A phenotypic image feature extraction method for screening of cinnamomum camphora characters

Abstract

The present application relates to the technical field of image pattern recognition, and discloses a phenotypic image feature extraction method for screening of Lagerstroemia indica characters, comprising: converting an original phenotypic image into a luminance feature matrix and a chroma component matrix, processing the luminance feature matrix by using an anisotropic diffusion filter to output a continuous gradient tensor field, calculating directional curvature alignment eigenvalues between spatial local topological nodes as topological weight values to be introduced into the chroma component matrix, determining multi-dimensional joint probability topological connectivity to construct a non-rigid manifold mesh skeleton spanning overlapping occlusion areas, and reducing an image sampling step length when the skeleton is distorted, wherein the present application eliminates target overlapping occlusion and light and shadow interference through geometric topological reconstruction and connectivity determination, realizes complete contour deduction of overlapping annihilation boundaries, and effectively improves feature recognition accuracy of dense aggregated character screening.

Description

Technical Field

[0001] This invention belongs to the field of image pattern recognition technology, and in particular relates to a method for extracting phenotypic image features for screening traits of Crape Myrtle. Background Technology

[0002] In current digital image analysis systems, extracting the morphological edge features of targets constitutes a fundamental step in image recognition and understanding. Conventional solutions employ fixed-scale operators to determine entity boundaries by solving local spatial pixel gradient abrupt changes. These solutions rely on the pre-defined conditions of independent targets and uniform external illumination. By setting a fixed threshold to segment the gray-level tensor, discretization of planar features is achieved. However, when multiple targets are densely clustered and there are non-rigid overlapping occlusions and non-uniform scattering of light radiation, the pixel spatial differential tensor field produces a monotonically decreasing first derivative amplitude in the overlapping boundary region. When encountering surface sections with densely overlapping leaves and inflorescences, the contrast at adjacent boundaries is locally annihilated, and the mottled tree shadows and target textures intertwine spatially, causing the differential tensor field to lose its topological continuity. Conventional operators only extract local step features and lack spatiotemporal dimensional linkage constraints, leading to the misjudgment of shadow noise as morphological boundaries, resulting in feature fragmentation and missed detection of features.

[0003] To mitigate feature fracture defects, conventional improvement approaches tend to add external multi-channel sensors or use global resampling gain to improve resolution. However, this linear improvement path, which relies on increasing input computation, not only triggers severe edge oversegmentation due to the introduction of high-frequency non-feature scattering noise, but also causes an exponential increase in the time consumption of the weight update path in multi-object matrix search, making it impossible to achieve stable convergence of data flow in standard equipment. Some improvement schemes improve physical transmission conditions by optimizing specific mechanical structure designs, adjusting roller morphology, or relying on external multi-channel sensor hardware reinforcement, but they do not change the underlying solution mechanism and still cannot overcome the inherent limitations of software control methods and algorithms. For example, Chinese invention patent CN114219953B discloses a method based on anisotropy The multi-scale detection image feature matching method of diffusion operation improves the edge preservation of anisotropic diffusion by introducing the transmission formula and combines the Hessian matrix for feature point detection. The core premise of this existing technology is that there are explicit regular geometric edges and relatively stable rigid textures in the image. This is not compatible with the phenotypic screening conditions of outdoor real plants with multiple targets densely clustered and prone to non-rigid deformation accompanied by mottled light and shadow non-uniform scattering. Under the non-ideal boundary conditions of dense crape myrtle leaves or inflorescences overlapping each other over a large area and local contrast annihilation, the first and second normalized derivatives on which this technology relies are prone to floating-point overflow or edge numerical blunting due to the nonlinear interweaving of image gradients, resulting in instability of the spatial nonlinear scale spatial topology and causing the feature matching matrix weight update path to diverge.

[0004] Therefore, how to maintain low-contrast overlapping boundary gradients in situ and reconstruct a spatial manifold network skeleton with strong noise resistance in high-density intensive screening scenarios has become the technical problem to be solved by this invention. Summary of the Invention

[0005] This invention aims to solve the problem of losing geometric topological consistency in high-density intensive screening scenarios due to target overlap and occlusion and light and shadow interference, which makes it impossible to maintain a weak contour gradient in situ.

[0006] In this technical solution, a phenotypic image feature extraction method for screening trait in Crape Myrtle includes the following steps: Step S1: Obtain the original phenotypic image containing overlapping occluded crape myrtle targets, convert the original phenotypic image into luminance channel data and color channel data, and construct the luminance feature matrix and color metric component matrix based on the luminance channel data and color channel data respectively. Step S2: Input the brightness feature matrix into the spatial anisotropic diffusion filter, calculate the diffusion coefficient boundary in the neighborhood of the current pixel to perform gradient smoothing, and output a continuous gradient tensor field. Step S3: In the continuous gradient tensor field, taking the spatial local pixel cluster as the topological node, calculate the directional curvature alignment feature value between adjacent topological nodes based on the maximum first-order gradient vector magnitude, the minimum first-order gradient vector magnitude, the arithmetic mean of the gradient vectors, and the preset boundary stability constant within the preset local sliding window. Step S4: The directional curvature alignment feature value is used as the topological connectivity weight and introduced into the color edge gradient distribution of the color degree component matrix. Multidimensional joint probability topological connectivity determination is performed to construct a non-rigid manifold mesh skeleton that spans the overlapping occlusion region in the original phenotypic image. When the non-rigid manifold mesh skeleton deforms in the overlapping occlusion region, the image sampling step size is reduced, and the crape myrtle phenotypic image features are output.

[0007] Preferably, step S1 includes the following sub-steps: step S11, converting the acquired original phenotypic image into luminance channel data and color channel data; step S12, extracting the pixel-level intensity distribution corresponding to the luminance channel data to construct a luminance feature matrix, and extracting the spectral wavelength energy distribution corresponding to the color channel data to construct a colorimetric component matrix.

[0008] Preferably, step S2 includes the following sub-steps: step S21, calculate the gray-level gradient difference between the current pixel and its neighboring pixels in the brightness feature matrix, and determine the diffusion coefficient boundary based on the gray-level gradient difference; step S22, control the brightness feature matrix to perform anisotropic diffusion smoothing along the diffusion coefficient boundary, and output a continuous gradient tensor field.

[0009] Preferably, the multidimensional joint probabilistic topological connectivity determination in step S4 includes the following sub-steps: Step S41, establish a joint probabilistic connectivity matrix and convert the directional curvature alignment feature value into the basic connectivity weights in the joint probabilistic connectivity matrix; Step S42, extract the color edge gradient features in the color intensity component matrix, perform conditional probability correction on the basic connectivity weights based on the color edge gradient features, and obtain the final joint connectivity probability value; Step S43, determine the connectivity of adjacent topological nodes based on the joint connectivity probability value.

[0010] Preferably, the reduction of the image sampling step size in step S4 includes the following sub-steps: Step S44, when constructing the non-rigid manifold mesh skeleton, measuring the mesh distortion rate of the non-rigid manifold mesh skeleton in a local space; Step S45, when the mesh distortion rate exceeds the set distortion threshold, determining that the non-rigid manifold mesh skeleton has entered an overlapping occlusion region, and reducing the current image sampling step size to 0.5 of the original sampling step size.

[0011] Preferably, the method further includes the following sub-steps in the pre-step stage of step S1: step S101, calculating the global hyperspectral energy variance of the input original phenotypic image; step S102, when the global hyperspectral energy variance is lower than the sharpness reference threshold of 0.65, retrieving temporally adjacent phenotypic image frames for similarity geometric reconstruction and alignment repair until the global hyperspectral energy variance of the reconstructed image frame is not lower than 0.65, and providing the reconstructed image frame as the original phenotypic image to step S1.

[0012] Preferably, the method further includes the following sub-steps in the post-stage of step S4: Step S46, when the edge contrast of the overlapping occlusion area decreases to below the contrast threshold, feedforward deduction is performed based on the consistency of the first-order differential geometric flow direction between adjacent topological nodes in the non-rigid manifold mesh skeleton to complete the topological flow direction response on both sides of the missing boundary and restore the occlusion contour of the crape myrtle target in the original phenotypic image.

[0013] Preferably, the method further includes the following sub-steps in the post-stage of step S4: Step S47, extracting phenotypic structural trait indicators of the Lagerstroemia indica target from the original phenotypic image based on the closed edge region locked by the non-rigid manifold mesh skeleton, the phenotypic structural trait indicators include leaf area, flower cluster density and branching angle; Step S48, screening Lagerstroemia indica traits based on the phenotypic structural trait indicators.

[0014] Preferably, the calculation of the diffusion coefficient boundary in the neighborhood of the current pixel in step S2 includes the following sub-steps: Step S23, in the filtering process of the anisotropic diffusion filter of the brightness feature matrix, a spatial neighborhood decoupling operator is introduced to calculate the second-order differential convection terms of the current pixel in the horizontal and vertical directions; Step S24, the diffusion coefficient boundary is corrected in real time according to the second-order differential convection terms to suppress the error accumulation of the shadow edge in the spatiotemporal evolution.

[0015] Compared with existing technologies, the phenotypic image feature extraction method for screening crape myrtle traits of this invention has the following advantages: 1. In the phenotypic image feature extraction for crape myrtle trait screening, the digital image frame is decoupled into a brightness feature matrix and a color component matrix by the image separation module. This cuts off the direct interference of ambient light fluctuations on color space features. The multi-level feature decoupling operator uses the spatial anisotropic mass transfer flow field to process the brightness feature matrix. While smoothing high-frequency mottled light and shadow noise, it preserves and strengthens the edge step gradient of the local overlapping area in situ, thereby outputting a continuous gradient tensor field. This changes the dependence of traditional fixed-scale edge detection operators on the continuity of explicit pixel contrast, allowing weak edge data in the sorting scene to be completely preserved, providing a data foundation with strong noise resistance for high-dimensional spatial geometric topology reconstruction.

[0016] 2. The manifold reconstruction operator, within a continuous gradient tensor field, uses spatial local pixel clusters as network nodes to calculate the directional curvature alignment feature value between adjacent nodes. It uses the difference between the maximum and minimum magnitudes of the gradient vectors within the local sliding window as the numerator and the sum of the arithmetic mean of the gradient vectors and the stable correction term as the denominator to obtain the ratio. This provides a steady-state numerical balance in regions where edge contrast is annihilated by environmental scattering, effectively simplifying the weight update path of high-dimensional topology matching. It enables sensitive topological flow responses on both sides of the boundaries of densely aggregated and highly self-similar entities, reduces the time consumption caused by invalid matrix search, and solves the problems of edge misclassification and computational divergence in densely overlapping regions.

[0017] 3. The spatial topology alignment operator uses the directional curvature alignment feature value as the connection weight and introduces the color edge gradient distribution of the color degree component matrix. It completes the joint probabilistic connectivity arbitration at the topology layer and constructs a non-rigid manifold mesh skeleton that spans the overlapping occlusion boundary. The manifold mesh skeleton simultaneously applies real-time reverse boundary constraints to the local sliding window range. When the topology mesh deforms in the occlusion core area, it adaptively narrows the sampling step size. This not only establishes a deterministic closed loop where the underlying control parameters are interdependent, but also enables the system to implicitly feedforward to deduce the complete contour when the overlap boundary is annihilated. This solves the dilemma of misjudging phenotypic parameters in high-density groups and improves the accuracy of feature recognition and matching. Attached Figure Description

[0018] Figure 1This is a flowchart of the output of the phenotypic image features of the Lagerstroemia indica target by overlapping occlusion according to the present invention; Figure 2 This is a screening diagram of the morphology of Lagerstroemia indica using a non-rigid manifold mesh framework, as described in this invention. Figure 3 This is a schematic diagram of the data processing and closed-loop sorting system structure of the present invention; Figure 4 This is a schematic diagram illustrating the principle of calculating the directional curvature alignment characteristic value of this invention. Figure 5 This is a graph showing the relationship between the occlusion area ratio and the feature misclassification rate of this invention. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0020] A method for extracting phenotypic image features for screening traits in Crape Myrtle includes the following steps: Step S1: Obtain the original phenotypic image containing overlapping occluded crape myrtle targets, convert the original phenotypic image into luminance channel data and color channel data, and construct the luminance feature matrix and color metric component matrix based on the luminance channel data and color channel data respectively. Step S2: Input the brightness feature matrix into the spatial anisotropic diffusion filter, calculate the diffusion coefficient boundary in the neighborhood of the current pixel to perform gradient smoothing, and output a continuous gradient tensor field. Step S3: In the continuous gradient tensor field, taking the spatial local pixel cluster as the topological node, calculate the directional curvature alignment feature value between adjacent topological nodes based on the maximum first-order gradient vector magnitude, the minimum first-order gradient vector magnitude, the arithmetic mean of the gradient vectors, and the preset boundary stability constant within the preset local sliding window. Step S4: The directional curvature alignment feature value is used as the topological connectivity weight and introduced into the color edge gradient distribution of the color degree component matrix. Multidimensional joint probability topological connectivity determination is performed to construct a non-rigid manifold mesh skeleton that spans the overlapping occlusion region in the original phenotypic image. When the non-rigid manifold mesh skeleton deforms in the overlapping occlusion region, the image sampling step size is reduced, and the crape myrtle phenotypic image features are output.

[0021] Preferably, step S1 includes the following sub-steps: step S11, converting the acquired original phenotypic image into luminance channel data and color channel data; step S12, extracting the pixel-level intensity distribution corresponding to the luminance channel data to construct a luminance feature matrix, and extracting the spectral wavelength energy distribution corresponding to the color channel data to construct a colorimetric component matrix.

[0022] Preferably, step S2 includes the following sub-steps: step S21, calculate the gray-level gradient difference between the current pixel and its neighboring pixels in the brightness feature matrix, and determine the diffusion coefficient boundary based on the gray-level gradient difference; step S22, control the brightness feature matrix to perform anisotropic diffusion smoothing along the diffusion coefficient boundary, and output a continuous gradient tensor field.

[0023] Preferably, the multidimensional joint probabilistic topological connectivity determination in step S4 includes the following sub-steps: Step S41, establish a joint probabilistic connectivity matrix and convert the directional curvature alignment feature value into the basic connectivity weights in the joint probabilistic connectivity matrix; Step S42, extract the color edge gradient features in the color intensity component matrix, perform conditional probability correction on the basic connectivity weights based on the color edge gradient features, and obtain the final joint connectivity probability value; Step S43, determine the connectivity of adjacent topological nodes based on the joint connectivity probability value.

[0024] Preferably, the reduction of the image sampling step size in step S4 includes the following sub-steps: Step S44, when constructing the non-rigid manifold mesh skeleton, measuring the mesh distortion rate of the non-rigid manifold mesh skeleton in a local space; Step S45, when the mesh distortion rate exceeds the set distortion threshold, determining that the non-rigid manifold mesh skeleton has entered an overlapping occlusion region, and reducing the current image sampling step size to 0.5 of the original sampling step size.

[0025] Preferably, the method further includes the following sub-steps in the pre-step stage of step S1: step S101, calculating the global hyperspectral energy variance of the input original phenotypic image; step S102, when the global hyperspectral energy variance is lower than the sharpness reference threshold of 0.65, retrieving temporally adjacent phenotypic image frames for similarity geometric reconstruction and alignment repair until the global hyperspectral energy variance of the reconstructed image frame is not lower than 0.65, and providing the reconstructed image frame as the original phenotypic image to step S1.

[0026] Preferably, the method further includes the following sub-steps in the post-stage of step S4: Step S46, when the edge contrast of the overlapping occlusion area decreases to below the contrast threshold, feedforward deduction is performed based on the consistency of the first-order differential geometric flow direction between adjacent topological nodes in the non-rigid manifold mesh skeleton to complete the topological flow direction response on both sides of the missing boundary and restore the occlusion contour of the crape myrtle target in the original phenotypic image.

[0027] Preferably, the method further includes the following sub-steps in the post-stage of step S4: Step S47, extracting phenotypic structural trait indicators of the Lagerstroemia indica target from the original phenotypic image based on the closed edge region locked by the non-rigid manifold mesh skeleton, the phenotypic structural trait indicators include leaf area, flower cluster density and branching angle; Step S48, screening Lagerstroemia indica traits based on the phenotypic structural trait indicators.

[0028] Preferably, the calculation of the diffusion coefficient boundary in the neighborhood of the current pixel in step S2 includes the following sub-steps: Step S23, in the filtering process of the anisotropic diffusion filter of the brightness feature matrix, a spatial neighborhood decoupling operator is introduced to calculate the second-order differential convection terms of the current pixel in the horizontal and vertical directions; Step S24, the diffusion coefficient boundary is corrected in real time according to the second-order differential convection terms to suppress the error accumulation of the shadow edge in the spatiotemporal evolution.

[0029] Example 1: When the system faces a high-density, intensive sorting environment, the non-rigid overlapping and occlusion within densely clustered flower heads and the spatial self-similarity between petals of the same type constitute interference factors for image feature extraction. Furthermore, the interplay of direct strong light and dappled light and shadow causes the gray-level gradient at the junction of overlapping petals to exhibit a first-order derivative amplitude attenuation. The gray-level gradient at the overlapping edges is annihilated by ambient scattered light. Traditional fixed-scale edge detection operators lack the ability to analyze the geometric and topological evolution of the target space, resulting in edge feature breaks at overlapping edges. This merges the overlapping separation edges into pixel color blocks, leading to the problem of missed detection of key phenotypic traits. The image separation module acquires a two-dimensional color image frame. By performing orthogonal projection transformation on the color channels of the two-dimensional color image frame, it converts the two-dimensional color image frame into a brightness feature matrix representing the energy spatial distribution and a chromaticity component matrix representing the spectral wavelength distribution, thereby reducing the interference of light radiation fluctuations on color texture components.

[0030] The spectral wavelength energy distribution is constructed based on the inverse approximation estimation of the standard colorimetric matching function. When processing the integrated photoelectric response data of red, green, and blue in conventional two-dimensional digital images, this invention utilizes a three-row, three-column linear reconstruction matrix pre-embedded in the image separation module to transform discrete color channel values ​​into a normalized combination of relative energy weights of the dominant wavelength. This orthogonal projection transformation completely separates the mixed illuminance components, making each element in the output colorimetric component matrix directly correspond to the relative spectral power distribution ratio of that pixel in the visible light band. Thus, at the algorithm level, a physical transformation is achieved across the color space to an approximate spectral wavelength energy distribution. Before obtaining the colorimetric component matrix, the image acquisition unit acquires a preset uniform diffuse reflection calibration plate digital image, and the data processing unit calculates the average gray value of each spatial pixel channel under uniform diffuse reflection conditions to determine the channel gain deviation correction coefficient of the current photosensitive chip. When processing the crape myrtle sample image, the pixel values ​​of the original red, green, and blue channels are subtracted by the corresponding channel gain deviation correction coefficients to filter out external diffuse reflection. The baseline zero-point drift caused by irradiation and high-frequency shadow noise is addressed by the feature decoupling module receiving the brightness feature matrix and processing it through a spatial anisotropic mass transfer control loop. This module smooths high-frequency light and shadow noise using the thermal conductivity diffusion boundary within the current pixel's neighborhood, preserving the gradient boundary of locally overlapping areas and outputting a continuous gradient tensor field. During this anisotropic mass transfer control process, the spatial neighborhood decoupling operator specifically performs a 5×5 pixel range discretized window convolution operation. This operator reads the gray values ​​of two adjacent pixels symmetrically distributed in the horizontal and vertical directions of the current center pixel. By adding the gray values ​​of the symmetrical pixels on both sides and subtracting twice the gray value of the center pixel, it independently calculates the second-order difference term in the horizontal and vertical directions. The difference between these two second-order differences is then introduced as a convection constraint term into the thermal conductivity diffusion boundary equation, achieving real-time dynamic position correction of the diffusion coefficient boundary and preventing the gradient smoothing flow from blurring the true edge line. The manifold reconstruction operator uses spatially local pixel clusters as topological network nodes within the continuous gradient tensor field.

[0031] The directional curvature alignment characteristic value between adjacent topological network nodes is calculated. This characteristic value characterizes the first-order differential geometric flow consistency of the superimposed non-rigid edges during spatial evolution. The calculation formula is as follows: ,in, This represents the directional curvature alignment characteristic value between adjacent nodes in the topology network. Let $\mathbf{a}$ be the maximum vector magnitude of the first-order gradient of a node in the topological network within a preset local sliding window. Let be the minimum vector magnitude of the first-order gradient of a node in the topological network within a preset local sliding window. The arithmetic mean of the gradient vectors within a preset local sliding window. The preset boundary stability constant is fixed at 0.004. It is used to provide numerical balance in the denominator when the sliding window slides to the edge region of the overlapping boundary and the difference between the magnitude of the first-order gradient maximum vector and the magnitude of the first-order gradient minimum vector approaches 0, suppressing computational overflow and making the output directional curvature alignment characteristic value show a monotonically decreasing trend, thus representing the geometric evolution law of the boundary region.

[0032] The spatial topology alignment operator uses the directional curvature alignment feature value as the topological connection weight and introduces it into the color edge gradient distribution of the chromaticity component matrix to construct a non-rigid manifold mesh skeleton that spans the overlapping occlusion region in the original phenotypic image. At the spatial mesh topology layer, the system uses multi-dimensional joint probability topological connectivity determination to convert the directional curvature alignment feature value into the basic connection weight in the joint probability connection matrix. The system then uses the color edge gradient feature in the chromaticity component matrix to provide conditional probability correction for the basic connection weight to obtain the joint connectivity probability value. Based on the joint connectivity probability value, the connectivity state of adjacent topological network nodes is determined. Furthermore, the transformation and dimensionality reduction calculation path of the multi-dimensional joint probability adopts the following deterministic mapping rule: the dimensionless directional curvature alignment feature value is normalized to a numerical range between 0 and 1 using a logistic mapping function, serving as the basic connection weight in the joint probability connection matrix; the color edge gradient magnitude at the corresponding position of the current node is extracted, and a geometric structure weight allocation coefficient of 0.6 and a color distribution weight allocation coefficient of 0.4 are used to adjust the basic connection weight and the normalized value. The color edge gradient features are weighted and summed. When the calculated comprehensive joint connectivity probability value is not lower than the set threshold of 0.5, the two adjacent topological network nodes are judged to be in a fully connected state in terms of topological logic. The multi-dimensional feature connectivity judgment is based on the principle of multi-feature linear fusion. The gradient annihilation effect of single color or local geometric contour in the occluded area is eliminated by the algebraic superposition of orthogonal spatial features. The geometric structure weight allocation coefficient of 0.6 is used to quantify the control weight of spatial geometric mesh topology on contour continuity inference. The color distribution weight allocation coefficient of 0.4 is used to bear the edge step response caused by color difference. The weighted sum value is used as the final connectivity index between topological network nodes. The matrix element flags of the corresponding row and column in the topological adjacency matrix are directly rewritten. When the non-rigid manifold mesh skeleton deforms in the overlapping occluded area and the measured spatial local mesh distortion rate exceeds the set distortion threshold, the system controls the narrowing of the sampling step size of the spatial pixel cluster, reducing the current image sampling step size to 0.5 times the original sampling step size. The high-dimensional phenotypic feature pattern vector is output using the closed edge region locked by the non-rigid manifold mesh skeleton.

[0033] When the real-time wind speed in the field exceeds 3.5 m / s and the original phenotypic image captured by the acquisition unit exhibits pixel dynamic blurring, the sharpness detection operator calculates the global hyperspectral energy variance of the input original phenotypic image. When the global hyperspectral energy variance is lower than the sharpness baseline threshold of 0.65, the system retrieves temporally adjacent phenotypic image frames and performs similarity geometric reconstruction and alignment repair based on temporal neighbor frame similarity alignment logic until the global hyperspectral energy variance of the reconstructed image frame is not lower than 0.65. This ensures that under dynamic blurring interference, the feature decoupling operator and the directional gradient clustering operator obtain the spatial topology matrix, completing the data flow transmission against motion blur. The global hyperspectral energy variance... Based on the quadratic statistical calculation of frequency domain energy, its mathematical formula is as follows: In the formula, For image resolution, For frequency domain response, It is a high-pass filter operator. The high-spectral energy mean and the sharpness benchmark threshold of 0.65 are empirical extreme values ​​derived from statistical analysis of 500 Lagerstroemia indica sample images under various complex conditions, including wind swaying and strong light scattering, during the pre-calibration stage of the sorting system. Experimental measurements show that when the normalized global high-spectral energy variance is... When the value is below 0.65, according to the formula for the attenuation ratio of the local step gradient amplitude: in, To clarify the gradient magnitude at the edges of the image, Calculations of gradient magnitudes at the same location in a motion-blurred image show that the local step gradient magnitude at the image edge experiences an overall attenuation of over 35% due to motion blur. This causes the topological network to diverge when searching for edges. However, by using similarity geometric recombination to forcibly restore and stabilize this variance at a level not lower than 0.65, it can be ensured that the high-frequency features of the edges are restored to the value range that the algorithm can accurately capture. The pattern recognition system processes the brightness feature matrix and the color component matrix to deduce the outline of the target boundary in the area of ​​overlapping and occlusion by plants, according to the edge misclassification probability formula: ,in, This represents the number of misclassified edge nodes within the overlapping occlusion region. Given the total number of edge nodes within the overlapping occlusion region, the probability of misclassification of image feature edges within the overlapping occlusion region is calculated. The correlation coefficient is 1.2%, and the total convergence time for feature extraction of a single image is 45ms. Meanwhile, according to the Spearman rank correlation coefficient formula: ,in, For the first The algorithm extracts the grade difference between trait indicators and manually measured values. Given the total number of test samples, the monotonic correlation coefficients between the extracted geometric cross-sectional parameters, texture distribution index, and artificially measured physical values ​​were calculated. The value is 0.98, and the system outputs a high-dimensional phenotypic feature pattern vector for trait sorting. Although this invention uses the screening of phenotypic traits of Lagerstroemia indica as an example for specific illustration, since the anisotropic mass transfer flow field and non-rigid manifold mesh skeleton calculated by this invention are based on the optical and geometric topological continuity that is universally present in plant tissues, this method is also applicable to the extraction of phenotypic features and trait screening of overlapping occlusion phenotypic images of other densely aggregated plants, such as lilac inflorescences, hydrangea corollas, fruit tree canopies, and dense spikes of crops.

[0034] Example 2: When the system faces a feature fracture condition caused by non-rigid overlapping occlusion within densely clustered flower patterns, the experimental test device includes an image acquisition unit with a resolution of 3840×2160 pixels and a sampling frame rate of 30Hz, used to reproduce the physical boundary scene of overlapping contour grayscale gradient amplitude attenuation in an outdoor environment with ambient scattered light interference; to test the actual stability of feature decoupling and manifold reconstruction, it is necessary to determine the size of the preset local sliding window and the preset boundary stability constant in the calculation formula. The mathematical values ​​of the preset local sliding window are determined by the relationship between the smoothing requirements of high-frequency noise in the local signal and the spatial positioning accuracy of the local edge, while the preset boundary stability constant is... The value determines the balance between the numerical stability boundary of the low-contrast boundary region and the directional curvature contrast sensitivity. If a preset boundary stability constant is used... Setting the value too small risks numerical overflow in the edge region where the gradient magnitude approaches zero. If a boundary stability constant is preset... Setting it too high will compress the directional curvature alignment feature value. The dynamic response range leads to numerical passivation of the geometric flow direction of the overlapping edges. According to the spatial power spectral density distribution rule of image signals, when the amplitude of the high-frequency noise component in the pixel neighborhood is greater than 0.15 times the gray-level step amplitude, in order to maintain the statistical stability of the directional differential calculation, the sampling side length of the preset local sliding window is set to 15 pixels. When the local geometric curvature radius of the target overlapping area is less than 5 pixels, the sampling side length of the preset local sliding window is set to 5 pixels. For the superimposed crape myrtle sample image in this test, the size of the preset local sliding window is set to a square matrix of 9×9 pixels, and the preset boundary stability constant is set to... Set within the range of 0.001 to 0.010.

[0035] To verify the feature extraction separation rules, the test system divided the acquired image sequences into the following groups: the sample group of this invention, the comparison sample group using the fixed-scale edge mapping operator, the partially missing control group with the spatial anisotropic mass transfer control loop removed, and the control group with the preset boundary stability constant changed. The out-of-range control group, in the first test scenario, inputs original image data containing a signal-to-noise ratio of 20dB and 50Hz power frequency interference fluctuations, resulting in a non-linear, jagged interlacing of grayscale gradients at the overlapping boundary. In the second test scenario, the original image data is converted into a luminance feature matrix and a chromaticity component matrix, and the orientation features are solved using a manifold reconstruction operator. The specific calculation formula is as follows: ,in, This represents the directional curvature alignment characteristic value between adjacent nodes in the topology network. Let $\mathbf{a}$ be the maximum vector magnitude of the first-order gradient of a node in the topological network within a preset local sliding window. Let be the minimum vector magnitude of the first-order gradient of a node in the topological network within a preset local sliding window. The arithmetic mean of the gradient vectors within a preset local sliding window. The preset boundary stability constant is used to provide a boundary numerical stable equilibrium; under the aforementioned calculation formula, the sample group of the present invention will use the preset boundary stability constant. Set to 0.004, measure the magnitude of the first-order gradient maximum vector within a sliding window at the overlapping boundary edge point. The first-order gradient minimum vector magnitude is 1.24 pixels of grayscale difference. The grayscale difference is 0.11 pixels, and the arithmetic mean of the gradient vector is... The directional curvature alignment feature value was calculated based on a grayscale difference of 0.43 pixels. The value is 2.604. This intermediate parameter exhibits geometric continuity within the spatial neighborhood, thus forming mesh connection weights in the mesh layer. However, for the partially missing control group, the lack of a spatial anisotropic mass transfer control loop to suppress high-frequency interference leads to the introduction of pseudo-edge noise within the local sliding window. Measurements show that... Increase the grayscale difference to 4.15 pixels. The grayscale difference is 0.08 pixels. The calculated directional curvature alignment feature value is the result of a 1.82 pixel grayscale difference. The value is changed to 2.231, disrupting the mathematical monotonicity of the edge flow; simultaneously, in the out-of-range control group, when the preset boundary stability constant is changed... When the value is reduced to 0.0005, measurements are taken in the same overlapping edge region. The grayscale difference is 0.02 pixels. The grayscale difference is 0.00 pixels. When the grayscale difference is 0.00 pixels, the denominator drops to 0.0005, triggering the directional curvature alignment feature value. A sudden change to 40.000 causes a local floating-point overflow, while when the preset boundary stability constant is set... When increased to 0.025, in the same overlapping edge region as above... , , Calculate the directional curvature alignment eigenvalues ​​of the output. Compared to the preferred value The feature value output by the weak edge The excessively high numerical setting introduces an overly strong smoothing penalty term, which causes the eigenvalue response to decay significantly, compresses the dynamic response range of the directional curvature contrast, and leads to numerical passivation at the edges of weak gradients.

[0036] In the third test scenario, the test system was configured with a gradient control system for problem intensity, including low-level occlusion, medium-level occlusion, and high-overlap occlusion. The contour inference feature output was quantitatively measured when the occlusion area ratio was 20%, 50%, and 80%, respectively. Experimental data showed that during the interference process where the occlusion ratio increased from 20% to 80%, the edge misclassification probability of the comparison sample using the fixed-scale edge mapping operator increased dramatically from 4.6% to 28.4%, exhibiting local failure due to manifold network breakdown and divergence. In contrast, the sample of this invention, relying on the deep coupling of the spatial anisotropic mass transfer control loop and the joint connectivity probability criterion, maintained an edge misclassification probability of 0.3% at an occlusion ratio of 20% and controlled it at 0.7% at an occlusion ratio of 50%. Even under severe occlusion conditions with an occlusion rate of 80%, the misclassification probability of feature edges remained stable at 1.2%. Furthermore, the total convergence time for feature extraction of a single image remained consistently between 43.6 ms and 45.2 ms across all gradients, without exponentially increasing with increasing interference intensity. Moreover, the monotonic correlation coefficient between the extracted geometric cross-sectional parameters, texture distribution index, and artificially measured physical values ​​was 0.982. This directly confirms that the technical approach of replacing explicit texture extraction with geometric topological manifold reconstruction solves the technical problem of feature structure fragmentation caused by overlapping and interweaving of boundary information in high-density sorting environments. While maintaining a deterministic closed loop for the entire system's data flow, it provides an image pattern recognition data foundation for the automated sorting of phenotypic traits in densely aggregated plants.

[0037] Example 3: This example combines Figures 1 to 5 This paper describes a phenotypic image feature extraction method for screening traits in crape myrtle, such as... Figure 1As shown, step S1 involves acquiring the original phenotypic image containing overlapping occluded crape myrtle targets, converting the original phenotypic image into luminance channel data and color channel data, and constructing the luminance feature matrix and chrominance component matrix based on the luminance channel data and color channel data, respectively. Step S2 then proceeds to step S2, where the luminance feature matrix is ​​input into a spatial anisotropic diffusion filter, and the diffusion coefficient boundary in the neighborhood of the current pixel is calculated for gradient smoothing, outputting a continuous gradient tensor field. Step S2 then proceeds to step S2, where, within the continuous gradient tensor field, using spatial local pixel clusters as topological nodes, the first-order gradient maximum vector within a preset local sliding window is used. Step S3 calculates the directional curvature alignment feature value between adjacent topological nodes using the magnitude, the first-order gradient minimum vector magnitude, the arithmetic mean of the gradient vector, and the preset boundary stability constant. Step S3 then proceeds to step S4, which uses the directional curvature alignment feature value as the topological connection weight and introduces it into the color edge gradient distribution of the color degree component matrix to perform multidimensional joint probabilistic topological connectivity determination. This constructs a non-rigid manifold mesh skeleton that spans the overlapping occlusion region in the original phenotypic image. When the non-rigid manifold mesh skeleton deforms in the overlapping occlusion region, the image sampling step size is reduced, and the crape myrtle phenotypic image features are output.

[0038] like Figure 2 As shown, the sharpness detection step, based on the trigger condition that the global high-spectral energy variance is lower than the sharpness reference threshold, flows to the similarity geometric reconstruction and alignment repair step. The similarity geometric reconstruction and alignment repair step, based on the conditions provided by the reconstructed image frame, flows to the sharpness detection step. The sharpness detection step, based on the trigger condition that the global high-spectral energy variance is not lower than the sharpness reference threshold, flows to the non-rigid manifold mesh skeleton construction step. The non-rigid manifold mesh skeleton construction step, based on the condition that the mesh distortion rate exceeds the set distortion threshold, flows to the image sampling step size reduction step. The image sampling step size reduction step, based on the condition that the edge contrast decreases to below the contrast threshold, flows to the first-order differential geometric flow feedforward deduction step. The first-order differential geometric flow feedforward deduction step, based on the condition of locking the closed edge region, flows to the crape myrtle morphology screening step.

[0039] like Figure 3As shown, the image acquisition unit captures a digital image and inputs it to the affinity matrix circuit. The calibration module generates a bias and inputs it to the affinity matrix circuit. Inside the affinity matrix circuit, the digital image is input to the row buffer queue unit, and the bias is input to the pixel gain configuration register group. The data from the row buffer queue unit and the pixel gain configuration register group are input to the parallel multiply-add interleaved operation array for calculation and output of the consistency correction matrix. The consistency correction matrix is ​​stored in the static random access memory and output as a correction coefficient to the data processing unit. The sharpness detection operator in the data processing unit outputs the original phenotypic image to the image separation module. The image separation module, combined with the input correction coefficient, decouples the original phenotypic image and outputs the luminance feature matrix and the chrominance component matrix. The luminance feature matrix is ​​input to the feature decoupling module, and the chrominance component matrix is ​​input to the spatial topology alignment operator. The spatial neighborhood decoupling operator in the data processing unit calculates the second-order difference convection term and inputs it into the feature decoupling module. The feature decoupling module processes the brightness feature matrix in combination with the second-order difference convection term and outputs a continuous gradient tensor field to the manifold reconstruction operator. The manifold reconstruction operator calculates the directional curvature alignment feature value based on the continuous gradient tensor field and inputs it into the spatial topology alignment operator. The spatial topology alignment operator performs multidimensional joint probabilistic topological connectivity operation in combination with the chromaticity component matrix and the directional curvature alignment feature value, and outputs the grid distortion rate and phenotypic structural trait index to the sorting device, respectively. The sorting device contains a closed-loop adjustment loop and a trait sorting module. The closed-loop adjustment loop receives the grid distortion rate and outputs a reduced image sampling step size feedback signal to the spatial topology alignment operator when the distortion exceeds the standard. The trait sorting module receives the phenotypic structural trait index and performs target crape myrtle trait screening.

[0040] like Figure 4 As shown, a preset local sliding window is defined within the continuous gradient tensor field. This preset local sliding window locks the central pixel cluster as the topological node. The data of the topological node is transmitted in parallel to the local feature vector extraction stage. In the local feature vector extraction stage, the maximum magnitude of the first-order gradient vector, the minimum magnitude of the first-order gradient vector, and the arithmetic mean of the gradient vectors are extracted. The data output from the local feature vector extraction stage is input to the feature fusion and balance calculation stage. The maximum magnitude and the minimum magnitude of the first-order gradient vector are input to the vector difference operation to solve for the difference between the maximum and minimum magnitudes. The arithmetic mean of the gradient vectors and the preset boundary stability constant are input to the numerical balance constraint to solve for the sum of the average value and the preset boundary stability constant. The preset boundary stability constant is simultaneously input to the numerical balance constraint. The calculation results output from the vector difference operation and the calculation results output from the numerical balance constraint are input to the ratio calculation stage. After the ratio is solved, the directional curvature alignment feature value is output to characterize the edge geometric evolution law.

[0041] like Figure 5As shown, the graph uses the occlusion area ratio as the horizontal axis with a scale range of 20% to 80%, and the image feature edge misclassification probability as the vertical axis with a scale range of 0% to 30%. The graph shows the comparison of the response curves of the image feature edge misclassification probability changes of the manifold reconstruction sample group of the present invention and the fixed-scale edge mapping comparison sample group under different occlusion area ratios. When the occlusion area ratio increases from 20% to 80%, the image feature edge misclassification probability of the fixed-scale edge mapping comparison sample group increases monotonically and sharply from 4.6% to 28.4%. However, in the entire process of the occlusion area ratio increasing from 20% to 80%, the image feature edge misclassification probability of the manifold reconstruction sample group of the present invention is 0.3% at a 20% occlusion ratio, 0.7% at a 50% occlusion ratio, and stably controlled at 1.2% at an 80% occlusion ratio, significantly reducing the feature misclassification rate under high occlusion conditions.

[0042] Example 4: When the sorting device faces the combined effects of wind load and solar radiation in the field sorting process, the densely clustered plant flowers undergo irregular deformation and leaf swaying, causing local geometric and topological distortions in the non-rigid manifold mesh skeleton constructed in the previous steps. This results in spatial positional shifts in the node topological correspondences between adjacent image frames, causing the extracted high-dimensional phenotypic feature pattern vectors to diverge in tracking values. Consequently, a technical problem of spatial mesh structure instability arises under field sorting conditions. The image processor retrieves the mesh node coordinate sequence stored in the memory and calculates the rate of change of the local mesh area of ​​the current non-rigid manifold mesh skeleton relative to the initial reference state. This determines the mesh distortion rate of the spatial mesh topology layer. The specific calculation formula is as follows: ,in, For mesh distortion rate, The area of ​​the local mesh after deformation. The reference grid area is the initial reference state.

[0043] To determine the distortion threshold that triggers sampling step size adjustment, the sorting device runs a deformation gradient test program in offline testing mode to measure the node feature matching error under different mesh distortion rates. The test data shows that when the mesh distortion rate... When the value is less than or equal to 0.15, the node spatial alignment residual of the local topology alignment operator is stable within 0.5 pixels, while when the mesh distortion rate... When the value exceeds 0.15, the geometric topological nonlinear distortion of the local mesh increases, leading to a mismatch in the conditional probability correction calculation of adjacent nodes. The topological matching error of the nodes shows a nonlinear upward trend, and the system generates a node matching incompatible state. Therefore, 0.15 is determined as the distortion threshold for closed-loop control.

[0044] During real-time sorting, when the image processor measures the spatial local grid distortion rate... When the value reaches 0.18 and exceeds the determined distortion threshold of 0.15, the closed-loop adjustment circuit adjusts the sampling frequency and initiates the image sampling step size reduction action, automatically reducing the current image sampling step size from 1.0 times to 0.5 times the original sampling step size, thereby narrowing the feature search window of the spatial pixel cluster, locally decomposing and compensating for the spatial position deviation caused by deformation, so that the reconstructed non-rigid manifold mesh skeleton converges to a stable state, maintaining the continuous flow direction of spatial topological features under continuous wind interference, keeping the image feature edge misclassification probability in the occlusion and overlap area at 1.2%, controlling the total convergence time of single image feature extraction within 45ms, and maintaining the monotonic correlation coefficient between the high-dimensional phenotypic feature pattern vector and the physical measured value at 0.98, completing the phenotypic image feature extraction of densely aggregated plant traits. In actual operation, although the overall displacement of the plant induced by wind load changes the physical spatial coordinates of the leaf petals, its deformation trajectory between adjacent sequence frames satisfies the geometric continuity constraint in the spatiotemporal dimension.

[0045] To achieve a closed-loop bridge between overall deformation and surface pixel flow direction prediction, the first-order differential geometric flow direction deduction mechanism of this invention treats the local strain field of the non-rigid manifold mesh skeleton as a constraint boundary. By extracting the tangent direction vector field of nodes in the unoccluded region, it performs first-order tensor extrapolation along the mesh topological connectivity path of the non-rigid manifold network towards the overlapping dead zone where contrast annihilation occurs. This internalizes the overall physical deformation law into the prediction of the evolution trend of discrete pixel gradients, realizing a closed-loop constraint between physical deformation and algorithm deduction. The first-order differential geometric flow feedforward deduction is based on the gradual curvature change law of the continuous manifold, where the edge contrast of the overlapping occluded region decreases to below the contrast threshold. At that time, the data processing unit reads the tangent direction deflection angle of the neighboring topological network nodes in the unshaded area, and uses this as the initial recursive flow direction. It then performs first-order numerical extrapolation along this direction with a spatial step of 1 pixel. At each extrapolation step, it calculates the local curvature change rate of the current pixel position. When the local curvature change rate does not exceed the 15-degree abrupt change limit for 3 consecutive pixels, it maintains the current extrapolation flow direction until it smoothly intersects with the nodes in the unshaded area on the other side. The first-order tensor extrapolation trajectory coordinates generated by the feedforward inference are directly written into the boundary index data structure of the memory, driving the subsequent trait sorting module to directly lock the integral accumulation boundary of leaf area and flower cluster density.

[0046] Example 5: When the system faces the deployment of image acquisition units in heterogeneous manufacturing batches, the gain differences of different photoelectric detection chips and lens edge attenuation cause hardware system deviations in the first-order gradient maximum vector amplitude of the original image. Before phenotypic extraction, the system runs a pre-calibration procedure to control the image acquisition unit to acquire a preset uniform diffuse reflection calibration plate digital image. The digital image is input into the affinity matrix circuit to calculate the consistency correction matrix of each spatial pixel channel. By mapping the consistency correction matrix with the pixel grayscale through matrix multiplication, the edge illumination variation caused by manufacturing tolerances is offset, and a unified physical spatial response baseline is established. In actual hardware deployment, the affinity matrix circuit adopts... Constructed using an application-specific integrated circuit (ASIC) or a field-programmable gate array (FPGA), the circuit internally includes row buffer queue units, pixel gain configuration register groups, and a parallel multiply-accumulate interleaving array. During data flow, the circuit reads the original grayscale response values ​​of each spatial channel obtained through diffuse reflection calibration in real time. It then calls the least squares matrix solver through internally hard-coded microcode control logic to calculate the scaling factors of each spatial channel covering the independent position of each pixel in the entire layer within 45 milliseconds. This generates a two-dimensional consistency correction matrix and stores it in static random access memory. The multiply-accumulate array performs point-by-point multiplication operations on the grayscale value of each input pixel of the current frame image with the corresponding correction coefficient in memory.

[0047] After aligning the physical space response baseline, the calibration module detects the magnitude of the maximum first-order gradient vector within the pixel neighborhood. With the magnitude of the first-order gradient minimum vector The initial distribution is then finely adjusted based on the noise covariance, using a preset boundary stability constant. The bias value is calculated and written into a register. This calibration procedure corrects for noise introduced by the heterogeneity of the photosensitive device, ensuring the mesh distortion rate during subsequent feature reconstruction. The judgment result does not deviate from the set threshold of 0.15, and the probability of misclassification of image feature edges in the occluded area is maintained at 1.2%, outputting a stable high-dimensional phenotypic feature pattern vector.

[0048] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.

Claims

1. A method for extracting phenotypic image features for screening traits of Crape Myrtle, characterized in that, Includes the following steps: Step S1: Obtain the original phenotypic image containing overlapping occluded crape myrtle targets, convert the original phenotypic image into luminance channel data and color channel data, and construct the luminance feature matrix and color metric component matrix based on the luminance channel data and color channel data respectively. Step S2: Input the brightness feature matrix into the spatial anisotropic diffusion filter, calculate the diffusion coefficient boundary in the neighborhood of the current pixel to perform gradient smoothing, and output a continuous gradient tensor field. Step S3: In the continuous gradient tensor field, taking the spatial local pixel cluster as the topological node, calculate the directional curvature alignment feature value between adjacent topological nodes based on the maximum first-order gradient vector magnitude, the minimum first-order gradient vector magnitude, the arithmetic mean of the gradient vectors, and the preset boundary stability constant within the preset local sliding window. Step S4: The directional curvature alignment feature value is used as the topological connectivity weight and introduced into the color edge gradient distribution of the color degree component matrix. Multidimensional joint probability topological connectivity determination is performed to construct a non-rigid manifold mesh skeleton that spans the overlapping occlusion region in the original phenotypic image. When the non-rigid manifold mesh skeleton deforms in the overlapping occlusion region, the image sampling step size is reduced, and the crape myrtle phenotypic image features are output.

2. The phenotypic image feature extraction method for screening crape myrtle traits according to claim 1, characterized in that, Step S1 includes the following sub-steps: Step S11, convert the acquired original phenotypic image into luminance channel data and color channel data; Step S12, extract the pixel-level intensity distribution corresponding to the luminance channel data to construct a luminance feature matrix, and extract the spectral wavelength energy distribution corresponding to the color channel data to construct a colorimetric component matrix.

3. The phenotypic image feature extraction method for screening crape myrtle traits according to claim 1, characterized in that, Step S2 includes the following sub-steps: Step S21, calculate the gray-level gradient difference between the current pixel and its neighboring pixels in the brightness feature matrix, and determine the diffusion coefficient boundary based on the gray-level gradient difference; Step S22, control the brightness feature matrix to perform anisotropic diffusion smoothing along the diffusion coefficient boundary, and output a continuous gradient tensor field.

4. The phenotypic image feature extraction method for screening crape myrtle traits according to claim 1, characterized in that, The multidimensional joint probabilistic topological connectivity determination in step S4 includes the following sub-steps: Step S41, establish a joint probabilistic connectivity matrix and convert the directional curvature alignment eigenvalues ​​into the basic connectivity weights in the joint probabilistic connectivity matrix; Step S42: Extract color edge gradient features from the color degree component matrix, and perform conditional probability correction on the basic connection weights based on the color edge gradient features to obtain the final joint connectivity probability value; Step S43: Determine the connectivity of adjacent topological nodes based on the joint connectivity probability value.

5. The phenotypic image feature extraction method for screening crape myrtle traits according to claim 1, characterized in that, The reduction of image sampling step size in step S4 includes the following sub-steps: Step S44, when constructing the non-rigid manifold mesh skeleton, the mesh distortion rate of the non-rigid manifold mesh skeleton in the local space is measured; Step S45, when the mesh distortion rate exceeds the set distortion threshold, it is determined that the non-rigid manifold mesh skeleton has entered the overlapping occlusion area, and the current image sampling step size is reduced to 0.5 of the original sampling step size.

6. The phenotypic image feature extraction method for screening crape myrtle traits according to claim 1, characterized in that, The method further includes the following sub-steps in the pre-step stage of step S1: Step S101, calculate the global hyperspectral energy variance of the input original phenotypic image; Step S102, when the global hyperspectral energy variance is lower than the sharpness benchmark threshold of 0.65, retrieve temporally adjacent phenotypic image frames for similarity geometric reconstruction and alignment repair until the global hyperspectral energy variance of the reconstructed image frame is not lower than 0.65, and provide the reconstructed image frame as the original phenotypic image to step S1.

7. The phenotypic image feature extraction method for screening crape myrtle traits according to claim 1, characterized in that, The method further includes the following sub-steps in the post-stage of step S4: Step S46, when the edge contrast of the overlapping occlusion area decreases to below the contrast threshold, feedforward deduction is performed based on the consistency of the first-order differential geometric flow direction between adjacent topological nodes in the non-rigid manifold mesh skeleton to complete the topological flow direction response on both sides of the missing boundary and restore the occlusion contour of the crape myrtle target in the original phenotypic image.

8. The phenotypic image feature extraction method for screening crape myrtle traits according to claim 1, characterized in that, The method further includes the following sub-steps in the post-stage of step S4: Step S47, extracting phenotypic structural trait indices of the Lagerstroemia indica target from the original phenotypic image based on the closed edge region locked by the non-rigid manifold mesh skeleton. The phenotypic structural trait indices include leaf area, flower cluster density and branching angle; Step S48, screening Lagerstroemia indica traits based on the phenotypic structural trait indices.

9. The phenotypic image feature extraction method for screening crape myrtle traits according to claim 1, characterized in that, Step S2, calculating the diffusion coefficient boundary in the neighborhood of the current pixel, includes the following sub-steps: Step S23, during the filtering process of the anisotropic diffusion filter of the brightness feature matrix, a spatial neighborhood decoupling operator is introduced to calculate the second-order difference convection terms of the current pixel in the horizontal and vertical directions. Step S24: The diffusion coefficient boundary is corrected in real time according to the second-order differential convection term to suppress the accumulation of error in the spatiotemporal evolution of the shadow edge.

Images