Field quadrat box recognition and vectorization method and system based on unmanned aerial vehicle image

Abstract

The present application belongs to the field of unmanned aerial vehicle image intelligent processing technology, and particularly relates to a field sample frame recognition and vectorization method and system based on unmanned aerial vehicle images. First, the unmanned aerial vehicle orthographic image with coordinate reference is subjected to feature extraction, and is divided into four types of geometric features, i.e. sample main body, edge line, center point and corner point. After the main body response and entity edge line are kept consistent in spatial position through bidirectional cross-task gating, the four types of geometric features are subjected to fusion prediction to obtain four task probability maps. Based on the four task probability maps, the center point and corner point are topologically associated to constrain the spatial relationship, then geometric matching is carried out by using a matching cost function value, and then global evaluation is carried out by using comprehensive scoring, so as to filter out false rectangles (ridges, crop rows and shadows) with low scores, abnormal sizes and aspect ratio deviating from the standard, and to complete instance-level accurate sample frame recognition and vectorization.

Description

Technical Field

[0001] This invention belongs to the field of intelligent processing technology of UAV images, specifically relating to a method and system for field quadrat recognition and vectorization based on UAV images. Background Technology

[0002] Quadrats are the core spatial units for field trials, remote sensing ground validation, and quantitative analysis of crop phenotypes in crops such as wheat, rice, and maize. After standardized physical quadrats are deployed in the field, the ground survey results need to be spatially correlated with UAV RGB imagery, multispectral imagery, vegetation index imagery, and quadrat-level phenotypic indicators such as cover, biomass, and canopy structure parameters. In quadrat-level agricultural remote sensing data processing, the accuracy of quadrat boundary locations, the consistency of quadrat numbering, and the completeness of quadrat silhouette images directly affect the reliability of subsequent remote sensing indicator statistics, model validation, phenotypic analysis, and agronomic interpretation.

[0003] Existing methods for processing field quadrats have inherent limitations. Manual visual drawing and RTK field measurement are labor-intensive and inefficient. Methods that generate plot boundaries based on experimental design drawings or regular grids are only suitable for experimental scenarios with known layouts and regular arrangement, and cannot identify physical quadrats set up in the field or obscured by crops or shadows. Conventional semantic segmentation algorithms extract target regions from images, but usually focus on foreground / background segmentation, making it difficult to support stable quadrat geometric regularization and geographic vectorization. Summary of the Invention

[0004] To address the problems in the background technology, this invention provides a method and system for field quadrat recognition and vectorization based on UAV imagery.

[0005] The technical solution of the present invention is as follows: This invention provides a method for field quadrat recognition and vectorization based on UAV imagery, comprising: S1: After extracting features from the UAV orthophoto with coordinate reference, decode to obtain the main features, edge features, center point features and corner features; S2: The main features are processed by convolution and activation functions to generate a main body guiding edge gating map, and the edge features are modulated to generate optimized edge features; the optimized edge features are then used to generate an edge guiding main body gating map, and the main features are modulated to generate optimized main body features. S3: Optimize the fusion of main features, edge features, center point features and corner features to predict four task probability maps, namely the main region probability map, entity edge probability map, center point probability map and corner point probability map; S4: Based on the four-task probability map, extract the set of candidate center points and the set of candidate corner points; Traverse the set of candidate center points, and use each candidate center point as an anchor point to search for candidate corner points in a preset neighborhood to obtain multiple combinations of four corner points. S5: Filter the target four-corner point combination by matching cost function, as follows: Based on the combination of four corner points, calculate the symmetry error value of the distance from the corresponding anchor point to the four corner points; use the perpendicular error value of adjacent sides and the parallel error value of opposite sides to calculate the rectangular structure error value; sample the edge line probability value to obtain the edge line response value; extract the corner point prediction value to obtain the four corner response value; Based on the distance symmetry error value, rectangular structure error value, edge response value, and four-corner response value, calculate the matching cost function value, and take the four-corner point combination whose matching cost function value is less than the preset function value as the target four-corner point combination. S6: The pixel coordinates of the target's four corner points are converted into projected coordinates to obtain quadrat vector data.

[0006] Based on the above-described method for field quadrat identification and vectorization using UAV imagery, step S5, calculating the symmetry error value of the distance from the corresponding anchor point to the four corner points, specifically involves: Calculate the straight-line distances from the anchor point to the four corner points respectively, and obtain the average value of the four straight-line distances; then calculate the discrete deviation between the four distances and the average value, and combine the numerical stability constant to avoid the denominator being zero to obtain the original symmetric error; finally, normalize and compress the original error.

[0007] Based on the above-described method for field quadrat identification and vectorization using UAV imagery, step S5, which calculates the rectangular structure error value using the perpendicular error value of adjacent sides and the parallel error value of opposite sides, specifically involves: Sort the corner points cyclically according to their polar angles relative to the anchor points, and connect the four corners in cyclic order to obtain four side vectors; calculate the perpendicular error value of adjacent sides and the parallel error value of opposite sides; take the average of the perpendicular error value and the parallel error value as the rectangular structure error value.

[0008] Based on the above-described method for field quadrat recognition and vectorization using UAV imagery, the sampling edge probability value described in S5 is used to obtain the edge response value, specifically: Extract the predicted probabilities of all pixels on the entity edge probability map along the four edges, and calculate the average of all pixel probabilities as the edge response value.

[0009] Based on the above-described method for field quadrat identification and vectorization using UAV imagery, the extraction of corner prediction values ​​in step S5 yields the four-corner response values, specifically as follows: Extract the predicted probabilities of the four corner points on the corner probability map, and calculate the average of the predicted probabilities of the four corner points.

[0010] Based on the above-described method for field quadrat identification and vectorization using UAV imagery, step S5 involves calculating the matching cost function value based on the distance symmetry error value, rectangular structure error value, edge response value, and four-corner response value. Specifically: The edge penalty value is generated by subtracting the edge response value from the numerical value 1, and the corner penalty value is generated by subtracting the four corner response values ​​from the numerical value 1. The distance symmetry error value, rectangular structure error value, edge penalty value and corner penalty value are weighted and summed to obtain the matching cost function value.

[0011] Based on the above-described method for field quadrat identification and vectorization using UAV imagery, S5 further includes combining the four corner points of each target using a formula. To calculate the overall score, eliminate the target four-corner point combination whose overall score is lower than the score threshold; In the formula, , , , , , These represent the combinations of the four corner points of the target. The comprehensive score, normalized response value of the main area, normalized response value of the entity edge, normalized response value of the center point, normalized response value of the corner point, and physical scale penalty value are given by α, β, γ, δ, and λ, which represent the non-negative weights of each item.

[0012] Based on the above-described method for field quadrat recognition and vectorization using UAV imagery, S2 specifically comprises: The main features are processed by convolutional mapping and the Sigmoid function to generate a gating map of the main body's guiding edge lines; The edge features are multiplied element-wise using the subject-guided edge gating graph to generate optimized edge features. The optimized edge features are processed by convolutional mapping and the Sigmoid function to generate an edge-guided gating map of the main body. By using the edge lines to guide the gating graph of the subject, the subject features are multiplied element by element to generate optimized subject features.

[0013] Based on the above-described method for field quadrat identification and vectorization using UAV imagery, the preset neighborhood mentioned in S4 is defined by the formula... To determine; In the formula, , These represent the reference width and reference height at the pixel scale, respectively. This represents the theoretical distance from the anchor point to the corner point.

[0014] This invention also provides a field quadrat recognition and vectorization system based on UAV imagery, comprising: Feature extraction module: After extracting features from the UAV orthophoto with coordinate reference, it decodes the main features, edge features, center point features, and corner features. Feature optimization module: The main features are processed by convolution and activation functions to generate a main guiding edge gating map, and the edge features are modulated to generate optimized edge features; the optimized edge features are then used to generate an edge guiding main gating map, and the main features are modulated to generate optimized main features. The four-task prediction module integrates optimized main features, optimized edge features, center point features, and corner features to predict four task probability maps, namely the main region probability map, entity edge probability map, center point probability map, and corner point probability map. Combination module: Based on the four-task probability map, extract the set of candidate center points and the set of candidate corner points; Traverse the set of candidate center points, and use each candidate center point as an anchor point to search for candidate corner points in a preset neighborhood to obtain multiple combinations of four corner points. Filtering module: Filters target four-corner point combinations by matching cost function, as follows: Based on the combination of four corner points, calculate the symmetry error value of the distance from the corresponding anchor point to the four corner points; use the perpendicular error value of adjacent sides and the parallel error value of opposite sides to calculate the rectangular structure error value; sample the edge line probability value to obtain the edge line response value; extract the corner point prediction value to obtain the four corner response value; Based on the distance symmetry error value, rectangular structure error value, edge response value, and four-corner response value, calculate the matching cost function value, and take the four-corner point combination whose matching cost function value is less than the preset function value as the target four-corner point combination. Vectorization module: The pixel coordinates of the target's four corner points are converted into projected coordinates to obtain quadrat vector data.

[0015] Beneficial effects This invention first extracts features from UAV orthophotos with coordinate reference, classifying them into four geometric feature categories: subject, edge, center point, and corner points. Through bidirectional cross-task gating, the subject response and the entity's edge are kept spatially consistent. Then, the four geometric feature categories are fused and predicted to obtain a four-task probability map. Based on this four-task probability map, the center point and corner points are topologically associated to constrain spatial relationships. Next, geometric matching is performed using a matching cost function value, followed by a comprehensive evaluation using a comprehensive score. Low-scoring, abnormally sized, and aspect ratio-deviation-deviated false rectangles (field ridges, crop rows, shadows) are filtered out, achieving instance-level accurate quadrat box recognition and vectorization. Attached Figure Description

[0016] Figure 1The diagram shows the probability maps and regularized quadrat vectors for four tasks, where (a) is an orthophoto of the UAV, (b) is the probability map of the main area, (c) is the probability map of the entity edge, (d) is the probability map of the center point, (e) is the probability map of the corner point, and (f) is the regularized quadrat vector.

[0017] Figure 2 This is a schematic diagram of the quadrat size regularization process, where (a) is a local orthophoto, (b) is a probability map of the main quadrat region, (c) is the quadrat outline, and (d) is the size regularization vector. Detailed Implementation

[0018] The following examples are intended to illustrate the present invention, and not to further limit the invention.

[0019] Example 1 This embodiment provides a method for field quadrat recognition and vectorization based on UAV imagery, including: S1: After extracting features from the UAV orthophoto with coordinate reference, decode to obtain the main features, edge features, center point features, and corner features.

[0020] Among them, the UAV orthophoto is preferably an RGB orthophoto, and the feature extraction network is based on a U-Net-like encoder-decoder network, which is performed through three parallel decoding branches.

[0021] In the specific implementation process, the UAV orthophoto with coordinate reference is first subjected to progressive convolution and downsampling by a shared encoder to obtain encoded features at multiple scales. Then, in the decoding stage, a main region decoding branch, an entity edge decoding branch, and an anchor point decoding branch are set up. Each decoding branch upsamples the high-level features progressively and performs skip connections to fuse them with the corresponding low-level features in the shared encoder. After convolution, normalization, and activation function processing, the spatial resolution corresponding to the input image is restored.

[0022] The main region decoding branch outputs main features to determine the approximate location and internal extent of the sample plot; the entity edge decoding branch outputs edge features to enhance thin borders, broken borders, and rotated boundaries; the anchor point (center point / corner point) decoding branch outputs shared anchor point features, and further obtains center point features and corner point features through center point mapping and corner point mapping.

[0023] By using branched decoding, interference between planar regions, fine line boundaries, and sparse point targets can be avoided during the same decoding process.

[0024] S2: The main features are processed by convolution and activation functions to generate a main subject guiding edge gating map. The modulated edge features generate optimized edge features. The optimized edge features are then used to generate an edge-guided main subject gating map, and the modulated main features generate optimized main subject features. Specifically: The main features are processed by convolutional mapping and the Sigmoid function to generate a gating map of the main body's guiding edge lines; The edge features are multiplied element-wise using the subject-guided edge gating graph to generate optimized edge features. The optimized edge features are processed by convolutional mapping and the Sigmoid function to generate an edge-guided gating map of the main body. By using the edge lines to guide the gating graph of the subject, the subject features are multiplied element by element to generate optimized subject features.

[0025] The edges of the field sample boxes are thin, making them easy to confuse with crop leaves, shadow edges, and planting row textures; at the same time, the main area prediction is prone to expansion, contraction, or boundary drift.

[0026] To this end, the present invention uses bidirectional cross-task gating to modulate the edge features using the features of the main body region, so that the edge recognition is concentrated in the candidate range of the sample plot; then the edge features are used to reverse correct the boundary of the main body region, so that the main body response and the edge of the solid frame are consistent in spatial position.

[0027] S3: Optimize the fusion of main features, edge features, center point features, and corner features to predict four task probability maps, namely the main region probability map, entity edge probability map, center point probability map, and corner point probability map.

[0028] This invention combines optimized main features, optimized edge features, center point features, and corner point features to form a geometric structure feature that simultaneously includes information about the interior of the sample plot, its boundaries, the center of the instance, and the four corner points. In other words, this geometric structure feature contains a shared geometric representation of the surface-line-point relationship of the same instance.

[0029] The geometric structure features were predicted using four task-specific prediction heads, resulting in a four-task probability map consistent with the original image grid, coordinate reference system, and affine geographic transformation. (See...) Figure 1 Because each task exchanges geometric information of the same instance before making predictions, the structural relationships between the main body, edges, center, and corners are strengthened.

[0030] S4: Based on the four-task probability map, extract the set of candidate center points and the set of candidate corner points; Traverse the set of candidate center points, using each candidate center point as an anchor point, and search for candidate corner points in the preset neighborhood. Enumerate closed and non-self-intersecting quadrangle combinations to obtain multiple quadrangle combination sets.

[0031] This invention establishes topological associations of disordered corner points using the center point as the instance anchor point, which is used to reduce directional instability and endpoint drift caused by relying solely on the mask contour.

[0032] The specific implementation steps are as follows: S41: Based on the four-task probability map, perform peak extraction and repeated response suppression to obtain the set of candidate center points and the set of candidate corner points.

[0033] S42: The preset neighborhood is determined by the formula... To determine.

[0034] In the formula, , These represent the reference width and reference height of the spline frame at the pixel scale, respectively; This represents the theoretical distance from the anchor point to the corner point, serving as the benchmark for corner point search.

[0035] The reference width and reference height are determined by converting the actual physical width and height of the quadrat frame into the pixel ground dimensions of the UAV orthophoto. Let the actual width and actual height of the quadrat frame be respectively... , The ground pixel dimensions in the column and row directions of the orthophoto are respectively , ,but: .

[0036] S43: Traverse the set of candidate center points, using each candidate center point as an anchor point, search for candidate corner points in the preset neighborhood, enumerate closed and non-self-intersecting quadrangle combinations, and obtain multiple quadrangle combination sets.

[0037] S5: Filter the target four-corner point combination by matching cost function, as follows: Based on the combination of four corner points, calculate the symmetry error value of the distance from the corresponding anchor point to the four corner points; use the perpendicular error value of adjacent sides and the parallel error value of opposite sides to calculate the rectangular structure error value; sample the edge line probability value to obtain the edge line response value; extract the corner point prediction value to obtain the four corner response value; Based on the distance symmetry error value, rectangular structure error value, edge response value, and four-corner response value, the matching cost function value is calculated, and the four-corner point combination whose matching cost function value is less than the preset function value is taken as the target four-corner point combination.

[0038] The specific implementation steps are as follows: S51: Regarding the calculation of the symmetry error value of the distance from the corresponding anchor point to the four corner points, the specific details are as follows: Calculate the straight-line distances from the anchor point to the four corner points respectively, and obtain the average value of the four straight-line distances; then calculate the discrete deviation between the four distances and the average value, and combine the numerical stability constant to avoid the denominator being zero to obtain the original symmetric error; finally, normalize and compress the original error.

[0039] The calculation formula is as follows.

[0040] , .

[0041] In the formula, , These represent the original symmetry error and the distance symmetry error values, respectively. This represents the average distance of four straight lines; Indicates anchor point To the corner The straight-line distance; This represents the numerical stability constant.

[0042] S52: Regarding the calculation of the rectangular structure error value using the perpendicularity error value of adjacent sides and the parallelism error value of opposite sides, the specific steps are as follows: Sort the corner points cyclically according to their polar angles relative to the anchor points, and connect the four corners in cyclic order to obtain four side vectors; calculate the perpendicular error value of adjacent sides and the parallel error value of opposite sides; take the average of the perpendicular error value and the parallel error value as the rectangular structure error value.

[0043] The calculation formula is as follows.

[0044] Let the indices of corner points and edge vectors cycle modulo 4, i.e. , and define the edge vectors Calculate the perpendicular error value of adjacent sides Parallelism error value of opposite sides .

[0045] In the formula, ||·|2 represents the Euclidean norm; Represents edge vectors The transpose of .

[0046] S53: Regarding the edge response value obtained from the sampled edge probability value, specifically: Extract the predicted probabilities of all pixels on the entity edge probability map along the four edges, and calculate the average of all pixel probabilities as the edge response value.

[0047] S54: Regarding the extraction of corner prediction values ​​to obtain the four-corner response values, specifically: Extract the predicted probabilities of the four corner points on the corner probability map, and calculate the average of the predicted probabilities of the four corner points.

[0048] S55: Regarding the calculation of the matching cost function value based on the distance symmetry error value, the rectangular structure error value, the edge response value, and the four corner response value, the specific steps are as follows: The edge penalty value is generated by subtracting the edge response value from the numerical value 1, and the corner penalty value is generated by subtracting the four corner response values ​​from the numerical value 1. The distance symmetry error value, rectangular structure error value, edge penalty value and corner penalty value are weighted and summed to obtain the matching cost function value.

[0049] The edge response value and corner response value of the present invention are mainly used for intra-instance topology matching between center point candidates and corner point candidates, that is, to determine whether several corner points can belong to the same center point and form a target four-corner point combination.

[0050] In addition, it also includes the combination of the four corner points of each target, using the formula To calculate the overall score, remove those with an overall score below the scoring threshold. τS The combination of the four corner points of the target, such as Figure 2 .

[0051] In the formula, , , , , , These represent the combinations of the four corner points of the target. The comprehensive score, normalized response value of the main area, normalized response value of the entity edge, normalized response value of the center point, normalized response value of the corner point, and physical scale penalty value are given by α, β, γ, δ, and λ, which represent the non-negative weights of each item.

[0052] In the comprehensive scoring calculation formula, the determination of the scoring threshold is carried out in the following manner in this embodiment.

[0053] A calibration sample set is constructed using calibration images with manually labeled sample boxes. Given a set of five non-negative weights, feature extraction, four-task probabilistic map prediction, center point and corner point topological association, and matching cost screening are performed on each image in the calibration sample set according to the method described in this invention to obtain target quadrilateral combinations, which are then used to construct candidate quadrilaterals. The intersection-union ratio (IoU) is calculated for each candidate quadrilateral and the manually labeled sample box. All candidate correspondences with IoU greater than 0 are sorted from highest to lowest IoU. When no correspondence is established between a candidate quadrilateral and a manually labeled sample box, a one-to-one correspondence is established; otherwise, the candidate correspondence is skipped. For correspondences with the same IoU, the processing order is determined by the ascending order of the candidate quadrilateral number and the manually labeled sample box number. This one-to-one correspondence is used to determine the correspondence between candidate quadrilaterals and manually labeled sample boxes without changing the target quadrilateral combination construction and vectorization steps of this invention.

[0054] Under this set of weights, calculate the comprehensive score for each candidate quadrilateral in the calibration sample set, sort them according to the comprehensive score from smallest to largest, and remove duplicate values ​​to obtain... s (1) < s (2) < ……< s ( m Construct a set of candidate scoring thresholds: take the boundary values ​​that are lower than the lowest scoring value. s (1)- e Take the midpoint of every two adjacent score values ​​( s ( j ) + s ( j + 1)) / 2, and take the boundary value higher than the highest score value. s ( m ) + e ;in, e It is a positive number and less than the difference between any two adjacent distinct score values. The above boundary values ​​and midpoints form the candidate score threshold set corresponding to this set of weights.

[0055] For each candidate score threshold corresponding to this set of weights tc , retain satisfaction S ( q ) ≥ tc Candidate quadrilaterals, eliminate those that satisfy the following conditions. S ( q ) < tc The candidate quadrilaterals are identified; then, for the retained candidate quadrilaterals, the correspondence is re-established according to the above one-to-one IoU correspondence rule. Candidate quadrilaterals that have completed the correspondence are counted as true positives (TP), candidate quadrilaterals that have not completed the correspondence but are retained are counted as false positives (FP), and manually labeled sample boxes that have not completed the correspondence are counted as false negatives (FN). Precision, recall, and F1 score are calculated.

[0056] For this set of weights, the candidate scoring threshold that maximizes the F1 score is selected as the optimal threshold corresponding to this set of weights; when multiple candidate scoring thresholds have the same F1 score, the threshold with higher precision and higher recall is selected in sequence.

[0057] In the comprehensive scoring formula, the determination of the non-negative weights for each item is carried out in the following manner in this embodiment.

[0058] To determine the weights, a candidate weight set is constructed. Each weight is chosen from the set {0, 0.05, 0.10, …, 1.00}, satisfying the following condition: α ≥ 0, β ≥ 0, c ≥ 0, d ≥ 0, l ≥ 0; α + β + c + d + l = 1.

[0059] A candidate weight set is constructed from weight combinations that satisfy the above conditions. For each candidate weight set, the optimal scoring threshold corresponding to that weight set is calculated, and the target-level precision, recall, and F1 score corresponding to that weight set and its optimal scoring threshold are obtained. Among all candidate weight combinations, the weight combination with the largest F1 score is selected as the final weight combination, and the optimal scoring threshold corresponding to this weight combination is determined as the scoring threshold. τS .

[0060] When multiple weight combinations have the same F1 score, select the weight combination with higher precision and then higher recall in that order; if the above metrics are still the same, select the weight combination with higher precision and then higher recall in that order. α , β , c , d , l The weight combination with the smallest lexicographical order. The determined weight combination and scoring threshold. τS The image resolution, sample frame size range, and imaging conditions remain constant.

[0061] Specifically, , , , It is a combination of the four corner points of the target. The four types of region-level normalized responses for the overall evaluation of the constructed quadrilaterals (denoted as candidate quadrilaterals) are derived from the main region probability map, entity edge probability map, center point probability map, and corner point probability map, respectively.

[0062] in, The normalized response of the internal region of the candidate quadrilateral to the probability map of the main region can be obtained by averaging or weighted averaging the main probabilities of the pixels inside the candidate quadrilateral. This represents the normalized response of the candidate quadrilateral boundary on the entity edge probability map; This represents the normalized response of the candidate quadrilateral center on the center point probability map, where... This indicates the center point response, not the number of feature channels; This represents the normalized response of the four corner points of the candidate quadrilateral on the corner probability map.

[0063] The above four items are used for comprehensive scoring. The purpose is to evaluate whether a complete candidate quadrilateral as a whole conforms to the geometric response of the face, line, and point of a real sample plot.

[0064] This invention calculates the main body region, entity edge, center point, and corner point response of each target quadrilateral composed of its four corner points, and scores it in conjunction with real geographical scales such as area, side length, and aspect ratio.

[0065] The formula for calculating the physical scale penalty value is as follows.

[0066] Reference area Length of two adjacent reference sides (j∈{1,2}) and reference aspect ratio ,definition , and And calculate the physical scale penalty value. .

[0067] In the formula, The quadrilateral formed by the four corner points of the target. The area; Indicates the area deviation value; Represents a quadrilateral The The length of each adjacent side; This indicates the deviation value of the side length; Represents a quadrilateral The aspect ratio; This indicates the aspect ratio deviation value.

[0068] Considering that the constructed quadrilaterals may have significant errors compared to the actual quadrats due to factors such as shadows and crop occlusion, this invention first uses the matching cost function value for geometric matching, and then uses the comprehensive score for global evaluation. The two methods are respectively applied to intra-instance topological matching and inter-instance candidate evaluation.

[0069] In other words, the matching cost function value is used to determine that corner points belong to the same center point and form the target quadrilateral, but it only evaluates the geometric matching rationality of the four corner points. The comprehensive score is used to evaluate the entire domain of the complete closed quadrilateral, filter the constructed target quadrilateral, distinguish between real quadrilaterals and interference rectangles that have reasonable geometric matching but no internal matching, filter out false rectangles (field ridges, crop rows, shadows) with low scores, abnormal size, and aspect ratios that deviate from the standard, and complete the instance-level accurate quadrat box recognition.

[0070] S6: The pixel coordinates of the target's four corner points are converted into projected coordinates to obtain quadrat vector data.

[0071] The pixel coordinates of the target's four corner points are converted to projected coordinates through an orthophoto affine geographic transformation, resulting in SHP, GeoPackage, or other GIS vector files. The attribute table is then used to input the center coordinates, area, side length, rotation angle, four-task score, overall score, and optional calibration confidence level. This yields an editable, measurable, numberable GIS quadrat object that can be directly used for spatial overlay and multi-source clipping.

[0072] This invention first extracts features from UAV orthophotos with coordinate reference, classifying them into four geometric feature categories: subject, edge, center point, and corner points. Through bidirectional cross-task gating, the subject response and the entity's edge are kept spatially consistent. Then, the four geometric feature categories are fused and predicted to obtain a four-task probability map. Based on this four-task probability map, the center point and corner points are topologically associated to constrain spatial relationships. Next, geometric matching is performed using a matching cost function value, followed by a comprehensive evaluation using a comprehensive score. Low-scoring, abnormally sized, and aspect ratio-deviation-deviated false rectangles (field ridges, crop rows, shadows) are filtered out, achieving instance-level accurate quadrat box recognition and vectorization.

[0073] Example 2 Based on Example 1, this example provides a field quadrat recognition and vectorization system based on UAV imagery, including: Feature extraction module: After extracting features from the UAV orthophoto with coordinate reference, it decodes the main features, edge features, center point features, and corner features. Feature optimization module: The main features are processed by convolution and activation functions to generate a main guiding edge gating map, and the edge features are modulated to generate optimized edge features; the optimized edge features are then used to generate an edge guiding main gating map, and the main features are modulated to generate optimized main features. The four-task prediction module integrates optimized main features, optimized edge features, center point features, and corner features to predict four task probability maps, namely the main region probability map, entity edge probability map, center point probability map, and corner point probability map. Combination module: Based on the four-task probability map, extract the set of candidate center points and the set of candidate corner points; Traverse the set of candidate center points, and use each candidate center point as an anchor point to search for candidate corner points in a preset neighborhood to obtain multiple combinations of four corner points. Filtering module: Filters target four-corner point combinations by matching cost function, as follows: Based on the combination of four corner points, calculate the symmetry error value of the distance from the corresponding anchor point to the four corner points; use the perpendicular error value of adjacent sides and the parallel error value of opposite sides to calculate the rectangular structure error value; sample the edge line probability value to obtain the edge line response value; extract the corner point prediction value to obtain the four corner response value; Based on the distance symmetry error value, rectangular structure error value, edge response value, and four-corner response value, calculate the matching cost function value, and take the four-corner point combination whose matching cost function value is less than the preset function value as the target four-corner point combination. Vectorization module: The pixel coordinates of the target's four corner points are converted into projected coordinates to obtain quadrat vector data.

[0074] 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 changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for field quadrat recognition and vectorization based on UAV imagery, characterized in that, include: S1: After extracting features from the UAV orthophoto with coordinate reference, decode to obtain the main features, edge features, center point features and corner features; S2: The main features are processed by convolution and activation functions to generate a main body guiding edge gating map, and the edge features are modulated to generate optimized edge features; the optimized edge features are then used to generate an edge guiding main body gating map, and the main features are modulated to generate optimized main body features. S3: Optimize the fusion of main features, edge features, center point features and corner features to predict four task probability maps, namely the main region probability map, entity edge probability map, center point probability map and corner point probability map; S4: Based on the four-task probability map, extract the set of candidate center points and the set of candidate corner points; Traverse the set of candidate center points, and use each candidate center point as an anchor point to search for candidate corner points in a preset neighborhood to obtain multiple combinations of four corner points. S5: Filter the target four-corner point combination by matching cost function, as follows: Based on the combination of four corner points, calculate the symmetry error value of the distance from the corresponding anchor point to the four corner points; use the perpendicular error value of adjacent sides and the parallel error value of opposite sides to calculate the rectangular structure error value; sample the edge line probability value to obtain the edge line response value; extract the corner point prediction value to obtain the four corner response value; Based on the distance symmetry error value, rectangular structure error value, edge response value, and four-corner response value, calculate the matching cost function value, and take the four-corner point combination whose matching cost function value is less than the preset function value as the target four-corner point combination. S6: The pixel coordinates of the target's four corner points are converted into projected coordinates to obtain quadrat vector data.

2. The method for field quadrat recognition and vectorization based on UAV imagery according to claim 1, characterized in that, S5 describes the calculation of the symmetry error value of the distance from the corresponding anchor point to the four corner points, specifically as follows: Calculate the straight-line distances from the anchor point to the four corner points respectively, and obtain the average value of the four straight-line distances; then calculate the discrete deviation between the four distances and the average value, and combine the numerical stability constant to avoid the denominator being zero to obtain the original symmetric error; finally, normalize and compress the original error.

3. The method for field quadrat recognition and vectorization based on UAV imagery according to claim 1, characterized in that, S5 describes calculating the rectangular structure error value using the perpendicularity error value of adjacent sides and the parallelism error value of opposite sides, specifically as follows: Sort the corner points cyclically according to their polar angles relative to the anchor points, and connect the four corners in cyclic order to obtain four side vectors; calculate the perpendicular error value of adjacent sides and the parallel error value of opposite sides; take the average of the perpendicular error value and the parallel error value as the rectangular structure error value.

4. The method for field quadrat recognition and vectorization based on UAV imagery according to claim 1, characterized in that, The edge response value obtained from the sampling edge probability value described in S5 is as follows: Extract the predicted probabilities of all pixels on the entity edge probability map along the four edges, and calculate the average of all pixel probabilities as the edge response value.

5. The method for field quadrat recognition and vectorization based on UAV imagery according to claim 1, characterized in that, S5 describes the extraction of corner prediction values ​​to obtain four-corner response values, specifically as follows: Extract the predicted probabilities of the four corner points on the corner probability map, and calculate the average of the predicted probabilities of the four corner points.

6. The method for field quadrat recognition and vectorization based on UAV imagery according to claim 1, characterized in that, S5 describes the calculation of the matching cost function value based on the distance symmetry error value, the rectangular structure error value, the edge response value, and the four corner response values, specifically as follows: The edge penalty value is generated by subtracting the edge response value from the numerical value 1, and the corner penalty value is generated by subtracting the four corner response values ​​from the numerical value 1. The distance symmetry error value, rectangular structure error value, edge penalty value and corner penalty value are weighted and summed to obtain the matching cost function value.

7. The method for field quadrat recognition and vectorization based on UAV imagery according to claim 1, characterized in that, S5 also includes combining the four corner points of each target using the formula. To calculate the overall score, eliminate the target four-corner point combination whose overall score is lower than the score threshold; In the formula, , , , , , These represent the combinations of the four corner points of the target. The comprehensive score, normalized response value of the main area, normalized response value of the entity edge, normalized response value of the center point, normalized response value of the corner point, and physical scale penalty value are given by α, β, γ, δ, and λ, which represent the non-negative weights of each item.

8. The method for field quadrat recognition and vectorization based on UAV imagery according to claim 1, characterized in that, S2 specifically refers to: The main features are processed by convolutional mapping and the Sigmoid function to generate a gating map of the main body's guiding edge lines; The edge features are multiplied element-wise using the subject-guided edge gating graph to generate optimized edge features. The optimized edge features are processed by convolutional mapping and the Sigmoid function to generate an edge-guided gating map of the main body. By using the edge lines to guide the gating graph of the subject, the subject features are multiplied element by element to generate optimized subject features.

9. The method for field quadrat recognition and vectorization based on UAV imagery according to claim 1, characterized in that, The preset neighborhood mentioned in S4 is determined by the formula. To determine; In the formula, , These represent the reference width and reference height at the pixel scale, respectively. This represents the theoretical distance from the anchor point to the corner point.

10. A field quadrat recognition and vectorization system based on UAV imagery, characterized in that, include: Feature extraction module: After extracting features from the UAV orthophoto with coordinate reference, it decodes the main features, edge features, center point features, and corner features. Feature optimization module: The main features are processed by convolution and activation functions to generate a main guiding edge gating map, and the edge features are modulated to generate optimized edge features; the optimized edge features are then used to generate an edge guiding main gating map, and the main features are modulated to generate optimized main features. The four-task prediction module integrates optimized main features, optimized edge features, center point features, and corner features to predict four task probability maps, namely the main region probability map, entity edge probability map, center point probability map, and corner point probability map. Combination module: Based on the four-task probability map, extract the set of candidate center points and the set of candidate corner points; Traverse the set of candidate center points, and use each candidate center point as an anchor point to search for candidate corner points in a preset neighborhood to obtain multiple combinations of four corner points. Filtering module: Filters target four-corner point combinations by matching cost function, as follows: Based on the combination of four corner points, calculate the symmetry error value of the distance from the corresponding anchor point to the four corner points; use the perpendicular error value of adjacent sides and the parallel error value of opposite sides to calculate the rectangular structure error value; sample the edge line probability value to obtain the edge line response value; extract the corner point prediction value to obtain the four corner response value; Based on the distance symmetry error value, rectangular structure error value, edge response value, and four-corner response value, calculate the matching cost function value, and take the four-corner point combination whose matching cost function value is less than the preset function value as the target four-corner point combination. Vectorization module: The pixel coordinates of the target's four corner points are converted into projected coordinates to obtain quadrat vector data.

Images