An adaptive 3D printing insole intelligent generation method based on monocular projection semantic recognition

Abstract

The application discloses a kind of self-adapting 3D printing insole intelligent generation method based on monocular projection semantic recognition, belongs to personalized footwear intelligent design and additive manufacturing field;It collects insole overhead image, axial reference length and user weight, obtains high-precision contour by pretreatment such as multi-scale Retinex light balance, SLIC superpixel segmentation;By aspect ratio constraint, second-order central moment solution is completed posture correction and regional division;Adopt ICP+TPS two-stage registration to realize two-dimensional-three-dimensional model adaptation;Rely on Gaussian mixture clustering prior model to combine weight correction pressure field, drive adaptive concave-convex surface and partition variable stiffness lattice design, output STL file after Laplace smoothing for 3D printing.The application does not need professional equipment, full-process automation, finished product size deviation ≤±0.2mm, fit degree ≥99%, solve traditional customization high cost, poor adaptability and other problems, suitable for daily wear, sports protection type insole mass customization.

Description

Technical Field

[0001] This invention belongs to the field of personalized footwear intelligent design and additive manufacturing technology, and particularly relates to an adaptive 3D printing insole intelligent generation method based on monocular projection semantic recognition. Background Technology

[0002] With the continuous development of additive manufacturing technology and ergonomic design theory, personalized custom insoles, with their excellent foot fit, shock absorption and cushioning capabilities, and gait correction effects, are widely used in daily wear, sports protection, and foot health protection. Currently, the mainstream methods for manufacturing custom insoles in the industry are mainly divided into two categories.

[0003] The first category is the high-precision professional customization model. This model requires the use of professional plantar pressure acquisition equipment and 3D scanning instruments to obtain complete plantar biomechanical data and foot contour data from the user, and then completes the model design through mechanical deformation algorithms. This method has high molding accuracy, but the equipment purchase and use costs are high, and the data acquisition process is complex. It can only be applied to the fields of medical rehabilitation and professional sports, and cannot be widely promoted to the general consumer market. Existing solutions under this technical approach all have significant shortcomings: Pure geometric modeling solutions (CN110001056A, CN105711091B, CN106942837B) rely solely on 3D scanning and image recognition to collect foot contours, discarding plantar pressure data. They rely on group average data and human experience to complete structural design, failing to reproduce the real dynamic force experienced by the human body during walking. Orthopedic and support designs lack quantitative mechanical basis and can only achieve basic shape fit. Pressure and image fusion solutions (CN107114861B) combine plantar pressure and image data for modeling, but have not established a standardized registration algorithm. They use traditional rigid coordinate transformation methods, which easily cause misalignment between the 2D pressure field and the 3D model, surface distortion, and insufficient accuracy of multimodal data fusion. Lattice structure solutions (CN110370647A) have vague data acquisition standards. Lattice structure optimization relies solely on qualitative judgment of force morphology, lacking quantitative driving force for individual plantar pressure, resulting in poor targeted structural design.

[0004] The second category is the mass production model of generic insoles. Factories process insoles according to fixed sizes and specifications, resulting in a completely fixed shape and internal structure. This makes them unsuitable for different shoe styles and fails to optimize the mechanical structure based on the weight and foot pressure characteristics of different groups of people. Common problems include poor fit, unreasonable soft-hard zoning, and insufficient wearing comfort. Traditional standardized footwear relies on linear scaling of fixed national standard shoe sizes, which cannot match the actual three-dimensional physiological parameters of the user's foot length, width, and instep height. This easily leads to problems such as squeezing, looseness, and inconsistent fit. Currently, the foot-supporting surface of mass-produced insoles uses industry experience or average foot data to construct a fixed concave surface. The position, depth, and curvature of the concave surface cannot be adjusted according to individual differences. Due to significant differences in the foot bone contours, pressure concentration areas, and foot convexity / concave shape among different groups of people, the fixed surface easily causes localized stress concentration. Long-term wear can lead to problems such as foot chafing, abrasion, foot pain, unsupported arches, and gait deviation, failing to meet the requirements for optimal biomechanical fit.

[0005] Current visual-assisted insole customization technology still faces several common bottlenecks in the industry: First, home photography is prone to problems such as uneven lighting distribution, shadow reflections, and background noise artifacts. Meanwhile, long-term use of insoles can lead to edge wear, incomplete contours, and burrs, causing traditional edge extraction algorithms to easily identify interfering information as valid contours, resulting in oversized contours, distorted boundaries, and reduced assembly accuracy. Second, insoles can be placed at any angle from 0 to 360°, and current technology cannot automatically determine orientation and distinguish between the forefoot and heel areas without feature markings; it must rely on... First, manual correction results in low automation. Second, conventional 3D registration technology relies on multi-view images to complete stereo reconstruction, which involves a large amount of data collection and high computational complexity. However, commercially available insoles have a symmetrical structure with no significant difference in projection features between the front and back sides, making traditional 3D reconstruction solutions resource-intensive and impractical. Third, most visual solutions rely on public foot data to build prior models, only combining weight to correct force characteristics. This is a group experience fitting mode, which cannot match the unique foot force and walking gait of an individual. Therefore, the ability to achieve high-precision personalized adaptation and gait correction is significantly lacking.

[0006] From the perspective of orthopedic adaptation, most existing customized solutions can only achieve static fit between the insole and the sole of the foot. For foot deformities such as flat feet, pronation, and supination, they can only adjust the shape passively through material partitioning and local thickening, and cannot combine gait design with dynamic mechanical guidance structure. At the same time, it is difficult to flexibly adjust the support stiffness according to the user's weight and sports scenario, and the scenario adaptability is insufficient.

[0007] In terms of mechanical structure design, the current mainstream lattice and dot matrix insole technologies lack a complete theoretical system. Most solutions adopt a uniform lattice structure with a fixed rod diameter, and variable lattice designs are only roughly adjusted based on a single parameter of body weight. They do not establish a quantitative mathematical model that combines the elastic-plastic mechanics of materials and the compression deformation law of human walking. The lattice support strength has a low degree of matching with the actual stress conditions, and it is impossible to achieve both high-strength support and flexible cushioning in the same product. Some multi-material composite structure insoles have complex manufacturing processes, making it impossible to achieve 3D printing integrated molding. The interfaces between different materials are prone to delamination and cracking, resulting in poor product durability.

[0008] Furthermore, the industry currently lacks a fully closed-loop, fully automated customization system. Data acquisition, parameter calculation, model deformation, mechanical design, and finished product processing are fragmented, requiring extensive manual calculations, adjustments, and verifications. This results in long customization cycles, high production costs, low standardization, and difficulty in large-scale production. Existing publicly available technical approaches are relatively singular, focusing only on foot pressure-driven processes, pure geometric modeling, or visual image analysis. There is currently no integrated customization solution that combines macroscopic contour adaptive deformation, microscopic pressure-driven concave deformation of the foot bearing surface, and nonlinear mechanical variable stiffness lattice structure, leaving significant gaps in related core technologies. Summary of the Invention

[0009] To address the problems in the prior art, the present invention proposes the following technical solution: An adaptive intelligent generation method for 3D printed insoles based on monocular projection semantic recognition, including... Step 1: Multi-source basic information collection Establish a unified information input port, collect three types of customized basic parameters, and construct an input parameter set:

[0010] In the formula: The original two-dimensional image of the insole taken in the user's home environment, where L is the axial reference length of the insole manually measured by the user, and W is the user's body weight parameter; Commercially available insoles adopt a symmetrical integrated structure with no distinguishing features between the front and back projection outlines. There is no need to determine the left or right category or identify the front and back. The entire shape design can be completed by collecting the top-view projection information.

[0011] Step 2: Global Preprocessing Optimization of Complex Scene Images To address multiple image defects caused by uneven lighting in household natural light, such as shadow reflections, background artifacts, edge wear and burrs, and incomplete contours, a multi-level joint processing algorithm framework is built to complete image standardization correction. 2.1 Multi-scale Retinex Illumination Equalization Processing Constructing a mathematical model for natural illumination imaging:

[0012] The original imaging image, These are pixel coordinates on the image. The inherent reflective characteristic component of the insole itself, This is the ambient light interference component; Global illumination components are estimated using a multi-scale Gaussian function:

[0013] Two-dimensional Gaussian kernel expression:

[0014] Where * denotes a convolution operation, For the k-th scale, a two-dimensional Gaussian kernel, The standard deviation of the Gaussian kernel at the k-th scale ranges from 0.8 to 6.0 pixels. In this embodiment, the standard deviations corresponding to the three scales are used. , , The larger the standard deviation value, the larger the illumination smoothing scale; k is the scale index, which is a positive integer. This scheme sets k=1, 2, 3 for three scales. The weight coefficients are for the corresponding scale, satisfying the normalization constraint. Each coefficient has a value range of (0,1) and a sum of 1.0. This embodiment preferably uses the following weights. , , By employing multi-scale weighted fusion, robust estimation of illumination changes at different scales is achieved, providing a foundation for subsequent image enhancement and feature extraction.

[0015] By separating illumination interference through logarithmic transformation, a standardized feature image without lighting deviation is restored:

[0016] Eliminate the problem of outline displacement caused by differences in light and dark colors, and local highlights and shadows.

[0017] 2.2 Superpixel segmentation and intelligent background artifact removal The SLIC superpixel clustering algorithm is used to divide the image into regions, and a region feature distance discrimination formula is established:

[0018] in, The normalized comprehensive feature distance between a pixel and its superpixel cluster center is used to determine the superpixel affiliation of a pixel. The Euclidean distance between the pixel and the cluster center in the CIELAB color space represents the difference in color features. The coordinates of the pixel and the cluster center in the image plane represent the Euclidean distance, which characterizes the spatial position difference; The compactness balancing parameter is used to adjust the weight ratio of color similarity and spatial proximity, and to control the boundary fit and shape regularity of superpixels. The value range is 10~40; this scheme preferably has m=20: the smaller the value of m, the closer the segmentation result is to the image edge and the more irregular the superpixel shape; the larger the value of m, the more regular the superpixel geometry is and the easier it is to ignore the fine texture boundary. The initial grid step size for superpixels is calculated from the preset total number of superpixels K and the total number of pixels N in the image, satisfying S= The function is used to normalize the spatial distance term, eliminating the scale influence of different image resolutions on distance calculation. The preset total number of superpixels, K, is adaptively set according to the input image resolution, forming a hierarchical granularity partitioning rule: Low-resolution images (total pixels ≤ 250,000): K is 100~300, suitable for coarse-grained region partitioning, quickly removing large areas of uniform background artifacts; Medium-resolution images (250,000 < total pixels ≤ 1,000,000): K is 300~600, the preferred range for this scheme, balancing segmentation accuracy and computational efficiency, suitable for target subject contour extraction; High-resolution images (total pixels > 1,000,000): K is 600~1000, achieving fine segmentation of subtle textures and small-sized target regions, eliminating local fragmented background noise artifacts. Based on this, according to geometric feature constraints such as region area, rectangularity, and ellipticity, the target connected components of the insole are automatically filtered, eliminating invalid artifact regions such as desktop clutter, edge halos, and extended edges, fundamentally avoiding the problem of excessively large contour recognition width, laying the foundation for accurate extraction of the insole contour and size subsequently.

[0019] 2.3 Repair and Reconstruction of Worn Edge Contours To address issues such as edge burrs, localized defects, and contour deformation that occur during the use of insoles, this method employs morphological opening and closing combination operations and cubic B-spline curves to complete contour repair and reconstruction, restoring the factory-standard smooth contour.

[0020] First, a combination of erosion and expansion operations is used to eliminate small burrs and protrusions at the edges and fill local wear gaps in the contour. The operation is defined as follows: Erosion calculation:

[0021] Dilation operation:

[0022] In the formula: The image is a binary image of the insole outline to be processed. The pre-defined structural element is used to define the shape of the neighborhood operation. This scheme uses a 3×3 square structural element. The size of the structural element can be adaptively adjusted according to the image resolution: a 2×2 size is used for low-resolution insole images to reduce the amount of computation; a 5×5 size is used for high-definition detailed images to enhance the neatness and repair effect of small burrs and gaps. structural element Translate to coordinates The position after the erosion operation removes small burrs and isolated noise points by shrinking the contour boundary; the dilation operation fills local defects and fracture gaps by expanding the contour boundary. The combination of the two achieves the initial regularization of the contour and the repair of defects.

[0023] Secondly, to eliminate the jagged edges of the contour after morphological processing, a three-stage process was employed. The mathematical expression for smoothing and reconstructing a discrete contour point set using spline curves is:

[0024] In the formula: For parameters The corresponding coordinates of the reconstructed contour points; Let i be the i-th control point in the contour point set; For the i-th kth time Spline basis functions, in this method =3 (three times) (Spline); n is the total number of control points. By interpolating and fitting discrete contour points using this formula, the standard smooth contour of the insole can be restored, eliminating shape errors caused by wear and tear during use, thus laying the foundation for subsequent high-precision edge detection.

[0025] 2.4 Subpixel High-Precision Edge Detection Based on a grayscale moment sub-pixel edge localization algorithm, this method breaks through the limitations of traditional pixel-level detection accuracy, achieving precise locking of the true physical boundary of the insole and reducing the error of the contour recognition system. The core formula of the algorithm is:

[0026] In the formula: ( , () represents the sub-pixel level edge point coordinates obtained from the solution; , () represents the pixel center coordinates of the local neighborhood of the edge; , , These are the zero-order moment, first-order horizontal moment, and first-order vertical moment of the edge neighborhood, calculated from the grayscale values ​​of the pixels within the neighborhood. By leveraging the statistical properties of the grayscale moments, the edge localization accuracy is improved from the integer pixel level to the sub-pixel level, effectively reducing the systematic error in contour recognition and providing high-precision boundary data for subsequent dimensional measurements.

[0027] Step 3: Adaptive effective region selection and arbitrary pose semantic recognition 3.1 Multi-scale sliding window candidate region extraction A global pixel traversal sliding window algorithm is used to generate a set of all candidate bounding boxes, achieving preliminary screening of potential target regions. The set is defined as follows:

[0028] In the formula: For the set of candidate bounding boxes; Let i be the i-th candidate rectangle, where ( ) represents the coordinates of the top left corner of the box. () represents the coordinates of the bottom right corner of the bounding box; n represents the total number of candidate bounding boxes. By traversing the image through sliding windows of different scales, all rectangular regions that may contain the insole target are generated, providing a candidate set for subsequent optimal region selection.

[0029] 3.2 Optimal Region Selection Based on Physiological Priors Based on a massive foot sample database, a foot contour aspect ratio constraint model was established to select the unique optimal bounding box that conforms to the physiological size characteristics of the foot, accurately defining the main area of ​​the insole. The model is defined as follows:

[0030] In the formula: The aspect ratio of the candidate region; This represents the vertical length of the candidate box; The horizontal width of the candidate box; Based on the reasonable range of length-to-width ratio obtained from statistical analysis of foot physiological characteristics, this scheme is statistically calibrated using a foot type database. =2.2、 =3.6; For different population groups, the following ranges are used for fitting: women's narrow-foot insoles use the range [2.8, 3.6], and men's wide-foot insoles use the range [2.2, 3.0]. Any candidate bounding boxes whose aspect ratios deviate from the corresponding range are considered invalid and directly removed. Using a Bayesian probability criterion, the optimal region satisfying the physiological prior constraints is selected from the candidate bounding box set. Invalid candidate boxes such as background clutter and local artifacts are removed, thus locking in the smallest bounding region containing the main body of the insole.

[0031] 3.3 Calculation of principal axis attitude angles of profile inertia Calculate the second-order central moment of the contour, construct the inertia tensor matrix, solve the global deflection azimuth angle of the contour, and realize the standardized correction of arbitrary posture images.

[0032] First, calculate the second central moments of the profile:

[0033] In the formula: , , These are the horizontal second moment, the vertical second moment, and the cross second moment, respectively; , ) represents the centroid coordinates of the contour; I(x,y) represents the pixel value of the binary image of the contour at (x,y).

[0034] Solving the global deflection azimuth angle of the profile based on second-order moments:

[0035] In the formula: The angle between the principal axis of contour inertia and the horizontal axis is the global deflection azimuth angle of the image. Based on this deflection angle, a global rotation transformation is performed on the image, uniformly correcting insole images in arbitrary orientations to the system's standard reference orientation, eliminating the influence of rotational deviations on subsequent contour registration and dimensional measurement.

[0036] 3.4 Hierarchical Semantic Location Intelligent Judgment The global centroid coordinates of the contour are calculated. Combining massive prior statistical features of plantar biomechanics, and using the contour radius of curvature and regional eccentricity as criteria, a fuzzy comprehensive evaluation model is constructed to autonomously divide the forefoot and heel regions without manual feature labeling. The formula for calculating the centroid coordinates is:

[0037] In the formula: The coordinates are the centroid coordinates of the contour. Based on the centroid coordinates and contour curvature features, combined with prior knowledge of foot biomechanics distribution, automatic semantic segmentation of the forefoot, arch, and heel areas of the insole is achieved, providing a regional positioning basis for subsequent pressure field mapping and surface design.

[0038] Step 4: Registration and deformation of orthogonal projection-constrained 3D model 4.1 Mathematical Modeling of Orthogonal Projection Mapping Establish a spatial Cartesian coordinate system O-xyz, setting the z-axis as the vertical top-view projection normal and xOy as the horizontal reference projection plane. Define the orthogonal projection transformation operator:

[0039] In the formula: X=(x,y,z) are the coordinates of the 3D mesh vertices; P(X) are the 2D coordinates of the vertices on the horizontal reference projection plane. Only the horizontal projection coordinates of the 3D mesh vertices are extracted as the registration basis, and the thickness direction and lateral warping structure are not included in the constraint calculation, so as to achieve the alignment of the projection domain of the 2D contour and the 3D model.

[0040] 4.2 Two-stage contour precise registration This method employs a two-stage strategy of "coarse registration + fine registration" to achieve precise alignment and adaptive deformation between the three-dimensional master model and the two-dimensional insole outline.

[0041] Coarse registration: Improved ICP Iterative Closest Point Algorithm Construct the optimal rigid body transformation objective function, solve for the rotation and translation matrices, and complete the initial contour alignment:

[0042] In the formula: It is a two-dimensional contour point set; Let R be the set of projection points of the 3D model; T be the rotation matrix to be solved; n be the translation matrix to be solved; and n be the total number of contour points. The rigid body transformation parameters are solved iteratively to achieve initial coarse alignment between the 3D model and the 2D contour. The ICP algorithm of this invention sets up dual iteration termination conditions, and the iteration stops when either condition is met: Error convergence threshold condition: The difference between the objective function values ​​of two consecutive iterations is less than the preset convergence threshold ε. In this scheme, ε = 10. -3 ; Maximum iteration limit: Set the maximum number of iterations N. max =100, to prevent infinite iteration when the point set fails to converge.

[0043] Iteration Termination Logic: After each iteration, the nearest point pairs are rematched and the rotation matrix R and translation matrix T are updated. The objective function loss value of the current iteration is calculated. If the loss difference between two adjacent iterations is less than ε, the contour rigid body alignment is determined to have reached a stable state, and the coarse registration iteration is terminated. If the cumulative number of iterations reaches N... max If the convergence threshold is not yet met, the iteration is forcibly terminated and the current optimal rigid body transformation parameters are output. Fine registration: TPS thin plate spline non-rigid deformation. With the constraint of minimizing deformation energy, the TPS thin plate spline algorithm is used to complete the non-rigid fine registration, driving the adaptive elastic deformation of the standard 3D STL mesh. The deformation energy function is:

[0044] In the formula: For thin plate spline transformation function; The bending energy represents the smoothness of the deformation. By minimizing the bending energy, smooth elastic deformation of the 3D mesh is achieved while ensuring complete contact with the contour boundaries, avoiding local distortion and mesh tearing.

[0045] 4.3 Layered Constraint Deformation Control A dual-layer differential constraint mechanism is set up to achieve layered control of contour and thickness: The horizontal projection plane is a strongly constrained region: it forces the fitting of the contour boundary to ensure that the two-dimensional contour and the three-dimensional model projection completely coincide, thus ensuring dimensional accuracy; The vertical thickness direction and the vertically warped edge region are relaxed constraint regions: constraint weight ω(z)→0, preserving the original structural shape of the template, without the need for forced matching, avoiding excessive deformation that could damage the structural mechanical properties.

[0046] A global scale scaling model is established based on the user input length L:

[0047] In the formula: These are the horizontal projection coordinates of the vertices of the 3D model. The coordinates are scaled; k is the scaling factor, determined by the ratio of the user-input length to the standard template length, to complete accurate dimension calibration.

[0048] 4.4 Registration accuracy verification Establish an evaluation index for contour overlap error to quantify registration accuracy:

[0049] In the formula: e is the contour coincidence error; The target outline is two-dimensional; S represents the projected contour of the 3D model; S is the total area of ​​the contour region. This scheme presets a registration qualification error threshold e. th =0.008, when the calculation error e is less than the preset threshold e th At this point, registration is considered complete. After the above operations, a 3D insole that perfectly matches the customer's original insole outline can be obtained.

[0050] Step 5: Construction of massive plantar prior models and correction of body weight load 5.1 Preprocessing of Plantar Pressure Dataset The original plantar pressure samples, numbering in the tens of millions, were processed by removing outliers, normalizing, and spatially registering to construct a standardized sample set.

[0051] Where: P is the standardized plantar pressure sample set; Let be the i-th preprocessed plantar pressure distribution sample; N is the total number of samples. Preprocessing eliminates acquisition noise and individual differences, providing a high-quality data foundation for subsequent mechanical prior modeling.

[0052] 5.2 Gaussian Mixture Clustering Mechanical Prior Modeling A Gaussian mixture model is used to perform force feature clustering and generate a general plantar pressure standard field. The mixture probability density function is:

[0053] In the formula: is the mixture probability density of plantar pressure characteristics; K is the number of Gaussian components; Let be the weighting coefficient of the k-th Gaussian component, satisfying =1, and after iterative solving using the EM algorithm on a large number of foot pressure samples, the optimal weight values ​​for the three components are: =0.42 (heel component), =0.23 (foot arch component) =0.35 (forefoot weight); Let k be the k-th Gaussian distribution, where The mean vector represents the central distribution of this type of force characteristic; The covariance matrix represents the degree of dispersion of the force distribution. This model is used to cluster massive samples, dividing the heel, arch, and forefoot into standard force-bearing regions, generating a universal plantar pressure standard field. .

[0054] 5.3 Global Load Correction for Body Weight Parameters Construct a weight-related correction function to achieve global adaptive adjustment of the pressure field, adapting to the force characteristics of different populations:

[0055]

[0056] In the formula: For the corrected personalized plantar pressure field; For general standard pressure fields; Here, W represents the weight correction coefficient; a and b are linear correction parameters obtained through fitting a large number of samples. The fitting rule adopts the least squares method, and linear regression is performed on samples with weight W∈[30kg, 120kg]. In this scheme, the parameter values ​​are a=0.012 and b=0.36. The larger the weight value, the higher the overall compressive strength, achieving a precise match between the pressure field and the user's weight, and providing a real mechanical basis for subsequent topographic surface design.

[0057] Step 6: Integrated design of topographic surfaces and lattice stiffness 6.1 Adaptive Concave-Convex Surface Generation The pressure-gradient coupled deformation formula is used to calculate the insole's concave displacement, enabling adaptive generation of concave-convex surfaces based on pressure distribution. For example... Figure 4 As shown, the green surface is the foot-supporting surface of the original 3D model of the insole, and the magenta surface is an adaptive concave-convex surface driven by foot pressure. The overall effect is that the greater the pressure, the more obvious the indentation.

[0058] The specific generation process is as follows:

[0059] In the formula: Let (u,v) be the displacement of the surface concavity at coordinates (u,v). The corrected plantar pressure field; The pressure gradient represents the rate of pressure change; α and β are positive coefficients, fitted and calibrated using a grid search method based on tens of thousands of foot-pressure surface topography samples. Both coefficients have a value range of [0, 1]. In this embodiment, the parameters α = 0.065 and β = 0.022 are preferably fixed. α controls the influence of the pressure value on the depression depth, and β controls the influence of the pressure gradient on the smooth extension range of the edge. The pressure value determines the depth of the foundation depression; the larger the pressure gradient, the larger the area of ​​the depression extension, achieving a continuous and smooth transition of the surface and avoiding local stress concentration.

[0060] 6.2 Partitioned Lattice Stiffness Regulation Establish a mapping relationship between pressure parameters and 3D printed lattice structures to achieve differentiated stiffness design for different stress regions. For example... Figure 5 As shown, the beam diameter is larger in areas of higher pressure, and smaller in areas of lower pressure. This achieves the effect of pressure-driven adaptive hardness adjustment.

[0061] The specific implementation of pressure-based adaptive control of partitioned lattice stiffness is as follows:

[0062] In the formula: ρ is the lattice porosity; t is the lattice unit wall thickness; , This is the mapping function from the pressure field to the lattice parameters; This represents the upper limit of porosity, corresponding to a basic state where 70% of the lattice space is hollow and only 30% is a solid supporting structure; the wall thickness constraint range... ,Right now , The mapping slope is calibrated as follows: =0.0022MPa -1 , =0.0018MPa -1 In high-pressure support areas (such as the heel and forefoot pressure points), porosity is reduced and wall thickness is increased to improve structural rigidity and enhance support performance; in low-pressure buffer areas (such as the arch and edge areas), porosity is increased and wall thickness is reduced to enhance shock absorption flexibility, achieving a balance between support and comfort.

[0063] Step 7: Model Smoothing Optimization and Format Output The Laplacian mesh smoothing algorithm is used to perform global optimization of the 3D model, eliminating mesh wrinkles, stress concentrations, and surface defects, generating a closed, complete, and directly printable standard STL format insole model. The algorithm iteration formula is as follows:

[0064] In the formula: These are the original coordinates of the i-th grid vertex; These are the coordinates of the smoothed vertices; As vertices The set of adjacent vertices; n is the number of adjacent vertices; λ is the smoothing coefficient, controlling the smoothing intensity. This scheme limits the value range of λ∈(0,0.5], and preferably uses a fixed smoothing coefficient λ=0.25; the algorithm sets the maximum number of iterations T. max =15, reaching the maximum number of iterations or the average offset of the vertex coordinates of two adjacent iterations is less than the convergence threshold of 10. -4 When the vertex coordinates reach mm, the smoothing iteration is terminated. By iteratively updating the vertex coordinates, local mesh distortions and wrinkles are eliminated, ensuring a smooth and continuous surface, and finally outputting a standard STL format model that meets the requirements of 3D printing.

[0065] The beneficial effects of this invention are as follows: 1. This invention addresses issues such as light and shadow interference, background clutter, and worn or damaged insoles in home photography. It sequentially performs illumination equalization, artifact removal, and contour restoration, then uses sub-pixel-level edge detection to pinpoint the true physical boundaries, ensuring accurate contour recognition. Combining the aspect ratio constraints of foot physiological characteristics, it filters effective areas, automatically calculates the insole's deflection angle and corrects its posture. Finally, it performs initial alignment through coarse registration, achieves elastic deformation through fine registration, and calibrates the dimensions according to the measured length, ensuring a precise fit between the finished product and the shoe cavity. Dimensional deviations are controlled within ±0.2mm, completely avoiding problems such as squeezing or looseness.

[0066] 2. A general mechanical model is built based on a massive amount of foot pressure samples. This model is then adjusted for user weight to obtain a personalized pressure distribution, ensuring the insole's mechanical design conforms to individual stress characteristics. An adaptive concave-convex surface is generated based on the pressure distribution, with deeper depressions in high-pressure areas and smoother surfaces in low-pressure areas. Simultaneously, a zoned lattice structure is designed—high-density, thick-walled lattices are used to reinforce support in high-pressure areas such as the heel and forefoot, while low-density, thin-walled lattices are used in low-pressure areas such as the arch and edges to enhance cushioning. This balances support and flexibility, preventing localized stress concentration, and ensuring that even after two hours of continuous wear, there will be no soreness or discomfort.

[0067] 3. Low customization cost, simple process, and easy scalability: Breaking away from the reliance on professional equipment in traditional customization, users only need to take a photo of the insole with a regular smartphone, manually measure the length, and input their weight to complete all information collection. The entire process requires no manual intervention, automatically completing image processing, posture correction, region division, model registration, mechanical optimization, and format output. The generated files can be directly used for regular 3D printing. It also adapts to most shoe size ranges, and the symmetrical nature of the insole simplifies the design process, significantly reducing the customization threshold and production costs, making it suitable for the large-scale production of everyday wear and sports protective insoles. Attached Figure Description

[0068] Figure 1 This diagram illustrates the global preprocessing optimization effect of the insole image. Figure 2 The diagram shown is a vertical rotation in any orientation. Figure 3 The diagram shown illustrates the registration and deformation process of a 3D insole model based on customer insole information. Figure 4 The diagram shown is a schematic of an adaptive concave-convex surface; Figure 5 The diagram shown is a schematic of partitioned lattice stiffness control based on pressure distribution. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments.

[0070] Example 1

[0071] 1. Application scenario: Adult men's daily commuting and casual leather shoes, size 42, flat low-top style, standard internal depth of 18mm and maximum internal width of 98mm.

[0072] 2. Basic User Input Parameters Top view of the insole: Taken by the user with the mobile phone placed flat in natural indoor light, at a shooting distance of 25cm, with an image resolution of 4096x3072 pixels; The axial reference length L of the insole: 270mm (measured manually). User weight W: 75kg.

[0073] 1. Preset general algorithm fixed parameters Retinex multiscale number n=3, weighting coefficients , , , corresponding to the Gaussian kernel standard deviation. , , ; SLIC superpixel compactness parameter m=20, preset number of superpixels per image K=800; Morphological structural element B: Square structure, 15x15 pixels in size; Foot length-to-width ratio constraint range [2.2,3.0]; ICP algorithm iteration termination condition: The average error of the contour point set in two consecutive iterations is <10. -3 mm; Registration error acceptable threshold e th =0.008mm 2 ; Gaussian mixture clustering component number K=3; Weight correction factors a=0.012, b=0.36; The surface deformation coefficients are α = 0.065 and β = 0.022. Laplace smoothness coefficient =0.25, number of iterations = 15.

[0074] 1. Anomaly Handling Rules: In this embodiment, the image is clear, the size measurement error is <0.5mm, and the weight is within the model sample range (40kg-120kg), so there is no need to activate the fault tolerance mechanism.

[0075] II. Step-by-step execution process and measured data of the entire process Step 1: Multi-source basic information collection Users upload a top-down image of the insole taken with their mobile phones via the system front end, and manually enter the measured insole length L=270mm and their own weight W=75kg. The system then generates a set of input parameters.

[0076] The system verifies the validity of the parameters: the length of 270mm and the weight of 75kg are both within the normal sample range, proceeding to the next step.

[0077] Step 2: Global Preprocessing Optimization of Complex Scene Images 1. Multi-scale Retinex illumination equalization processing Indoor natural light presents localized shadows and tabletop reflections. The system employs a three-scale Retinex algorithm to separate the light and remove ambient light and shadow interference. The processed image exhibits uniform brightness and darkness, fully preserving the insole's features without any highlights or shadows.

[0078] 2. SLIC superpixel segmentation + background artifact removal The modified distance formula (sum of squares) is used to complete the region segmentation. Combined with the region area and rectangularity features, artifacts such as desktops, clutter, and edge halos are removed, and only the effective connected components of the insole are retained.

[0079] 3. Repair and reconstruction of worn edge contours The target insole has slight edge burrs and a small wear gap. The erosion and dilation morphological operation was completed using a 15*15 pixel square structural element, and the contour was smoothed using a cubic B-spline curve. After repair, the contour is smooth and continuous, without defects or burrs.

[0080] 4. Subpixel high-precision edge detection Subpixel-level edges are extracted based on the gray-scale moment algorithm, resulting in a closed, high-precision insole outline with an overall pixel size error of <0.3 pixels.

[0081] Step 3: Adaptive effective region selection and arbitrary pose semantic recognition 1. Multi-scale sliding window candidate region extraction: The system generates 128 sets of candidate bounding boxes at different scales.

[0082] 2. Optimal Region Selection: Calculate the aspect ratio of each candidate box. In this embodiment, the effective region aspect ratio is... =2.76, which is within the reasonable range of [2.2, 3.0]. Based on the Bayesian criterion, the unique optimal insole region is selected.

[0083] 3. Attitude Angle Calculation and Rotation Correction The insole placement angle in the original image =28.6°. The system solves the deflection angle by the second-order central moment and completes the global rotation transformation. After correction, the main axis of the insole is completely coincident with the standard reference direction.

[0084] 4. Hierarchical semantic location discrimination The centroid coordinates of the contour are calculated, and the contour curvature, eccentricity, and a priori fuzzy evaluation model of plantar mechanics are combined to automatically divide the region: the rear end is the heel area, the middle is the arch area, and the front end is the forefoot area. The boundary of the region division is consistent with the physiological distribution of the human foot and no manual marking is required.

[0085] Step 4: Orthogonal projection constrained 3D model registration and deformation 1. Basic template selection: Use the 3D STL template of standard insole for size 42 casual leather shoes. The axial length of the standard template is 265mm.

[0086] 2. Scale scaling calculation The scaling factor k = 270 / 265 ≈ 1.0189, and the system scales the horizontal projection coordinates of the model as a whole.

[0087] 3. Two-stage contour registration Coarse registration (improved ICP algorithm): Iterates 18 times in total, reaches the termination condition, and completes the initial alignment of the two-dimensional contour and the projection points of the three-dimensional model. Precise registration (TPS thin plate spline deformation): Non-rigid elastic deformation is completed with the minimum deformation energy as a constraint. The horizontal projection area is strongly constrained to fit the contour, while the thickness (18mm) and the vertical warping area at the edge are relaxed to retain the original structure of the template.

[0088] 4. Registration accuracy verification The calculated contour overlap error is e = 0.005 mm. 2 It is less than the acceptable threshold of 0.008mm. 2 The registration was deemed successful.

[0089] After registration, the shape and size of the 3D insole model are completely consistent with the user's original insole.

[0090] Step 5: Construction of a large number of foot prior models and correction of body weight load. 1. Call up the standard pressure field: retrieve the standard plantar pressure field P_0(u,v) for the general adult population.

[0091] 2. Weight Load Correction Calculation Weight correction factor: =aW+b=0.012×75+0.36, calculated to be 1.26, the personalized pressure field after correction: P(u,v)=1.26×P_0(u,v) The overall plantar pressure intensity is adaptively adjusted upwards according to the user's weight, with high-pressure areas concentrated in the heel and forefoot metatarsal region, and lower pressure in the arch area, which is consistent with the force distribution characteristics of normal adult walking.

[0092] Step 6: Integrated design of topography, surface features, and lattice stiffness 1. Adaptive Concave-Convex Surface Generation The indentation displacement is calculated using the formula Z(u,v): High pressure zone on the heel and forefoot: depression depth 1.0~5.0mm; Low-pressure area of ​​the arch: depression depth 0.8~1.5mm; The surface is continuous and smooth, without local abrupt changes, thus avoiding stress concentration.

[0093] Designed to meet the needs of casual leather shoes, the overall structure features a low-deformation planar design with gently undulating curves.

[0094] 1. Partitioned lattice stiffness modulation Based on pressure field mapping lattice parameters, a general lattice mapping rule is adopted: Heel region (high pressure): lattice porosity 75%, unit wall thickness 2.0mm, high support stiffness; (the higher the porosity, the more severe the hollowness). Forefoot region (medium-high pressure): lattice porosity 81%, unit wall thickness 1.8mm, balancing support and bending; Foot arch region (low pressure): lattice porosity 85%, unit wall thickness 1.2mm, high flexibility buffer.

[0095] Step 7: Model Smoothing Optimization and Format Output The Laplacian mesh smoothing algorithm, with a smoothing factor of 0.25, is employed in 15 iterations to eliminate mesh wrinkles, broken surfaces, and localized stress concentrations. The final result is a closed, complete standard STL model, 838MB in size, which can be directly imported into a photopolymer 3D printer.

[0096] III. 3D Printing Molding and Assembly Inspection 1. Printing parameters Printing equipment: Consumer-grade photopolymer 3D printer; Consumables: Flexible photosensitive resin; Printing layer thickness: 0.04mm; Single-layer exposure time: 2.8s; Sub-layer hardening exposure time: 15s; Printing time: approximately 4 hours and 10 minutes.

[0097] 2. Finished product dimensions remeasurement Total axial length: 270.1 mm (error +0.1 mm); Maximum width: 97.9mm (error -0.1mm); Overall thickness: 18.0mm (no error); Dimensional deviations are all controlled within ±0.2mm, meeting assembly requirements.

[0098] 1. Shoe cavity assembly test The finished insole was embedded into the target size 42 casual leather shoe. The insole fits perfectly with the bottom and side walls of the shoe cavity, without any looseness, curling, or squeezing.

[0099] 2. Wearable Sensation Test After two hours of continuous walking, the heel and forefoot are well supported, and there is no issue of the arch of the foot being unsupported or uncomfortable. The overall softness and firmness are moderate, and the shock absorption and support performance are well-suited for daily commuting.

[0100] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it.

Claims

1. An adaptive 3D printed insole intelligent generation method based on monocular projection semantic recognition, characterized in that, Includes the following steps: Step 1: Multi-source basic information collection: Collect top-view 2D images of the insoles taken by the user. The input parameter set is constructed using manually measured axial reference length L of the insole and user weight W. ; Step 2: Global preprocessing and optimization of complex scene images: The image reflection component and illumination component are separated by multi-scale Retinex illumination equalization algorithm, background artifacts are removed by SLIC superpixel segmentation, and the wear contour is repaired by morphological opening and closing operation + cubic B-spline curve. Combined with gray-scale moment subpixel edge detection algorithm, a standardized high-precision insole contour is obtained. Step 3, Adaptive effective region selection and arbitrary pose semantic recognition: based on the foot aspect ratio constraint model Optimal regions are selected, and the deflection angle is calculated using the second-order central moments of the contour. Complete attitude correction, combined with contour centroid Automatically divides the forefoot, arch, and heel areas based on curvature characteristics; Step 4: Orthogonal projection constrained 3D model registration and deformation: Establish orthogonal projection transformation operator Using the horizontal projection plane as a strong constraint and thickness and edge warping as relaxed constraints, the ICP coarse registration algorithm is employed. By combining the TPS thin plate spline non-rigid deformation algorithm, adaptive deformation of two-dimensional contours and three-dimensional templates is achieved, and size calibration is completed according to the scale scaling factor. Step 5: Construction of a massive plantar prior model and correction of body weight load: Constructing a Gaussian mixture clustering prior model based on tens of millions of plantar pressure samples. Through the weight correction function Generate personalized pressure fields ; Step 6: Integrated design of topographic surfaces and lattice stiffness: based on the pressure-gradient coupling formula Generate adaptive concave-convex surfaces and establish the relationship between pressure and lattice porosity. Wall thickness mapping relationship This enables partitioned variable stiffness design; Step 7, Model Smoothing Optimization and Format Output: Using the Laplacian Mesh Smoothing Algorithm Optimize the model and output a standard STL format file for 3D printing.

2. The method according to claim 1, characterized in that, In step 2, the multi-scale Retinex algorithm satisfies the following conditions: number of scales n=3, weight coefficients... , , Corresponding Gaussian kernel standard deviation , , The formula for light separation is: Morphological operations use a circular structuring element with a radius of 3 pixels, and the erosion operation is... The expansion operation is .

3. The method according to claim 1, characterized in that, Step 3 Foot length-to-width ratio constraint range ∈[2.2,3.6], the formula for calculating the centroid coordinates of the profile is: Second-order central moments .

4. The method according to claim 1, characterized in that, In step 4, the ICP algorithm iteration termination condition is that the average error of the contour point set in two consecutive iterations is <0.02mm, and the registration error meets the acceptable threshold. <0.008mm 2 The TPS deformation is constrained by minimizing bending energy.

5. The method according to claim 1, characterized in that, In step 5, the number of Gaussian mixture cluster components K=8, and the weight coefficients satisfy... =1, weight correction factors a=0.012, b=0.36, general standard pressure field It was obtained through cluster analysis.

6. The method according to claim 1, characterized in that, In step 6, the surface deformation coefficients α = 0.065 and β = 0.022; the high-pressure areas (heel and forefoot) satisfy... [70%, 85%]、 [1.5mm, 2.5mm], low-pressure area (arch) meets requirements [80%, 90%]、 [1.0mm, 1.5mm].

7. The method according to claim 1, characterized in that, In step 7, the Laplacian smoothness coefficient λ=0.25, the number of iterations is 15, and the optimized model has no mesh wrinkles or broken surface defects. The STL file is compatible with photopolymer 3D printers and photosensitive resin flexible consumables.

8. The method according to claim 1, characterized in that, The insole has a symmetrical integrated structure with no distinguishing features between the front and back projection outlines. There is no need to classify left and right feet or identify the front and back sides. The shape design is completed in all dimensions only through the top-view two-dimensional projection information.

9. The method according to claim 1, characterized in that, The image acquisition device is a smartphone with 12 megapixels or higher, a shooting distance of 25cm, an image resolution of ≥2048×4096 pixels, and an indoor natural light or conventional lighting environment.

10. The method according to any one of claims 1 to 9, characterized in that, Suitable for everyday wear and sports protective insoles for shoe sizes 35-46. The finished product has an axial length deviation of ≤±0.2mm and a width deviation of ≤±0.2mm. The fit of the shoe cavity is ≥99%. There is no local pressure soreness or foot discomfort after wearing for 2 hours.

Images