypic feature maps. On the basic phenotypic feature maps, establish the ornamental judgment path, growth judgment path, and stress resistance judgment path to form a set of judgment paths.

[0010] Step 4: Perform separation and extraction on the basic phenotypic feature maps along the decision path set to form ornamental features, growth features, and stress resistance features;

[0011] Step 5: Map the ornamental characteristics and growth characteristics to the whole plant view determination channel, and map the ornamental characteristics and stress resistance characteristics to the local organ view determination channel to form a dual-view corresponding characterization.

[0012] Step 6: Construct a whole-plant relationship matrix and an organ relationship matrix for the dual-view correspondence representation, and use Gromov-Wasserstein optimal transmission execution structure matching to form a dual-view coupled representation;

[0013] Step 7: Perform joint determination on the dual-view coupled representation to generate a single-plant breeding value representation;

[0014] Step 8: Calculate the probability of a superior breeding single tree, the probability of a potentially superior single tree, and the probability of an ordinary single tree based on the single tree breeding value characterization. Determine the screening results of superior forest trees based on the maximum value among the three probabilities.

[0015] Optionally, step one specifically includes:

[0016] Individual trees in the seedling stage, rapid growth stage, flowering stage, and leaf color viewing stage of the new ornamental greening tree breeding nursery were selected, and a collection object set was formed according to the individual tree identifiers corresponding to the individual trees;

[0017] Collect full-tree images from the top view, front view, and side view of individual trees in the collection object set to form a full-tree image set;

[0018] Close-up images of the flower area of ​​individual trees in the flowering period of the collection object group were collected, and close-up images of the leaf area of ​​individual trees in the leaf color viewing period of the collection object group were collected to form a set of close-up images;

[0019] The whole-plant image set and the partial close-up image set are merged according to the individual plant identifier, and the pairing relationship between the whole-plant image and the partial close-up image is established to form a candidate single-plant image set.

[0020] Optionally, step two specifically includes:

[0021] Foreground localization is performed on the top-view, front-view, and side-view full-tree images in the candidate single-tree image set to extract the main boundary of the forest single tree and form a full-tree candidate region set.

[0022] Perform region clipping, size regularization, and coordinate unification on the whole plant candidate region set to form the whole plant region;

[0023] Organ localization is performed on local close-up images of flower regions and leaf regions in the candidate single plant image set to extract the main boundaries of flower and leaf bodies and form a candidate organ region set.

[0024] Perform region clipping, size regularization, and coordinate unification on the candidate organ regions to form organ regions.

[0025] Optionally, step three specifically includes:

[0026] The whole-plant region and organ region are introduced into the improved AlexNet shared convolutional backbone according to a unified input scale. The improved AlexNet shared convolutional backbone sets up convolutional mapping layers, pooling mapping layers, and attention focusing layers.

[0027] The whole plant region and organ region are subjected to texture response extraction, color response extraction and boundary response extraction along the convolutional mapping layer, and response compression and dominant position preservation are performed along the pooling mapping layer to complete the hierarchical mapping in a unified feature space.

[0028] Within a unified feature space, attentional focus enhancement is applied to the flower color location, leaf color location, flower shape location, plant shape location, and lesion location to maintain the distinct distribution of flower color response, flower shape response, leaf color response, plant shape response, and lesion response;

[0029] The enhanced response is unfolded and arranged according to its spatial and channel positions in a unified feature space to form a basic phenotypic feature map.

[0030] An ornamental evaluation path was established based on the correlation distribution of flower color response, flower shape response, leaf color response, and plant shape response in the basic phenotypic characteristic diagram. A growth evaluation path was established based on the correlation distribution of plant height response, ground diameter response, branch response, and cover response in the basic phenotypic characteristic diagram. A stress resistance evaluation path was established based on the correlation distribution of thickness response, disease-free rate response, and lesion response in the basic phenotypic characteristic diagram, thus forming a set of evaluation paths.

[0031] Optionally, step four specifically includes:

[0032] In the basic phenotypic feature map, the response positions are located along the ornamental judgment path, growth judgment path, and stress resistance judgment path to form ornamental response set, growth response set, and stress resistance response set, respectively.

[0033] The flower color response, flower shape response, leaf color response, and plant shape response in the observation response set are expanded and associated and aggregated according to the connection order of the observation determination path to form an observation representation;

[0034] Based on the growth determination path, the plant height response, ground diameter response, branch response, and cover response in the growth response set are expanded and associated and aggregated to form growth characteristics;

[0035] Based on the coverage location of the stress resistance determination path, the thickness response, disease-free rate response, and lesion response of the stress resistance response set are expanded and associated to form a stress resistance characterization.

[0036] Separate and organize the observation, growth, and stress resistance characteristics according to their correspondence with the observation, growth, and stress resistance response sets, so that the observation, growth, and stress resistance characteristics maintain an independent correspondence.

[0037] Optionally, step five specifically includes:

[0038] A whole-plant view determination channel is established based on the spatial coverage of the whole plant region in the basic phenotypic feature map, and a local organ view determination channel is established based on the spatial coverage of the organ region in the basic phenotypic feature map.

[0039] The plant shape response in the ornamental representation and the plant height response, ground diameter response, branch response and cover response in the growth representation are mapped to the whole plant view judgment channel according to the positional relationship of the whole plant area, so that the ornamental representation and the growth representation form a positional correspondence in the whole plant view judgment channel.

[0040] The flower color response, flower shape response, and leaf color response in the ornamental characterization are mapped to the local organ view judgment channel according to the positional relationship of the organ region, so that the ornamental characterization and the stress resistance characterization form a positional correspondence relationship within the local organ view judgment channel.

[0041] The positional correspondences in the whole-plant view determination channel and the positional correspondences in the local organ view determination channel are matched and organized according to the single-plant identifier to form a dual-view correspondence representation.

[0042] Optionally, step six specifically includes:

[0043] Extract the correspondence between the positions of the whole plant view determination channels in the dual-view correspondence representation, arrange them according to the relative distance and adjacent connection of the response positions, and construct the whole plant relationship matrix;

[0044] The correspondence between the local organ view determination channels in the dual-view correspondence representation is arranged according to the relative distance and adjacent connection of the response positions to construct an organ relationship matrix;

[0045] The whole-plant relation matrix and the organ relation matrix are introduced into the Gromov-Wasserstein optimal transport solution process to calculate the structural matching relationship between the matrix positions in the whole-plant relation matrix and the matrix positions in the organ relation matrix, thus forming the structural matching result;

[0046] Based on the structural matching results, joint alignment is performed on the correspondence between the determination channel positions in the whole plant view and the determination channel positions in the local organ view to form a dual-view coupled representation.

[0047] Optionally, step seven specifically includes:

[0048] The matching position, response order, and connection relationship of the whole plant view determination channel and the local organ view determination channel after alignment are extracted from the dual-view coupling representation to form a joint determination record set.

[0049] Based on the matching positions in the joint judgment record set, the observation characteristics, growth characteristics, and stress resistance characteristics are associated with the execution positions to form a three-trait association record set;

[0050] Based on the positional adjacency and connectivity relationships in the three-trait association record set, joint aggregation is performed on ornamental characteristics, growth characteristics, and stress resistance characteristics to form a joint judgment sequence;

[0051] Based on the positional arrangement and response correspondence in the joint determination sequence, the ornamental characteristics, growth characteristics, and stress resistance characteristics are fused and arranged to form a single plant breeding value characterization.

[0052] Optionally, step eight specifically includes:

[0053] The positional arrangement results and response correspondence results in the characterization of single plant breeding value are mapped to the scoring channels of superior breeding single plants, potential superior single plants, and ordinary single plants, respectively, to obtain the scoring values ​​of superior breeding single plants, potential superior single plants, and ordinary single plants.

[0054] Normalization was performed based on the proportional relationship between the breeding superior single plant score, the potential superior single plant score, the ordinary single plant score, and the sum of the three scores to obtain the probability of breeding superior single plant, the probability of potential superior single plant, and the probability of ordinary single plant.

[0055] Sort the plants according to the numerical values ​​of the probability of superior breeding plants, the probability of potential superior breeding plants, and the probability of ordinary breeding plants, and determine the category label corresponding to the highest probability.

[0056] The category of superior breeding single plants, potential superior single plants, and ordinary single plants corresponding to the category identifier will be determined as the screening result of superior forest trees.

[0057] The beneficial effects of this invention are:

[0058] Compared with existing screening methods that rely on manual visual inspection and single-view analysis, this invention establishes a pairing relationship between whole-plant images and local close-up images of the same tree during the breeding stage. Based on region regularization, it introduces an improved AlexNet shared convolutional backbone to construct ornamental, growth, and stress resistance determination paths. This allows ornamental, growth, and stress resistance characteristics to be separated, extracted, and expressed in a unified feature space, thereby effectively reducing interference caused by the mixing of plant type, organ, and abnormal information. It improves the completeness and accuracy of identification of key phenotypic information such as flower color, flower shape, leaf color, plant height, diameter at breast height, and lesions, solving the problems of fragmented whole-plant and local information, inconsistent screening standards, and strong subjectivity of manual judgment in existing technologies.

[0059] Meanwhile, this invention forms a dual-view correspondence representation based on the whole-plant view determination channel and the local organ view determination channel. By constructing a whole-plant relationship matrix and an organ relationship matrix, and using Gromov-Wasserstein optimal transmission execution structure matching, a dual-view coupled representation is formed. Then, combined with joint determination and probability calculation, the screening results of superior individual trees are determined. This enables the structural alignment and collaborative determination of whole-plant morphological characteristics and local organ characteristics in the same screening process. This not only improves the ability to distinguish between superior individual trees, potential superior individual trees and ordinary individual trees, but also enhances the stability and breeding adaptability of the screening results. It can meet the application needs of batch, precise and intelligent screening in the breeding process of new varieties of ornamental greening trees. Attached Figure Description

[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0061] Figure 1 This is a flowchart of the image recognition-based method for screening superior individual trees proposed in this invention;

[0062] Figure 2 This is a schematic diagram illustrating the determination path construction of the image recognition-based method for screening superior individual trees proposed in this invention.

[0063] Figure 3 This is a schematic diagram of the dual-view coupled characterization generation of the image recognition-based method for screening superior individual trees proposed in this invention. Detailed Implementation

[0064] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0065] refer to Figures 1-3 The image recognition-based method for screening superior individual trees includes the following steps:

[0066] Step 1: Collect whole-tree images and close-up images of individual trees in the same forest during the breeding stage, establish pairing relationships according to individual tree identifiers, and form a candidate individual tree image set;

[0067] Step 2: Perform region normalization on the candidate single-plant image set to determine the whole-plant region and organ region;

[0068] Step 3: Map the whole plant region and organ region to the improved AlexNet shared convolutional backbone to generate basic phenotypic feature maps. On the basic phenotypic feature maps, establish the ornamental judgment path, growth judgment path, and stress resistance judgment path to form a set of judgment paths.

[0069] Step 4: Perform separation and extraction on the basic phenotypic feature maps along the decision path set to form ornamental features, growth features, and stress resistance features;

[0070] Step 5: Map the ornamental characteristics and growth characteristics to the whole plant view determination channel, and map the ornamental characteristics and stress resistance characteristics to the local organ view determination channel to form a dual-view corresponding characterization.

[0071] Step 6: Construct a whole-plant relationship matrix and an organ relationship matrix for the dual-view correspondence representation, and use Gromov-Wasserstein optimal transmission execution structure matching to form a dual-view coupled representation;

[0072] Step 7: Perform joint determination on the dual-view coupled representation to generate a single-plant breeding value representation;

[0073] Step 8: Calculate the probability of a superior breeding single tree, the probability of a potentially superior single tree, and the probability of an ordinary single tree based on the single tree breeding value characterization. Determine the screening results of superior forest trees based on the maximum value among the three probabilities.

[0074] In this embodiment, step one specifically includes:

[0075] When selecting individual trees in the seedling, rapid growth, flowering, and leaf color viewing stages from the breeding nursery of new ornamental greening tree species, the individual trees are classified and organized according to their growth stages in the breeding nursery. The individual tree identifier corresponding to each individual tree is registered with the corresponding growth stage, so that each individual tree entering the collection process has a unique individual tree identifier and a clear breeding stage. The collection objects are then gathered according to the individual tree identifiers.

[0076] When collecting top-view, front-view, and side-view full-tree images of individual trees in the collection target set, top-view imaging, front-view imaging, and side-view imaging are completed around the same individual tree to form a full-tree image record covering the tree shape outline, crown distribution, and branch morphology of the same individual tree. The top-view, front-view, and side-view full-tree images corresponding to the same individual tree are then collected to form a full-tree image set.

[0077] When collecting close-up images of the flower area on individual trees in the flowering period of the collection target group, and when collecting close-up images of the leaf area on individual trees in the leaf color viewing period of the collection target group, the location of the close-up images is determined according to the breeding stage. The flower area is located on the individual trees in the flowering period to complete the close-up imaging of the flower area, and the leaf area is located on the individual trees in the leaf color viewing period to complete the close-up imaging of the leaf area. The close-up images record and reflect the changes in flower color, flower shape, leaf color, and leaf texture, and are collected to form a set of close-up images.

[0078] When merging the whole plant image set and the partial close-up image set according to the single plant identifier, the whole plant image set is searched for top view, front view, and side view full plant images corresponding to the same single plant identifier. The partial close-up image set is searched for flower area partial close-up images or leaf area partial close-up images corresponding to the same single plant identifier. The retrieved image records are then combined according to the same single plant identifier to establish a pairing relationship between the whole plant image and the partial close-up image, forming a candidate single plant image set.

[0079] In this embodiment, step two specifically involves:

[0080] When performing foreground localization on the top-view full-tree images, front-view full-tree images, and side-view full-tree images in the candidate single-tree image set, the boundary position between the main body of the single tree and the background area is identified according to the pixel distribution difference in the image, and the main body outline is continuously unfolded along the boundary position, so that the top-view full-tree images, front-view full-tree images, and side-view full-tree images form the corresponding main body boundary of the single tree, and are collected to form a set of candidate regions for the whole tree.

[0081] When performing region cropping, size regularization, and coordinate unification on the whole tree candidate region set, the main body coverage area is cropped around the main body boundary of the single tree. The cropping results are size adjusted according to the unified image width and unified image height, and the main body position is aligned according to the unified coordinate reference. This ensures that the top view whole tree image, front view whole tree image, and side view whole tree image corresponding to the same single tree identifier maintain a consistent spatial scale and consistent position reference, thus forming a whole tree region.

[0082] When performing organ localization on local close-up images of flower regions and leaf regions in the candidate single plant image set, the main range of the flower and the main range of the leaf are identified according to the flower shape edge and the leaf shape edge, and continuous organ contours are extracted along the main range of the flower and the main range of the leaf, so that the local close-up images of the flower region form the main boundary of the flower, and the local close-up images of the leaf region form the main boundary of the leaf, and are combined to form a set of candidate organ regions;

[0083] When performing region cropping, size regularization, and coordinate unification on the candidate organ region set, the organ coverage area is cropped around the main boundary of the flower and the main boundary of the leaf. The cropped result is resized according to a unified image width and a unified image height, and the organ position is aligned according to a unified coordinate reference. This ensures that the close-up images of the flower region and the close-up images of the leaf region maintain a consistent spatial scale and a consistent position reference, thus forming the organ region.

[0084] In this embodiment, step three specifically includes:

[0085] When the whole plant region and organ region are introduced into the improved AlexNet shared convolutional backbone with a unified input scale, the whole plant region and organ region are adjusted to have the same image width, image height and channel arrangement, and the whole plant region and organ region are mapped to the improved AlexNet shared convolutional backbone in the same batch. The convolutional mapping layer in the improved AlexNet shared convolutional backbone is used to extract feature responses within the local receptive range, the pooling mapping layer is used to compress the response distribution and retain the dominant response position, and the attention focusing layer is used to adjust the response intensity at a specified position.

[0086] The convolutional mapping layer extracts texture response, color response, and boundary response, while the pooling mapping layer performs response compression and dominant position preservation. When performing hierarchical mapping within a unified feature space, the convolutional mapping layer continuously scans the branch texture, crown edge, and plant outline in the whole plant region, and continuously scans the petal texture, leaf vein texture, and lesion boundary in the organ region. The local responses obtained from the scan are aggregated by position in the pooling mapping layer, so that the unified feature space simultaneously carries the overall phenotypic information of the whole plant region and the local phenotypic information of the organ region.

[0087] Within a unified feature space, attention-focusing enhancement is applied to the flower color location, leaf color location, flower shape location, plant shape location, and lesion location. While maintaining the distinct distribution of flower color response, flower shape response, leaf color response, plant shape response, and lesion response, the attention-focusing layer increases the proportion of responses corresponding to flower color location, leaf color location, flower shape location, plant shape location, and lesion location according to the positional relationship, and suppresses the proportion of responses corresponding to background area and irrelevant area, so that color change, morphological change, and abnormal change form a distinguishable response concentration area in the unified feature space.

[0088] When expanding and arranging the augmented responses according to their spatial and channel positions in the unified feature space to form the basic phenotypic feature map, the augmented responses corresponding to each position in the unified feature space are arranged in spatial and channel order so that the color response, boundary response, and texture response at the same position maintain the correspondence in the basic phenotypic feature map, and the augmented responses at different positions maintain the arrangement order in the basic phenotypic feature map.

[0089] Based on the correlation distribution of flower color response, flower shape response, leaf color response, and plant shape response in the basic phenotypic feature diagram, an ornamental judgment path is established. Based on the correlation distribution of plant height response, ground diameter response, branch response, and cover response in the basic phenotypic feature diagram, a growth judgment path is established. Based on the correlation distribution of thickness response, disease-free rate response, and lesion response in the basic phenotypic feature diagram, a stress resistance judgment path is established. When forming the judgment path set, the response positions in the basic phenotypic feature diagram used to characterize ornamental features are connected according to the correlation strength to form an ornamental judgment path. The response positions in the basic phenotypic feature diagram used to characterize growth features are connected according to the correlation strength to form a growth judgment path. The response positions in the basic phenotypic feature diagram used to characterize stress resistance features are connected according to the correlation strength to form a stress resistance judgment path. The ornamental judgment path, growth judgment path, and stress resistance judgment path are then combined to form a judgment path set.

[0090] This invention addresses the shortcomings of traditional AlexNet in classifying natural images. It struggles to simultaneously consider the overall morphological information of trees and the details of their organs, fails to consistently highlight key breeding criteria such as flower color, leaf color, flower shape, tree shape, and disease spots, and cannot support the collaborative modeling of ornamental, growth, and stress resistance characteristics. The invention provides a targeted improvement to AlexNet for selecting superior individual trees. The improvements are primarily in three areas: input organization, feature mapping, and decision organization. Regarding input organization, instead of directly feeding a single image into the network, the entire tree region and organ regions are introduced into a shared convolutional backbone at a unified input scale. This allows the network to simultaneously receive information on overall tree shape, crown width, branch distribution, petal texture, leaf vein texture, and disease spot boundaries within the same feature space. This solves the problem of traditional AlexNet only making overall judgments on single visual objects and failing to consider phenotypic information at different granularities. In terms of feature mapping, the ordinary convolution and pooling calculation process is enhanced into a collaborative mapping structure of convolutional mapping layers, pooling mapping layers, and attention focusing layers, enabling continuous convolutional mapping layers... By extracting texture, color, and boundary responses, the pooling mapping layer retains dominant positions while compressing redundant responses. The attention-focusing layer provides targeted enhancement to flower color, leaf color, flower shape, plant shape, and lesion location, significantly increasing the response proportion of key breeding traits in a unified feature space and suppressing interference from background and irrelevant regions. In terms of decision organization, this invention goes beyond the single-path classification output of ordinary AlexNet. Instead, it establishes ornamental, growth, and stress resistance decision paths based on the correlation distribution of different responses in the basic phenotypic feature map. This elevates the network output from general category discrimination to a structured set of decision paths oriented towards breeding evaluation logic. The improvement not only enhances the discriminative power and aggregation ability of key phenotypic features but also strengthens feature consistency, decision targeting, and result stability when analyzing whole-plant and organ regions together. This allows the selection of superior individual trees to maintain higher recognition accuracy, stronger anti-interference ability, and better breeding adaptability even under complex backgrounds, fine-grained differences, and the coexistence of multi-dimensional breeding indicators.

[0091] In this embodiment, step four specifically includes:

[0092] When locating response positions along the observation determination path, growth determination path, and stress resistance determination path in the basic phenotypic feature map, the response values ​​at the corresponding positions in the basic phenotypic feature map are read according to the path direction in the determination path set. Response positions within the coverage area of ​​the observation determination path are assigned to the observation response set, response positions within the coverage area of ​​the growth determination path are assigned to the growth response set, and response positions within the coverage area of ​​the stress resistance determination path are assigned to the stress resistance response set. This allows the response positions in the basic phenotypic feature map to be classified and organized according to their determination purpose, forming the observation response set, growth response set, and stress resistance response set respectively.

[0093] When expanding and aggregating the flower color response, flower shape response, leaf color response, and plant shape response in the ornamental response set according to the connection order of the ornamental judgment path, the flower color response, flower shape response, leaf color response, and plant shape response are read according to the arrangement order of the ornamental judgment path in the basic phenotypic feature map, and the correlation strength between adjacent positions is continuously merged so that the responses reflecting color changes, morphological changes, and overall plant shape distribution form a continuous representation on the same path, thus forming an ornamental representation.

[0094] When performing expansion and association aggregation on the plant height response, ground diameter response, branch response, and cover response in the growth response set according to the growth determination path, the plant height response, ground diameter response, branch response, and cover response are read according to the response position covered by the growth determination path, and the response connectivity and distribution relationship between the corresponding positions are sorted out in order so that the responses reflecting the growth status form a continuous aggregation result on the growth determination path, forming a growth characterization.

[0095] When expanding and aggregating the thickness response, disease-free rate response, and lesion response in the stress resistance assessment path according to the coverage location of the stress resistance assessment path, the thickness response, disease-free rate response, and lesion response are extracted based on the coverage range of the stress resistance assessment path. The correspondence between the thickness response, disease-free rate response, and lesion response within the path range is then centrally arranged so that the responses reflecting the stress resistance status form a continuous aggregation result on the stress resistance assessment path, thus forming a stress resistance characterization.

[0096] When performing separation and sorting according to the correspondence between the observation judgment path, growth judgment path, and stress resistance judgment path and the observation response set, growth response set, and stress resistance response set, the aggregation result corresponding to the observation response set is retained in the observation representation, the aggregation result corresponding to the growth response set is retained in the growth representation, and the aggregation result corresponding to the stress resistance response set is retained in the stress resistance representation. Furthermore, the responses at the intersection positions are stripped according to the belonging range of the observation judgment path, growth judgment path, and stress resistance judgment path, so that the observation representation, growth representation, and stress resistance representation maintain an independent correspondence.

[0097] In this embodiment, step five specifically includes:

[0098] A whole-plant view determination channel is established based on the spatial coverage of the whole plant region in the basic phenotypic feature map. When establishing a local organ view determination channel based on the spatial coverage of the organ region in the basic phenotypic feature map, the region boundary corresponding to the whole plant region is mapped to the spatial coordinates of the basic phenotypic feature map to determine the feature position range occupied by the whole plant region. The occupied feature position range is then connected to form the whole-plant view determination channel. Similarly, the region boundary corresponding to the organ region is mapped to the spatial coordinates of the basic phenotypic feature map to determine the feature position range occupied by the organ region. The occupied feature position range is then connected to form the local organ view determination channel.

[0099] The plant shape response in the ornamental representation and the plant height response, ground diameter response, branch response, and canopy response in the growth representation are mapped to the whole plant view judgment channel according to the positional relationship of the whole plant area. When the ornamental representation and the growth representation form a positional correspondence in the whole plant view judgment channel, the plant shape response is configured to the corresponding plant shape position, the plant height response is configured to the corresponding height position, the ground diameter response is configured to the corresponding base position, the branch response is configured to the corresponding branch distribution position, and the canopy response is configured to the corresponding crown coverage position according to the longitudinal and lateral positions in the whole plant area. The plant shape response, plant height response, ground diameter response, branch response, and canopy response are aligned in the channel according to the positional adjacency relationship, so that the ornamental representation and the growth representation form a one-to-one corresponding positional arrangement in the whole plant view judgment channel.

[0100] The flower color response, flower shape response, and leaf color response in the ornamental characteristics are mapped to the local organ view judgment channel according to the positional relationship of the organ regions, along with the thickness response, disease-free rate response, and lesion response in the stress resistance characteristics. When the ornamental characteristics and stress resistance characteristics form a positional correspondence in the local organ view judgment channel, the flower color response is configured to the corresponding color distribution position, the flower shape response to the corresponding outline position, the leaf color response to the corresponding leaf surface position, the thickness response to the corresponding thick area of ​​the leaf, the disease-free rate response to the corresponding healthy area position, and the lesion response to the corresponding abnormal area position, based on the flower position and leaf position in the organ region. Furthermore, the flower color response, flower shape response, leaf color response, thickness response, disease-free rate response, and lesion response are aligned within the channel according to the spatial adjacency relationship in the organ region, so that the ornamental characteristics and stress resistance characteristics form a one-to-one corresponding positional arrangement in the local organ view judgment channel.

[0101] When the positional correspondences in the whole-tree view determination channel and the local organ view determination channel are matched and organized according to the single-tree identifier to form a dual-view correspondence representation, the positional arrangement of the whole-tree view determination channel and the positional arrangement of the local organ view determination channel corresponding to the same single-tree identifier are extracted, merged according to the same single-tree identifier, and then the positional correspondences in the whole-tree view determination channel and the local organ view determination channel are jointly arranged according to the spatial positional correspondence, so that the whole-tree positional correspondences and organ positional correspondences of the same forest tree are gathered to form a dual-view correspondence representation.

[0102] In this embodiment, step six specifically includes:

[0103] Extract the position correspondence of the whole-plant view determination channel in the dual-view correspondence representation, arrange them according to the relative distance and adjacency connection of the response positions, and when constructing the whole-plant relationship matrix, read the spatial coordinates and connectivity order corresponding to each response position in the whole-plant view determination channel, determine the relative distance based on the distance between any two response positions, determine the adjacency connection based on whether there is a direct connectivity relationship between the response positions, and write the relative distance and adjacency connection into the same matrix structure in the order of matrix rows and columns, so that the position distribution relationship and connection relationship in the whole-plant view determination channel are centrally expressed as the whole-plant relationship matrix;

[0104] The positional correspondence of the local organ view determination channel in the dual-view correspondence representation is arranged according to the relative distance and adjacent connection of the response positions. When constructing the organ relationship matrix, the positional distribution corresponding to flower color response, flower shape response, leaf color response, thickness response, disease-free rate response, and lesion response is read in the local organ view determination channel. The relative distance is determined according to the spatial interval between the response positions, and the adjacent connection is determined according to the connectivity between the response positions. The relative distance and adjacent connection are written into the same matrix structure in the order of matrix rows and columns, so that the positional distribution and connection relationships in the local organ view determination channel are centrally expressed as the organ relationship matrix.

[0105] The whole-plant relation matrix and the organ relation matrix are introduced into the Gromov-Wasserstein optimal transport solution process. The structural matching relationship between the matrix positions in the whole-plant relation matrix and the matrix positions in the organ relation matrix is ​​calculated. When forming the structural matching result, a matching search range corresponding to the matrix positions is established between the whole-plant relation matrix and the organ relation matrix. Within the matching search range, the relative distance structure between the matrix positions in the whole-plant relation matrix and the relative distance structure between the matrix positions in the organ relation matrix are compared. The matrix positions with close distance structures and consistent connection relationships are determined as matching positions, and the structural matching results are collected.

[0106] Based on the structural matching results, joint alignment is performed on the positional correspondence of the whole plant view determination channel and the positional correspondence of the local organ view determination channel to form a dual-view coupled representation. When the matching position in the structural matching result is mapped back to the whole plant view determination channel and the local organ view determination channel, position alignment and relationship integration are performed according to the response order corresponding to the matching position, so that the positional correspondence of the whole plant view determination channel and the positional correspondence of the local organ view determination channel remain consistent in the same arrangement structure, thus forming a dual-view coupled representation.

[0107] In this embodiment, step seven specifically includes:

[0108] When extracting the matching position, response order, and connection relationship after aligning the whole-plant view determination channel and the local organ view determination channel from the dual-view coupling representation to form a joint determination record set, each set of alignment positions is read along the arrangement order in the dual-view coupling representation. The response position of each set of alignment positions in the whole-plant view determination channel, the response position in the local organ view determination channel, the order of arrangement and connection relationship between the two types of response positions are recorded. This allows the alignment results in the whole-plant view determination channel and the alignment results in the local organ view determination channel to be aggregated in the same positional framework to form a joint determination record set.

[0109] According to the matching positions in the joint judgment record set, when performing positional association on the ornamental character, growth character, and stress resistance character, forming a three-trait association record set, based on each set of matching positions in the joint judgment record set, the ornamental response falling within the matching position range is extracted in the ornamental character, the growth response falling within the matching position range is extracted in the growth character, and the stress resistance response falling within the matching position range is extracted in the stress resistance character. The ornamental response, growth response, and stress resistance response are then arranged according to the same matching position to maintain the association relationship of the three types of responses under the same matching position, thus forming a three-trait association record set.

[0110] Based on the positional adjacency and connectivity relationships in the three-trait association record set, joint aggregation is performed on the ornamental characteristics, growth characteristics, and stress resistance characteristics to form a joint decision sequence. When forming a joint decision sequence, the ornamental response, growth response, and stress resistance response are read sequentially according to the positional adjacency order in the three-trait association record set. Based on the adjacency strength and connectivity order between positions, the three types of responses at consecutive positions are combined into the same decision unit, and then continuously expanded according to the arrangement order of the decision units. This allows the ornamental characteristics, growth characteristics, and stress resistance characteristics to complete joint aggregation in a unified order, forming a joint decision sequence.

[0111] According to the positional arrangement and response correspondence in the joint judgment sequence, the ornamental characteristics, growth characteristics, and stress resistance characteristics are integrated and arranged to form a single-tree breeding value characterization. When the ornamental response, growth response, and stress resistance response in the joint judgment sequence are written into the same fusion structure according to their positional arrangement, the responses reflecting the ornamental state, the responses reflecting the growth state, and the responses reflecting the stress resistance state are centrally arranged according to their response correspondence, so that the same forest tree maintains consistent position, consistent response, and consistent relationship in the unified fusion structure, thus forming a single-tree breeding value characterization.

[0112] In this embodiment, step eight specifically includes:

[0113] The positional arrangement results and response correspondence results in the single-plant breeding value representation are mapped to the excellent breeding single plant scoring channel, the potential excellent single plant scoring channel, and the ordinary single plant scoring channel, respectively. When obtaining the excellent breeding single plant score value, the potential excellent single plant score value, and the ordinary single plant score value, the fusion response of ornamental characteristics, growth characteristics, and stress resistance characteristics is extracted according to the arrangement order of each position in the single-plant breeding value representation. The fusion response reflecting the degree of excellence is written into the excellent breeding single plant scoring channel, the fusion response reflecting the cultivation potential is written into the potential excellent single plant scoring channel, and the fusion response reflecting the general state is written into the ordinary single plant scoring channel. Then, the fusion response in the three types of scoring channels is cumulatively calculated to form the excellent breeding single plant score value, the potential excellent single plant score value, and the ordinary single plant score value.

[0114] Normalization is performed based on the proportional relationship between the scores of superior breeding plants, potential superior breeding plants, and ordinary breeding plants and the sum of the three scores to obtain the probabilities of superior breeding plants, potential superior breeding plants, and ordinary breeding plants. The scores of superior breeding plants, potential superior breeding plants, and ordinary breeding plants are added together to form the sum of the three scores. The score of superior breeding plants is divided by the sum of the three scores to obtain the probability of superior breeding plants. The score of potential superior breeding plants is divided by the sum of the three scores to obtain the probability of potential superior breeding plants. The score of ordinary breeding plants is divided by the sum of the three scores to obtain the probability of ordinary breeding plants.

[0115] The probability of a superior breeding plant, the probability of a potentially superior breeding plant, and the probability of an ordinary breeding plant are sorted according to their numerical values. When determining the category identifier corresponding to the highest probability, the probability of a superior breeding plant, the probability of a potentially superior breeding plant, and the probability of an ordinary breeding plant are compared item by item, and the category name corresponding to the scoring channel with the highest numerical probability is extracted as the category identifier.

[0116] When determining one of the three categories—superior breeding single trees, potentially superior single trees, and ordinary single trees—as the screening result for superior forest trees, the category identifier is matched with the three screening categories of superior breeding single trees, potentially superior single trees, and ordinary single trees, and the matching result is written into the classification result record of forest trees to form the screening result for superior forest trees.

[0117] Example 1: To verify the feasibility of this invention in practice, it was applied to the screening of superior individual trees in a breeding nursery for a new variety of ornamental greening tree. The nursery contained individual trees at various stages: seedling stage, rapid growth stage, flowering stage, and leaf color ornamental stage. With a large number of candidate trees, breeders commonly encountered two types of problems during manual screening: First, the overall tree shape, growth vigor, and local flower color, leaf color, and lesion information were scattered across different images, making it easy to overlook certain aspects during manual comparison. This led to individual trees with outstanding local ornamental value but poor overall shape being misjudged as superior, or individuals with good overall growth but defects in local organs being missed. Second, different breeders had inconsistent criteria for judging flower color uniformity, leaf color uniqueness, tree compactness, and lesion severity. The conclusions of the same batch of candidate trees fluctuated significantly during repeated screenings, making it difficult to meet the requirements for stable screening results in new variety breeding.

[0118] In this embodiment, full-tree images from top view, front view, and side view are collected for each individual tree. Close-up images of the flower and leaf regions are also collected. A candidate tree image set is formed by establishing pairing relationships through tree identifiers. After region normalization to obtain the full-tree and organ regions, an improved AlexNet shared convolutional backbone is introduced. Texture, color, and boundary responses are extracted within a unified feature space to form basic phenotypic feature maps. Then, observational, growth, and stress-resistance determination paths are established to obtain observational, growth, and stress-resistance representations. Subsequently, plant type response, plant height response, ground diameter response, branch response, and cover response are mapped to the whole-plant view judgment channel, while flower color response, flower shape response, leaf color response, fruit thickness response, disease-free rate response, and lesion response are mapped to the local organ view judgment channel, forming a dual-view correspondence representation. Then, through structural matching of the whole-plant relationship matrix and the organ relationship matrix, a dual-view coupled representation is obtained, completing the joint judgment and calculating the probability of superior breeding plants, the probability of potentially superior breeding plants, and the probability of ordinary breeding plants. Finally, the screening results are output. This entire process can place the judgment of whole-plant growth and the judgment of organ details within the same judgment framework, making it suitable for batch screening in breeding scenarios.

[0119] In this embodiment, 600 candidate trees were selected for verification, resulting in 1800 full-tree images and 1200 close-up images, totaling 3000 images. To verify the effectiveness, the method of this invention was compared with manual screening and ordinary convolutional recognition methods, with the conclusions of manual review serving as the control benchmark. The comparison results are shown in Table 1.

[0120] Table 1. Comparison of the identification effects of superior single plants using different screening methods

[0121] Filtering methods Number of candidate individual plants (plants) Number of images (pieces) Total time (hours) Overall accuracy (%) Accuracy rate of identifying superior single plants (%) Incorrect selection rate (%) Manual screening 600 3000 10.8 83.5 80.7 9.8 Ordinary convolutional recognition method 600 3000 1.4 91.2 89.6 5.1 Method of the present invention 600 3000 1.1 96.7 96.1 2.3

[0122] As shown in Table 1, under the same conditions of the number of candidate trees and the number of images, the total time consumed by the method of the present invention is significantly lower than that of manual screening and slightly lower than that of ordinary convolutional recognition methods, indicating that the present invention has high processing efficiency in batch screening scenarios. More importantly, the overall accuracy of the present invention reaches 96.7%, and the accuracy of identifying superior trees reaches 96.1%, both significantly higher than that of manual screening and ordinary convolutional recognition methods; the false selection rate is reduced to 2.3%, indicating that the present invention can effectively reduce misjudgments caused by the separation of local ornamental features and overall growth characteristics, and can also reduce the fluctuations in results caused by differences in subjective judgment among different screening personnel. It can be seen that the present invention not only solves the problem of the difficulty in unifying the use of whole-plant information and local organ information in the existing superior tree screening, but also takes into account screening efficiency, identification accuracy and result stability in real breeding applications, and can meet the actual needs of intelligent screening of superior trees in the breeding of new varieties of ornamental greening trees.

[0123] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for screening superior individual trees in forests based on image recognition, characterized in that, Includes the following steps: Step 1: Collect whole-tree images and close-up images of individual trees in the same forest during the breeding stage, establish pairing relationships according to individual tree identifiers, and form a candidate individual tree image set; Step 2: Perform region normalization on the candidate single-plant image set to determine the whole-plant region and organ region; Step 3: Map the whole plant region and organ region to the improved AlexNet shared convolutional backbone to generate basic phenotypic feature maps. On the basic phenotypic feature maps, establish the ornamental judgment path, growth judgment path, and stress resistance judgment path to form a set of judgment paths. Step 4: Perform separation and extraction on the basic phenotypic feature maps along the decision path set to form ornamental features, growth features, and stress resistance features; Step 5: Map the ornamental characteristics and growth characteristics to the whole plant view determination channel, and map the ornamental characteristics and stress resistance characteristics to the local organ view determination channel to form a dual-view corresponding characterization. Step 6: Construct a whole-plant relationship matrix and an organ relationship matrix for the dual-view correspondence representation, and use Gromov-Wasserstein optimal transmission execution structure matching to form a dual-view coupled representation; Step 7: Perform joint determination on the dual-view coupled representation to generate a single-plant breeding value representation; Step 8: Calculate the probability of a superior breeding single tree, the probability of a potentially superior single tree, and the probability of an ordinary single tree based on the single tree breeding value characterization. Determine the screening results of superior forest trees based on the maximum value among the three probabilities.

2. The method for screening superior individual trees based on image recognition according to claim 1, characterized in that, Step one specifically involves: Individual trees in the seedling stage, rapid growth stage, flowering stage, and leaf color viewing stage of the new ornamental greening tree breeding nursery were selected, and a collection object set was formed according to the individual tree identifiers corresponding to the individual trees; Collect full-tree images from the top view, front view, and side view of individual trees in the collection object set to form a full-tree image set; Close-up images of the flower area of ​​individual trees in the flowering period of the collection object group were collected, and close-up images of the leaf area of ​​individual trees in the leaf color viewing period of the collection object group were collected to form a set of close-up images; The whole-plant image set and the partial close-up image set are merged according to the individual plant identifier, and the pairing relationship between the whole-plant image and the partial close-up image is established to form a candidate single-plant image set.

3. The method for screening superior individual trees based on image recognition according to claim 1, characterized in that, Step two specifically involves: Foreground localization is performed on the top-view, front-view, and side-view full-tree images in the candidate single-tree image set to extract the main boundary of the forest single tree and form a full-tree candidate region set. Perform region clipping, size regularization, and coordinate unification on the whole plant candidate region set to form the whole plant region; Organ localization is performed on local close-up images of flower regions and leaf regions in the candidate single plant image set to extract the main boundaries of flower and leaf bodies and form a candidate organ region set. Perform region clipping, size regularization, and coordinate unification on the candidate organ regions to form organ regions.

4. The method for screening superior individual trees based on image recognition according to claim 1, characterized in that, Step three specifically involves: The whole-plant region and organ region are introduced into the improved AlexNet shared convolutional backbone according to a unified input scale. The improved AlexNet shared convolutional backbone sets up convolutional mapping layers, pooling mapping layers, and attention focusing layers. The whole plant region and organ region are subjected to texture response extraction, color response extraction and boundary response extraction along the convolutional mapping layer, and response compression and dominant position preservation are performed along the pooling mapping layer to complete the hierarchical mapping in a unified feature space. Within a unified feature space, attentional focus enhancement is applied to the flower color location, leaf color location, flower shape location, plant shape location, and lesion location to maintain the distinct distribution of flower color response, flower shape response, leaf color response, plant shape response, and lesion response; The enhanced response is unfolded and arranged according to its spatial and channel positions in a unified feature space to form a basic phenotypic feature map. An ornamental evaluation path was established based on the correlation distribution of flower color response, flower shape response, leaf color response, and plant shape response in the basic phenotypic characteristic diagram. A growth evaluation path was established based on the correlation distribution of plant height response, ground diameter response, branch response, and cover response in the basic phenotypic characteristic diagram. A stress resistance evaluation path was established based on the correlation distribution of thickness response, disease-free rate response, and lesion response in the basic phenotypic characteristic diagram, thus forming a set of evaluation paths.

5. The method for screening superior individual trees based on image recognition according to claim 1, characterized in that, Step four specifically involves: In the basic phenotypic feature map, the response positions are located along the ornamental judgment path, growth judgment path, and stress resistance judgment path to form ornamental response set, growth response set, and stress resistance response set, respectively. The flower color response, flower shape response, leaf color response, and plant shape response in the observation response set are expanded and associated and aggregated according to the connection order of the observation determination path to form an observation representation; Based on the growth determination path, the plant height response, ground diameter response, branch response, and cover response in the growth response set are expanded and associated and aggregated to form growth characteristics; Based on the coverage location of the stress resistance determination path, the thickness response, disease-free rate response, and lesion response of the stress resistance response set are expanded and associated to form a stress resistance characterization. Separate and organize the observation, growth, and stress resistance characteristics according to their correspondence with the observation, growth, and stress resistance response sets, so that the observation, growth, and stress resistance characteristics maintain an independent correspondence.

6. The method for screening superior individual trees based on image recognition according to claim 1, characterized in that, Step five specifically involves: A whole-plant view determination channel is established based on the spatial coverage of the whole plant region in the basic phenotypic feature map, and a local organ view determination channel is established based on the spatial coverage of the organ region in the basic phenotypic feature map. The plant shape response in the ornamental representation and the plant height response, ground diameter response, branch response and cover response in the growth representation are mapped to the whole plant view judgment channel according to the positional relationship of the whole plant area, so that the ornamental representation and the growth representation form a positional correspondence in the whole plant view judgment channel. The flower color response, flower shape response, and leaf color response in the ornamental characterization are mapped to the local organ view judgment channel according to the positional relationship of the organ region, so that the ornamental characterization and the stress resistance characterization form a positional correspondence relationship within the local organ view judgment channel. The positional correspondences in the whole-plant view determination channel and the positional correspondences in the local organ view determination channel are matched and organized according to the single-plant identifier to form a dual-view correspondence representation.

7. The method for screening superior individual trees based on image recognition according to claim 1, characterized in that, Step six specifically involves: Extract the correspondence between the positions of the whole plant view determination channels in the dual-view correspondence representation, arrange them according to the relative distance and adjacent connection of the response positions, and construct the whole plant relationship matrix; The correspondence between the local organ view determination channels in the dual-view correspondence representation is arranged according to the relative distance and adjacent connection of the response positions to construct an organ relationship matrix; The whole-plant relation matrix and the organ relation matrix are introduced into the Gromov-Wasserstein optimal transport solution process to calculate the structural matching relationship between the matrix positions in the whole-plant relation matrix and the matrix positions in the organ relation matrix, thus forming the structural matching result; Based on the structural matching results, joint alignment is performed on the correspondence between the determination channel positions in the whole plant view and the determination channel positions in the local organ view to form a dual-view coupled representation.

8. The method for screening superior individual trees based on image recognition according to claim 1, characterized in that, Step seven specifically involves: The matching position, response order, and connection relationship of the whole plant view determination channel and the local organ view determination channel after alignment are extracted from the dual-view coupling representation to form a joint determination record set. Based on the matching positions in the joint judgment record set, the observation characteristics, growth characteristics, and stress resistance characteristics are associated with the execution positions to form a three-trait association record set; Based on the positional adjacency and connectivity relationships in the three-trait association record set, joint aggregation is performed on ornamental characteristics, growth characteristics, and stress resistance characteristics to form a joint judgment sequence; Based on the positional arrangement and response correspondence in the joint determination sequence, the ornamental characteristics, growth characteristics, and stress resistance characteristics are fused and arranged to form a single plant breeding value characterization.

9. The method for screening superior individual trees based on image recognition according to claim 1, characterized in that, Step eight specifically involves: The positional arrangement results and response correspondence results in the characterization of single plant breeding value are mapped to the scoring channels of superior breeding single plants, potential superior single plants, and ordinary single plants, respectively, to obtain the scoring values ​​of superior breeding single plants, potential superior single plants, and ordinary single plants. Normalization was performed based on the proportional relationship between the breeding superior single plant score, the potential superior single plant score, the ordinary single plant score, and the sum of the three scores to obtain the probability of breeding superior single plant, the probability of potential superior single plant, and the probability of ordinary single plant. Sort the plants according to the numerical values ​​of the probability of superior breeding plants, the probability of potential superior breeding plants, and the probability of ordinary breeding plants, and determine the category label corresponding to the highest probability. The category of superior breeding single plant, potential superior single plant, or ordinary single plant corresponding to the category identifier is determined as the screening result of superior forest single plant.

Images