A method and device for tracing perpetrators based on deep learning

Abstract

The present invention proposes a method and device for tracing an assailant based on deep learning, which is characterized by comprising: collecting impact damage images in a real knife stabbing scene; predicting the impact damage images based on a two-stage impact damage prediction model to obtain key puncture parameters; wherein the two-stage impact damage prediction model is obtained through data set training, and the data set includes: key puncture parameters of the sample at different initial speeds and impact damage images of the sample under different dynamic puncture test conditions; the two-stage impact damage prediction model includes: a first-stage TraceNet network and a second-stage shallow CNN network; the key puncture parameters are input into a classifier to trace the assailant. The present invention can realize the rapid and accurate detection of key parameters in the puncture process based on puncture damage, and trace the behavioral characteristics of the assailant based on the key parameters, providing a new means for hunting down the assailant in military activities such as counter-reconnaissance.

Description

Technical Field

[0001] The present invention belongs to the field of machine vision and artificial intelligence technology, and in particular, relates to a method and device for tracing an assailant based on deep learning. Background Art

[0002] In real life, police usually rely on fingerprints, footprints and other information to solve cases, but in many murder cases, it is often difficult to extract effective information about the perpetrators, such as fingerprints and footprints. And due to gun control, cold weapons such as stabbings have become the main weapons of fatal injuries, so obtaining information based on stabbing injuries has become a breakthrough.

[0003] Although a large number of studies have focused on injuries caused by low-speed impacts such as stabbing, they mainly rely on theoretical, experimental and numerical simulation studies. Traditional methods of studying impact injuries are limited to analyzing the degree of injury and exploring the mechanism of injury, but these methods are highly limited and the information obtained about the perpetrator is very limited.

[0004] Machine learning has made breakthrough achievements in recent years. It has also been successfully applied in the field of impact damage. Through machine learning methods, more relevant physical quantities can be discovered, such as the relationship between different impact conditions and impact damage characteristics. More information on images can also be obtained, such as the relationship between internal damage and surface damage. Combining machine learning methods with impact damage images, the impact damage suffered by stab-proof materials can be quantitatively restored. It is also necessary to establish a database of perpetrators and trace back the characteristics of the perpetrators. Summary of the invention

[0005] The purpose of the present invention is to propose a method and device for tracing perpetrators based on deep learning, combining machine learning methods and impact damage images to quantitatively restore the impact damage suffered by stab-proof materials. A perpetrator database is established to trace the characteristics of the perpetrators, which can play a huge role in solving cases and tracing perpetrators.

[0006] To achieve the above objectives, the present invention provides a method for tracing an offender based on deep learning, comprising:

[0007] Collect impact damage images in real knife stabbing scenarios;

[0008] Based on the two-stage impact damage prediction model, the impact damage image is predicted to obtain the key puncture parameters; wherein the two-stage impact damage prediction model is obtained by training a data set, and the data set includes: the key puncture parameters of the sample at different initial speeds of the tool and the impact damage image of the sample under different dynamic puncture test conditions; the two-stage impact damage prediction model includes: a first-stage TraceNet network and a second-stage shallow CNN network;

[0009] The key puncture parameters are input into a classifier to perform perpetrator tracing.

[0010] Optionally, the key puncture parameters include: the initial speed of the cutter when contacting the sample in the dynamic puncture test, the initial kinetic energy of the cutter, the peak puncture force, and the maximum number of penetration layers for the sample.

[0011] Optionally, the first TraceNet network is used to extract damage features in the impact damage image;

[0012] The first segment of the TraceNet network includes: a feature extraction module and a Traceflow module;

[0013] The feature extraction module is used to extract initial features of a preset dimension in the impact damage image;

[0014] The Traceflow module is used to recombine the initial features to obtain the damage features.

[0015] Optionally, the feature extraction module includes: a convolution layer, a maximum pooling layer, a convolution block and a marking block connected in sequence;

[0016] The image is input into the feature extraction module, first passes through the convolution layer, the output of the convolution layer passes through the ReLU activation function and then input into the maximum pooling layer, the output of the maximum pooling layer is used as the input of the convolution block, and the features extracted by the convolution block are input into the identification block to extract features again.

[0017] The convolution layer is used to perform a convolution operation on the input image;

[0018] The maximum pooling layer is used to downsample the output of the convolutional layer while retaining preset image features;

[0019] The identification block is used to learn the identity mapping and directly transfer the input information to the output;

[0020] The convolution block introduces a convolution layer with a preset convolution kernel of 1x1 as a residual connection in the feature extraction module.

[0021] Optionally, the Traceflow module includes: a plurality of flow layers of preset sizes, the flow layers include: a flow layer of a 3x3 convolution kernel and a flow layer of a 1x1 convolution kernel, each flow layer of a 3x3 convolution kernel and each flow layer of a 1x1 convolution kernel are used as a flow module, and the Traceflow module is provided with a total of 8 flow submodules; the input of the first flow submodule is the output of the feature extraction module, the output of the first flow submodule is used as the input of the second flow submodule, the output of the second flow submodule is used as the input of the third flow submodule, and so on; wherein, an activation normalization layer and an affine coupling layer are used in each flow layer;

[0022] The activation normalization layer is used to independently normalize each channel of the input information of each flow layer;

[0023] The affine coupling layer uses a reversible neural network structure to ensure that the training results of each flow layer do not deviate from the original features while being able to extract more complex features.

[0024] Optionally, the second shallow CNN network is used to characterize the relationship between the injury characteristics and the key puncture parameters;

[0025] The second shallow CNN network includes: a plurality of two-dimensional convolutional layers, a plurality of preset maximum pooling layers, and two preset fully connected layers;

[0026] The two-dimensional convolutional layers are used to extract features of the input image and generate feature maps in several steps;

[0027] The plurality of preset maximum pooling layers are used to continuously reduce the spatial size of the feature map;

[0028] The first fully connected layer is used to learn the nonlinear relationship between features, and the second fully connected layer is used to output the prediction results of the second shallow CNN network on the key puncture parameters.

[0029] Optionally, training the two-stage impact damage prediction model using a data set includes:

[0030] Setting a minimization loss function for the first segment of the TraceNet network; wherein the minimization loss function is minimizing negative log-likelihood;

[0031] Based on a preset epoch, the first segment of the TraceNet network is trained using the minimized loss function;

[0032] The second shallow CNN network is trained based on a preset training cycle.

[0033] Optionally, the minimization loss function is:

[0034]

[0035] Among them, L is the loss value, N is the number of data points, i is the sequence number of the data point, and p X (x i ) is the data point x i The probability density in the original space, x i is the i-th data point, df is the value of the data point x i The mapping transformation, is the transformation f with respect to x i The Jacobian matrix, p Z (f(x i )) is f(x) in the latent space i ), f(x i ) is x i Mapping in latent space.

[0036] Optionally, inputting the puncture key parameters into a classifier to perform perpetrator tracing includes: training the classifier;

[0037] Training the classifier includes:

[0038] A real experimental data set and a model prediction data set are constructed respectively; wherein the real experimental data set includes the stabbing distance, the initial speed of the tool, and the number of penetration layers, and the participant number is used as a label; and the model prediction data set includes the initial kinetic energy of the tool, the peak puncture force, and the number of penetration layers, and the participant number is used as a label;

[0039] Training the classifier using the real experimental data set and the model prediction data set;

[0040] The trained classifier is used to classify the characteristics of different personnel, and an unpaired T test is performed on the evaluation indicators obtained from the three-month real experimental data set and the model prediction data set.

[0041] To achieve the above-mentioned purpose, the present invention also provides a perpetrator tracing device based on deep learning, the device comprising: a data set construction module, an impact damage prediction module,

[0042] The data set construction module is used to construct a data set according to the key puncture parameters of the sample at different initial speeds and the impact damage images of the sample under different dynamic puncture test conditions; wherein the key puncture parameters include: the initial speed of the tool when it contacts the sample in the dynamic puncture test, the initial kinetic energy of the tool, the puncture peak force, and the maximum number of penetration layers for the sample.

[0043] An impact damage prediction module, used to use the data set to train the two-stage impact damage prediction model to predict key puncture parameters;

[0044] A real scene prediction module is used to predict key puncture parameters using a two-stage impact damage prediction model for impact damage images generated in real scenes;

[0045] The classifier training module classifies the perpetrators according to the key puncture parameters predicted in real scenarios and traces the characteristics of the perpetrators;

[0046] An operation processor module is used to operate the two-stage impact injury prediction model and the classifier in the assailant tracing method;

[0047] The memory module is used to store the data set, the training results and prediction results of the two-stage impact damage prediction model, and the training results and prediction results of the classifier in the device.

[0048] The present invention has the following beneficial effects:

[0049] The present invention obtains the key puncture parameters of the sample and the impact damage images of the tool at different initial speeds through dynamic puncture testing and image acquisition, and then constructs a data set corresponding to the image-puncture key parameters; a two-stage impact damage prediction model is trained based on the data set, and the model realizes the ability to predict the key puncture parameters based on the impact damage image. The first stage TraceNet realizes the segmentation of the impact damage image under unsupervised learning, and the second stage shallow CNN realizes the prediction of the key puncture parameters; the impact damage images of multiple experimenters in real knife stabbing scenes are collected, and the two-stage impact damage prediction model is used to predict the key puncture parameters to obtain the perpetrator database; the classifier is trained to realize the function of tracing the perpetrator based on the key puncture parameters.

[0050] The present invention can realize rapid and accurate detection of key parameters in the puncture process according to the puncture damage, and trace the behavioral characteristics of the attacker based on the key parameters, providing a new means for hunting down the attacker in military actions such as counter-reconnaissance. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0052] Figure 1 It is a flowchart of a method for tracing an assailant based on deep learning according to an embodiment of the present invention;

[0053] Figure 2It is a schematic diagram of the structure of a two-stage impact damage prediction model according to an embodiment of the present invention;

[0054] Figure 3 It is a schematic diagram of the structure of a perpetrator tracing device based on deep learning according to an embodiment of the present invention;

[0055] Figure 4 This is an example diagram of an image obtained by the image acquisition system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0056] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0057] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0058] like Figure 1 As shown, this embodiment proposes a method for tracing an assailant based on deep learning, including:

[0059] Collect impact damage images in real knife stabbing scenarios;

[0060] Based on the two-stage impact damage prediction model, the impact damage image is predicted to obtain the key puncture parameters; the two-stage impact damage prediction model is obtained through data set training, and the data set includes: the key puncture parameters of the sample at different initial speeds and the impact damage images of the sample under different dynamic puncture test conditions; the two-stage impact damage prediction model includes: the first stage TraceNet network and the second stage shallow CNN network; the first stage TraceNet realizes the segmentation of the impact damage image under unsupervised learning, and the second stage shallow CNN realizes the prediction of the key puncture parameters;

[0061] The key puncture parameters are input into the classifier to trace the perpetrator.

[0062] Furthermore, the key puncture parameters include: the initial speed of the cutter when it contacts the sample in the dynamic puncture test, the initial kinetic energy of the cutter, the peak puncture force, and the maximum number of penetration layers for the sample.

[0063] In this embodiment, constructing a data set corresponding to image-puncture key parameters includes:

[0064] A standard tool is fixed on the impactor with an overall weight of 2 kg. The impactor carries the tool and can slide vertically on a track within a height range of 210 cm and the sliding resistance is negligible. The sample is fixed directly below the impactor. During a single experiment, the impactor carries the tool and falls freely from a certain height. When the tool stands upright on the sample stably, the single experiment ends. Set 8 different drop heights, from 45 cm to 185 cm and at intervals of 20 cm. In addition, a photoelectric gate is placed at the height where the tool is about to contact the sample to record the initial velocity of the tool when it pierces the sample. The initial kinetic energy of the tool is obtained according to the kinetic energy calculation formula:

[0065]

[0066] Where m is the total mass of the tool and impactor, 2 kg, and v is the initial velocity of the tool when it penetrates the sample.

[0067] In order to collect physical quantities in the dynamic puncture test, a data acquisition system was established. The force sensor placed between the bottom of the tool and the impactor collects the impact force during the tool puncture process. A 10μm thick aluminum film is laid on the 1st, 20th, 30th, 40th, and 50th layers of 50-layer PE. Each layer of aluminum film is connected to the data acquisition system through a wire (range 10V, acquisition frequency 2MHz). When the tool penetrates a layer of aluminum film, the closed loop formed will enable the data acquisition system to capture an electrical signal. The position of the tool at different times can be determined according to the time when the signal is generated, so the number of layers penetrated by the tool can be determined. In order to reduce errors and avoid the contingency of the experiment, the experiment was repeated 6 times to obtain reliable data. The average of the initial speed of the tool collected from the 6 sets of experimental data was calculated to obtain the average speed at 8 heights, and the initial impact kinetic energy of the tool at different heights was obtained according to the kinetic energy calculation formula. At the same time, the peak force and the number of layers penetrated during the puncture process were also collected in the experiment.

[0068] After the dynamic puncture test was completed, 6 sets of images at 8 heights were collected, and the images were enlarged 5 times through basic image processing methods such as random cropping and rotation. Finally, 30 images of each drop height were obtained, a total of 240 impact damage images. Correspondingly, 240 non-damaged images were also collected as part of the data set. In order to take into account the training efficiency and accuracy of the network model, the resolution of all images was set to 512x512.

[0069] Furthermore, the first segment of the TraceNet network is used to extract damage features from the impact damage image;

[0070] The first section of the TraceNet network includes: feature extraction module and Traceflow module;

[0071] A feature extraction module, used to extract initial features of preset dimensions in the impact damage image;

[0072] Traceflow module is used to recombine the initial features to obtain damage features.

[0073] The feature extraction module includes: a convolutional layer, a maximum pooling layer, a convolutional block and a marking block connected in sequence;

[0074] Convolutional layer, used to perform convolution operations on the input image;

[0075] The max pooling layer is used to downsample the output of the convolutional layer while retaining the preset image features;

[0076] The identification block is used to learn the identity mapping and directly pass the input information to the output;

[0077] The preset convolutional layer is introduced in the convolutional block as the residual connection in the feature extraction module.

[0078] The Traceflow module includes: several convolutional layers of preset sizes, activation normalization layers, and affine coupling layers, where each convolutional layer takes the initial features as input;

[0079] Activate the normalization layer to normalize each channel independently;

[0080] Affine coupling layers for performing reversible changes.

[0081] Furthermore, the second shallow CNN network is used to characterize the relationship between injury characteristics and key puncture parameters;

[0082] The second shallow CNN network includes: several two-dimensional convolutional layers, several preset maximum pooling layers, and two preset fully connected layers;

[0083] Several two-dimensional convolutional layers are used to extract features of the input image and generate feature maps in several steps;

[0084] Several preset max pooling layers are used to continuously reduce the spatial size of feature maps;

[0085] The first fully connected layer is used to learn the nonlinear relationship between features, and the second fully connected layer is used to output the prediction results of the second shallow CNN network on the key puncture parameters.

[0086] Specifically, the structure of the two-stage impact damage prediction model of this embodiment is as follows: First, in the first stage, an unsupervised TraceNet model is independently built. Accurate segmentation and feature extraction of the damaged area in the impact damage image generated by different tool drop heights are achieved. Then, in the second stage, a shallow CNN network is trained to characterize the relationship between the TraceNet segmentation results and the experimentally measured key damage parameters. Finally, the two-stage model as a whole realizes the function of accurately predicting the key damage parameters of any untrained impact damage image. The schematic diagram of the two-stage model is shown in the figure. Figure 2 shown.

[0087] The feature extraction module in TraceNet consists of convolutional layers, convolutional blocks, and identification blocks. The first layer is the convolutional layer, which performs convolution operations on the input image. Then comes the maximum pooling layer, which downsamples the output of the convolutional layer while retaining the most important image features. The output of the maximum pooling layer then passes through a series of convolutional blocks and identification blocks. Convolutional blocks and identification blocks are key components. The main purpose of the identification block is to learn the identity mapping and pass the input information directly to the output. Compared with the identification block, the convolutional block introduces an additional convolutional layer as a residual connection. The core idea of ​​the residual module is to introduce a skip connection so that the output of one layer of the network not only depends on its immediate previous layer, but also can skip one or more layers. This residual structure can help the network better capture the complex features in the image, thereby improving the performance of image recognition and segmentation. The main structure of the Traceflow module in TraceNet is 8 flow modules, and the flow layer of each 3x3 convolution kernel and the flow layer of each 1x1 convolution kernel are used as a flow module. The input of the first flow module is the HxWxC (Height x Weight x Channel) dimension feature extracted by the feature extractor. The output of the first flow module is used as the input of the second flow module, and so on. In each flow layer, an activation normalization layer is used to independently normalize each channel of the input information of each flow layer; an affine coupling layer is used to ensure that the training results of each flow layer do not deviate from the original features while extracting more complex features. The role of segmentation is to split the input features into two parts, which can increase the flexibility of the model. Because one part of the features can be directly passed to the output, while the other part of the features need to be transformed. This allows Traceflow to gradually increase its attention to the details in the data. The parameters s1 and s2 in the affine coupling layer are scale parameters used to control the degree of feature scaling. b1 and b2 are offset parameters used to control the amount of feature translation. These parameters are learned in Traceflow. The exponential function exp is used to ensure that the scale parameters s1 and s2 are both positive values, so as to ensure the reversibility of the transformation. Finally, the transformed features and the untransformed features are recombined through the merging layer. This reorganization ensures that the model can integrate all information. Therefore, each step of Traceflow output will obtain the fused HxWxC features for the next operation.

[0088] The shallow CNN network is mainly composed of 3 two-dimensional convolutional layers, 3 maximum pooling layers, and 2 fully connected layers. The 3 convolutional layers extract the features of the input image in three steps and generate feature maps. The 3 maximum pooling layers reduce the computational cost and extract key features by continuously reducing the spatial size of the feature map. The first fully connected layer learns the nonlinear relationship between features by flattening the feature map into a vector and connecting it with the fully connected layer. The last fully connected layer finally outputs the prediction results of the shallow CNN model for the key damage parameters.

[0089] Furthermore, the two-stage impact damage prediction model is trained through the data set including:

[0090] Set the minimization loss function of the first segment of the TraceNet network; the minimization loss function is to minimize the negative log-likelihood;

[0091] Based on the preset epoch, the first segment of the TraceNet network is trained by minimizing the loss function;

[0092] The second shallow CNN network is trained based on the preset training cycle.

[0093] Specifically, in this embodiment, the training process of the two-stage impact damage prediction model is as follows:

[0094] First, the TraceNet model is trained. The input image size is set to 512x512 resolution, and the batch size of each input image is 8. The learning rate is set to 0.001.

[0095] In TraceNet, the calculation of log-likelihood is based on the model's evaluation of the probability density of the data point. Specifically, given a data point x, it is mapped to a point z in the latent space through the transformation f. Where z = f(x). Since f is reversible, it can be obtained by f -1 Calculate the inverse mapping of z to get x. According to the transformed probability density function, we can get:

[0096]

[0097] Among them, p X (x) is the probability density of data point x in the original space, p Z (z) is the probability density of z in the latent space. is the Jacobian matrix of the transformation f with respect to x, is the determinant of the Jacobian matrix.

[0098] Therefore, the log-likelihood is:

[0099]

[0100] For all data For , the learning goal is to maximize the sum of the log-likelihoods of all data points:

[0101]

[0102] Therefore, minimizing the loss function can be expressed as minimizing the negative log-likelihood:

[0103]

[0104] Using this as the loss function ensures that the model can not only learn the potential representation of the data points, but also ensures that the distribution in the latent space is as simple as possible. At the same time, it also retains the ability to recover from the latent space to the data space, ensuring the conversion capability of the data mapping.

[0105] After 100 epochs of training, the loss value has approached 0 and remained stable. This shows that the TraceNet model has fully learned the image features. Therefore, the shallow CNN network is trained next. 200 training cycles are set. It can be found that when it is close to 200 training cycles, the loss value of the shallow CNN regression model has approached 0 and remained stable. This shows that the model has achieved the ideal training effect. The relationship between the impact damage image and the initial kinetic energy of the tool, the peak penetration force and the number of penetration layers has been successfully established.

[0106] Furthermore, the key puncture parameters are input into a classifier, and the perpetrator tracing includes: training the classifier;

[0107] Training a classifier involves:

[0108] A real experimental data set and a model prediction data set were constructed respectively; the real experimental data set used the knife stabbing distance, the initial speed of the knife, and the number of penetration layers as data, and the participant number as label; the model prediction data set used the knife initial kinetic energy, the puncture peak force, and the number of penetration layers as data, and the participant number as label;

[0109] Use real experimental data sets and model prediction data sets to train the classifier;

[0110] The trained classifier is used to classify the characteristics of different personnel, and an unpaired T-test is performed on the evaluation indicators obtained from the three-month real experimental data set and the model prediction data set.

[0111] Specifically, in this embodiment, the process of training the classifier is as follows: recruit participants to perform puncture tests in a real knife stabbing scenario. Use a high-speed camera to record the puncture process of the knife and calculate the initial puncture speed of the knife. The real experimental data set and the model prediction data set are constructed respectively. The real experimental data set uses the knife stabbing distance, the initial speed of the knife, and the number of penetration layers as data, and the participant number as a label. The model prediction data set uses the initial kinetic energy of the knife, the peak puncture force, and the number of penetration layers as data, and the participant number as a label. Use a machine learning classifier to classify the characteristics of different people. Perform an unpaired T test on the evaluation indicators obtained under the real experimental data set and the model prediction data set. The unpaired T test is a statistical method used to compare whether there is a significant difference between the means of two independent sample groups. The calculation formula is:

[0112]

[0113] in and are the means of the two sets of data samples, and are the variances of the two data sets, and n1 and n2 are the sample sizes of the two data sets. The confidence level p of each set of evaluation indicators is calculated to be greater than 0.05, so it can be considered that there is no significant difference between the real experimental data set and the model prediction data set. Therefore, it can be concluded that the constructed two-stage impact damage prediction model can effectively predict the key parameters in impact damage based on the impact damage image. The classification results of the classifier can also distinguish the characterization capabilities of different personnel characteristics.

[0114] like Figure 3 As shown, this embodiment also provides a perpetrator tracing device based on deep learning, the device comprising: a data set construction module, an impact damage prediction module, a real scene prediction module, a classifier training module, an operation processor module and a memory module;

[0115] A data set construction module, used to form an image-puncture key parameter corresponding data set from images and data obtained from dynamic puncture test and image acquisition test;

[0116] Impact damage prediction module, used to train the two-stage impact damage prediction model to predict key puncture parameters such as initial kinetic energy, peak force, and number of penetration layers of the tool;

[0117] The real scene prediction module is used to predict the key puncture parameters such as the initial kinetic energy, peak force and number of penetration layers of the tool using a two-stage impact damage prediction model for the impact damage images generated in the real scene;

[0118] The classifier training module classifies the perpetrators according to the key puncture parameters predicted in real scenarios and traces back to the specific characteristics of the perpetrators;

[0119] A running processor module is used to run a two-stage impact injury prediction model and a classifier in the assailant tracing method;

[0120] The memory module is used to store the image-puncture key parameter corresponding data set, the training results and prediction results of the two-stage impact damage prediction model, and the training results and prediction results of the classifier in the device.

[0121] An embodiment of the present application also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method or module when executing the computer program.

[0122] The embodiment of the present application further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above method are executed.

[0123] The present embodiment is further described below with specific experimental data;

[0124] A method for tracing an assailant based on deep learning provided in an embodiment includes:

[0125] Step 1: Dynamic puncture test and image acquisition test to obtain the image-puncture key parameter corresponding data set.

[0126] During a single test, the impactor carries the tool and falls freely from a certain height. When the tool stands upright on the sample stably, the single test ends. 8 different drop heights are set, ranging from 45 cm to 185 cm and at intervals of 20 cm. In addition, the photoelectric gate is placed at the height where the tool is about to contact the sample to record the initial speed of the tool when it pierces the sample. The initial kinetic energy of the tool is obtained according to the kinetic energy calculation formula:

[0127]

[0128] Where m is the total mass of the tool and impactor, 2 kg, and v is the initial velocity of the tool when it penetrates the sample.

[0129] In order to collect physical quantities in the dynamic puncture test, a data acquisition system was established. The force sensor placed between the bottom of the tool and the impactor collects the impact force during the tool puncture process. A 10μm thick aluminum film is laid on the 1st, 20th, 30th, 40th, and 50th layers of 50-layer PE. Each layer of aluminum film is connected to the data acquisition system through a wire (range 10V, acquisition frequency 2MHz). When the tool penetrates a layer of aluminum film, the closed loop formed will enable the data acquisition system to capture an electrical signal. The position of the tool at different times can be determined according to the time when the signal is generated, so the number of layers penetrated by the tool can be determined. In order to reduce errors and avoid the contingency of the experiment, the experiment was repeated 6 times to obtain reliable data. The average of the initial speed of the tool collected from the 6 sets of experimental data was calculated to obtain the average speed at 8 heights, and the initial impact kinetic energy of the tool at different heights was obtained according to the kinetic energy calculation formula. At the same time, the peak force and the number of layers penetrated during the puncture process were also collected in the experiment. The key puncture parameters are shown in Table 1.

[0130] Table 1

[0131]

[0132]

[0133] After the dynamic puncture test was completed, 6 sets of images at 8 heights were collected, and the images were enlarged 5 times through basic image processing methods such as random cropping and rotation. Finally, 30 images of each drop height were obtained, a total of 240 impact damage images. Correspondingly, 240 non-damaged images were also collected as part of the data set. In order to take into account the training efficiency and accuracy of the network model, the resolution of all images was set to 512x512. The image examples are as follows: Figure 4 shown.

[0134] Step 2: Train the two-stage impact damage prediction model. 80% of the 240 undamaged images are used as the training set. The remaining 20% ​​of the undamaged images and the 240 impact damaged images are divided into the test set and the validation set in a 1:1 ratio. Therefore, the training set images are 192, the validation set images are 144, and the test set images are 144. In the training phase, only 192 undamaged images are input. The feature extraction module is used to extract features and Traceflow is used to generate the corresponding data distribution from the original pixel distribution. Through training, TraceNet can fully understand the diverse texture features of the PE image surface and avoid the interference of the PE background when detecting impact damage. In the validation phase, 24 undamaged images and 120 impact damaged images are processed at the same time. The features extracted by the feature extraction module based on the impact damaged images are also converted into corresponding data distributions through Traceflow. If the data distribution of a certain area of ​​the impact damaged image cannot match the data distribution of the undamaged image, this area is considered to be the impact damaged area and is labeled. In the testing phase, 144 images that have never been trained will be input to observe the model effect. In this way, pixel-level mapping can be applied to accurately locate the impact damage area in the image. More importantly, only the undamaged image needs to be learned to enable the TraceNet model to flexibly identify various impact damages, greatly improving the model's efficiency in segmenting damaged areas. After training for 100 epochs, the loss value has approached 0 and remained stable, indicating that the TraceNet model has fully learned the image features.

[0135] After TraceNet achieves accurate segmentation of the damaged area in the impact damage image, the shallow CNN network is used as a regression model to characterize the relationship between the TraceNet segmentation results and the experimentally measured key damage parameters. The shallow CNN model is trained. The tool initial kinetic energy, puncture peak force and number of penetration layers in the key damage parameters corresponding to the image are used as labels for training in 240 impact damage images. 200 training cycles are set. When the training cycle is close to 200, the loss value of the shallow CNN regression model is close to 0 and remains stable. This shows that the model has achieved the ideal training effect. The relationship between the impact damage image and the tool initial kinetic energy, puncture peak force and number of penetration layers has been successfully established.

[0136] In the third step, the classifier was trained to realize the function of tracing the puncture experimenter based on the key puncture parameters. 10 participants (6 males and 4 females, aged between 22 and 30 years old) were recruited to participate in the puncture test. The information of the 10 participants is shown in Table 2. In order to truly restore the knife stabbing scene, the participants controlled the distance between the knife and the backing to be 10 cm and 20 cm respectively in each experiment. Each participant repeated the experiment 10 times at the same distance. A high-speed camera was used to record the process of the knife puncturing the sample at a sampling rate of 700 frames per second. Setting a ring light source not only increases the amount of light entering the high-speed camera and improves the quality of the photo, but also locates the knife puncture height through the central ring of the light source.

[0137] After the real puncture test is completed, the impact damage image of the sample is collected using an image acquisition system and the two-stage impact damage prediction model is used to predict the impact key parameters to obtain the perpetrator database. In order to prove the characterization ability of the impact key parameters on the characteristics of the puncture test participants, the classification model in machine learning is used for analysis. The real experimental data set and the model prediction data set are constructed respectively. The real experimental data set uses the knife stabbing distance, the initial speed of the knife, and the number of penetration layers as data, and the participant number as the label. The model prediction data set uses the initial kinetic energy of the knife, the peak puncture force, and the number of penetration layers as data, and the participant number as the label.

[0138] Table 2

[0139] Participant Number gender age Height (cm) Weight(kg) 1 male 29 180 78 2 male 25 178 70 3 male 24 178 70 4 male 24 175 70 5 male 23 180 75 6 male 23 175 77 7 female 26 158 56 8 female 25 160 47 9 female 24 172 60 10 female 23 165 68

[0140] The classifier can achieve an accuracy of 70.8% and 71.4% for the real experimental data set and the model prediction data set, respectively. In addition, precision, recall, and F1 score can all obtain values ​​above 68%, and ROCAUC can reach above 90%. Unpaired T tests were performed on the evaluation indicators obtained under the real experimental data set and the model prediction data set, respectively. The confidence level of each group of evaluation indicators calculated was greater than 0.05, so it can be considered that there is no significant difference between the real experimental data set and the model prediction data set. Therefore, it can be concluded that the two-stage impact damage prediction model can effectively predict the key parameters in impact damage based on the impact damage image. The classification accuracy of the classifier of more than 70% also shows that the key parameters predicted using this model have the ability to characterize the characteristics of different personnel.

[0141] The equipment used in this embodiment mainly includes the following main modules, including:

[0142] A data set construction module, used to form an image-puncture key parameter corresponding data set from images and data obtained from dynamic puncture test and image acquisition test;

[0143] Impact damage prediction module, used to train the two-stage impact damage prediction model to predict key puncture parameters such as initial kinetic energy, peak force, and number of penetration layers of the tool;

[0144] The real scene prediction module is used to predict the key puncture parameters such as the initial kinetic energy, peak force and number of penetration layers of the tool using a two-stage impact damage prediction model for the impact damage images generated in the real scene;

[0145] The classifier training module classifies the perpetrators according to the key puncture parameters predicted in real scenarios and traces back to the specific characteristics of the perpetrators;

[0146] A running processor module is used to run a two-stage impact injury prediction model and a classifier in the assailant tracing method;

[0147] The memory module is used to store the image-puncture key parameter corresponding data set, the training results and prediction results of the two-stage impact damage prediction model, and the training results and prediction results of the classifier in the device.

[0148] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for tracing an assailant based on deep learning, characterized in that: include: Collect impact damage images in real knife stabbing scenarios; Based on the two-stage impact damage prediction model, the impact damage image is predicted to obtain the key puncture parameters; wherein the two-stage impact damage prediction model is obtained by training a data set, and the data set includes: the key puncture parameters of the sample at different initial speeds of the tool and the impact damage image of the sample under different dynamic puncture test conditions; the two-stage impact damage prediction model includes: a first-stage TraceNet network and a second-stage shallow CNN network; The first TraceNet network is used to extract damage features in the impact damage image; The first segment of the TraceNet network includes: a feature extraction module and a Traceflow module; The feature extraction module is used to extract initial features of a preset dimension in the impact damage image; The Traceflow module is used to recombine the initial features to obtain the damage features; The second shallow CNN network is used to characterize the relationship between the injury characteristics and the key puncture parameters; The second shallow CNN network includes: a plurality of two-dimensional convolutional layers, a plurality of preset maximum pooling layers, and two preset fully connected layers; The two-dimensional convolutional layers are used to extract features of the input image and generate feature maps in several steps; The plurality of preset maximum pooling layers are used to continuously reduce the spatial size of the feature map; The first fully connected layer is used to learn the nonlinear relationship between the features, and the second fully connected layer is used to output the prediction result of the second shallow CNN network on the key puncture parameters; The key puncture parameters are input into a classifier to perform perpetrator tracing.

2. The method for tracing the perpetrator based on deep learning according to claim 1 is characterized in that: The key puncture parameters include: the initial speed of the cutter when contacting the sample in the dynamic puncture test, the initial kinetic energy of the cutter, the peak puncture force, and the maximum number of penetration layers for the sample.

3. The method for tracing the perpetrator based on deep learning according to claim 1 is characterized in that: The feature extraction module includes: a convolution layer, a maximum pooling layer, a convolution block and a marking block connected in sequence; The image is input into the feature extraction module, firstly passes through the convolution layer, the output of the convolution layer passes through the ReLU activation function and then input into the maximum pooling layer, the output of the maximum pooling layer is used as the input of the convolution block, and the features extracted by the convolution block are input into the identification block to extract features again; Wherein, the convolution layer is used to perform convolution operation on the input image; The maximum pooling layer is used to downsample the output of the convolutional layer while retaining preset image features; The identification block is used to learn the identity mapping and directly transfer the input information to the output; A convolution layer with a convolution kernel of 1x1 is introduced into the convolution block as a residual connection in the feature extraction module.

4. The method for tracing the perpetrator based on deep learning according to claim 1, characterized in that: The Traceflow module includes: a plurality of flow layers of preset sizes, the flow layers include: a flow layer of a 3x3 convolution kernel and a flow layer of a 1x1 convolution kernel, each flow layer of a 3x3 convolution kernel and each flow layer of a 1x1 convolution kernel are used as a flow module, and the Traceflow module is provided with 8 flow submodules in total; the input of the first flow submodule is the output of the feature extraction module, the output of the first flow submodule is used as the input of the second flow submodule, the output of the second flow submodule is used as the input of the third flow submodule, and so on; wherein, an activation normalization layer and an affine coupling layer are used in each flow layer; The activation normalization layer is used to independently normalize each channel of the input information of each flow layer; The affine coupling layer uses a reversible neural network structure to ensure that the training results of each flow layer do not deviate from the original features while being able to extract more complex features.

5. The method for tracing the perpetrator based on deep learning according to claim 1, characterized in that: Training the two-stage impact damage prediction model using the data set includes: Setting a minimization loss function for the first segment of the TraceNet network; wherein the minimization loss function is minimizing negative log-likelihood; Based on a preset epoch, the first segment of the TraceNet network is trained using the minimized loss function; The second shallow CNN network is trained based on a preset training cycle.

6. The method for tracing the perpetrator based on deep learning according to claim 5 is characterized in that: The minimization loss function is: Among them, L is the loss value, N is the number of data points, i is the sequence number of the data point, and p X (x i ) is the data point x i The probability density in the original space, x i is the i-th data point, df is the value of the data point x i The mapping transformation, is the transformation f with respect to x i The Jacobian matrix, p Z (f(x i )) is f(x) in the latent space i ), f(x i ) is x i Mapping in latent space.

7. The method for tracing the perpetrator based on deep learning according to claim 1, characterized in that: Inputting the puncture key parameters into a classifier, and tracing the perpetrator includes: training the classifier; Training the classifier includes: A real experimental data set and a model prediction data set are constructed respectively; wherein the real experimental data set includes the stabbing distance, the initial speed of the tool, and the number of penetration layers, and the participant number is used as a label; and the model prediction data set includes the initial kinetic energy of the tool, the peak puncture force, and the number of penetration layers, and the participant number is used as a label; Training the classifier using the real experimental data set and the model prediction data set; The trained classifier is used to classify the characteristics of different personnel, and an unpaired T test is performed on the evaluation indicators obtained from the three-month real experimental data set and the model prediction data set.

8. A perpetrator tracing device based on deep learning, characterized in that: Used to implement the deep learning-based perpetrator tracing method as described in any one of claims 1 to 7, the device comprises: a data set construction module, an impact damage prediction module, a real scene prediction module, a classifier training module, an operation processor module, and a memory module; A data set construction module is used to construct a data set according to the key puncture parameters of the sample at different initial speeds of the tool and the impact damage images of the sample under different dynamic puncture test conditions; wherein the key puncture parameters include: the initial speed of the tool when contacting the sample in the dynamic puncture test, the initial kinetic energy of the tool, the puncture peak force, and the maximum number of penetration layers for the sample; An impact damage prediction module, used to use the data set to train the two-stage impact damage prediction model to predict key puncture parameters; A real scene prediction module is used to predict key puncture parameters using a two-stage impact damage prediction model for impact damage images generated in real scenes; The classifier training module classifies the perpetrators according to the key puncture parameters predicted in real scenarios and traces the characteristics of the perpetrators; An operation processor module is used to operate the two-stage impact injury prediction model and the classifier in the assailant tracing method; The memory module is used to store the data set, the training results and prediction results of the two-stage impact damage prediction model, and the training results and prediction results of the classifier in the device.

Images