CN112381109B - Line trace comparison system applied to single-point laser detection - Google Patents
Line trace comparison system applied to single-point laser detection Download PDFInfo
- Publication number
- CN112381109B CN112381109B CN202010346738.9A CN202010346738A CN112381109B CN 112381109 B CN112381109 B CN 112381109B CN 202010346738 A CN202010346738 A CN 202010346738A CN 112381109 B CN112381109 B CN 112381109B
- Authority
- CN
- China
- Prior art keywords
- data
- comparison
- trace
- value
- result
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Abstract
The invention discloses a line trace comparison system applied to single-point laser detection, which belongs to the technical field of criminal investigation. The invention uses the laser displacement sensing technology and the computer control technology to carry out 360-degree rapid data scanning on rifling trace on the surface of the bullet head along the circumferential direction; after the scan data is noise-reduced, the scan data is analyzed and compared by a computer. Thereby, the correlation of different warheads and guns can be rapidly and accurately judged.
Description
Technical Field
The invention belongs to the technical field of criminal investigation, and particularly relates to a line trace comparison system applied to single-point laser detection.
Background
Bullet trace inspection is an important criminal investigation technique used to detect gun-related cases. The main task of the bullet trace test is the identification of the species of bullet and the identity of the gun, i.e. the determination of which gun the bullet, the case, found in the field was fired.
To solve these two problems, the trace left on the bullet and the case is inspected. The relationship between the bullet and the firearm is determined from these marks, some of which are recurring and some of which are occasional. Accidental traces are of no value to the determination of the firearm from which the bullet was fired, and the repeated trace is of great value because it reflects the inherent characteristics of the firearm with respect to the machine. The guns produced by two identical tools in turn have many similarities (e.g. identical rifling) which are the basis for identifying the gun species. Traces on the bullet or case of the same gun can reflect the detailed characteristics of the components associated with the firing gun. The identity of a particular firearm can be determined by inspecting the trace on the bullet or case, provided that the firearm's parts are not substantially changed.
Conventional methods of bullet rifling trace alignment include comparison microscopy, sectioning, stylus detection, etc. The methods have the defects of long comparison time consumption, low efficiency, large influence by subjective factors, high accuracy of quantitative analysis and the like, and cannot meet the requirements of rapidness and accuracy of bullet trace detection. Therefore, there is a need for a method for rapid, accurate, and automatic identification of the trace of a warhead.
In addition, in order to further enhance the management of the gun, a bullet shell and bullet head sample is manufactured, and a bullet trace file is established. Because no bullet rifling trace digitizing equipment exists at present, bullet trace files of each unit are difficult to digitize, and quick retrieval and comparison are not possible by utilizing a computer technology.
Disclosure of Invention
Firstly, carrying out 360-degree rapid data scanning on rifling marks on the surface of a bullet head of a bullet along the circumferential direction by utilizing a laser displacement sensing technology and a computer control technology; after the scan data is noise-reduced, the scan data is analyzed and compared by a computer. Trace data generated on the bullet warhead surface by firing gun rifling can be obtained rapidly and accurately by utilizing a laser displacement sensing technology, and the similarity of different bullet trace curves is calculated, analyzed and compared by utilizing computer analysis software, so that objective judgment is rapidly and accurately made on the correlation of different bullets and guns. The collected data can be used for one-to-one data comparison between warheads and one-to-many data comparison.
In order to achieve the above purpose, the present invention is realized by the following technical scheme: the comparison system is applied to criminal investigation, bullet trace detection and other scenes needing trace comparison, the line trace comparison system applied to single-point laser detection comprises system hardware and system software, the system software comprises detection data exception handling, data noise reduction, rotation correction and comparison model establishment, and the system hardware is used for data acquisition of system detection and bullet fixation.
Preferably, the data abnormality processing is performed by using a corrected box diagram, which is an improved diagram obtained by correcting an abnormal value based on the box diagram.
Preferably, the data noise reduction aims at the remarkable characteristics of strong randomness of trace laser detection signals, easiness in interference of background noise and the like, and adopts an exponential moving average method to smooth time domain data so as to eliminate tiny saw teeth in the data to the greatest extent and obtain a relatively stable coarse coherent peak trend;
in the actual noise reduction process, multiple layers of decomposition are needed to be carried out on the original signal f (t), and the more the layers of decomposition are, the more the detail data are processed, the more the noise can be eliminated, but the more the detail can be smoothed. Therefore, finding a balanced number of decomposition layers is necessary;
in the decomposition process, the original signal is generally decomposed into two parts, and assuming n layers of decomposition are performed, the composition of the original signal can be described as follows:
wherein, the layered information can be represented by the correspondence of each node and the original formula:
a i an approximation of the nth layer, denoted as (I, 0);
d i detail data for the i-th layer, denoted as (i.1);
f is the original data, denoted (0, 0).
Preferably, the rotation correction is defined according to the actual situation:
assume that the trace signal with length n is scanned to be S= { S 1 ，s 2 ，…，s mid ，…，s n Where mid is the site therein, then the tilt may be defined as follows:
in general, the RotateRange is not 0, so for each input trace signal, corresponding rotation correction is generally required, the rotation correction is performed on the input signal according to the value of RotateRange, the rotation correction is performed on the basis of the known RotateRange according to different trace positions, and the correction manner for each point is as follows:
wherein news is i For the corrected value, the value far from the middle end is corrected to a larger extent, and the correction amplitude received by the value close to the middle is limited.
Preferably, the comparison model establishment comprises the steps of comparison pretreatment in step 1, parameter training in step 2, similarity settlement in step 3 and merging and outputting in step 4; the step 1 of comparison pretreatment is that before similarity comparison is carried out on signals of traces after noise reduction, the problems of length uncertainty and partial overlapping are firstly required to be treated: 1) Setting input data as A and B respectively, wherein the data are required data; 2) Setting a minimum length L for comparison, namely selecting a part from the A to the B from the longest length to the shortest length for comparison if the two have to meet the minimum overlapping length, and selecting different positions for comparison for multiple times; 3) Iteratively executing comparison of each position, wherein each comparison is used for comparing the variance of the difference degree (of the corresponding position) of the corresponding position of each comparison, and if the variance is minimum, recording the current state; 4) After the function of 3) is completed, the roles of A, B are exchanged, and 2 and 3 are continuously completed once; 5) And calculating the difference degree with the minimum variance, and outputting a comparison result.
Preferably, the parameter training (1) ensures that the sample library establishment step is completed, the samples have enough representativeness and discrimination, the machine learning is carried out by adopting a graph convolution neural network algorithm, the graph convolution neural network algorithm comprises 1) establishing a training set, 2) adjusting parameters and establishing a graph convolution neural network model, (3) finishing the parameter training after acquiring the learned parameters, and (4) carrying out the parameter training by using a more targeted training set when the sample library is changed or the use scene is changed; sample library establishment its establishment generally follows the following steps: the method comprises the steps of (1) determining the type, range and type of tools to be identified, numbering each tool by using a unified rule, recording parameter information in detail, (2) detecting a single broken end trace at least twice in order to eliminate the accident in detection, determining qualified data when the coincidence of signal data of the two times can reach more than 99%, recording different data of a plurality of positions by the same broken end, wherein each position adopts the mode as above, (3) taking data of the sample library as test data after all data are acquired, testing each feature, wherein the test result is that the data similarity degree of each sample data and own group data is obviously higher than that of other groups, and if the sample data are mixed together, the data are invalid, and repeating the steps (1) - (2).
Preferably, the specific method for 2) parameter tuning and graph convolution neural network model establishment is that G= (V, E) and V represent node sets, namely
E represents the edge set, i.e.)>
The training model consists of two parts: 1) A GCN component responsible for sampling all node information in the K-order neighborhood, 2) a self-encoder (AE) component for extracting hidden features of the activation value matrix A learned by the GCN component, and combiningThe Laplacian Eigenmap (LE) preserves node cluster structure, and the GCN component uses graph convolution neural network in order to node +.>
For the central sampling of the structure and characteristic information of all nodes in the step K, namely encoding K-order neighborhood information, generating an activation value matrix A as an input of a self-encoder component by combining with the label training of the nodes, enabling a GCN to encode the local structure and characteristic information of a network at the same time by supervised learning based on the node labels, omitting secondary structure information with smaller influence on the low-dimensional vector of the generated node outside the K-order neighborhood, utilizing the activation value matrix A obtained by the GCN as the input of the self-encoder, further extracting the characteristic information from the A by the self-encoder in an unsupervised learning mode, and mapping the original network to a lower-dimensional space by combining with Laplace characteristic mapping.
Preferably, in the step 3, different algorithms in the similarity calculation are calculated, and the calculated degree, degree of difference or similarity may be calculated, and the obtained result values of the different algorithms do not all fall between 0 and 1, so that relevant result mapping methods are formulated for different types of algorithms;
wherein it is assumed that the distance d of the normalized unit vector is calculated i，j Taking into account d i，j The value of (2) is between 0 and 1, and the distance is 0 when the values are identical, then:
H＝1-d i，j
provided that it is, for example, the overlap ratio p i，j The same value falls before 0 and 1, and is the most similar value as 1
H＝1-p i，j
If the result of the calculation is v which cannot be distributed between 0 and 1 i，j Then there are:
where w is an empirical weight and b is a bias, the specific parameters are determined by the specific procedure.
Preferably, the step 4 is combined and output, provided that n algorithm strategies are provided, and the result vector obtained by calculation is calculated
Then we also give here a weight of each policy +.>
The end result is then:
/>
how to determine the weight
There are mainly two methods
(1) Expert knowledge is combined with an actual experimental result to determine an empirical value, and the method is generally quick and close to optimal, and does not need too much data for verification;
(2) A correlation method of machine learning is used to calculate the optimal weights, such as a gradient descent method.
Preferably, the at least moving average process is:
assuming that each observation is noisy, and we expect the mean value of the noise to be 0, the variance to be
The relationship between the observed value and the true value is as follows:
g t ＝x t +ε t
wherein x is t G is the observed value t Is true value, epsilon t In the event of a noise occurrence,
the formula of the exponential moving average method is as follows:
p t ＝ω*x t +(1-ω)*p t -1 (1)
p t represents the predicted value, ω represents the decay weight, typically we set to a fixed value of 0.9, x t Representing observations, which is a recursive formula, the weights of the exponential moving average method are exponentially decaying over time, extending the above formula (1):
p t -1＝ω*x (t-1) +(1-ω)*p t-2 (2)
combining equations (1) (2) can be obtained:
p t ＝ω*x t +(1-ω)*(ω*x (t-1) +(1-ω)*p t-2 ) (3)
the following relationship exists at the initial time:
p 0 ＝x 0 (4)
from this relationship and the above recursive formula we can get the formula of the whole algorithm, the error formula of the exponential sliding average method is that t=2 is set, and g will be at the same time t ＝x t +ε t And equation (4) brings into equation (3), the error term is obtained as:
ε＝ω*ε 2 +(1-ω)*(ω*ε 1 +(1-ω)*ε 0 ) (5)。
the invention has the beneficial effects that:
firstly, carrying out 360-degree rapid data scanning on rifling marks on the surface of a bullet head of a bullet along the circumferential direction by utilizing a laser displacement sensing technology and a computer control technology; after the scan data is noise-reduced, the scan data is analyzed and compared by a computer. Trace data generated on the bullet warhead surface by firing gun rifling can be obtained rapidly and accurately by utilizing a laser displacement sensing technology, and the similarity of different bullet trace curves is calculated, analyzed and compared by utilizing computer analysis software, so that objective judgment is rapidly and accurately made on the correlation of different bullets and guns. The collected data can be used for one-to-one data comparison between warheads and one-to-many data comparison.
Drawings
FIG. 1 is a diagram of the relative positions of the bullet center and the scanning platform center of rotation;
FIG. 2 is a graph of the distance between the laser displacement sensor and the warhead surface;
FIG. 3 is a schematic diagram of the comparative pretreatment.
Detailed Description
The present invention will be further described in detail with reference to the drawings and examples, which are not intended to limit the invention, in order to make the objects, technical solutions and advantages of the present invention more apparent.
The comparison system is applied to criminal investigation, bullet trace detection and other scenes needing trace comparison, the line trace comparison system applied to single-point laser detection comprises system hardware and system software, the system software comprises detection data exception handling, data noise reduction, rotation correction and comparison model establishment, and the system hardware is used for data acquisition of system detection and bullet fixation.
The data abnormality processing adopts a correction box diagram for processing, and the correction box diagram is an improved diagram after correcting abnormal values on the basis of the box diagram. First, before the correction of the box map is introduced, it is necessary to specify that the correction of the box map is an improvement pattern obtained by correcting the abnormal value on the basis of the box map. The box line graph is a graph formed by boxes and straight lines, and can intuitively reflect the distribution trend of sample data in the statistical field of data. The box diagram mainly comprises 5 parts, which are respectively: minimum value Min, first quartile Q1, median M, third quartile Q3 and maximum value Max.
Wherein the remaining three parts are all based on the data set, except the minimum and maximum values, and the definition of the figures is as follows:
quantiles: if the probability 0 < p < 1, the random variable X or the quantile Z of its probability distribution a Means that the condition P (X > Z) is satisfied a ) Real number of =a.
After having these five statistics, we can do an image based on these 5 quantities:
1. the drawing number axis, the measurement unit size is consistent with the unit of the data batch, the starting point is slightly smaller than the minimum value, and the length is slightly longer than the full distance of the data batch.
2. A rectangular box is drawn, and the positions of two end edges respectively correspond to the upper quartile and the lower quartile (Q1 and Q3) of the data batch. A line segment is drawn at the median (Xm) position inside the rectangular box as a median.
This resulting image can be referred to as a box plot, by which we can intuitively compare the properties of multiple datasets. However, when one observed value in the data set is unusually larger or smaller than other data in the data set, the observed value can cause the maximum value end or the minimum value end of the box diagram to be abnormal and far away from the middle box, and therefore, a set of rules is specifically proposed to correct the problem, and the correction process is as follows:
1. the first quartile is known as Q1, the third quartile is known as Q3, and the distance between them is noted as iqr=q3-Q1, which becomes the quartile spacing.
2. All data less than Q1-1.5IQR and greater than Q3+1.5IQR are marked as outliers and are marked individually as outliers when drawing the graph and are no longer drawn among the subjects of the box graph.
The values of most anomalies are filtered from a statistical point of view by this correction step. In our algorithm, the same detection method as the corrected box map will be used to label outliers. In our device, when some reflection points or other interfering objects are scanned, obvious anomalies will occur in the scanned trace signal. This anomaly is distinct from the usual fine noise interference that is quite noticeable and generally difficult to remove using filtering noise reduction algorithms. By adopting the detection method of the correction box diagram, the abnormal position can be found from the statistical angle and corrected.
Aiming at the remarkable characteristics of strong randomness of trace laser detection signals, easiness in interference of background noise and the like, the data noise reduction adopts an exponential moving average method to carry out time domain data smoothing so as to eliminate tiny saw teeth in the data to the greatest extent and obtain a relatively stable coarse coherent peak trend.
Assuming that each observation is noisy, and we expect the mean value of the noise to be 0, the variance to be
The relationship between the observed value and the true value is as follows:
g t ＝x t +ε t
wherein x is t G is the observed value t Is true value, epsilon t In the event of a noise occurrence,
the formula of the exponential moving average method is as follows:
p t ＝ω*x t +(1-ω)*p t -1 (1)
p t represents the predicted value, ω represents the decay weight, typically we set to a fixed value of 0.9, x t Representing observations, which is a recursive formula, the weights of the exponential moving average method are exponentially decaying over time, extending the above formula (1):
p t -1＝ω*x (t-1) +(1-ω)*p t-2 (2)
combining equations (1) (2) can be obtained:
p t ＝ω*x t +(1-ω)*(ω*x (t-1) +(1-ω)*p t-2 ) (3)
the following relationship exists at the initial time:
p 0 ＝x 0 (4)
from this relationship and the above recursive formula we can get the formula of the whole algorithm, the error formula of the exponential sliding average method is that t=2 is set, and g will be at the same time t ＝x t +ε t And equation (4) brings into equation (3), the error term is obtained as:
ε＝ω*ε 2 +(1-ω)*(ω*ε 1 +(1-ω)*ε 0 ) (5)。
in the actual noise reduction process, multiple layers of decomposition are needed to be carried out on the original signal f (t), and the more the layers of decomposition are, the more the detail data are processed, the more the noise can be eliminated, but the more the detail can be smoothed. It is therefore necessary to find a balanced number of decomposition levels. In the decomposition process, the original signal is generally decomposed into two parts, and assuming n layers of decomposition are performed, the composition of the original signal can be described as follows:
wherein, the layered information can be represented by the correspondence of each node and the original formula:
a i an approximation of the nth layer, denoted as (I, 0);
d i detail data for the i-th layer, denoted as (i.1);
f is the original data, denoted (0, 0).
In practical use, although leveling operation is performed during laser scanning, an image actually obtained by the user still has a certain inclination, and the inclination can cause great interference to subsequent comparison work, so that rotation correction is required during preprocessing.
In the engineering field, the inclination refers to the ratio of the sedimentation difference of the inclination directions of two end points of a foundation to the distance of the two end points, and in our algorithm, we redefine the inclination according to the actual situation.
Assume that the trace signal with length n is scanned to be S= { S 1 ，s 2 ，…，s mid ，…，s n Where mid is the site therein, then the tilt may be defined as follows:
in general, the RotateRange is not 0, so for each input trace signal, corresponding rotation correction is generally required, the rotation correction is performed on the input signal according to the value of RotateRange, the rotation correction is performed on the basis of the known RotateRange according to different trace positions, and the correction manner for each point is as follows:
wherein news is i For the corrected value, the value far from the middle end is corrected to a larger extent, and the correction amplitude received by the value close to the middle is limited. This correction method is only suitable for correcting the situation of the tool placement inclination, if the tool placement inclination is completely uncorrectable due to shearing or the tool itself, but the subsequent calculation is not affected.
The comparison model establishment comprises the steps of comparison pretreatment in step 1, parameter training in step 2, similarity settlement in step 3 and merging and outputting in step 4; the step 1 of comparison pretreatment is that before similarity comparison is carried out on signals of traces after noise reduction, the problems of length uncertainty and partial overlapping are firstly required to be treated: 1) Setting input data as A and B respectively, wherein the data are required data; 2) Setting a minimum length L for comparison, namely selecting a part from the A to the B from the longest length to the shortest length for comparison if the two have to meet the minimum overlapping length, and selecting different positions for comparison for multiple times; 3) Iteratively executing comparison of each position, wherein each comparison is used for comparing the variance of the difference degree (of the corresponding position) of the corresponding position of each comparison, and if the variance is minimum, recording the current state; 4) After the function of 3) is completed, the roles of A, B are exchanged, and 2 and 3 are continuously completed once; 5) And calculating the difference degree with the minimum variance, and outputting a comparison result.
As shown in fig. 3, trace a and trace B are both part of a common trace, and after trace B is aligned to 1/4 of trace a, matching with trace a can be started until matching of a is completed. After alignment AB, the matching of the two is continued until one end.
By default we choose the overlapping mode to be 70% of the smallest trace length, and choose 70% because if no clipping is performed, a larger difference will be caused to the final result due to the local difference, and after clipping, the most valuable 70% of the trace length can be automatically chosen to perform similarity measurement. Also, note that the 70% position is not fixed, with 70% meaning that only 70% of the length is selected each time the ratio is made, and the 70% length trace may be any position on the break and the final output is only the most similar 70% position.
How does this 70% of the locations chosen? We also assume that the test sample is a and its signal is S A While the identified samples in the sample library are B, the signal of which is S B Then 70% of the signal at a certain position of A is S Ai And 70% of B is S Bj 。
At this time, it is assumed that the algorithm for similarity calculation is a function H (x, y) ∈0, 1.
The way in which i and j are determined is to have the function with the greatest value of i, j.
Maximum i，j H(a*S Ai +b，S Bj )
That is, 70% of the data S in A should be found for any similarity calculation mode Aimax Data S of 70% of B Bimax At this time S Aimax After linear transformation, H (x, y) can be obtained to be considered as S in B Bimax The result of the greatest degree of similarity, provided that the selected position in A, B is not S Aimax ，S Bimax The end result is not greater than it.
The parameter training (1) ensures that the sample library establishment step is completed, the samples have enough representativeness and discrimination, the machine learning is carried out by adopting a graph rolling neural network algorithm, the graph rolling neural network algorithm comprises 1) establishing a training set, 2) adjusting parameters and establishing a graph rolling neural network model, (3) finishing the parameter training after acquiring the learned parameters, and (4) carrying out the parameter training by using a more targeted training set when the sample library is changed or the scene is changed; sample library establishment its establishment generally follows the following steps: the method comprises the steps of (1) determining the type, range and type of tools to be identified, numbering each tool by using a unified rule, recording parameter information in detail, (2) detecting a single broken end trace at least twice in order to eliminate the accident in detection, determining qualified data when the coincidence of signal data of the two times can reach more than 99%, recording different data of a plurality of positions by the same broken end, wherein each position adopts the mode as above, (3) taking data of the sample library as test data after all data are acquired, testing each feature, wherein the test result is that the data similarity degree of each sample data and own group data is obviously higher than that of other groups, and if the sample data are mixed together, the data are invalid, and repeating the steps (1) - (2).
2) The concrete method for parameter tuning and graph convolution neural network model establishment is that G= (V, E) and V represent node sets, namely
E represents the edge set, i.e.)>
The training model consists of two parts: 1) A GCN component responsible for sampling all node information in the K-order neighborhood, 2) a self-encoder (AE) component for extracting hidden features of the activation value matrix A learned by the GCN component and preserving node cluster structure in combination with Laplacian feature map (LE), the GCN component utilizing a graph convolution neural network to segment->
For the central sampling of the structure and characteristic information of all nodes in the K steps, namely encoding K-order neighborhood information, generating an activation value matrix A as the input of a self-encoder component by combining with the label training of the nodes, wherein the GCN can encode the local structure and characteristic information of a network at the same time by the supervised learning based on the node labels, omitting the secondary structure information with smaller influence on the low-dimensional vector of the generated nodes outside the K-order neighborhood, further extracting the characteristic information from the A by using the activation value matrix A obtained by the GCN as the input of the self-encoder by using an unsupervised learning mode and combining with the Laplace characteristic mappingThe original network is mapped to a lower dimensional space.
The two components are linearly combined and combined with the training set by using a Stacking method (Stacking) in ensemble learning, so that the node low-dimensional vector representation obtained by the whole model can retain the characteristic information of the node and the structure, the GCN component and the AE component are linearly combined by means of Stacking, and the two component loss functions are controlled by using two super parameters alpha and beta,
the loss function of the node sampling component is as follows:
The loss function from the encoder component AE is:
beta is the weight of the self encoder component AE loss function.
Finally, the loss function of the training model is defined as:
wherein y is i For the node to be truly a tag,
predictive tag for GCN, < >>
Is an active value matrix, K is node v i Neighborhood order,/->
A reconstructed activation value matrix,/->
The hidden layer of the first layer of the self-encoder is AE, and L is the hidden layer number of AE.
Model training with graphics card (GPU) acceleration using TensorFlow framework model optimization part updates model parameters using AdamOptimezer optimizer provided by TensorFlow, improves traditional gradient descent by using momentum (i.e. moving average of parameters), facilitates super-parameter dynamic adjustment, and enables fast and efficient training of models. Only one batch is used for updating the model parameters at a time, so that the memory occupation during model training is further reduced.
Different algorithms in the similarity calculation in the step 3 are calculated, wherein the calculated possibly is the degree of difference, the degree of difference or the similarity, and the obtained result values do not all fall between 0 and 1, so that related result mapping methods are formulated for different types of algorithms;
wherein it is assumed that the distance d of the normalized unit vector is calculated i，j Taking into account d i，j The value of (2) is between 0 and 1, and the distance is 0 when the values are identical, then:
H＝1-d i，j
provided that it is, for example, the overlap ratio p i，j The same value falls before 0 and 1, and is the most similar value as 1
H＝1-p i，j
If the result of the calculation is v which cannot be distributed between 0 and 1 i，j Then there are:
where w is an empirical weight and b is a bias, the specific parameters are determined by the specific procedure.
Preferably, the step 4 is combined and output, provided that n algorithm strategies are provided, and the result vector obtained by calculation is calculated
Then we also give here a weight of each policy +.>
The final result is:
how to determine the weight
There are mainly two methods
(1) Expert knowledge is combined with an actual experimental result to determine an empirical value, and the method is generally quick and close to optimal, and does not need too much data for verification;
(2) A correlation method of machine learning is used to calculate the optimal weights, such as a gradient descent method.
Example 2:
nine-two type pistol bullet trace comparison test
The nine-two pistol warhead participating in the comparison test is mainly provided by the public security hall of Yunnan province, the public security bureau of Guangxi Liuzhou and a certain weapon factory. Wherein the public security hall of Yunnan province provides 78 warheads triggered by 13 guns; the Guangxi Liuzhou public security office provides 4 warheads fired by one gun; a weapon factory provides 6 warheads fired by one gun, and a total of 15 gun-fired 88 9mm warheads. And scanning, collecting and processing rifling marks on the surface of the warhead according to the method, and performing one-to-many comparison calculation.
Here we agree that the comparison of 6 warhead scan data from the same gun is all at the first 10%, i.e. the first 9 bits, the comparison is excellent; the comparison and sequencing of 4 warhead scanning data is 10% before, and the comparison result is recorded as good; the comparison and sequencing of 3 warhead scanning data is 10% before, and the comparison result is marked as pass; the rest is the failed case. The statistical results are shown in Table 1-1.
TABLE 1-1 comparison result statistics
| Number of guns | Percentage of | Gun numbering | |
| Excellent and excellent properties | 4 | 27％ | dai、NN、yG、yI |
| Good quality | 8 | 53％ | GX、yA、yB、yC、yE、yH、yK、yL |
| Pass the lattice | 1 | 7％ | yJ |
| Failing to check | 2 | 13％ | yD、yF |
Penny-type pistol 9mm warhead trace comparison test
Warheads that participated in the comparison test were provided by the Shijia public security bureau. Each fifty-nine pistol fires 4 bullets, with a total of 30 gun-fired 120 9mm bullets. The gun numbers are S01-S30, and the bullet number fired by each gun is 1-4. And scanning, collecting and processing rifling marks on the surface of the warhead according to the method, and performing one-to-many comparison calculation.
The comparison and sequencing of the 4 warhead scanning data shot by the same gun is all positioned at the first 10%, namely the first 12 bits, so that the comparison result is excellent; the comparison and sequencing of 3 warhead scanning data is 10% before, and the comparison result is recorded as good; the comparison and sequencing of 2 warhead scanning data is 10% before, and the comparison result is marked as pass; the rest is the failed case. The statistical results are shown in tables 1-2.
TABLE 1-2 comparison result statistics
7.62mm warhead trace comparison test
7.62mm warheads involved in the comparison test were provided by the Shijia public security agency and Guangxi Liuzhou public security agency. Wherein the stone house police office provides 89 warheads fired by 20 seven-seven mobile phone guns (wherein 11 guns fire 4 bullets each and the rest 9 guns fire 5 bullets each); the Guangxi Liuzhou public security office provides 12 warheads fired by two seven-seven handguns; two six four-gun firing 12 warheads. The 7.62mm warheads involved in the comparison test total 24 gun fired 113 warheads. And scanning, collecting and processing rifling marks on the surface of the warhead according to the method, and performing one-to-many comparison calculation.
Here we agree that the comparison result is excellent if all the bullet scan data comparison sequences fired by the same gun are all at the first 10%, i.e. the first 10 bits; the comparison and sequencing of 75% of bullet scanning data of the same gun is 10% before, and the comparison result is recorded as good; comparing and sequencing 50% of bullet scanning data to the first 10%, and marking the comparison result as pass; the rest is the failed case. The statistical results are shown in tables 1-3.
Tables 1-3 comparison result statistics
According to the invention, rifling trace data scanning and acquisition are carried out on 321 warheads shot by 69 different guns, and comparison calculation analysis experiments are carried out on the acquired data, so that the detection comparison fine rate reaches more than 80%, and the actual comparison effect is good.
The method has the advantages of high measurement precision, high analysis and comparison speed, no damage to the texture of the warhead surface and the like, and meets the requirement of measuring the three-dimensional texture information of the bullet mark. In the firearm filing process, the digital storage and automatic comparison of the bullet data can be realized by scanning and collecting the rifling trace data on the surface of the bullet. Meets the working requirements of realizing rapid, accurate and automatic comparison of the trace of the bullet rifling. The gun and bullet rifling trace comparison system based on the laser measurement technology comprises the fields of optics technology, image processing technology, bullet trace comparison and other relevant knowledge.
Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention and not for limiting the same; although the present application has been described in detail with reference to preferred embodiments, it should be understood by those of ordinary skill in the art that: the specific embodiments of the present disclosure may be modified or some technical features may be replaced equivalently; without departing from the spirit of the present technical solution, it should be covered in the scope of the technical solution claimed in the present application.
Claims (1)
1. Be applied to single-point laser detection's line trace contrast system, its characterized in that: the comparison system is applied to criminal investigation, bullet trace detection and other scenes needing trace comparison, the line trace comparison system applied to single-point laser detection comprises system hardware and system software, the system software comprises detection data exception processing, data noise reduction, rotation correction and comparison model establishment, and the system hardware is used for data acquisition of system detection and bullet fixing;
the detection data abnormality processing adopts a correction box diagram for processing, wherein the correction box diagram is an improved diagram after correcting an abnormal value on the basis of the box diagram;
smoothing time domain data by adopting an exponential moving average method so as to eliminate tiny saw teeth in the data to the greatest extent and obtain a relatively stable coarse coherent peak trend;
in the actual noise reduction process, multiple layers of decomposition are required for the original signal f (t), and in the decomposition process, the original signal is decomposed into two parts, and assuming that n layers of decomposition are performed, the composition of the original signal can be described as follows:
wherein, the layered information can be represented by the correspondence of each node and the original formula:
a i is an approximation of the nth layer;
d i detail data of the ith layer;
f is original data;
the rotation correction is defined according to the actual conditions:
assume that the trace signal with length n is scanned to be S= { S 1 ,s 2 ,…,S mid ,…,s n And mid is the site therein, then the tilt is defined as follows:
the value of RotateRange is not 0, so that corresponding rotation correction work is needed for each input trace signal, the rotation correction work carries out corresponding rotation processing work on the input signal according to the value of RotateRange, and the rotation correction is carried out on the basis of the known RotateRange according to different trace positions, wherein the correction mode for each point is as follows:
wherein news is i For the corrected value, the value far away from the middle end is corrected to a large extent, and the value close to the middle is limited in correction amplitude;
the comparison model establishment comprises the steps of comparison pretreatment in step 1, parameter training in step 2, similarity settlement in step 3 and merging and outputting in step 4; the step 1 of comparison pretreatment is that before similarity comparison is carried out on signals of traces after noise reduction, the problems of length uncertainty and partial overlapping are firstly required to be treated: 1) Setting input data as A and B respectively, wherein the data are required data; 2) Setting a minimum length L for comparison, namely selecting a part from the A to the B from the longest length to the shortest length for comparison if the two have to meet the minimum overlapping length, and selecting different positions for comparison for multiple times; 3) Iteratively executing comparison of each position, comparing the variance of the difference degrees of the corresponding positions of each comparison, and recording the current state if the variance is the smallest; 4) After the function of 3) is completed, the roles of A, B are exchanged, and 2 and 3 are continuously completed once; 5) Calculating the difference degree with the minimum variance, and outputting a comparison result;
the parameter training (1) ensures that the sample library establishment step is completed, the samples have enough representativeness and discrimination, the machine learning is carried out by adopting a graph rolling neural network algorithm, the graph rolling neural network algorithm comprises 1) establishing a training set, 2) adjusting parameters, establishing a graph rolling neural network model, 3) finishing the parameter training after obtaining learned parameters, and 4) carrying out the parameter training by using a more targeted training set when the sample library is changed or the scene is changed; sample library creation follows the following steps: the method comprises the steps of (1) determining the types, the ranges and the types of tools to be identified, numbering each tool by using a unified rule, recording parameter information in detail, (2) detecting a single broken trace at least twice in order to eliminate the accidental during detection, determining qualified data when the coincidence degree of signal data of the two times can reach more than 99%, recording different data of a plurality of positions by the same broken trace, wherein each position adopts the mode as above, (3) testing by adopting the data of the sample library as test data after the acquisition of all the data is completed, wherein the test result is that the data similarity degree of each sample data and own group data is obviously higher than that of other groups, if the sample data and the own group data are mixed together, the data are invalid, and repeating the steps (1) - (2);
the specific method for setting up the graph convolution neural network model is that G= (V, E) and V represent node sets, namely
E represents the edge set, i.e.)>
The graph convolution neural network model consists of two parts: 1) A GCN component responsible for sampling all node information in the K-order neighborhood, 2) a self-encoder (AE) component for extracting hidden features of the activation value matrix A learned by the GCN component and preserving node cluster structure in combination with Laplacian feature map (LE), the GCN component utilizing a graph convolution neural network to segment->
For central sampling of structure and characteristic information of all nodes in K steps, namely encoding K-order neighborhood information, generating an activation value matrix A as an input of a self-encoder component by combining with label training of the nodes, enabling GCN to encode local structure and characteristic information of a network simultaneously through supervised learning based on node labels, omitting secondary structure information with small influence on low-dimensional vectors of the generated nodes outside the K-order neighborhood, utilizing the activation value matrix A obtained by GCN as an input of the self-encoder, further extracting the characteristic information from the A by the self-encoder through an unsupervised learning mode, and mapping an original network to one by combining with Laplace characteristic mappingA low dimensional space;
different algorithms in the similarity calculation in the step 3 calculate the degree of difference, the degree of difference or the similarity, and the obtained result values do not all fall between 0 and 1, so that relevant result mapping methods are formulated for different types of algorithms;
wherein it is assumed that the distance d of the normalized unit vector is calculated i,j Taking into account d i,j The value of (2) is between 0 and 1, and the distance is 0 when the values are identical, then:
H＝1-d i,j
if it is the overlap ratio p i,j The same value falls before 0 and 1, and is the most similar value as 1
H＝1-p o,j
If the result of the calculation is v which cannot be distributed between 0 and 1 i,j Then there are:
wherein w is an empirical weight, b is a bias, and specific parameters are determined by specific programs;
the step 4 is combined and output, provided with n algorithm strategies, and the result vector obtained by calculation
Then we also give here a weight of each policy +.>
/>
The end result is then:
how to determine the weight
There are mainly two methods
(1) Expert knowledge is combined with an actual experimental result to determine an experience value;
(2) Gradient descent method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010346738.9A CN112381109B (en) | 2020-04-27 | 2020-04-27 | Line trace comparison system applied to single-point laser detection |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010346738.9A CN112381109B (en) | 2020-04-27 | 2020-04-27 | Line trace comparison system applied to single-point laser detection |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112381109A CN112381109A (en) | 2021-02-19 |
| CN112381109B true CN112381109B (en) | 2023-05-05 |
Family
ID=74586363
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010346738.9A Active CN112381109B (en) | 2020-04-27 | 2020-04-27 | Line trace comparison system applied to single-point laser detection |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112381109B (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105674887A (en) * | 2016-01-19 | 2016-06-15 | 昆明理工大学 | Characteristic coincidence degree comparison method for onsite defect linear trace |
| CN107421463A (en) * | 2017-06-09 | 2017-12-01 | 李刚 | A kind of automatic computer recognition system for bullet marks and its shoot mark comparison method |
| CN108921875A (en) * | 2018-07-09 | 2018-11-30 | 哈尔滨工业大学（深圳） | A kind of real-time traffic flow detection and method for tracing based on data of taking photo by plane |
| CN110555389A (en) * | 2019-08-09 | 2019-12-10 | 南京工业大学 | bullet line-bore trace identification method based on ridgelet transformation and rotation matching |
Family Cites Families (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105243454A (en) * | 2015-11-06 | 2016-01-13 | 广州威沃电子有限公司 | Big data-based electrical load prediction system |
| CN105259178B (en) * | 2015-11-20 | 2019-03-15 | 云南卡索实业有限公司 | A kind of shearing class linear trace laser detection system |
| CN105787487B (en) * | 2016-02-24 | 2019-04-12 | 昆明理工大学 | A kind of Similarity Match Method of cutting implement picture |
| CN106709527B (en) * | 2017-01-09 | 2020-03-10 | 钟少童 | Tool trace storage and automatic comparison system and method |
| CN107730044A (en) * | 2017-10-20 | 2018-02-23 | 燕山大学 | A kind of hybrid forecasting method of renewable energy power generation and load |
| CN108095708A (en) * | 2018-01-19 | 2018-06-01 | 动析智能科技有限公司 | A kind of physiology monitoring and analysis method, system based on mixing sensing |
| US11269943B2 (en) * | 2018-07-26 | 2022-03-08 | JANZZ Ltd | Semantic matching system and method |
-
2020
- 2020-04-27 CN CN202010346738.9A patent/CN112381109B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105674887A (en) * | 2016-01-19 | 2016-06-15 | 昆明理工大学 | Characteristic coincidence degree comparison method for onsite defect linear trace |
| CN107421463A (en) * | 2017-06-09 | 2017-12-01 | 李刚 | A kind of automatic computer recognition system for bullet marks and its shoot mark comparison method |
| CN108921875A (en) * | 2018-07-09 | 2018-11-30 | 哈尔滨工业大学（深圳） | A kind of real-time traffic flow detection and method for tracing based on data of taking photo by plane |
| CN110555389A (en) * | 2019-08-09 | 2019-12-10 | 南京工业大学 | bullet line-bore trace identification method based on ridgelet transformation and rotation matching |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112381109A (en) | 2021-02-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109903282B (en) | Cell counting method, system, device and storage medium | |
| CN110906875B (en) | Visual processing method for aperture measurement | |
| CN110928237B (en) | Vibration signal-based numerical control machining center flutter online identification method | |
| CN108550166B (en) | Spatial target image matching method | |
| CN111126134A (en) | Radar radiation source deep learning identification method based on non-fingerprint signal eliminator | |
| CN114091606A (en) | Tunnel blasting blast hole half-hole mark identification and damage flatness evaluation classification method | |
| JP5023238B2 (en) | How to select the best trace area for shell case-based automatic region segmentation and shell case comparison | |
| CN114663373A (en) | Point cloud registration method and device for detecting surface quality of part | |
| CN112381109B (en) | Line trace comparison system applied to single-point laser detection | |
| Zhang et al. | Convergence‐improved congruent matching cells (CMC) method for firing pin impression comparison | |
| CN113111706B (en) | SAR target feature unwrapping and identifying method for azimuth continuous deletion | |
| Smith et al. | Optical imaging techniques for ballistics specimens to identify firearms | |
| CN110765668B (en) | Concrete penetration depth test data abnormal point detection method based on deviation index | |
| CN112652003A (en) | Three-dimensional point cloud registration method based on RANSAC measure optimization | |
| CN110990383A (en) | Similarity calculation method based on industrial big data set | |
| CN116311391A (en) | High-low precision mixed multidimensional feature fusion fingerprint retrieval method | |
| CN111738288B (en) | Bullet warhead fast matching method based on single-point laser detection | |
| CN113362296B (en) | Tunnel crack extraction method and system | |
| CN114139772B (en) | Method for predicting workpiece surface quality based on optimized generalized regression neural network | |
| CN113674174B (en) | Line scanning cylinder geometric correction method and device based on significant line matching | |
| Kong et al. | A firearm identification system based on neural network | |
| CN111639664A (en) | Line trace batch comparison system based on multi-strategy mode | |
| CA3200944A1 (en) | Method and system for determining a similarity or distance measure between ballistic specimens | |
| CN113820311B (en) | Water quality heavy metal monitoring method and device based on remote sensing combination inversion formula | |
| CN107862126B (en) | System reliability assessment method under component-level information diversity condition |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |