A method for automatically obtaining phenotypic parameters of three-dimensional structure of bread slice layers

Abstract

The present invention discloses a method for automatically obtaining phenotypic parameters of the three-dimensional structure of a bread slice layer, comprising the steps of: S1, obtaining a three-dimensional point cloud of a bread slice layer; performing straight-through filtering on the three-dimensional point cloud to obtain a bread slice surface point cloud; S2, performing interpolation segmentation and plane fitting to obtain a data set; S3, constructing a 3D bread segmentation network to perform point cloud segmentation on the data set to obtain a bread plane point cloud; then performing projection processing on the bread plane point cloud to obtain each point cloud boundary; S4, traversing each point cloud boundary to restore the bread slice layer pore point cloud; S5, extracting a first phenotypic parameter and a second phenotypic parameter. The present invention uses a three-axis motion platform and a line laser profile sensor for scanning, which is cost-controllable and can be used for a long time; it also uses a 3D bread segmentation network and architecture restoration to accurately reconstruct the three-dimensional structure of the bread slice layer; it also obtains a large number of phenotypic parameters from the two perspectives of surface and pores, making the evaluation more comprehensive.

Description

Technical Field

[0001] The present invention relates to the technical field of bread physical property measurement, and in particular to a method for automatically acquiring three-dimensional structural phenotype parameters of bread slice layers. Background Art

[0002] The pores on the surface of bread are formed by the gas produced during the fermentation process. The fermentation and baking processes jointly determine the pore morphology of the bread surface. Generally, the amount of pores in bread can greatly affect the taste of bread. Therefore, by analyzing the pore structure on the surface of bread and giving quantitative indicators of pores, the fermentation process of bread can be evaluated, so as to make bread with better taste, soft and hard texture and visual effect.

[0003] At present, the evaluation of bread porosity mainly adopts the two-dimensional image processing method, which collects RGB images of bread slices and uses the two-dimensional image analysis software ImageJ to analyze simple pore parameter indicators such as porosity, pore diameter, and pore area. However, the pore structure is often three-dimensional, and the accuracy of evaluating the pore structure through two-dimensional images is low.

[0004] Therefore, evaluating the porosity through three-dimensional structure has become an urgent problem to be solved. At present, some technicians use X-ray scanners to obtain the three-dimensional structure of bread pores, and use a visualized three-dimensional view to show the bread pores, thereby assisting relevant technicians in evaluation. However, there are still the following three problems when using this method:

[0005] 1) The imaging equipment used for scanning is expensive and difficult for conventional baking companies to afford;

[0006] 2) There is a lack of pore parameter extraction algorithms and application analysis software for the 3D images obtained by scanning, and data processing is difficult and complex;

[0007] 3) When conducting an assessment, people often only focus on the number and volume of pores, but lack attention to many other related parameters, which makes the final assessment less accurate. Summary of the invention

[0008] In view of the above three problems, the purpose of the present invention is to propose a method for automatically acquiring the three-dimensional structural phenotypic parameters of the bread slice layer, which uses a three-axis motion platform and a line laser profile sensor for scanning. The method has controllable costs and can be used for a long time, and can effectively complete the scanning of the bread slice layer; at the same time, combined with the relevant processing of the 3D bread segmentation network and the point cloud boundary, the three-dimensional structure of the bread pores can be accurately and efficiently reconstructed; a large number of phenotypic parameters such as height, width, surface area, volume, maximum pore diameter and maximum pore area are also calculated. From the two perspectives of the surface and pores of the bread slice layer, all data related to the bread slice layer are fully known, which also makes subsequent related evaluations more accurate.

[0009] This is achieved through the following technical solutions:

[0010] A method for automatically obtaining three-dimensional structural phenotype parameters of bread slice layers comprises the following steps:

[0011] S1. Using a line laser profile sensor to obtain a three-dimensional point cloud of a bread slice layer; performing a through-filter on the three-dimensional point cloud to filter out interference objects in the three-dimensional point cloud, and obtaining a point cloud of the bread slice surface;

[0012] S2, performing interpolation segmentation on the surface point cloud of the bread slice in step S1 to obtain a plurality of segmentation blocks; and then performing plane fitting by specifying the same search radius for each of the plurality of segmentation blocks to obtain a data set;

[0013] S3, constructing a 3D bread segmentation network to perform point cloud segmentation on the data set in step S2 to obtain a bread plane point cloud; then projecting the bread plane point cloud to obtain each point cloud boundary where each hole of the bread slice layer is located;

[0014] S4, traversing each point cloud boundary in step S3, and using difference expansion to restore the comprehensive pore structure of the bread slice layer in three dimensions to obtain the pore point cloud of the bread slice layer;

[0015] S5. Extract the first phenotypic parameter from the bread slice surface point cloud of step S1, extract the second phenotypic parameter from the bread slice layer pore point cloud of step S4, and end the operation.

[0016] Combining the three-axis motion platform and the line laser profile sensor can effectively scan the data of the bread slice layer, which is cost-controlled and can be used for a long time. Then, the scanned bread slice surface point cloud is interpolated, segmented, plane fitted, projected and traversed to accurately and efficiently restore the three-dimensional structure of the bread slice layer, and then two types of phenotypic parameters are extracted separately, so that all data related to the slice layer can be fully known.

[0017] Preferably, when the line laser profile sensor is used to obtain the three-dimensional point cloud of the bread slice layer in step S1, based on the three-dimensional Cartesian coordinate system, the three-axis motion platform is first used to control the line laser profile sensor to obtain the position information of the bread slice layer in the X direction, and at the same time, the position information in the Z direction corresponding to the position information in the X direction is determined according to the laser triangulation reflection principle; then the three-axis motion platform is used to control the line laser profile sensor to move at a uniform speed along the Y direction to obtain the position information in the Y direction; based on the position information in the X direction, the position information in the Y direction and the position information in the Z direction, the three-dimensional point cloud of the bread slice layer is obtained using 3D synthesis software. The laser triangulation reflection principle can accurately determine the Z direction position information of each scanning position without the need for the three-axis motion platform to move in the Z direction, which greatly improves the simplicity and speed of scanning; at the same time, the uniform motion in the Y direction ensures the stability and consistency of data acquisition during the scanning process, and reduces the adverse effects such as scanning errors caused by speed changes.

[0018] Preferably, when performing interpolation segmentation in step S2, first set the corresponding maximum and minimum values ​​in the X direction and Y direction of the bread plane point cloud respectively and make a corresponding difference to obtain the difference in the X direction and the difference in the Y direction; then set a common number of interpolation points, and divide the difference in the X direction and the difference in the Y direction by the number of interpolation points to obtain the corresponding interpolation step lengths; set an interpolation point every other interpolation step length between the corresponding maximum value and the corresponding minimum value in the X direction, and set an interpolation point every other interpolation step length between the corresponding maximum value and the corresponding minimum value in the Y direction; then connect each interpolation point with the same Y value in the X direction and the same X value in the Y direction to obtain multiple segmentation blocks. Using interpolation segmentation, the bread plane point cloud can be processed efficiently and accurately, the difference setting of the maximum and minimum values ​​ensures that the data coverage range is uniform and complete, and the precise control of the step length also improves the final accuracy.

[0019] Preferably, in step S2, after obtaining multiple segmentation blocks, a RANSAC plane fitting with a first search radius of 1 mm is performed on each segmentation block to obtain the fitting plane equations of the multiple segmentation blocks; based on the multiple fitting plane equations, each point outside the first search radius of each fitting plane is used as each pore point of the bread slice layer, and then the data set is integrated. Using RANSAC plane fitting and integrating the data set not only improves the accuracy and reliability of pore detection, but also provides high-quality basic data for subsequent data processing.

[0020] Preferably, when constructing a 3D bread segmentation network to perform point cloud segmentation on a data set in step S3, the 3D bread segmentation network calculates the data set according to a 70% training ratio, a 20% test ratio, and a 10% verification ratio, and determines the bread plane point cloud based on the calculation results. The 70% training ratio facilitates improving the accuracy of the calculation, the 20% test ratio provides an effective prediction estimate for the final calculation, and the 10% verification ratio performs the final verification, which can effectively verify the accuracy of the training and prediction processes, thereby ensuring the accuracy of the 3D bread segmentation network during operation.

[0021] Preferably, when the bread plane point cloud is projected in step S3, a two-dimensional projection method is used to orthogonally project the point cloud where the bread slice surface is located along the Z direction, generate a two-dimensional image in the form of one pixel per 1 mm, and map the two-dimensional point cloud to the two-dimensional image; then, the mapped two-dimensional image is subjected to two-dimensional edge extraction based on an edge detection operator to obtain two-dimensional edge points; then, according to the inverse transformation of the orthogonal projection, the two-dimensional edge points are restored to a three-dimensional point cloud to obtain the three-dimensional coordinates of each pore contour, and each point cloud boundary where each pore of the corresponding bread slice layer is located is determined based on the three-dimensional coordinates of each pore contour. The projection processing method combined with two-dimensional edge detection and inverse transformation not only simplifies data processing and improves computational efficiency, but also enhances the recognition accuracy and reliability of the pore structure.

[0022] Preferably, when the three-dimensional coordinates of each pore contour are obtained, the centroid of each pore contour is obtained, and a second search radius of 5 mm is performed on the bread slice surface point cloud with each centroid as the center to obtain the point cloud where each pore is located and each point cloud boundary where each pore of the bread slice layer is located. By calculating the centroid and implementing the radius search method, the accuracy and efficiency of pore feature recognition can be improved.

[0023] Preferably, when traversing each point cloud boundary in step S4, first use the nearest neighbor-based patching algorithm to patch the edge of each pore point, traverse each element of each pore point on the first search radius and within the first search radius, take each element point of each pore point on the first search radius as the set where the bread slice layer is located, and take each element point of each pore point within the first search radius as the set where the corresponding pore point is located; connect each element point in the set where the bread slice layer is located to each nearest element point in the set where the corresponding pore point is located within the first search radius, and after completing the pairwise corresponding lines, set an interpolation step of 1mm, perform interpolation expansion in each pairwise corresponding line, and obtain the restored three-dimensional pore point cloud as the pore point cloud of the bread slice layer. The use of the nearest neighbor-based patching algorithm and interpolation expansion improves the effect of pore edge patching and data integrity, and accurately obtains the pore point cloud of the bread slice layer.

[0024] Preferably, the first phenotypic parameters of the bread slice surface point cloud include height, width, length, surface area, volume, symmetry index, and uniformity index; the second phenotypic parameters of the bread slice layer pore point cloud include maximum pore diameter, maximum pore area, pore number, maximum pore depth, maximum pore volume, pore growth, pore distribution, and pore layer distribution. Obtaining a large number of phenotypic parameters from the two perspectives of bread slice surface and pores can effectively evaluate the structure of bread from multiple perspectives and reduce the occurrence of evaluation bias and incomplete evaluation.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] The technical solution of the present invention utilizes a three-axis motion platform and a line laser profile sensor for scanning, which is cost-controllable and can be used for a long time, and can effectively complete the scanning of the bread slice layer; at the same time, combined with the relevant processing of the 3D bread segmentation network and the point cloud boundary, the three-dimensional structure of the bread pores can be accurately and efficiently reconstructed; a large number of phenotypic parameters such as height, width, surface area, volume, maximum pore diameter and maximum pore area are also calculated, and from the two perspectives of the surface and pores of the bread slice layer, various data related to the bread slice layer are fully known, which also makes subsequent related evaluations more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A flowchart of a method for automatically obtaining phenotypic parameters of three-dimensional structure of bread slice layers;

[0028] Figure 2 It is a schematic diagram of plane fitting;

[0029] Figure 3 A schematic diagram of the structure of a 3D bread segmentation network;

[0030] Figure 4 A schematic diagram of extracting a pore 3D contour by transforming a 2D edge into a 3D pore edge;

[0031] Figure 5 A schematic diagram of restoration of a pore three-dimensional structure;

[0032] Figure 6 A schematic diagram is shown for one phenotypic parameter. DETAILED DESCRIPTION

[0033] The following will be combined with the attached embodiment of the present invention Figure 1 To Attachment Figure 6 , the technical solutions in the embodiments of the present invention are described in detail.

[0034] like Figure 1The figure shows a flowchart of a method for automatically obtaining the three-dimensional structural phenotypic parameters of a bread slice layer. The method uses a three-axis motion platform and a line laser profile sensor to scan, and then performs relevant processing on the scanned image by means of a 3D bread segmentation network and traversing the point cloud boundary, so as to effectively extract a large number of phenotypic parameters related to the surface and pores of the bread slice layer, thereby being able to accurately and comprehensively evaluate the bread slice layer. The method specifically includes the following steps:

[0035] S1. Use a line laser profile sensor to obtain a three-dimensional point cloud of a bread slice layer; perform through filtering on the three-dimensional point cloud to filter out interference objects in the three-dimensional point cloud, and obtain a point cloud of a bread slice surface.

[0036] In this embodiment, when the line laser profile sensor is used to obtain the three-dimensional point cloud of the bread slice layer in step S1, based on the three-dimensional Cartesian coordinate system, the three-axis motion platform is first used to control the line laser profile sensor to obtain the position information of the bread slice layer in the X direction, and at the same time, the position information in the Z direction corresponding to the position information in the X direction is determined according to the laser triangulation reflection principle; then the three-axis motion platform is used to control the line laser profile sensor to move at a uniform speed along the Y direction to obtain the position information in the Y direction; based on the position information in the X direction, the position information in the Y direction and the position information in the Z direction, the three-dimensional point cloud of the bread slice layer is obtained using 3D synthesis software. The laser triangulation reflection principle can accurately determine the Z direction position information of each scanning position without the need for the three-axis motion platform to move in the Z direction, which greatly improves the simplicity and speed of scanning; at the same time, the uniform motion in the Y direction ensures the stability and consistency of data acquisition during the scanning process, and reduces the adverse effects such as scanning errors caused by speed changes.

[0037] Specifically, the model of the line laser profile sensor can be UC3D500-165x140-B / R. This line laser profile sensor supports the acquisition of profiles, depth maps and three-dimensional point clouds, and can emit laser beams at a fixed frequency. The laser beam will be amplified into a laser line and projected onto the surface of the object to be measured. At the same time, there is a filter in front of the lens of the line laser profile sensor, which only allows light in the laser band to enter the sensor. In this way, the line laser is irradiated on the object, and the received light will form a line. The line received by the X-direction sensor is defined as a profile in the X direction of the bread slice layer surface. Because the sensor only receives light in the laser band, the profile is reflected in the image as a discrete non-zero pixel sequence. The pixel values ​​of other objects that are not illuminated by the laser are all zero. Then, the imaging matrix formed by the received light is processed using the laser triangulation reflection principle to obtain the distance from the sensor to the bread surface, that is, the position information in the Z direction, and the position information in the X direction along the laser line.

[0038] In this embodiment, the filtering method in step S1 is: perform through filtering in the Z direction to filter out interference objects in the three-dimensional point cloud. Since the bread slices are placed on the stage, the height of the stage is fixed, so through filtering can be performed in the Z direction, and the distance from the line laser profile sensor to the stage is used as the filtering threshold, which can effectively filter out the relevant interference in the background of the bread during scanning and improve data accuracy.

[0039] S2, performing interpolation segmentation on the surface point cloud of the bread slice in step S1 to obtain a plurality of segmentation blocks; and then specifying the same search radius for each of the plurality of segmentation blocks to perform plane fitting to obtain a data set.

[0040] In this embodiment, when interpolation segmentation is performed in step S2, the corresponding maximum and minimum values ​​are first set in the X direction and Y direction of the bread plane point cloud respectively and the difference is made accordingly to obtain the difference in the X direction and the difference in the Y direction; then a common number of interpolation points is set, and the difference in the X direction and the difference in the Y direction are divided by the number of interpolation points to obtain the corresponding interpolation step lengths; an interpolation point is set every interpolation step length between the corresponding maximum value and the corresponding minimum value in the X direction, and an interpolation point is also set every interpolation step length between the corresponding maximum value and the corresponding minimum value in the Y direction; then each interpolation point with the same Y value in the X direction and the same X value in the Y direction is connected accordingly to obtain multiple segmentation blocks. By using interpolation segmentation, the bread plane point cloud can be processed efficiently and accurately. The difference setting of the maximum and minimum values ​​ensures that the data coverage range is uniform and complete, and the precise control of the step length also improves the final accuracy.

[0041] like Figure 2 As shown, it is a schematic diagram of plane fitting. After obtaining multiple segments in step S2, a RANSAC plane fitting with a first search radius of 1mm is performed on each segment to obtain the fitting plane equations of multiple segments; based on multiple fitting plane equations, the points outside the first search radius of each fitting plane can be used as each pore point of the bread slice layer, and then the data set is integrated. Assuming that the number of interpolation points is 10, a total of 40 interpolation points are added in the maximum and minimum directions of the X-axis and Y-axis; by connecting the interpolation points with the same Y-axis value in the X direction and the same X-axis value in the Y direction, the entire bread point cloud can be divided into 100 segments, and then a RANSAC plane fitting with a search radius of 1mm is performed on each segment to obtain the fitting plane equations of 100 segments; then, the points outside the search radius of each fitting plane can be used as the points where the bread pores are located, and the points within the fitting plane can be used as the points where the bread surface is located. Using RANSAC plane fitting and integrating the data set not only improves the accuracy and reliability of pore detection, but also provides high-quality basic data for subsequent data processing.

[0042] Here, RANSAC stands for Random Sample Consensus, which is an iterative method of random sample consistency, used to estimate the parameters of a mathematical model from a data set containing multiple points, and is suitable for processing data sets that may contain outliers or erroneous values.

[0043] S3. Construct a 3D bread segmentation network to perform point cloud segmentation on the data set in step S2 to obtain a bread plane point cloud; then project the bread plane point cloud to obtain each point cloud boundary where each hole in the bread slice layer is located.

[0044] In this embodiment, when constructing a 3D bread segmentation network in step S3 to perform point cloud segmentation on a data set, the 3D bread segmentation network calculates the data set according to a 70% training ratio, a 20% test ratio, and a 10% verification ratio, and determines the bread plane point cloud based on the calculation results. The 70% training ratio facilitates improving the accuracy of the calculation, the 20% test ratio provides an effective prediction estimate for the final calculation, and the 10% verification ratio performs the final verification, which can effectively verify the accuracy of the training and prediction processes, thereby ensuring the accuracy of the 3D bread segmentation network during operation.

[0045] like Figure 4 As shown, it is a schematic diagram of pore three-dimensional contour extraction from two-dimensional edge transformation to three-dimensional pore edge, showing the effect of projecting from a two-dimensional image to a three-dimensional point cloud. Specifically, when the bread plane point cloud is projected in step S3, a two-dimensional projection method is used to orthogonally project the point cloud where the bread slice surface is located along the Z direction, and a two-dimensional image is generated in the form of one pixel per 1mm, and the two-dimensional point cloud is mapped to the two-dimensional image. Orthogonal projection is a basic geometric transformation method that has been made public, and will not be described in detail here. Then, the mapped two-dimensional image is subjected to two-dimensional edge extraction based on the edge detection operator to obtain two-dimensional edge points; then, according to the inverse transformation of the orthogonal projection, the two-dimensional edge points are restored to a three-dimensional point cloud to obtain the three-dimensional coordinates of each pore contour, and the boundary of each point cloud where each pore of the corresponding bread slice layer is located is determined according to the three-dimensional coordinates of each pore contour. The projection processing method combined with two-dimensional edge detection and inverse transformation not only simplifies data processing and improves calculation efficiency, but also enhances the recognition accuracy and reliability of pore structure.

[0046] In this embodiment, when the three-dimensional coordinates of each pore outline are obtained, the centroid of each pore outline is obtained, and a radius search with a second search radius of 5 mm is performed on the bread slice surface point cloud with each centroid as the center to obtain the point cloud where each pore is located and each point cloud boundary where each pore of the bread slice layer is located. By calculating the centroid and implementing the radius search method, the accuracy and efficiency of pore feature recognition can be improved.

[0047] like Figure 3 The figure shows a schematic diagram of the structure of a 3D bread segmentation network. Taking the processing of 100 sets of bread slice layer point cloud data in the data set as an example, 70 sets of point clouds are used for training, 20 sets of point clouds are used for testing, and 10 sets of point clouds are used for verification. During training, each point contains ten dimensions of information, including three-dimensional position information (x, y, z), color information (r, g, b), local normal vector information (N x , N y , N z ) and classification information, where 0 indicates the point where the bread slice layer plane is located, and 1 indicates the point where the pore is located. During training, the learning rate is set to 0.001, the learning rate decay value is 0.5, and it decreases every 20 iterations. The Adam optimizer is used for adjustment, and the batch size is set to 4; Adam stands for Adaptive MomentEstimation, and the Adam optimizer is a gradient descent optimization algorithm that makes training fast and smooth. The 3D bread segmentation network will take the dataset containing the bread point cloud to be segmented as input, and enter the spatial attention module, three abstract modules, and four upsampling modules in turn, and then undergo two convolution transformations, and finally output the binary segmentation result, which corresponds to the above-mentioned classification information of 0 and 1. Among them, the spatial attention module has a total of four sub-modules, which use three convolutions to form a multi-layer perceptron in the form of [32,32,64] to accurately input and output, specifically: a convolution whose output is a 32-dimensional tensor; a channel collaboration module with a dimension of 32 and 4 heads, which is also a shareable multi-semantic spatial attention module; a convolution whose output is a 32-dimensional tensor; and a convolution whose output is a 64-dimensional tensor.

[0048] In the sharable multi-semantic spatial attention module, first, the input is decomposed into the form of B×C×H×W along the height and width, and then two single sequences are created along the height and width to split the two sequences into 4 independent feature subsequences respectively. Subsequently, the four independent feature subsequences are convolved at four scales and aligned using lightweight shared convolutions to implicitly model the dependency between the two dimensions by learning consistent features in the two dimensions. Finally, the normalization module and activation function are used to generate spatial attention features. Then, a pooling operation is used to scale the resolution of length and width from (H×W) to , then, after the average pooling layer and activation function, the output is a tensor result of B×C×H×W, which is also the input of the subsequent abstract module.

[0049] The abstract module contains two submodules: the sampling grouping submodule and the PointNet submodule. The sampling grouping module selects sampling points by farthest point sampling. PointNet is a deep learning architecture for processing unordered point cloud data. In the PointNet submodule, the coordinates of the points in the local area are converted to the local coordinate system relative to the centroid of each spherical area, and the feature point set is extracted through the respective built-in multilayer perceptrons. Set the abstract modules 1, 2 and 3, the number of search points in the three modules are 256, 64 and 16 respectively; the search radius is 0.2, 0.4 and 0.8 respectively, and the corresponding built-in multilayer perceptrons are [64, 64, 128], [128, 128, 256], [256, 256, 521] respectively.

[0050] Upsampling modules 1, 2, 3, and 4 are all responsible for upsampling. They interpolate and upsample the features in the feature point set group to ensure feature consistency between point cloud data of different resolutions. In the four upsampling modules, the output of the shareable multi-semantic spatial attention module and the output of the abstract module 3 are interpolated and fused, and a multilayer perceptron of [256, 256] is used to further extract features. In upsampling module 3, the output of upsampling module 4 and the output of abstract module 2 are interpolated and fused, and a multilayer perceptron of [256, 256] is used to further extract features. In upsampling module 2, the output of upsampling module 3 and the output of abstract module 1 are interpolated and fused, and a multilayer perceptron of [256, 128] is used to further extract features. In upsampling module 1, the output of upsampling module 2 is interpolated and fused with itself, and a multilayer perceptron of [128, 128, 128] is used to further extract features and transmitted to the 128-bit input channel. Then, for a large number of further extracted features, it is necessary to perform a convolution operation + Bn batch normalization processing + Relu corrected linear unit + Dp regularization processing in sequence to correct the large number of features input by the input channel and then output them through the 128-bit output channel; finally, after another convolution operation with an input channel of 128 and an output channel of 2, the binary segmentation results are output to form a bread plane point cloud.

[0051] S4, traversing each point cloud boundary in step S3, and using difference expansion to restore the three-dimensional structure of the comprehensive pore structure of the bread slice layer to obtain the pore point cloud of the bread slice layer.

[0052] like Figure 5 As shown, it is a schematic diagram of the restoration of the three-dimensional structure of the pore, where X axis represents the X axis, Y axis represents the Y axis, and Zaxis represents the Z axis; Figure 5 The left side in the middle shows the structure of the pore point cloud before restoration. Figure 5The right side of the figure shows the structure of the restored bread slice layer pore point cloud; specifically, when traversing each point cloud boundary in step S4, the nearest neighbor-based patching algorithm is first used to patch the edge of each pore point, and each element of each pore point on and within the first search radius is traversed, and each element point of each pore point on the first search radius is used as the set where the bread slice layer is located, and each element point of each pore point within the first search radius is used as the set where the corresponding pore point is located; each element point in the set where the bread slice layer is located is connected to each nearest element point in the set where the corresponding pore point is located within the first search radius, and after completing the pairwise corresponding lines, an interpolation step of 1mm is set, and interpolation expansion is performed in each pairwise corresponding line to obtain the restored three-dimensional pore point cloud as the bread slice layer pore point cloud. The nearest neighbor-based patching algorithm and interpolation expansion are used to improve the effect of pore edge patching and data integrity, and accurately obtain the bread slice layer pore point cloud.

[0053] S5, extracting the first phenotypic parameter, namely the overall phenotypic parameter, from the bread slice surface point cloud of step S1; then extracting the second phenotypic parameter, namely the pore phenotypic parameter, from the bread slice layer pore point cloud of step S4, and then ending the operation.

[0054] In this embodiment, the first phenotypic parameters of the bread slice surface point cloud include height, width, length, surface area, volume, symmetry index, and uniformity index; the second phenotypic parameters of the bread slice layer pore point cloud include maximum pore diameter, maximum pore area, number of pores, maximum pore depth, maximum pore volume, pore growth, pore distribution, and pore layer distribution. Obtaining a large number of phenotypic parameters from the two perspectives of bread slice surface and pores can effectively evaluate the structure of bread from multiple perspectives and reduce the occurrence of evaluation bias and incomplete evaluation. The specific phenotypic parameters are introduced as follows:

[0055] Height: Take the bread plane point cloud segmented by the 3D bread segmentation network and calculate the average depth value of all plane points. Because the stage is a plane with a fixed height, the depth value of the stage minus the average depth value of the plane points can get an estimated value of the height.

[0056] Length: Take the bread plane point cloud segmented by the 3D bread segmentation network, calculate the difference between the maximum and minimum values ​​of all points in the Y direction, and use each difference as the estimated value of each length of the corresponding point.

[0057] Width: Take the bread plane point cloud segmented by the 3D bread segmentation network, and calculate the difference between the maximum and minimum values ​​of all points in the bread plane point cloud in the X direction as the estimated value of the corresponding length.

[0058] Surface area: Take the point cloud to be projected onto the two-dimensional image in step S3, extract the two-dimensional maximum boundary, i.e., the boundary where the outermost layer is located, and then use the polygon area algorithm to calculate the area of ​​the polygon enclosed by the outermost boundary; then, multiply the calculated polygon area by the scaling factor of the three-dimensional point cloud projected onto the two-dimensional image to obtain an estimated value of the surface area.

[0059] Volume: When the surface of the bread slice layer is parallel to the horizontal plane, the product of the corresponding estimates of height and volume is the estimate of volume.

[0060] Symmetry index: Take the bread plane point cloud segmented by the 3D bread segmentation network, calculate the difference between the maximum and minimum values ​​of all points in the bread plane point cloud in the Y direction, divide the difference by four as the interpolation step, and add three interpolation points between the maximum and minimum values ​​in the Y direction according to the interpolation step, and calculate the average height H of all points in the bread plane point cloud in the Y direction within the threshold range of 2mm from these three interpolation points. B , H C and H D , and using the formula 2H C -H B -H D Get an estimate of the symmetry index.

[0061] Uniformity index: Take H in the process of calculating the symmetry index B and H D , using formula H B -H D Obtain an estimate of the uniformity index.

[0062] Maximum pore diameter: Take the point cloud to be projected onto the two-dimensional image in step S3, extract the second largest boundary in the two dimensions, which corresponds to the boundary where the largest pore is located, and use the two-dimensional to three-dimensional transformation relationship to restore the two-dimensional point cloud of the largest pore into a three-dimensional point cloud, and process it with a minimum bounding box, and use the major axis value of the minimum bounding box as the estimated value of the maximum pore diameter.

[0063] Maximum pore area: Take the 3D point cloud where the largest pore is located, mesh the 3D point cloud using triangulation, and calculate the area sum of all triangular facets in the meshed 3D point cloud as the estimated value of the maximum pore area.

[0064] Number of Pores: Use the number of centroids as an estimate of the number of pores.

[0065] Maximum pore depth: Take the maximum void in the pores of the restored bread slice layer and the local plane position of the maximum pore, calculate the Euclidean distance from the lowest point of the maximum pore to the local plane, that is, the Euclidean distance from the maximum value of the depth direction of all points in the pore three-dimensional point cloud to the local plane, as the estimated value of the maximum pore depth.

[0066] Maximum pore volume: Take the pore point cloud of the bread slice layer after the nearest neighbor repair algorithm is repaired, and use triangulation to divide the point cloud into multiple non-intersecting tetrahedrons. The sum of the volumes of these tetrahedrons is calculated as the estimated value of the maximum pore volume.

[0067] Pore ​​growth E: Pore growth E is the main component of the pores. The maximum shape characteristic value Smax of the pores is divided by the calculated value of the minimum shape characteristic value Smin to estimate the pore growth E, that is, pore growth E = Smax / Smin.

[0068] Pore ​​distribution: Use the same interpolation method as step S2 to divide the bread plane point cloud into 100 regions, label the upper left region as 1, and label the 100 regions in sequence from left to right and from top to bottom diagonally. Then, the number of centroids corresponding to the pores in each region can be counted as the pore distribution.

[0069] Pore ​​layer distribution: The area of ​​the bread slice layer is stratified. Taking an n×n area as an example, the area of ​​the bread slice layer is divided into n / 2 layers, and the number of pores in each layer is counted as the pore layer distribution.

[0070] In summary, the present invention utilizes a three-axis motion platform and a line laser profile sensor for scanning, which is cost-controllable and can be used for a long time, and can effectively complete the scanning of the bread slice layer; at the same time, combined with the relevant processing of the 3D bread segmentation network and the point cloud boundary, the three-dimensional structure of the bread pores can be accurately and efficiently reconstructed; a large number of phenotypic parameters such as height, width, surface area, volume, maximum pore diameter and maximum pore area are also calculated, and various data related to the bread slice layer are known from all aspects from the two perspectives of the surface and pores of the bread slice layer, which also makes subsequent related evaluations more accurate and has significant progress.

[0071] The above embodiments are only for illustrating the technical idea of ​​the present invention, and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for automatically obtaining three-dimensional structural phenotype parameters of bread slice layers, characterized in that: The steps include: S1. Use a line laser profile sensor to obtain a three-dimensional point cloud of a bread slice layer. Based on a three-dimensional Cartesian coordinate system, first use a three-axis motion platform to control the line laser profile sensor to obtain the position information of the bread slice layer in the X direction. At the same time, according to the laser triangulation reflection principle, determine the position information in the Z direction corresponding to the position information in the X direction; then use the three-axis motion platform to control the line laser profile sensor to move at a uniform speed along the Y direction to obtain the position information in the Y direction; based on the position information in the X direction, the position information in the Y direction and the position information in the Z direction, use 3D synthesis software to obtain a three-dimensional point cloud of the bread slice layer; Perform straight-through filtering on the 3D point cloud to filter out interference objects in the 3D point cloud and obtain the surface point cloud of the bread slice; S2, performing interpolation segmentation on the surface point cloud of the bread slice in step S1 to obtain a plurality of segmentation blocks; and then performing plane fitting by specifying the same search radius for each of the plurality of segmentation blocks to obtain a data set; S3, construct a 3D bread segmentation network to perform point cloud segmentation on the data set in step S2 to obtain a bread plane point cloud; then project the bread plane point cloud, use a two-dimensional projection method to orthogonally project the point cloud where the bread slice surface is located along the Z direction, generate a two-dimensional image in the form of one pixel per 1 mm, and map the two-dimensional point cloud to the two-dimensional image; then perform two-dimensional edge extraction based on the edge detection operator on the mapped two-dimensional image to obtain two-dimensional edge points; then, according to the inverse transformation of the orthogonal projection, restore the two-dimensional edge points to a three-dimensional point cloud, obtain the three-dimensional coordinates of each pore contour, and determine each point cloud boundary where each pore of the corresponding bread slice layer is located based on the three-dimensional coordinates of each pore contour; S4, traversing each point cloud boundary in step S3, and using difference expansion to restore the comprehensive pore structure of the bread slice layer in three dimensions to obtain the pore point cloud of the bread slice layer; S5. Extract the first phenotypic parameter from the bread slice surface point cloud of step S1, extract the second phenotypic parameter from the bread slice layer pore point cloud of step S4, and end the operation.

2. The method for automatically obtaining the three-dimensional structural phenotype parameters of bread slice layers according to claim 1, characterized in that: When performing interpolation segmentation, first set the corresponding maximum and minimum values ​​in the X and Y directions of the bread plane point cloud and make corresponding differences to obtain the difference in the X direction and the difference in the Y direction; then set a common number of interpolation points, divide the difference in the X direction and the difference in the Y direction by the number of interpolation points, and obtain the corresponding interpolation step lengths; set an interpolation point every other interpolation step length between the corresponding maximum value and the corresponding minimum value in the X direction, and set an interpolation point every other interpolation step length between the corresponding maximum value and the corresponding minimum value in the Y direction; then connect each interpolation point with the same Y value in the X direction and the same X value in the Y direction to obtain multiple segmentation blocks.

3. The method for automatically obtaining the three-dimensional structural phenotype parameters of bread slice layers according to claim 1, characterized in that: In step S2, after obtaining multiple segments, a RANSAC plane fitting with a first search radius of 1 mm is performed on each segment to obtain the fitting plane equations of the multiple segments; based on the multiple fitting plane equations, the points outside the first search radius of each fitting plane are taken as each pore point of the bread slice layer, and then the data set is integrated.

4. The method for automatically obtaining the three-dimensional structural phenotype parameters of bread slice layers according to claim 1, characterized in that: When constructing a 3D bread segmentation network to perform point cloud segmentation on the data set in step S3, the 3D bread segmentation network calculates the data set according to a training ratio of 70%, a test ratio of 20%, and a verification ratio of 10%, and determines the bread plane point cloud based on the calculation results.

5. The method for automatically obtaining the three-dimensional structural phenotype parameters of bread slice layers according to claim 1, characterized in that: When the three-dimensional coordinates of each pore outline are obtained, the centroid of each pore outline is obtained, and a second search radius of 5 mm is performed on the bread slice surface point cloud with each centroid as the center to obtain the point cloud where each pore is located and each point cloud boundary where each pore of the bread slice layer is located.

6. The method for automatically obtaining the three-dimensional structural phenotype parameters of bread slice layers according to claim 3, characterized in that: When traversing the boundary of each point cloud, first use the nearest neighbor-based patching algorithm to patch the edge of each pore point, traverse each element of each pore point on and within the first search radius, and take each element point of each pore point on the first search radius as the set where the bread slice layer is located, and take each element point of each pore point within the first search radius as the set where the corresponding pore point is located; connect each element point in the set where the bread slice layer is located to each nearest element point in the set where the corresponding pore point is located within the first search radius, and after completing the pairwise corresponding lines, set an interpolation step of 1mm, and perform interpolation and expansion in each pairwise corresponding line to obtain the restored three-dimensional pore point cloud and use it as the pore point cloud of the bread slice layer.

7. The method for automatically obtaining the three-dimensional structural phenotype parameters of bread slice layers according to claim 1, characterized in that: The first phenotypic parameters of the bread slice surface point cloud include height, width, length, surface area, volume, symmetry index, and uniformity index; the second phenotypic parameters of the bread slice layer pore point cloud include maximum pore diameter, maximum pore area, number of pores, maximum pore depth, maximum pore volume, pore growth, pore distribution, and pore layer distribution.

Images