Bullet trace recognition algorithm of rotation invariant coordinate convolutional neural network

Abstract

The invention discloses a bullet trace identification algorithm based on a rotation-invariant coordinate convolutional neural network. The algorithm is used for realizing key point detection and feature matching of bullet traces. The algorithm comprises the following steps: collecting bullet traces of gun shooting; processing the bullet mark data by using a confocal microscope; reducing image noise through a three-dimensional spline filter; calling a coordinate generation function to generate coordinate points and establishing a coordinate system; carrying out image feature extraction on the trace image by using a self-supervised interest point detection and description algorithm; a non-maximum suppression algorithm is used for post-processing in feature point detection; and rotating the image and generating a new coordinate system, and performing feature matching on the trace image by using a rotation invariant coordinate convolutional neural network under the new coordinate system so as to judge the matching degree of the bullet. The method can further provide support for bullet trace matching and further provide guarantee for case detection.

Description

Technical Field

[0001] The present invention relates to the field of recognition and identification of bullet traces, and in particular to a bullet trace recognition algorithm of a rotationally invariant coordinate convolutional neural network. Background Art

[0002] When a bullet passes through a gun barrel after being fired, it is affected by the rifling, which is usually spiral and is designed to make the bullet rotate to increase its stability and accuracy. The bullet will produce friction when it contacts the metal surface on the inner wall of the gun barrel. This friction causes tiny scratches and grooves on the surface of the bullet, forming specific marks. These marks include rifling marks caused by the shape and structure of the rifling of the gun barrel and scratches caused by tiny unevenness of the inner wall of the gun barrel. The scratches and impressions left on the surface of the bullet and the shell can link the evidence bullet with the gun that fired it, and assist in case detection. At present, manual inspection dominates the vast majority of gun and bullet trace identification institutions at home and abroad. However, the manual inspection process is time-consuming and difficult to meet the needs of linking and combining cases. In addition, the objectivity and scientificity of the manual inspection method are widely questioned.

[0003] With the rapid development of computer technology and artificial intelligence, methods such as machine learning and deep learning have been applied to the field of bullet trace matching and have achieved remarkable results, such as CMC, ALIAS, etc. The feature matching of bullet traces is mainly divided into two stages: extracting feature key points and matching feature points. At present, the mainstream methods for extracting feature key points include SIFT, ORB, SURF, etc., but SIFT has a high computational complexity and is not friendly to real-time applications. ORB performs poorly in scenes with a small number of special points or complex textures, and SURF is sensitive to changes in feature points in the image. Summary of the invention

[0004] In order to solve the problem that most of the existing bullet trace image recognition algorithms are not rotationally invariant, the present invention provides a bullet trace recognition algorithm of a rotationally invariant coordinate convolutional neural network to achieve rotational invariance of image recognition. Compared with traditional feature extraction and matching methods, the SuperPoint algorithm it uses is more efficient and accurate. Moreover, with the support of the rotationally invariant coordinate convolutional neural network, even if the image is rotated and offset, feature matching can still be performed, which improves the upper limit of bullet trace recognition and provides more possibilities for solving cases.

[0005] The present invention proposes the following technical methods to achieve the above objectives:

[0006] A bullet trace recognition algorithm combining the SuperPoint model and the rotation-invariant coordinate convolutional neural network is proposed. The specific steps are as follows:

[0007] Step (1): collecting bullet traces and using a three-dimensional confocal microscope to image the bullet traces;

[0008] Step (2): filtering the image obtained in step (1) to obtain a preprocessed image data set;

[0009] Step (3): Use the GC function to batch generate coordinates and establish a coordinate system;

[0010] Step (4): Establish a SuperPoint model. SuperPoint consists of three parts: a VGG-style encoder, an interest point decoder, and a descriptor decoder. This method has improved the SuperPoint method and applied it to bullet mark recognition. The image features obtained in step (2) are extracted through the SuperPoint algorithm.

[0011] Step (5): Use the NMS algorithm to filter the feature points in step (4) to improve the matching accuracy;

[0012] Step (6): Rotate the image in step (3) to generate new feature points and a new coordinate system, calculate the coordinate system through a rotation-invariant coordinate convolutional neural network, and achieve matching of the bullet mark image in the new coordinate system after rotation, so as to achieve the purpose of bullet mark judgment;

[0013] Furthermore, in the step (1), the irrelevant areas of the pit mark are removed by three-dimensional confocal microscopy, and the microscopic detail areas are retained to prepare for subsequent filtering;

[0014] Furthermore, in step (2), the image is subjected to spline filtering to reduce image noise, highlight the microscopic features of bullet marks, and obtain a bullet mark data set, wherein the three-dimensional spline filtering formula is:

[0015] S i (x) = a i +b i (xx i )+c i (xx i ) 2 +d i (xx i ) 3 ;

[0016] Where (x 0 ,y 0 ),(x 1 ,y 1 ),…,(x n ,y n ) is a given point, x i ≤x≤x i+1,i=0,1,...,n-1。 Through spline filtering, image smoothing can be achieved and high-frequency noise can be reduced.

[0017] Furthermore, in step (3), a batch of specified number of coordinates is generated by using the GC function, the center position is calculated, the offsets in the X and Y directions are calculated respectively, the point coordinates are calculated, and the coordinate system is established. The specific steps are as follows:

[0018] 1) Define the image as a multi-channel function Where X = (x 1 ,x 2 ) T ;

[0019] 2) In the center of the image A Cartesian coordinate system is established at the position, so there is always X c =(0,0) T ;

[0020] 3) The positive integer m represents the channel dimension of F(x). For any k∈(1,2,...,m), the scalar function f k (X) as a mapping from the domain Ω to R;

[0021] 4) Generate coordinates using the GenerateCoordinates function;

[0022] 5) Calculate the center position, calculate the X and Y direction offset, and calculate the point coordinates;

[0023] 6) Establish coordinate center and coordinate system;

[0024] Furthermore, in the step (4), in the SuperPoint model, a neural network forward propagation is established, and a variable convolution layer is used to process input data, and convolution data is obtained after nonlinear transformation;

[0025] Furthermore, in step (5), the NMS algorithm is used to screen the feature points in step (4) to improve the matching accuracy, and the specific steps are as follows:

[0026] 1) Sort the corner point array from high to low according to the scores of the corner points;

[0027] 2) Create a Boolean mask to mark the selected corner points;

[0028] 3) Traverse each corner point in the sorted corner point array and compare it with the threshold;

[0029] 4) Return mask;

[0030] Furthermore, in step (6): the obtained image is rotated and a new coordinate system is generated. On this basis, a rotation-invariant coordinate convolutional neural network is used to match new feature points to achieve generation of a new coordinate system and calculation. The specific steps are as follows:

[0031] 1) Image F(X) around the origin X c Rotate θ degrees to We can get G(Y)=F(X), where:

[0032]

[0033] 2) For a given point Designing a new coordinate system and in Sampling is performed at (2n+1)·(2n+1) sampling points K, and the sum of the sampled values ​​weighted by a (2n+1)·(2n+1) convolution kernel W is calculated.

[0034] 3) To obtain rotation-invariant sampling, uniformly distribute the 0 is a circle with a center and a radius of r, where r = 1, 2, ..., n, i = 0, 1, ..., (8r-1), The first sampling point is along the X axis from the origin c To X 0 Take a sample. Get the value divided by X 0 Other than (4n 2 +4n) sample points Q are as follows:

[0035]

[0036] 4) Replace all conventional convolutions in the traditional convolution model with rotation-invariant coordinate convolutions, and rotate the image around the center X c Rotate to generate new coordinate points and coordinate systems;

[0037] 5) Calculate and match the coordinates of new feature points through the feature descriptor (MatchDescriptors) function in the new convolutional neural network;

[0038] 6) Calculate the Euclidean distance E of the new feature point W , when E W If the number of feature matching points of two images exceeds a certain number, they are considered homologous images. The Euclidean distance formula is:

[0039] min(E W (x 1 ,x 2 ))=||G W (x 1 )-G w(x 2 )||;

[0040] Where E W is the Euclidean distance, x 1 ,x 2 are two images to be matched, G W is the model mapping function;

[0041] 7) Use feature point function and feature line drawing function to display the correspondence between feature matching points and realize rotation matching of bullet trace images.

[0042] The technical advantage of the present invention is that the present invention provides an image rotation recognition method of a rotation-invariant coordinate convolutional neural network, generates coordinate feature points and establishes a coordinate system using a coordinate generation function, extracts feature information of the image using a SuperPoint algorithm, processes subsequent feature extraction steps using a non-maximum suppression algorithm, rotates the image and generates a new coordinate system, and uses a coordinate convolutional neural network to recalculate and match the rotated image. It increases the image matching angle, improves the upper limit of bullet trace image matching, can further provide support for bullet trace matching, and further provide guarantees for solving cases. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a full flow chart of the bullet trace detection provided by the present invention.

[0044] Figure 2 Schematic diagram of the SuperPoint algorithm used in the present invention.

[0045] Figure 3 Schematic diagram of the rotation invariant coordinate convolutional neural network algorithm of the present invention.

[0046] Figure 4(a) is a schematic diagram of feature matching of the traditional convolution algorithm when the bullet mark image is rotated 60°.

[0047] FIG4( b ) is a schematic diagram of feature matching of the invented rotation-invariant coordinate convolutional neural network algorithm when the bullet mark image is rotated 60°.

[0048] Figure 5(a) is a schematic diagram of feature matching of the traditional convolution algorithm when the bullet mark image is rotated 90°.

[0049] FIG5( b ) is a schematic diagram of feature matching of the invented rotation-invariant coordinate convolutional neural network algorithm when the bullet mark image is rotated 90°.

[0050] Figure 6(a) is a schematic diagram of feature matching of the traditional convolution algorithm when the bullet mark image is rotated 120°.

[0051] FIG6( b ) is a schematic diagram of feature matching of the invented rotation-invariant coordinate convolutional neural network algorithm when the bullet mark image is rotated 120°.

[0052] Figure 7(a) is a schematic diagram of feature matching of the traditional convolution algorithm when the bullet mark image is rotated 150°.

[0053] FIG7( b ) is a schematic diagram of feature matching of the invented rotation-invariant coordinate convolutional neural network algorithm when the bullet mark image is rotated 150°.

[0054] Figure 8(a) is a schematic diagram of feature matching of the traditional convolution algorithm when the bullet mark image is rotated 180°.

[0055] FIG8( b ) is a schematic diagram of feature matching of the invented rotation-invariant coordinate convolutional neural network algorithm when the bullet mark image is rotated 180°. DETAILED DESCRIPTION

[0056] The technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present application, not all of the embodiments. Based on the embodiment of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0057] The present invention provides a gunshot trace recognition algorithm based on a rotation-invariant coordinate convolutional neural network, which solves the problem that the traditional algorithm does not have rotation invariance in image recognition and improves the upper limit of gunshot trace image matching. The following is a detailed description with examples:

[0058] A bullet trace recognition algorithm combining the SuperPoint model and the rotation-invariant coordinate convolutional neural network is proposed. The specific steps are as follows:

[0059] Step (1): collecting bullet traces and using a three-dimensional confocal microscope to image the bullet traces;

[0060] Step (2): filtering the image obtained in step (1) to obtain a preprocessed image data set;

[0061] Step (3): Use the GC function to batch generate coordinates and establish a coordinate system;

[0062] Step (4): Establish a SuperPoint model. SuperPoint consists of three parts: a VGG-style encoder, an interest point decoder, and a descriptor decoder. This method has improved the SuperPoint method and applied it to bullet mark recognition. The image features obtained in step (2) are extracted through the SuperPoint algorithm.

[0063] Step (5): Use the NMS algorithm to filter the feature points in step (4) to improve the matching accuracy;

[0064] Step (6): Rotate the image in step (3) to generate new feature points and a new coordinate system, calculate the coordinate system through a rotation-invariant coordinate convolutional neural network, and achieve matching of the bullet mark image in the new coordinate system after rotation, so as to achieve the purpose of bullet mark judgment;

[0065] Furthermore, in the step (1), the irrelevant areas of the pit mark are removed by three-dimensional confocal microscopy, and the microscopic detail areas are retained to prepare for subsequent filtering;

[0066] Furthermore, in step (2), the image is subjected to spline filtering to reduce image noise, highlight the microscopic features of bullet marks, and obtain a bullet mark data set, wherein the three-dimensional spline filtering formula is:

[0067] S i (x) = a i +b i (xx i )+c i (xx i ) 2 +d i (xx i ) 3 ;

[0068] Where (x 0 ,y 0 ),(x 1 ,y 1 ),…,(x n ,y n ) is a given point, x i ≤x≤x i+1 ,i=0,1,...,n-1。 Through spline filtering, image smoothing can be achieved and high-frequency noise can be reduced.

[0069] Furthermore, in step (3), a batch of specified number of coordinates is generated by using the GC function, the center position is calculated, the offsets in the X and Y directions are calculated respectively, the point coordinates are calculated, and the coordinate system is established. The specific steps are as follows:

[0070] 1) Define the image as a multi-channel function Where X = (x 1 ,x 2 ) T ;

[0071] 2) In the center of the image A Cartesian coordinate system is established at the position, so there is always X c =(0,0) T;

[0072] 3) The positive integer m represents the channel dimension of F(x). For any k∈(1,2,...,m), the scalar function f k (X) as a mapping from the domain Ω to R;

[0073] 4) Generate coordinates using the GenerateCoordinates function;

[0074] 5) Calculate the center position, calculate the X and Y direction offset, and calculate the point coordinates;

[0075] 6) Establish coordinate center and coordinate system;

[0076] Furthermore, in the step (4), in the SuperPoint model, a neural network forward propagation is established, and a variable convolution layer is used to process input data, and convolution data is obtained after nonlinear transformation;

[0077] Furthermore, in step (5), the NMS algorithm is used to screen the feature points in step (4) to improve the matching accuracy, and the specific steps are as follows:

[0078] 1) Sort the corner point array from high to low according to the scores of the corner points;

[0079] 2) Create a Boolean mask to mark the selected corner points;

[0080] 3) Traverse each corner point in the sorted corner point array and compare it with the threshold;

[0081] 4) Return mask;

[0082] Furthermore, in step (6): the obtained image is rotated and a new coordinate system is generated. On this basis, a rotation-invariant coordinate convolutional neural network is used to match new feature points to achieve generation of a new coordinate system and calculation. The specific steps are as follows:

[0083] 2) Image F(X) around the origin X c Rotate θ degrees to We can get G(Y)=F(X), where:

[0084]

[0085] 2) For a given point Designing a new coordinate system and in Sampling is performed at (2n+1)·(2n+1) sampling points K, and the sum of the sampled values ​​weighted by a (2n+1)·(2n+1) convolution kernel W is calculated.

[0086] 3) To obtain rotation-invariant sampling, uniformly distribute the 0 is a circle with a center and a radius of r, where r = 1, 2, ..., n, i = 0, 1, ..., (8r-1), The first sampling point is along the X axis from the origin c To X 0 Take a sample. Get the value divided by X 0 Other than (4n 2 +4n) sample points Q are as follows:

[0087]

[0088] 4) Replace all conventional convolutions in the traditional convolution model with rotation-invariant coordinate convolutions, and rotate the image around the center X c Rotate to generate new coordinate points and coordinate systems;

[0089] 5) Calculate and match the coordinates of new feature points through the feature descriptor (MatchDescriptors) function in the new convolutional neural network;

[0090] 6) Calculate the Euclidean distance E of the new feature point W , when E W If the number of feature matching points of two images exceeds a certain number, they are considered homologous images. The Euclidean distance formula is:

[0091] min(E W (x 1 ,x 2 ))=||G W (x 1 )-G w (x 2 )||;

[0092] Where E W is the Euclidean distance, x 1 ,x 2 are two images to be matched, G W is the model mapping function;

[0093] 7) Use feature point function and feature line drawing function to display the correspondence between feature matching points and realize rotation matching of bullet trace images.

[0094] The specific steps of the bullet trace recognition algorithm of the present invention are as follows: Figure 1 As shown, this example uses the Fadul bullet hole dataset as the experimental sample; the sample is cropped by MATLAB, the feature effective area is retained, and spline filtering is performed to obtain the sample dataset;

[0095] Build the Pytorch framework in the Pycharm environment, and use the Pytorch framework to build the SuperPoint model and the rotation-invariant coordinate convolutional neural network. The SuperPoint model consists of three parts: the VGG-style encoder, the interest point decoder, and the descriptor decoder.

[0096] Use the GC function to batch generate coordinates and establish a coordinate system, import the sample data into the SuperPoint model in step (4), and use this model to extract the feature points of the bullet mark image.

[0097] After extraction, the NMS algorithm is used to screen the feature points in step (4) to improve the matching accuracy.

[0098] The image is rotated to generate new point coordinates, and a new coordinate system is established. The new feature points are matched using the rotation invariant coordinate convolutional neural network, and the corresponding relationship is depicted, as shown in Figure 4(a), Figure 4(b), Figure 5(a), Figure 5(b), Figure 6(a), Figure 6(b), Figure 7(a), Figure 7(b), Figure 8(a), and Figure 8(b). The improved algorithm is much better than the algorithm before the improvement in terms of rotation angle recognition.

[0099] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A gunshot trace recognition algorithm based on a rotation-invariant coordinate convolutional neural network, the characteristics of which include Follow these steps: Step (1): collecting bullet traces and using a three-dimensional confocal microscope to image the bullet traces; Step (2): filtering the image obtained in step (1) to obtain a preprocessed image data set; Step (3): Use the GC function to batch generate coordinates and establish a coordinate system; Step (4): Establish a SuperPoint model. SuperPoint consists of three parts: VGG-style encoder, interest point decoder and descriptor decoder. The image features obtained in step (2) are extracted through the SuperPoint algorithm. Step (5): Use the NMS algorithm to filter the feature points in step (4) to improve the matching accuracy; Step (6): Rotate the image in step (3) to generate new feature points and a new coordinate system, calculate the coordinate system through a rotation-invariant coordinate convolutional neural network, and achieve matching of the bullet mark image in the new coordinate system after rotation, so as to achieve the purpose of bullet mark judgment; In the step (1), the irrelevant areas of the pit mark are removed by three-dimensional confocal microscopy, and the microscopic detail areas are retained to prepare for subsequent filtering.

2. According to the rotation-invariant coordinate convolutional neural network bullet trace recognition algorithm described in claim 1, it is characterized in that In step (2), the image is subjected to spline filtering to reduce image noise and obtain a bullet mark data set, wherein the three-dimensional spline filtering formula is: S i (x)=a i +b i (x-x i )+c i (x-x i ) 2 +d i (x-x i ) 3 ; where (x0,y0),(x1,y1),…,(x n ,y n ) is a given point, x i ≤x≤x i+1 ,i=0,1,...,n-1。 Through spline filtering, image smoothing is achieved and high-frequency noise is reduced.

3. According to the rotation-invariant coordinate convolutional neural network bullet trace recognition algorithm described in claim 1, it is characterized in that In step (3), the GC function is used to generate a batch of specified number of coordinates, calculate the center position, calculate the offset in the X and Y directions respectively, calculate the point coordinates, and establish a coordinate system. The specific steps are as follows: 1) Define the image as a multi-channel function F(X): Where X = (x1, x2) T ; 2) In the center of the image A Cartesian coordinate system is established at the position, so there is always X c =(0,0) T ; 3) The positive integer m represents the channel dimension of F(x). For any k∈(1,2,...,m), the scalar function f k (X) as a mapping from the domain Ω to R; 4) Generate coordinates using the GenerateCoordinates function; 5) Calculate the center position, calculate the X and Y direction offset, and calculate the point coordinates; 6) Establish coordinate center and coordinate system; In the step (4): in the SuperPoint model, a neural network forward propagation is established, and a variable convolution layer is used to process input data, and convolution data is obtained after nonlinear transformation.

4. According to the rotation-invariant coordinate convolutional neural network bullet trace recognition algorithm described in claim 1, it is characterized in that: In the step (5), the feature points in step (4) are screened using the NMS algorithm, and the specific steps are as follows: 1) Sort the corner point array from high to low according to the scores of the corner points; 2) Create a Boolean mask to mark the selected corner points; 3) Traverse each corner point in the sorted corner point array and compare it with the threshold; 4) Returns the mask.

5. According to the rotation-invariant coordinate convolutional neural network bullet trace recognition algorithm described in claim 1, it is characterized in that In step (6), the obtained image is rotated and a new coordinate system is generated. On this basis, a rotation-invariant coordinate convolutional neural network is used to match new feature points to generate a new coordinate system and calculate. The specific steps are as follows: 1) Image F(X) around the origin X c Rotate θ degrees to G(Y): We can get G(Y)=F(X), where: 2) For a given point Designing a new coordinate system and in Sampling is performed at (2n+1)·(2n+1) sampling points K, and the sum of the sampled values ​​weighted by a (2n+1)·(2n+1) convolution kernel W is calculated. 3) To obtain rotation-invariant sampling, uniformly distribute the samples on a circle with radius r and X0 as the center, where r = 1, 2, ..., n, i = 0, 1, ..., (8r-1), The first sampling point is along the X axis from the origin c To X0 for sampling. Get other (4n 2 +4n) sample points Q are as follows: 4) Replace all conventional convolutions in the traditional convolution model with rotation-invariant coordinate convolutions, and rotate the image around the center X c Rotate to generate new coordinate points and coordinate systems; 5) Calculate and match the coordinates of new feature points through the feature descriptor (MatchDescriptors) function in the new convolutional neural network; 6) Calculate the Euclidean distance E of the new feature point W , when E W If the number of feature matching points of two images exceeds a certain number, they are considered homologous images. The Euclidean distance formula is: min(E W (x1,x2))=||G W (x1)-G w (x2)||; Where E W is the Euclidean distance, x1, x2 are two images to be matched, G W is the model mapping function; 7) Use feature point function and feature line drawing function to display the correspondence between feature matching points and realize rotation matching of bullet trace images.

Images