Personalized variable-density insole manufacturing method based on deep learning

Abstract

The invention discloses a personalized variable-density insole manufacturing method based on deep learning, and the method comprises the following steps: 1) collecting plantar pressure data of a volunteer, and enabling a network deep learning model to have the recognition capability for a plantar pressure imbalance case; 2) detecting plantar pressure data on a pressure detection plate by a testee, and carrying out regional distinguishing on an image generated by plantar pressure scanning; 3) calculating a pressure average value P; 4) calculating the basic thickness Hbase of the forefoot, the arch and the heel of the testee; 5) comparing the pressure average value P with a preset value T, and increasing a weight coefficient W to obtain a recommended key parameter H of the insole; (6) a foot contour scanner scans the feet of the testee; 7) superposing and fusing the 3D contour of the foot sole of the testee and the insole prototype; and 8) 3D printing. According to the method, the MobileNetV3 deep learning model is fused into the design and manufacturing process of the orthopedic insole, so that the individuation and the accuracy of the insole can be remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of variable density insole manufacturing, and specifically to a method for manufacturing personalized variable density insoles based on deep learning. Background Art

[0002] Scoliosis is a common spinal deformity characterized by an abnormal curvature of the spine on the side, forming a "C" or "S" shaped curve. Under normal circumstances, the spine should be a straight line when viewed from the back, but patients with scoliosis have a significant curvature of the spine in the thoracic or lumbar region. Scoliosis can be divided into many types, the most common of which is adolescent idiopathic scoliosis, which usually occurs in adolescents between the ages of 10 and skeletal maturity. Adolescent idiopathic scoliosis is a common adolescent bone disease that not only affects the patient's appearance, but may also cause problems such as cardiopulmonary dysfunction, nerve compression, and visceral dysfunction, seriously affecting the patient's mental health and quality of life. At present, the treatment methods for scoliosis mainly include observation and follow-up, physical therapy and exercise, brace treatment, and surgical treatment. For mild scoliosis, doctors usually recommend regular observation and follow-up to monitor the development of the disease, while cooperating with specific physical therapy and exercise to improve muscle strength and balance and relieve symptoms. For moderate or severe scoliosis, it may be necessary to wear a special brace to limit further curvature of the spine, or even surgery to correct severe spinal deformity. However, these methods all have their limitations to a certain extent. For example, the effects of physical therapy and brace therapy are limited by patient compliance and individual differences, and surgery has certain risks and recovery periods.

[0003] There is a mutually deteriorating relationship between plantar pressure imbalance and scoliosis symptoms. Early intervention of plantar pressure imbalance can effectively control or help improve scoliosis symptoms. Therefore, orthopedic insoles have gradually attracted attention as an auxiliary treatment method. Orthopedic insoles can assist in correcting scoliosis by adjusting the pressure distribution on the soles of the feet, improving body posture, and reducing the burden on the spine and core muscles.

[0004] The invention application with application number 202210746260.8 discloses a method for making an adaptive plantar pressure imbalance adjustment insole, including the following steps: collecting data, generating a plantar pressure distribution map, binarizing the plantar pressure distribution map, detecting feature points, feature matching, calculating homography matrix, twisting the image, and printing the insole. Although this technical solution can solve the problem of plantar pressure imbalance to a certain extent, the plantar pressure distribution of each person is not exactly the same. When the system generates orthopedic insoles, based on deep learning and 3D printing technology, a personalized variable density insole design can be achieved, providing a new solution to completely solve the problem of plantar pressure imbalance, which can better adapt to the patient's foot characteristics and provide more effective correction and support. Summary of the invention

[0005] The purpose of the present invention is to provide a personalized variable density insole manufacturing method based on deep learning, which accurately reflects the plantar pressure distribution of patients in daily activities and provides personalized, precise and comfortable orthopedic insole solutions according to the specific conditions of patients with adolescent idiopathic scoliosis, so as to overcome the shortcomings of the prior art.

[0006] The technical solution of the present invention is: a method for manufacturing a personalized variable-density insole based on deep learning, characterized by comprising the following steps:

[0007] 1) Host computer data preprocessing: Collect the plantar pressure data of volunteers, use data enhancement technology for data balance, and train the network of the improved MobileNetV3 network deep learning model to enable the network deep learning model to identify cases of plantar pressure imbalance;

[0008] 2) The subject detects the plantar pressure data on the pressure detection board, and the plantar pressure data is uploaded to the host computer and displayed in graphics; the image generated by the plantar pressure scan is divided into three parts: forefoot, arch, and heel;

[0009] 3) Calculate the average pressure of the forefoot, arch and heel, and after normalization, generate the average pressure P of the forefoot, arch and heel;

[0010] 4) The network deep learning model identifies the plantar pressure data of the subject, identifies whether the subject has the characteristics of plantar pressure imbalance, and calculates the basic thickness H of the subject's forefoot, arch, and heel base ;

[0011] 5) Compare the average pressure value P with the preset value T and add the weight factor W to obtain the recommended key parameter H of the insole:

[0012] H=W×(max(0,k×(PT))+H base );

[0013] 6) The foot contour scanner scans the subject's foot to obtain a three-dimensional foot model, which is imported into the host computer software;

[0014] 7) The 3D contour of the test subject's sole is superimposed and fused with the insole prototype, the thickness of the insole is adjusted, and a personalized insole model or an orthotic insole model is generated, and then exported to a compatible format file;

[0015] 8) Print the model using a 3D printer to produce an insole.

[0016] Step S2 includes:

[0017] S21) uploading the pressure matrix data obtained by the plantar pressure detector to the host computer through a custom protocol;

[0018] S22) obtaining the minimum and maximum values ​​of the pressure matrix data and mapping them into a pressure distribution diagram;

[0019] S23) filtering the data to remove noise, then binarizing the pressure data, performing image masking, and separating them from the central axis to generate left and right foot pressure data;

[0020] S24) Use the image correction algorithm in OpenCV to correct the pressure matrices of the left and right feet respectively, and generate a vertical plantar pressure correction map of the forefoot and heel. In the plantar pressure correction map, the pressure matrix is ​​evenly divided into three areas A, B, and C, corresponding to the forefoot, arch, and heel areas respectively.

[0021] In step S1, data enhancement is performed using one or more of the methods ad:

[0022] a) rotating one or more of the horizontally placed plantar pressure distribution cloud maps by 10-25 degrees on a horizontal plane to obtain a new plantar pressure distribution cloud map;

[0023] b) increasing or reducing the width and / or height of the plantar pressure distribution cloud map by less than 15% to obtain a new plantar pressure distribution cloud map;

[0024] c) cutting part of the edge of the plantar pressure distribution cloud map by less than 15% to obtain a new plantar pressure distribution cloud map;

[0025] d) Adjust the brightness of the plantar pressure distribution cloud map within the range of [0.9-1.5] to obtain a new plantar pressure distribution cloud map.

[0026] Furthermore, step 7 includes:

[0027] S701) setting a reference plane M, moving the sole plane determined by the forefoot and the heel in the foot 3D model to coincide with the reference plane M, and moving the foot 3D model downward so that the bottom of the foot 3D model partially sinks into the insole prototype;

[0028] S702) The partially sunken portion of the bottom of the foot 3D model in the insole prototype is removed to form a partially sunken first insole model.

[0029] Furthermore, step 7 also includes the step of adjusting the density of the insole by zoning: setting a reference plane M, drawing an area AP where the plantar pressure is greater than a preset value T, and an area BP where the plantar pressure is less than or equal to the preset value T according to the forefoot, arch, and heel areas on the reference plane M, satisfying ρ AP ≥ρ BP .

[0030] Furthermore, step 7 includes the step of generating a plantar contour adaptive pressure correction insole model, by editing and correcting the foot model using Meshmixer software, thereby generating a personalized correction insole:

[0031] S721) First, import the foot 3D model into Meshmixer;

[0032] S722) using the “deformation” tool in Meshmixer to perform nonlinear deformation on the model by dragging points or faces, so as to correct the abnormal area, such as lifting the arch of the foot or correcting the angle of the heel;

[0033] S723) After editing is completed, the edited foot model is saved and exported to a compatible format file.

[0034] Furthermore, the improved MobileNetV3 network deep learning model adopts fixed residual blocks and SE blocks. The point-level convolution layer follows the deconvolution component, and the point-level convolution changes the width; the output channel widening layer of the point-level convolution layer is adjusted by multiplying the total number of channels by the scaling factor. The fully connected layer (FCL), global average pooling (GAP) and element-level scaling (EMS) constitute the SE block, and the number of channels appearing in the input feature map is used as a criterion for determining the size of the SE block.

[0035] Furthermore, the improved MobileNetV3 network deep learning model uses the h-swish activation function. The h-swish is based on the h-sigmoid activation function. Formula (4) is generated. The swish activation function is replaced by h-swish, see formula (5), and the sigmoid activation function is replaced by the h-sigmoid activation function.

[0036] The present invention has the following beneficial effects: integrating the MobileNetV3 deep learning model into the design and manufacturing process of orthopedic insoles can significantly improve the personalization and accuracy of insoles. The deep learning model analyzes the three-dimensional morphological data and pressure distribution data of the sole, automatically extracts and learns the key features of the user's sole, such as high-pressure areas, low-pressure areas, arch morphology, etc., to generate a highly personalized insole design.

[0037] Combined with 3D printing technology, variable density insoles that meet the characteristics of the user's feet can be quickly manufactured to achieve efficient correction and support effects, making up for the shortcomings of traditional orthopedic insoles. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a whole flow chart of the method for manufacturing personalized variable density insoles based on deep learning described in the present invention.

[0039] Figure 2 This is a diagram showing the matrix update working principle of the improved MobileNetV3 network deep learning model of the present invention.

[0040] Figure 3 This is a schematic diagram of the fusion of the foot contour and the insole prototype. Figure 3 The middle grid area represents the sunken part.

[0041] Figure 4 Schematic diagram of insole density area.

[0042] Figure 5 This is the improved Mobile-netv3 network architecture diagram.

[0043] Figure 6 This is a pressure distribution diagram generated by the plantar pressure matrix data of the present invention.

[0044] Figure 7 This is a flow chart for processing plantar pressure data.

[0045] Figure 8 Schematic diagram of the regional division of the plantar pressure correction map. DETAILED DESCRIPTION

[0046] The technical solution of the present application will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments.

[0047] The components of the embodiments of the present application generally described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application claimed, but merely represents the selected embodiments of the present application.

[0048] Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of this application.

[0049] In the description of the present application, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present application. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0050] In the description of this application, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0051] See also Figure 1-Figure 8 , a method for manufacturing personalized variable-density insoles based on deep learning, comprising the following steps:

[0052] S1) Data preprocessing on the host computer: Collect the plantar pressure data of volunteers, use data enhancement technology for data balance, and train the network of the improved MobileNetV3 network deep learning model to enable the network deep learning model to identify cases of plantar pressure imbalance;

[0053] S2) The subject detects the plantar pressure data on the pressure detection board, and the plantar pressure data is uploaded to the host computer and displayed in a graphic; the image generated by the plantar pressure scan is divided into three parts: forefoot, arch, and heel;

[0054] S3) calculating the average pressure of the forefoot, arch and heel, and generating an average pressure coefficient after normalization;

[0055] S4) The network deep learning model identifies the plantar pressure data of the subject, identifies whether the subject has the characteristics of plantar pressure imbalance, and calculates the thickness H of the forefoot, arch, and heel of the subject. front , H arch , H heel ;

[0056] S5) Set the thickness of the forefoot, arch and heel H front , H arch , H heel After multiplying by the average pressure coefficient, and then by the body weight coefficient, the total thickness of the insole is adjusted to obtain the recommended key parameters of the insole to generate the insole prototype;

[0057] S6) scanning the subject's foot with a foot contour scanner to obtain a three-dimensional model of the foot, and importing it into the host computer software;

[0058] S7) superimposing and fusing the 3D contour of the testee's sole with the insole prototype to determine and generate a personalized insole model or an orthotic insole model, and exporting it to a compatible format file (such as STL);

[0059] S8) Printing the model with a 3D printer to generate an insole.

[0060] In step S1), the plantar pressure detection board generates a static plantar pressure distribution cloud map; the static plantar pressure distribution cloud map presents different pressure distribution areas with a variety of different visual identifiers (such as colors).

[0061] In step S1), the plantar pressure data of the volunteers are classified according to the symptom characteristics of plantar pressure imbalance, and are classified into five categories: bow legs, X-shaped legs, left foot disease, right foot disease, and healthy feet. The improved MobileNetV3 network deep learning model learns multiple static plantar pressure distribution cloud maps of the five categories respectively.

[0062] In step S1), when the overall data volume given to the network deep learning model is insufficient, or the information volume of a single category is insufficient, for example, the system requires 1,000 pieces of learning information for each category, when the number of volunteers with symptom characteristics of this category is insufficient, data balancing is needed to balance all categories before the training process of the network deep learning model. Data enhancement technology is used for data balancing. In this process, one or more of the following methods are used:

[0063] A) One or more of the horizontally placed plantar pressure distribution cloud maps are rotated 10-25 degrees on the horizontal plane to obtain a new plantar pressure distribution cloud map.

[0064] B) increasing or reducing the width and / or height of the plantar pressure distribution cloud map by less than 15% to obtain a new plantar pressure distribution cloud map.

[0065] C) Cutting part of the edge of the plantar pressure distribution cloud map by less than 15% to obtain a new plantar pressure distribution cloud map.

[0066] D) Adjust the brightness of the plantar pressure distribution cloud map within the range of [0.9-1.5] to obtain a new plantar pressure distribution cloud map.

[0067] Through the above steps, the data is enhanced to meet the requirements of the network deep learning model for the amount of learning data. See Table 1 below.

[0068] Table 1

[0069] Classification Mark name Actual quantity Quantity after processing 0 Bow Legs 124 1000 1 X-shaped legs 236 1000 2 Left foot disease 76 1000 3 Right foot disease 61 1000 4 Healthy legs 518 1000 total 1015 5000

[0070] The improved MobileNetV3 network is a classifier for cases of plantar pressure imbalance. After data enhancement, the plantar pressure distribution cloud map of the volunteers has sufficient information for learning and training the learning model. Through the training of the improved MobileNetV3 network, the network has the ability to identify cases of plantar pressure imbalance.

[0071] The improved MobileNetV3 network configuration is shown in Table 2 and Figure 2 As shown:

[0072] Table 2:

[0073]

[0074] In Table 2, "input" indicates the shape of the feature matrix of the current layer, i.e., the number of rows × the number of columns. In the "operation" column, bneck indicates matrix update, followed by the kernel size of the grouped convolution. bneck has the inverted residual with linear bottleneck structure of MobileNetV2, i.e., first use 1x1 convolution to increase the dimension, then perform the following operations, and have residual edges; depthwise separable convolutions of MobileNetV1; lightweight attention model. The calculation process of bneck is as follows: Figure 6 As shown in the figure, conv2d is a two-dimensional convolution. Pool represents a pooling operation.

[0075] "Convolution dimension" refers to the dimension of the n-th up-convolution output in bneck.

[0076] "Output channel" represents the output channel size.

[0077] “Whether Attention Mechanism” indicates whether to use the attention mechanism.

[0078] "Activation Function" indicates the type of nonlinear activation function currently used.

[0079] "Step length" is also called stride.

[0080] NBN (No BN) means that the convolution of the classifier part will not use the BN layer.

[0081] Compared with the traditional MobileNetV3 network, the improved MobileNetV3 network deep learning model mainly adopts fixed residual blocks and SE blocks. The point-level convolution layer follows the deconvolution component, and the point-level convolution changes the width (the number of channels). Adjusting the output channel of the point-level convolution layer can widen the layer, which is achieved by multiplying the total number of channels by the scaling factor. The SE module can achieve channel-dependent modulation. Although it cannot naturally adjust the layer width, it can be used on the layer to improve model performance and adapt to specific tasks. The fully connected layer (FCL), global average pooling (GAP) and element-level scaling (EMS) constitute the SE block, and the number of channels appearing in the input feature map is used as a criterion for determining the size of the SE block. The improved MobileNet-V3 allows us to modify the channel range within the starting convolution layer using the last classification layer to adjust the number of result models for plantar pressure imbalance case analysis.

[0082] The improved MobileNetV3 network deep learning model makes improvements in the attention mechanism and activation function:

[0083] 1) In the original MobileNet-V3 model, the attention mechanism (SE) performs pooling processing on each channel to obtain the number of channel elements, and then passes through two fully connected layers to obtain the output vector. The output obtained is equivalent to analyzing each channel of the original feature matrix and assigning corresponding weights to its importance. The second fully connected layer uses the Hard-Sigmoid activation function. Then the channel weights are multiplied back to the original feature matrix to obtain a new feature matrix.

[0084] In this paper, we replace the cross entropy loss function in the mathematical framework of Mobilenet-v3 with a bias loss function that can improve accuracy. The SE component of the model is replaced with the Enough Channel Attention (ECA) component to achieve parameter reduction. Cross-layer connections between mobile modules have been proposed to effectively utilize the comprehensive characteristics. Dilated convolutions are added to the model to increase the receptive field.

[0085] 2) Redesign the activation function

[0086] The old version of MobileNet uses swish, which is x multiplied by sigmoid, as the activation function, as shown in formula (2). The calculation and derivation are complex and not friendly to the quantization process.

[0087]

[0088] swish(x)=σ(x) (2)

[0089] The h-swish activation function is used in the improved Mobi 1eNetV3. h-swish draws on the h-sigmoid activation function. As shown in formula (4), the calculation formula and derivation of h-swish become very simple.

[0090] ReLU6(x)=min(max(x,0),6) (3)

[0091]

[0092] h-swish is x multiplied by h-sigmoid, as shown in formula (5). By replacing the swish activation function with h-swish and the sigmoid activation function with h-sigmoid activation function, it is helpful for the reasoning speed of the network and is beneficial to the quantization process.

[0093]

[0094] In step 2-4), the host computer obtains the plantar pressure distribution information of the subject through the 32*32 matrix of the array distributed flexible film pressure sensor. The host computer obtains the plantar pressure distribution data through the pressure sensor, and after normalization, generates a pressure distribution diagram based on the plantar pressure data, divides the pressure value range into multiple sections, and draws equal pressure areas based on different pressure sections. In the above steps, the collected data is transmitted to the host computer, and the data is processed using an algorithm, and the analysis algorithm focuses on the pressure distribution of the three key points of the forefoot, arch, and heel. Based on the plantar pressure peak and distribution, the algorithm will automatically generate insole parameters suitable for the user, and the forefoot thickness is determined according to the pressure distribution of the forefoot area to ensure sufficient cushioning and support. The arch height will be calculated according to the pressure distribution at the arch to provide appropriate support and prevent excessive collapse. The heel height is set according to the pressure distribution in the heel area to reduce impact and maintain stability.

[0095] See also Figure 7-8 , the steps include:

[0096] S21) uploading the pressure matrix data obtained by the plantar pressure detector to the host computer through a custom protocol;

[0097] S22) Obtain the minimum and maximum values ​​of the pressure matrix data, and map them into a pressure distribution image. Figure 6 As shown;

[0098] S23) filtering the data to remove noise, then binarizing the pressure data, performing image masking, and separating the pressure data from the central axis to generate left and right foot pressure data;

[0099] S24) using the image correction algorithm in OpenCV, respectively correcting the pressure matrices of the left and right feet, generating a plantar pressure correction map with the forefoot and heel vertically, and evenly dividing the pressure matrix into three areas A, B, and C according to the equidistant principle in the plantar pressure correction map, corresponding to the forefoot, arch, and heel areas, respectively;

[0100] S25) calculating the forefoot thickness, arch height, and heel height.

[0101] In step S25, the principles for assigning values ​​to the three parameters are:

[0102] Forefoot thickness: Based on the pressure distribution in the forefoot area, if the forefoot pressure is found to be too high, the forefoot thickness will be appropriately increased to provide better cushioning; if the forefoot pressure is evenly distributed, the standard thickness will be maintained.

[0103] Arch height: According to the pressure distribution in the arch area, if the arch pressure is found to be too large or insufficient, the arch height should be adjusted appropriately to provide sufficient support. If the arch pressure distribution is normal, maintain the standard height.

[0104] Heel height: Based on the pressure distribution in the heel area, if you find that the heel pressure is too great, increase the heel height to disperse the pressure; if the heel pressure is evenly distributed, maintain the standard height.

[0105] Forefoot thickness calculation:

[0106] 1. Calculate the average pressure P in the forefoot area avg .

[0107] 2. If P avg Exceeding the preset value T front =0.73±0.27kg / cm 2 (Reference [1]), then increase the forefoot thickness H front .

[0108] 3. Calculation formula:

[0109] H front =W×(max(0, k×(P avg -T front ))+H base1 ) (6)

[0110] Where k = 0.1 mm cm 2 / kg is the thickness coefficient, H base1 It is the basic thickness of the forefoot.

[0111] Among them, the literature [1]: Yuan Gang, Zhang Muxun, Wang Zhongqin, et al. Analysis of plantar pressure distribution and its influencing factors in normal people [J]. Chinese Journal of Physical Medicine and Rehabilitation, 2004, 26(3): 4. DOI: 10.3760 / j:issn:0254-1424.2004.03.008.

[0112] Arch height calculation:

[0113] 1. Calculate the minimum pressure P in the arch area min .

[0114] 2. If P min Below the preset threshold T arch =0.35±0.13kg / cm 2 (Reference [1]), then adjust the arch height to H arch .

[0115] 3. Calculation formula:

[0116] H arch =W×(max(0, k′×(T arch -P min ))+H base2 ) (7)

[0117] where k′ = 0.2 mm cm 2 / kg is the height increase coefficient, H base2 It is the basic heel height.

[0118] Heel height calculation:

[0119] 1. Calculate the average pressure P in the heel area heel .

[0120] 2. If P heel Exceeding the preset value T heel =1.10±0.34kg / cm 2 (Reference [1]), then increase the heel height H heel .

[0121] 3. Calculation formula:

[0122] H heel =W×(max(0, k″×(P heel -T heel ))+H base3 ) (8)

[0123] Where k” = 0.11 mm cm 2 / kg is the height increase coefficient, H base3 It is the base thickness of the heel.

[0124] After learning the input data, the improved MobileNetV3 deep learning model generates the basic thickness H corresponding to the classification of 5 preset plantar pressure imbalance symptoms (including healthy legs). base1 , H base2 , H base3 , respectively substituted into equations (6) to (8) to generate the key insole thickness parameter H front (forefoot), H heel (Heel), H areh (Arch of the foot), unit is mm.

[0125] In formula (6) to formula (8), W is the weight coefficient, which is related to the weight of the person being tested. The relationship diagram is shown in Table 3.

[0126] Table 3

[0127] Weight(kg) 25-50 50-65 65-75 75-90 ＞90 W 0.8 0.9 1 1.1 1.2

[0128] See also Figure 3 In step 7), there is a step of manually adjusting the relative position between the 3D model and the insole prototype, specifically:

[0129] S701) Set a reference plane M, move the plantar plane determined by the forefoot and the heel in the foot 3D model to coincide with the reference plane M, and move the foot 3D model downward so that the bottom of the foot 3D model is partially sunken into the insole prototype; the reference plane M is set to a horizontal plane with the forefoot thickness as a parameter, or a slope formed by the front end of the forefoot and the rear end of the heel.

[0130] S702) The partially sunken portion of the bottom of the foot 3D model in the insole prototype is removed to form a partially sunken first insole model.

[0131] See also Figure 4 In step 7), the step of adjusting the density of the insole by partition is also included, specifically:

[0132] S703) measuring the distance L between each point on the bottom surface of the foot 3D model and the reference plane M,

[0133] S704) Draw an interval contour diagram of L≤0, 0<L≤b1, b1<L≤b2, b2<L on the reference plane, which are b0 area, b1 area, b2 area, and b3 area respectively, and the corresponding densities are ρ0, ρ1, ρ2, and ρ3.

[0134] S705) The density of each zone of the first insole is set according to the interval contour map, wherein the density of each zone meets the following formula: ρ0≥ρ1≥ρ≥ρ3.

[0135] The step of adjusting the density of the partitions can also be divided according to the difference between the plantar pressure and the preset value T: on the reference plane M, the area AP where the plantar pressure is greater than the preset value T, the area BP where the plantar pressure is less than or equal to the preset value T, and the area ρ AP ≥ρ BP The preset values ​​T are respectively T of formula (6-8) front 、T arch 、T heel The numerical value of .

[0136] In step 7), there is also a step of generating a plantar contour adaptive pressure correction insole model, by editing and correcting the foot model using Meshmixer software, and then generating a personalized correction insole:

[0137] S721) First, import the foot 3D model into Meshmixer;

[0138] Meshmixer is a powerful open source software specifically designed for editing and processing 3D models. In Meshmixer, use the "Measurement" tool to perform a comprehensive analysis of the foot model to determine the pressure distribution in various parts of the foot and any possible abnormal areas, such as arch collapse, eversion or inversion, etc.

[0139] S722) using the “deformation” tool in Meshmixer to perform nonlinear deformation on the model by dragging points or faces, so as to correct the abnormal area, such as lifting the arch of the foot or correcting the angle of the heel;

[0140] S723) After the editing is completed, the edited foot model is saved and exported to a compatible format file;

[0141] S723) Return to step S701.

[0142] In addition, in S722), you can also use Meshmixer's "carving" tool to make fine modifications on the model surface, such as adding bumps or grooves where extra support is needed. In this way, the design of the insole can be customized to the specific needs of the patient. If necessary, you can also use the "cutting" tool to remove unnecessary parts or create a more ergonomic design.

[0143] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for manufacturing personalized variable-density insoles based on deep learning, characterized in that The following steps are involved: 1) Host computer data preprocessing: Collect the plantar pressure data of volunteers, use data enhancement technology for data balance, and train the network of the improved MobileNetV3 network deep learning model to enable the network deep learning model to identify cases of plantar pressure imbalance; 2) The subject detects the plantar pressure data on the pressure detection board, and the plantar pressure data is uploaded to the host computer and displayed in graphics; the image generated by the plantar pressure scan is divided into three parts: forefoot, arch, and heel; 3) Calculate the average pressure of the forefoot, arch and heel, and after normalization, generate the average pressure P of the forefoot, arch and heel; 4) The network deep learning model identifies the plantar pressure data of the subject, identifies whether the subject has the characteristics of plantar pressure imbalance, and calculates the basic thickness H of the subject's forefoot, arch, and heel base ; 5) Compare the average pressure value P with the preset value T and add the weight factor W to obtain the recommended key parameter H of the insole: H=W×(max(0,k×(P-T))+H base ); 6) The foot contour scanner scans the subject's foot to obtain a three-dimensional foot model, which is imported into the host computer software; 7) The 3D contour of the test subject's sole is superimposed and fused with the insole prototype, the thickness of the insole is adjusted, and a personalized insole model or an orthotic insole model is generated, and then exported to a compatible format file; 8) Print the model using a 3D printer to produce an insole.

2. The method for manufacturing a personalized variable density insole based on deep learning according to claim 1, characterized in that: Step S2 includes: S21) uploading the pressure matrix data obtained by the plantar pressure detector to the host computer through a custom protocol; S22) obtaining the minimum and maximum values ​​of the pressure matrix data and mapping them into a pressure distribution diagram; S23) filtering the data to remove noise, then binarizing the pressure data, performing image masking, and separating them from the central axis to generate left and right foot pressure data; S24) Use the image correction algorithm in OpenCV to correct the pressure matrices of the left and right feet respectively, and generate a vertical plantar pressure correction map of the forefoot and heel. In the plantar pressure correction map, the pressure matrix is ​​evenly divided into three areas A, B, and C, corresponding to the forefoot, arch, and heel areas respectively.

3. The method for manufacturing a personalized variable density insole based on deep learning according to claim 1, characterized in that: In step S1, data enhancement is performed using one or more of the methods ad: (a) rotating one or more of the horizontally placed plantar pressure distribution cloud maps by 10-25 degrees on a horizontal plane to obtain a new plantar pressure distribution cloud map; (b) increasing or reducing the width and / or height of the plantar pressure distribution cloud map by less than 15% to obtain a new plantar pressure distribution cloud map; (c) cutting the edge of the plantar pressure distribution cloud map by less than 15% to obtain a new plantar pressure distribution cloud map; Adjust the brightness of the plantar pressure distribution cloud map within the range of [0.9-1.5] to obtain a new plantar pressure distribution cloud map.

4. The method for manufacturing a personalized variable density insole based on deep learning according to claim 1, characterized in that: Step 7 includes: S701) setting a reference plane M, moving the sole plane determined by the forefoot and the heel in the foot 3D model to coincide with the reference plane M, and moving the foot 3D model downward so that the bottom of the foot 3D model partially sinks into the insole prototype; S702) The partially sunken portion of the bottom of the foot 3D model in the insole prototype is removed to form a partially sunken first insole model.

5. The method for manufacturing a personalized variable density insole based on deep learning according to claim 1, characterized in that: Step 7 also includes the step of adjusting the density of the insole by zoning: setting a reference plane M, drawing an area AP where the plantar pressure is greater than a preset value T, and an area BP where the plantar pressure is less than or equal to the preset value T according to the forefoot, arch, and heel areas on the reference plane M, satisfying ρ AP ≥ρ BP .

6. The method for manufacturing a personalized variable density insole based on deep learning according to claim 1, characterized in that: Step 7 includes the steps of generating a plantar contour adaptive pressure correction insole model. The foot model is edited and corrected using Meshmixer software to generate a personalized correction insole: S721) First, import the foot 3D model into Meshmixer; S722) using the "deformation" tool in Meshmixer to perform nonlinear deformation on the model by dragging points or faces, so as to correct the abnormal area, such as lifting the arch of the foot or correcting the angle of the heel; S723) After editing is completed, the edited foot model is saved and exported to a compatible format file.

7. The method for manufacturing a personalized variable density insole based on deep learning according to claim 1, characterized in that: The improved MobileNetV3 network deep learning model adopts fixed residual blocks and SE blocks. The point-level convolution layer follows the deconvolution component, and the point-level convolution changes the width. The output channel widening layer of the point-level convolution layer is adjusted by multiplying the total number of channels by the scaling factor. The fully connected layer, global average pooling and element-level scaling constitute the SE block. The number of channels appearing in the input feature map is used as the criterion for determining the size of the SE block.

8. The method for manufacturing a personalized variable density insole based on deep learning according to claim 7, characterized in that: The improved MobileNetV3 network deep learning model uses the h-swish activation function. h-swish draws on the h-sigmoid activation function. Formula (4) is generated. The swish activation function is replaced by h-swish, as shown in formula (5). The sigmoid activation function is replaced by the h-sigmoid activation function:

Images