A neural network-based anti-penetration performance detection method, system and device

Abstract

The application discloses a kind of based on neural network's anti-penetration performance detection method, system and device, method includes: by wrinkle experiment obtains the anti-penetration material sample set of different anti-penetration performance;Prepare data set, collect the image of the anti-penetration material sample set, obtain the anti-penetration performance thermodynamic diagram of the anti-penetration material sample set performance test, the image and its corresponding performance thermodynamic diagram jointly constitute data set;Based on the training set, neural network is trained, and neural network model is obtained;Collect the image of the anti-penetration material of performance damage in daily life, input the corresponding anti-penetration performance thermodynamic diagram of image in neural network model and obtain it.The protection performance of the anti-penetration material of wrinkle generation is detected, it is a kind of safety guarantee test to the application of anti-penetration material, with practical application value.

Description

Technical Field

[0001] This invention belongs to the fields of machine learning, computer vision and materials mechanics, and particularly relates to a method, system and device for detecting stab resistance performance based on neural networks. Background Technology

[0002] In the past, heavy metal equipment was often used for protection, but this was not only cumbersome and inconvenient, but also had some drawbacks in practical applications. Therefore, the new stab-resistant equipment currently in use, primarily made of dipped aramid fabric and foam, not only improves the flexibility and comfort of security personnel but also provides better safety assurance. The protective performance of stab-resistant materials is also an important component of bulletproof materials and protective equipment; improvements in their performance can further enhance the overall performance of these protective devices. Therefore, research on stab-resistant materials not only has practical application value but also promotes the development of the entire field of protective materials.

[0003] In some cases, stab-resistant materials may develop wrinkles during use, such as when torn or compressed, or when folded or rolled during manufacturing. These wrinkles cause the material surface to become uneven, thus reducing its protective performance.

[0004] Current research on the performance of wrinkle-resistant stab-resistant materials primarily relies on experimental methods. This involves simulating wrinkle formation and then using standardized testing methods in the laboratory to evaluate the material's performance. While this method is direct and reliable, it also has drawbacks such as high time and resource consumption and limitations in the experimental process. In contrast, machine learning methods can more quickly assess and predict the performance of wrinkle-resistant stab-resistant materials, and in some cases, provide more accurate predictions. Machine learning methods utilize large amounts of data to train models, enabling them to identify performance-related features and predict future performance.

[0005] Convolutional Neural Networks (CNNs), a special type of neural network structure, are designed to process data with a grid-like structure. They are primarily used in image recognition, image classification, and object detection. The emergence of CNNs has led to the rapid development of image processing technology and provided a powerful tool for achieving more accurate and efficient image recognition and segmentation tasks. With the continuous development of machine learning technology, the application of CNNs in the evaluation and prediction of the protective performance of wrinkle-resistant stab-resistant materials has become possible, enabling the effective implementation of image-based stab-resistant performance detection schemes. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention proposes a neural network-based method, system, and device for detecting stab resistance performance, which enables image-based stab resistance performance detection.

[0007] To achieve the above objectives, the present invention provides a method for detecting stab resistance performance based on neural networks, comprising: acquiring a sample set of stab-resistant materials with different stab resistance performances, and acquiring images of the sample set of stab-resistant materials;

[0008] The stab-resistant material sample set was tested to obtain a thermogram of the stab-resistant performance of the stab-resistant material sample set;

[0009] A dataset is constructed based on the stab-resistant material sample image and the stab-resistant performance heat map; a neural network is trained based on the dataset to obtain the trained neural network model;

[0010] The trained neural network model is used to detect damaged stab-resistant materials and obtain a heat map of the stab-resistant performance of the damaged stab-resistant materials, thereby realizing the detection of the protective performance of the stab-resistant materials.

[0011] Optionally, testing the stab-resistant material sample set and obtaining a stab-resistant performance thermogram of the stab-resistant material sample set includes:

[0012] The samples in the stab-resistant material sample set are divided to obtain several identical rectangular blocks;

[0013] Quasi-static puncture test, dynamic puncture test, hardness test and stiffness test are performed on several identical rectangular blocks respectively to obtain the quasi-static puncture performance, dynamic puncture performance, hardness performance and stiffness performance of each rectangular block;

[0014] The puncture resistance of each rectangular block is obtained based on its quasi-static puncture performance, dynamic puncture performance, hardness performance, and stiffness performance.

[0015] The stab resistance performance of each rectangular block is normalized to obtain a stab resistance performance thermogram of the stab resistance material sample.

[0016] Optionally, the method for obtaining the stab resistance N of each of the rectangular blocks includes:

[0017] N = a1*N1 + a2*N2 + a3*N3 + a4*N4

[0018] Where a1, a2, a3 and a4 are weighting coefficients, N1 is the quasi-static puncture performance, N2 is the dynamic puncture performance, N3 is the hardness performance, and N4 is the stiffness performance.

[0019] Optionally, training the neural network based on the dataset to obtain a neural network model includes:

[0020] Based on the dataset, train a generative adversarial network and obtain the trained generative adversarial network;

[0021] Using the trained generative adversarial network, first synthetic data is obtained, and the first synthetic data is added to the dataset to complete the first augmentation of the dataset;

[0022] The dataset after the first augmentation is subjected to rotation, translation, flipping, scaling, and cropping transformations to obtain the second synthetic data;

[0023] The second synthetic data is added to the dataset after the first amplification to obtain the training set;

[0024] The neural network model incorporates a deformable activation function and periodic oscillations. The neural network model is trained based on the training set to obtain the trained neural network model.

[0025] Optionally, the deformable activation function is:

[0026]

[0027] Where λ is the weighting coefficient and α is the variability parameter.

[0028] Optionally, the periodic oscillation is:

[0029]

[0030] Where, θ i， j represents the j-th weight parameter of the i-th layer of the neural network, ξ represents the regularization coefficient, T represents the training cycle length, and t represents the current training step.

[0031] Optionally, the detection of damaged stab-resistant materials using the trained neural network model includes:

[0032] The image of the damaged stab-resistant material is subjected to multi-scale convolution and pooling through a multi-scale pyramid to obtain several feature maps.

[0033] A new feature map is obtained by concatenating several feature maps using a feature cascade layer.

[0034] Feature extraction is performed on the new feature map based on the convolutional layer to obtain the features of the damaged stab-resistant material, and the stab-resistant performance heat map of the damaged stab-resistant material is output through the output layer to realize the detection of the protective performance of the stab-resistant material.

[0035] The present invention also provides a stab resistance performance detection system based on neural networks, comprising: a sample preparation module, an experimental testing module, an image acquisition module, a network training module, and an evaluation and detection module;

[0036] The sample preparation module is used to obtain a sample set of stab-resistant materials with different stab-resistant properties;

[0037] The experimental testing module is used to perform quasi-static puncture tests, dynamic puncture tests, hardness tests, and stiffness tests on the puncture-resistant material sample set, and to obtain a thermogram of the puncture resistance performance of the puncture-resistant material sample set.

[0038] The image acquisition module is used to acquire images of the stab-resistant material sample set;

[0039] The evaluation and detection module is used to train a neural network, and using the trained neural network model, to detect the damaged stab-resistant material, obtain a heat map of the stab-resistant performance of the damaged stab-resistant material, and realize the detection of the protective performance of the stab-resistant material.

[0040] The present invention also provides a stab resistance performance testing device based on neural network, comprising: a wrinkle testing machine and a machine vision device, wherein the machine vision device is vertically arranged above the wrinkle testing machine;

[0041] The wrinkle testing machine includes a first base and a sample preparation section; the base and the sample preparation section are fixedly connected.

[0042] The machine vision device includes an image acquisition unit and an image data storage and processing unit, which are wirelessly connected via a network.

[0043] The sample preparation unit includes an industrial computer, a reciprocating motor, a transmission rod, a moving plate, a fixing plate, and a sample holder;

[0044] The industrial control computer controls the magnitude and frequency of the output power of the reciprocating motor. The reciprocating motor is connected to the moving plate through the transmission rod, and the fixed plate is movably connected to the sample holder.

[0045] The image acquisition unit includes an industrial camera lens, a liftable light source control console, a light source, a stab-resistant material sample, and a second base.

[0046] A vertical support rod is fixedly installed on the second base plane. The industrial camera lens is fixedly connected to the support rod, and the liftable light source control console is movably connected to the support rod.

[0047] The light source is placed on the liftable light source control console, and the stab-resistant material sample is placed on the second base. The centers of the industrial camera lens, the liftable light source control console, the light source, and the stab-resistant material sample are aligned.

[0048] The technical effects of this invention are as follows: In the application of the neural network-based stab-resistant performance detection method, system, and device of this invention, a dataset required for a model to predict the protective performance of stab-resistant materials with wrinkles is constructed through wrinkling experiments and performance tests. The protective performance of stab-resistant materials with wrinkles is then detected through a neural network. On the one hand, detecting the protective performance of stab-resistant materials with wrinkles provides a safety assurance test for the application of stab-resistant materials and has practical application value. On the other hand, it proposes a new method for detecting stab-resistant performance based on images, effectively avoiding the problems of high experimental testing costs and limited conditions, and promoting the development of the field of stab-resistant performance evaluation and testing of stab-resistant materials. Attached Figure Description

[0049] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0050] Figure 1 A flowchart of a neural network-based stab resistance detection method provided in an embodiment of the present invention;

[0051] Figure 2 A schematic diagram of puncture-resistant material samples with different puncture-resistant properties obtained in the wrinkling experiment provided in an embodiment of the present invention;

[0052] Figure 3 A neural network structure diagram provided in an embodiment of the present invention;

[0053] Figure 4 A flowchart illustrating the training process of a neural network provided in an embodiment of the present invention;

[0054] Figure 5 This is a schematic diagram of the neural network performance detection principle provided in an embodiment of this application;

[0055] Figure 6 A schematic diagram of a neural network-based stab resistance detection system provided in an embodiment of this application;

[0056] Figure 7 This is a schematic diagram of the wrinkle testing machine device provided in the embodiment of this application, wherein 701 is the first base, 702 is the industrial control computer, 703 is the reciprocating motor, 704 is the transmission rod, 705 is the moving plate, 706 is the fixed plate, and 707 is the sample holder.

[0057] Figure 8This is a schematic diagram of a machine vision device provided in an embodiment of this application, wherein 801 is an industrial camera lens, 802 is a height-adjustable light source control console, 803 is a light source, 804 is a stab-resistant material sample, 805 is a second base, and 806 is an image storage and processor. Detailed Implementation

[0058] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0059] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0060] Example 1

[0061] like Figure 1 As shown, this embodiment provides a method for detecting stab resistance performance based on neural networks, including:

[0062] Prepare a sample set and obtain a sample set of stab-resistant materials with different stab-resistant properties through a folding experiment;

[0063] Prepare the dataset by collecting images of stab-resistant material sample sets and performing performance tests on the stab-resistant material sample sets to obtain stab-resistant performance heatmaps. The images and their corresponding performance heatmaps together constitute the dataset.

[0064] Train the neural network by training the neural network on the training set to obtain the neural network model;

[0065] Performance testing involves collecting images of stab-resistant materials whose performance has been impaired in daily life, and then inputting these images into a neural network model to obtain a corresponding heat map of stab-resistant performance.

[0066] like Figure 2 As shown, the wrinkling experiment involves subjecting stab-resistant material samples to wrinkling treatments of grade 0, grade 1, ... 100 to obtain stab-resistant material sample sets with different stab-resistant properties.

[0067] The steps for performance testing are as follows:

[0068] The samples in the stab-resistant material sample set are divided into rectangular blocks of equal area of ​​M×M (M is a constant). Specifically, a 1024x1024 pixel sample set image is acquired by a machine vision device and divided into 128x128 pixel samples as the dataset.

[0069] For each rectangular block, quasi-static puncture test, dynamic puncture test, hardness test and stiffness test were performed to obtain quasi-static puncture performance N1, dynamic puncture performance N2, hardness performance N3 and stiffness performance N4.

[0070] The puncture resistance N of each rectangular block is calculated, where N is defined as...

[0071] N = a1*N1 + a2*N2 + a3*N3 + a4*N4

[0072] Where a1, a2, a3, and a4 are weighting coefficients;

[0073] The stab resistance performance of each rectangular block is normalized and then mapped to a grayscale space of 0 to 255. Finally, the blocks are combined to obtain an N×N stab resistance performance heatmap.

[0074] Quasi-static puncture test, dynamic puncture test, hardness test, and stiffness test are testing methods for the quasi-static puncture force, dynamic puncture force, hardness, and stiffness performance of wrinkle-resistant puncture-proof materials. The quasi-static and dynamic puncture tests are conducted according to the GA68-2019 industry testing standard for police stab-proof vests. Five samples from each type are selected for the puncture test. The first puncture is placed at the geometric center, the second and third punctures are placed within a radius of 50-75 mm from the point of impact of the first puncture, and the fourth and fifth punctures are placed at relatively weaker locations. The hardness test is conducted according to the national standard GB / T 531-1999, "Indentation Hardness Test Method for Rubber Pocket Hardness Tester." For the aramid fabric, an LX-D type hardness tester is used, and five samples from each type are selected for hardness measurement. The stiffness test is conducted according to the national standard GB / T... The inclined plane method in the 18318.12009 method for determining the bending properties of textiles uses the inclined plane cantilever beam principle to test the bending stiffness of the specimen. Each type of sample is randomly cut into rectangular specimens 25 mm wide and placed on a horizontal platform. The specimen is advanced along the long axis of the platform, extending beyond the platform and bending under its own weight. When the extended end of the specimen passes the leading edge of the platform and reaches an angle of 41.5° with the horizontal line, the process stops and the elongation length L is recorded. The elongation length is generally the bending length C. The stiffness is calculated using the formula G = mC. 3 x10 -3 m is the areal density of the aramid fabric, and the unit of m is g / cm³. 2 .

[0075] The specific steps to obtain the heat map of stab resistance performance are as follows:

[0076] The Rectangular Pulse Distributed Interpolation (RPDI) method maps performance data obtained from quasi-static puncture tests, dynamic puncture tests, hardness tests, and stiffness tests onto a 128x128 pixel experimental sample image. Pixel values ​​are then taken horizontally from the experimental sample image, forming a discontinuous rectangular pulse-like pixel value distribution map along each horizontal line, representing the distribution of crease-bearing and crease-free locations. Based on the rectangular pulse-like pixel values, the tested performance data values ​​are assigned, with each segment of continuously distributed pulse signal-like region corresponding to the same experimental performance data value. This interpolation method can obtain a distribution matrix of four types of experimental test data values ​​within a 128x128 pixel experimental sample.

[0077] The normalized distance method calculates a 128x128 data distribution heatmap by analyzing the distribution matrices of the four experimental test data values. First, an empty 128x128 matrix is ​​constructed. Then, each row and column of the experimental test data matrix is ​​normalized, using the row number, column number, and the data value as labels. The rows and columns of the original data are mapped to 128x128 pixel images of samples with different stab-resistant properties. The distance from each pixel to the data point is calculated, and the values ​​of all distances to the data point are accumulated to obtain the value of a single pixel. After calculating the distance to all pixels in the 128x128 pixel image, mapping them to grayscale values ​​yields a 128x128 pixel heatmap.

[0078] like Figure 4 As shown, the training process of the neural network is as follows:

[0079] The generator and discriminator in the generative adversarial network are trained using real data from the dataset. The discriminator is used to distinguish whether the input data is real or synthetic data, and the generator is used to generate synthetic data that is similar to real data.

[0080] After the generative adversarial network is trained, some synthetic data is generated using the trained generative adversarial network model and added to the dataset to expand the size of the dataset, resulting in the first expanded training set.

[0081] The data in the training set after the first amplification is rotated, translated, flipped, scaled, and cropped to generate some synthetic data again, which is then added to the training set after the first amplification to obtain the training set after the second amplification.

[0082] The data in the training set after the secondary amplification are subjected to image filtering preprocessing to eliminate noise and redundant information in the data;

[0083] In each layer of the neural network, a deformable activation function is introduced, which is designed as follows:

[0084]

[0085] Where λ is the weighting coefficient and α is the variability parameter;

[0086] Traditional activation functions (such as sigmoid and ReLU) remain unchanged throughout the training process, which often leads to slow model convergence and a high risk of overfitting. Deformable activation functions, on the other hand, can adaptively adjust based on the distribution and characteristics of the data, allowing the model to reach a lower loss value more quickly with the same number of iterations, thus reducing the risk of overfitting.

[0087] A periodic regularization term is added to the loss function of the neural network to introduce periodic oscillations during the training process. The design of the periodic regularization term is as follows:

[0088]

[0089] Where, θ i， j represents the j-th weight parameter of the i-th layer of the neural network, ξ represents the regularization coefficient, T represents the training cycle length, and t represents the current training step.

[0090] Introducing a periodic regularization term can make the model less sensitive to small changes in the input data, increasing its robustness and enabling it to maintain stable prediction results better when faced with noise or interference. The periodic regularization term can also provide a regularization mechanism to help the model resist the risk of overfitting.

[0091] The statistical properties of the data in the training set after secondary amplification are used to initialize the weights of the neural network;

[0092] The training process of the neural network is parallelized by using a distributed computing framework to divide the training data into multiple batches for processing and perform parallel computation on multiple computing nodes.

[0093] like Figure 3 As shown, the neural network adopts a parallel pyramid structure, which performs multi-scale convolution and pooling operations on the input image, then concatenates the feature maps of multiple pyramids together, and then performs feature extraction and classification through a series of convolution and fully connected layers.

[0094] The input image is:

[0095] X∈R W×H×C

[0096] Where W and H represent the width and height of the image, respectively, and C represents the number of channels in the image. The neural network structure is as follows:

[0097] Multi-scale pyramids perform multi-scale convolution and pooling operations on the input image to obtain feature maps of multiple pyramids. The feature map of the i-th pyramid is...

[0098]

[0099] Among them W i and H i C represents the width and height of the pyramid, respectively. i This represents the number of passages in the pyramid. Here, three pyramids are used, i = 1, 2, 3.

[0100] The feature cascade layer concatenates the feature maps from each pyramid to obtain a new feature map.

[0101]

[0102] Among them W c =W1+W2+W3,H c =H1=H2=H3, C c =C1+C2+C3.

[0103] Following the feature cascade layers, a series of convolutional layers are used for feature extraction. Assume the input to the k-th convolutional layer is F. k-1 The output is F k Then it can be expressed as:

[0104] F k =σ(W k *F k-1 +b k )

[0105] Among them, W k Let b represent the weight matrix of the k-th convolutional layer. k σ represents the bias vector, σ represents the activation function, and * represents the convolution operation.

[0106] The output layer, the last layer is a convolutional layer, and the output is F. conv The heatmap representing the stab resistance performance of the input image can be expressed as:

[0107]

[0108] Among them, W conv and H conv C represents the width and height of the last convolutional layer, respectively. conv This represents the number of channels in the last convolutional layer, where C is... conv The value is 1.

[0109] The neural network uses 128x128 pixel sample images with different stab resistance properties as sample units, and a 128x128 pixel performance distribution heatmap as label units to segment and train on 1024x1024 pixel sample images. A 1024x1024 pixel sample image can be segmented into 64 image information units distributed in an 8x8 pattern. Each image information unit can be considered a training unit, allowing simultaneous training of 64 image information units on a single sample image based on their data labels. The neural network is used to train on 64 images for each sample image. The trained neural network extracts features using convolutional kernels of different scales and then reconstructs the data, generating a corresponding 1024x1024 pixel performance distribution heatmap. This enables the detection of the stab resistance performance of stab-resistant material samples with different levels of wrinkle treatment. The overall performance detection principle is as follows: Figure 5 As shown

[0110] Example 2

[0111] like Figure 6 As shown, this embodiment provides a stab resistance performance detection system based on neural networks. The system mainly includes: a sample preparation module, an experimental testing module, an image acquisition module, and an evaluation and detection module.

[0112] The sample preparation module is used to obtain samples with different puncture resistance properties;

[0113] The experimental testing module is used to perform four types of experimental tests: quasi-static puncture test, dynamic puncture test, hardness test, and stiffness test, and to obtain experimental test performance data.

[0114] The image acquisition module is used to acquire images of stab-resistant material sample sets with different stab-resistant properties prepared in the sample preparation module;

[0115] The evaluation and detection module is used to train a neural network model. By inputting images of stab-resistant materials whose performance is impaired in daily life into the trained neural network model, the corresponding stab-resistant performance heat map can be obtained.

[0116] The evaluation and detection module includes: a dataset construction submodule, a network training submodule, a storage submodule, and a detection submodule.

[0117] The dataset construction submodule uses 128x128 pixel sample images of different stab resistance performances acquired by the image acquisition module as samples, and obtains 128x128 pixel performance distribution heatmaps as labels using rectangular pulse distributed interpolation and normalized distance methods. The samples and labels constitute the training dataset.

[0118] The network training submodule is used to train neural networks based on a dataset;

[0119] The detection submodule is used by the neural network model to detect the protective performance of stab-resistant material samples with different levels of wrinkling treatment, which are 1024x1024 pixels in size.

[0120] The storage submodule is used to store experimental data from the experimental testing module and image information from the image acquisition module.

[0121] Example 3

[0122] This embodiment provides a puncture resistance testing device based on neural networks, including: a wrinkle testing machine and a machine vision device.

[0123] like Figure 7 As shown, a wrinkle testing machine is used to prepare wrinkle-resistant stab-proof material samples. The stab-proof material sample is fixed between a fixed plate 706 and a movable plate 705 using a sample holder 707. The fixed plate 706 and movable plate 705 are steel plates 35 cm long, 35 cm wide, and 0.4 cm thick. The force and frequency of the reciprocating motor 703 are set by an industrial control computer 702, and pressure is applied to the sample clamped between the fixed plate 706 and movable plate 705 via a transmission rod 704, thus generating wrinkles in the stab-proof material sample. The sample holder 707 is 180° rotatable, allowing adjustment of its angle to change the shape of the wrinkles. The fixed plate 706, reciprocating motor 703, and industrial control motor are all fixedly connected to the first base 701.

[0124] like Figure 8 The diagram shows a machine vision device, including an industrial camera lens 801, a height-adjustable light source control console 802, a light source 803, a second base 805, and an image storage and processor 806. After placing a sample on the second base 805, the brightness of the light source 803 and the focal length of the industrial camera lens 801 are adjusted so that the image of the stab-resistant material sample 804 can be clearly captured and transmitted to the image storage and processor 806 via a network cable. The height-adjustable light source control console 802 is used to provide high-brightness parallel light before acquiring the image, and the distance of the light source is adjusted according to the sample size.

[0125] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for detecting stab resistance performance based on neural networks, characterized in that, include: Obtain a sample set of stab-resistant materials with different stab-resistant properties, and acquire images of the stab-resistant material sample set; The images of the stab-resistant material sample set are tested to obtain a stab-resistant performance heat map corresponding to the images of the stab-resistant material sample set, including: The image of the stab-resistant material sample set is segmented to obtain several identical rectangular blocks; Quasi-static puncture test, dynamic puncture test, hardness test and stiffness test are performed on several identical rectangular blocks respectively to obtain the quasi-static puncture performance, dynamic puncture performance, hardness performance and stiffness performance of each rectangular block; The puncture resistance of each rectangular block is obtained based on its quasi-static puncture performance, dynamic puncture performance, hardness performance, and stiffness performance. The stab resistance performance of each rectangular block is normalized to obtain a stab resistance performance heat map corresponding to the stab resistance material sample set image. A dataset is constructed based on the images of the stab-resistant material sample set and the heat map of the stab-resistant performance; Training a neural network model based on the dataset to obtain the trained neural network model includes: Based on the dataset, train a generative adversarial network and obtain the trained generative adversarial network; Using the trained generative adversarial network, first synthetic data is obtained, and the first synthetic data is added to the dataset to complete the first augmentation of the dataset; The dataset after the first augmentation is subjected to rotation, translation, flipping, scaling, and cropping transformations to obtain the second synthetic data; The second synthetic data is added to the dataset after the first amplification to obtain the training set; The neural network model incorporates a deformable activation function and periodic oscillations. The neural network model is trained based on the training set to obtain a trained neural network model. Using the trained neural network model, the damaged stab-resistant material is detected, and a heat map of the stab-resistant performance of the damaged stab-resistant material is obtained, thereby realizing the detection of the protective performance of the stab-resistant material; The detection of damaged stab-resistant materials using the trained neural network model includes: The image of the damaged stab-resistant material is subjected to multi-scale convolution and pooling through a multi-scale pyramid to obtain several feature maps. A new feature map is obtained by concatenating several feature maps using a feature cascade layer. Feature extraction is performed on the new feature map based on the convolutional layer to obtain the features of the damaged stab-resistant material, and the stab-resistant performance heat map of the damaged stab-resistant material is output through the output layer to realize the detection of the protective performance of the stab-resistant material.

2. The stab resistance performance detection method based on neural networks as described in claim 1, characterized in that, The method for obtaining the stab resistance N of each of the rectangular blocks includes: N = a1*N1 + a2*N2 + a3*N3 + a4*N4 Where a1, a2, a3 and a4 are weighting coefficients, N1 is the quasi-static puncture performance, N2 is the dynamic puncture performance, N3 is the hardness performance, and N4 is the stiffness performance.

3. The stab resistance performance detection method based on neural networks as described in claim 1, characterized in that, The deformable activation function is: in, These are the weighting coefficients. This is a variable parameter.

4. The stab resistance detection method based on neural networks as described in claim 1, characterized in that, The periodic oscillation is: in, This represents the j-th weight parameter of the i-th layer of the neural network. represents the regularization coefficient, T represents the training period length, and t represents the current training step number.

5. A stab resistance performance detection system based on neural networks, characterized in that, include: Sample preparation module, experimental testing module, image acquisition module, and evaluation and detection module; The sample preparation module is used to obtain a sample set of stab-resistant materials with different stab-resistant properties; The image acquisition module is used to acquire images of the stab-resistant material sample set; The experimental testing module is used to perform quasi-static puncture tests, dynamic puncture tests, hardness tests, and stiffness tests on the images of the puncture-resistant material sample set, and to obtain the puncture resistance performance thermogram corresponding to the images of the puncture-resistant material sample set. Obtaining the stab resistance performance heatmap corresponding to the stab-resistant material sample set images includes: The image of the stab-resistant material sample set is segmented to obtain several identical rectangular blocks; Quasi-static puncture test, dynamic puncture test, hardness test and stiffness test are performed on several identical rectangular blocks respectively to obtain the quasi-static puncture performance, dynamic puncture performance, hardness performance and stiffness performance of each rectangular block; The puncture resistance of each rectangular block is obtained based on its quasi-static puncture performance, dynamic puncture performance, hardness performance, and stiffness performance. The stab resistance performance of each rectangular block is normalized to obtain a stab resistance performance heat map corresponding to the stab resistance material sample set image. The evaluation and detection module is used to construct a dataset based on the images of the stab-resistant material sample set and the stab-resistant performance heat map; train a neural network based on the dataset; use the trained neural network model to detect damaged stab-resistant materials; obtain the stab-resistant performance heat map of the damaged stab-resistant materials; and realize the detection of the protective performance of the stab-resistant materials. Obtaining the trained neural network model includes: Based on the dataset, train a generative adversarial network and obtain the trained generative adversarial network; Using the trained generative adversarial network, first synthetic data is obtained, and the first synthetic data is added to the dataset to complete the first augmentation of the dataset; The dataset after the first augmentation is subjected to rotation, translation, flipping, scaling, and cropping transformations to obtain the second synthetic data; The second synthetic data is added to the dataset after the first amplification to obtain the training set; The neural network model incorporates a deformable activation function and periodic oscillations. The neural network model is trained based on the training set to obtain the trained neural network model.

6. A stab resistance performance detection device based on neural networks, characterized in that, include: A wrinkle testing machine and a machine vision device, wherein the machine vision device is vertically positioned above the wrinkle testing machine; The wrinkle testing machine includes a first base (701) and a sample preparation section; the first base (701) and the sample preparation section are fixedly connected, and the wrinkle testing machine is used to obtain a sample set of stab-resistant materials with different stab-resistant properties; The machine vision device is used to implement the neural network-based stab resistance detection method according to any one of claims 1-4, and includes an image acquisition unit and an image data storage and processing unit, wherein the image acquisition unit and the image data storage and processing unit are wirelessly connected via a network.

7. The stab resistance detection device based on a neural network as described in claim 6, characterized in that, The sample preparation unit includes an industrial computer (702), a reciprocating motor (703), a transmission rod (704), a moving plate (705), a fixed plate (706), and a sample holder (707). The industrial computer (702) controls the magnitude and frequency of the output power of the reciprocating motor (703). The reciprocating motor (703) is connected to the moving plate (705) through the transmission rod (704). The fixed plate (706) and the sample holder (707) are movably connected. The image acquisition unit includes an industrial camera lens (801), a height-adjustable light source control console (802), a light source (803), a stab-resistant material sample (804), and a second base (805); The second base (805) is fixedly provided with a vertical support rod, the industrial camera lens (801) is fixedly connected to the support rod, and the liftable light source control console (802) is movably connected to the support rod; The light source (803) is placed on the liftable light source control console (802), and the stab-resistant material sample (804) is placed on the second base (805). The centers of the industrial camera lens (801), the liftable light source control console (802), the light source (803), and the stab-resistant material sample (804) are aligned.

Images