A method for recognizing bullet bottom dimple marks by combining manual features and learning features

Abstract

The present invention provides a trace recognition method combining manual features and learning features to achieve key point detection and feature description of key points. Compared with the end-to-end neural network algorithm, it not only utilizes the gradient information of the image, but also reduces the network structure parameters and parameter calculations, and can achieve the stability of the key point detection of bullet marks. Compared with the prior art, the present invention has the following technical effects: using a method combining manual features and learning features, the gradient information of the image is utilized, and the network structure parameters and parameter calculations are reduced, so that the stability of the key point detection of bullet marks can be achieved. The CNN structure of the Key.Net model is improved to make it more suitable for the feature learning and combination capabilities of bullet mark tasks; the Key.Net model is trained using a large-scale data set and migrated to a small sample of bullet mark instances, which improves the reliability of the model and the accuracy of detection.

Description

Technical Field

[0001] The invention relates to the field of bullet trace recognition and identification, and in particular to a bullet bottom dimple feature recognition method combining manual features and learning features. Background Art

[0002] When a bullet is fired, the chamber pressure causes the mechanical processing and other textures on the surface of the bullet base to be imprinted on the bottom surface of the cartridge case, forming a mark. The trace feature is relatively stable, and the processing features of individual guns are different, so the bullet base mark is distinguishable from other guns. Therefore, by comparing the bullet base marks, the bullet can be matched and traced, providing strong evidence for the public security organs to solve gun-related cases. The traditional bullet mark inspection method is compared by professional technicians, which has low identification efficiency, large workload, and large subjective influence;

[0003] With the advancement of computer vision recognition technology, countries around the world have been committed to the research of bullet mark automatic recognition systems and have achieved remarkable results, such as IBIS, ALIAS, etc. The key point feature matching method of bullet marks is mainly divided into two stages: calculating the key points of bullet mark images and describing the key point features through feature extraction methods, and matching the key point features to calculate the degree of consistency between images. However, existing feature extraction methods mostly use manual feature algorithms such as SIFT or end-to-end convolutional neural network technology. The former is unstable in extracting key points of images with unclear features such as bullet marks, and the latter is difficult to guarantee the repeatability of key point extraction in different scenarios. Summary of the invention

[0004] In order to solve the problem of robustness of feature point extraction of traditional bullet mark images, the present invention provides a trace recognition method that combines manual features and learning features to achieve key point detection and feature description of key points. Compared with the end-to-end neural network algorithm, it not only utilizes the gradient information of the image, but also reduces the network structure parameters and parameter calculation, and can achieve stability in the key point detection of bullet marks.

[0005] The present invention proposes the following technical solutions to achieve the above-mentioned purpose:

[0006] A trace recognition method combining manual features and learning features includes the following specific steps:

[0007] Step S1: The bullet mark image collected by the three-dimensional confocal microscope is cropped and filtered to highlight its roughness characteristics; the pre-processed bullet mark image samples are constructed into a total bullet mark data set, 80% of the sample set is taken as the algorithm training set, and 20% is taken as the verification set;

[0008] Step S2: Establish an improved feature point extraction network model Key.Net, which includes a manual feature point filter, a learning feature point filter, and a multi-scale spatial index suggestion layer. This method improves the original Key.Net learning feature point filter to adapt to the bullet mark task;

[0009] Step S3: Perform image affine transformation on the ImageNet large-scale dataset to generate an image pair dataset, which is input into the Key.Net twin model established in step S2 to obtain the optimal model parameters;

[0010] Step S4: Import the bullet mark training set in step S1 into the optimal parameter model in S3 to continue iterative training, and use the validation set in step S1 to test the model accuracy, freeze the Key.Net model parameters with the highest matching accuracy in the validation set, and use them as the final key point detection model;

[0011] Step S5: Collect the bullet mark image to be registered, input it into the Key.Net feature point detection model of step S4, obtain the key point response map, and then use HardNet to construct feature descriptions for the key points. By calculating the Euclidean distance of the feature descriptions between the key points, registration with the reference bullet mark image can be achieved.

[0012] Furthermore, in step S1, the sampling and filtering process includes: removing irrelevant areas around the dimple marks at the bottom of the bullet by cutting, retaining the intermediate feature concentration area, and then performing a two-dimensional Gaussian regression filter on the dimple marks at the bottom of the bullet to obtain a dimple image highlighting the roughness feature; wherein the two-dimensional Gaussian regression filter formula is:

[0013]

[0014] Wherein, t(ξ,η) is the input surface, ξ,η are the horizontal and vertical coordinates under the t surface respectively; s(x,y) is the output surface after filtering, x,y are the horizontal and vertical coordinates under the s surface respectively; ρ(r) is the square term;

[0015] Furthermore, in step S2, the Key.Net model includes a multi-layer structure, and the CNN module uses He weight initialization and L2 kernel regularization:

[0016] Step S2.1: To construct a three-level scale space, the input image (size is W×H) is downsampled at three scale levels, and the layer-by-layer sampling coefficient and Gaussian blur coefficient are 1.2;

[0017] Step S2.2: The manual feature point filter provides an anchor point structure for the CNN filter, locates, scores, and sorts the feature points with strong repetitiveness, and draws on the Harris and Hession corner detection algorithms: the corner points are in the area where the grayscale changes rapidly, and their gradients and gradient change rates are large. The prominent corner features are obtained by calculating the first and second order derivatives of the images at each scale, and the Soble operator is used to calculate d x , d y , d xx , d yy , d xy , d xx ×d yy , d x ×d y There are 10 parameters in total, constituting an output tensor of 10 channels;

[0018] Step S2.3: The learning feature filter consists of three basic learning blocks and one separate convolution block. This method is based on the original Key.Net learning module and improves it. The Inception module is constructed using 1×1, 3×3, and 5×5 convolution kernels to replace the original convolution block so that the model can adaptively select the convolution kernel size. The activation layer is changed to the Sigmoid function, and its function is:

[0019]

[0020] The basic learning block consists of an Inception module, a batch normalization layer, and an activation function layer. The input channel of the first learning block is 10, and the input and output channels of the remaining learning blocks are set to 8.

[0021] The individual convolutional blocks consist of a convolutional layer with a kernel size of 3×3 and a ReLU function. Its output channel is 1, that is, the output is the response feature map of the bullet mark image. The network scores each pixel.

[0022] Step S2.4: The output images obtained after the images at the three scale levels are processed by the manual feature layer and the learning feature layer are respectively upsampled, and each layer obtains a different scale feature response map (size is W×H), which are cascaded into a 3-channel scale space image;

[0023] Step S2.5: This method obtains the true value of the image based on affine transformation, which is used to train the entire model. The single index proposal layer divides the image feature response map into grids multiple times with a grid size of N×N, and calculates a key point of each grid based on the true value of the affine transformation. The loss function on a single scale is:

[0024]

[0025]

[0026] in and is an image With image The characteristic point response map, H ab Representing images and There is an affine transformation matrix between them, with parameter α i Determine the value of the response value when calculating the loss based on the position of each feature point;

[0027] The multi-scale loss function is expressed as:

[0028]

[0029] Where s is the scale factor, N s ∈[8, 16, 24, 32, 40] is the window size, λ s ∈[256, 64, 16, 4, 1] is the control parameter when the scale factor is s, N s and λ s Determined by performing a hyperparameter search on the validation set.

[0030] Furthermore, in step S4, the registration between the image to be registered and the reference image is performed by detecting and describing the key points of the key points using Key.Net+HardNet (obtaining the coordinates of the key points and the 128-dimensional feature vectors of the key points), and the Euclidean distance of the feature vectors of each key point between the two images is calculated to be less than 0.7, that is, it is determined to be a matching key point. If the number of matching key points of the two images exceeds a certain value, it can be determined as a correctly registered bullet mark image;

[0031] The Euclidean distance formula is:

[0032] min(E w (x1, x2)) = || G W (x1)-G w (x2)||

[0033] Among them, x1 is the reference bullet trace image, x2 is the image to be registered, G w is the model mapping function, E w is the Euclidean distance.

[0034] The present invention adopts the above technical means and has the following technical effects compared with the prior art:

[0035] 1. The combination of manual features and learning features utilizes the gradient information of the image, reduces the network structure parameters and parameter calculation, and can achieve the stability of key point detection of bullet marks.

[0036] 2. Improve the CNN structure of the Key.Net model to make it more suitable for feature learning and combination capabilities of bullet scar tasks;

[0037] 3. The Key.Net model is trained using a large-scale data set and migrated to a small sample of bullet marks, which improves the reliability of the model and the accuracy of detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is an overall flow chart of the bullet mark registration method provided by the present invention.

[0039] Figure 2 It is an improved Key.Net model and scale space flow chart implemented by the present invention.

[0040] Figure 3 It is a diagram of the improved Key.Net network structure implemented by the present invention.

[0041] Figure 4 It is a feature point detection map in a multi-scale space provided by the present invention.

[0042] FIG. 5( a ) is a diagram showing the key points of bullet marks of the model of the present invention.

[0043] Figure 5(b) is the bullet mark key point matching diagram described by SIFT. DETAILED DESCRIPTION

[0044] The embodiments of the present invention are described in detail below in conjunction with the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more definite definition of the protection scope of the present invention.

[0045] The present invention provides a bullet base dimple feature recognition method combining manual features and learning features, aiming to overcome the problem of poor repeatability of end-to-end key point detection network in different scenarios and improve the stability of bullet mark key point detection process.

[0046] In order to solve the above technical problems, the present invention proposes a bullet mark feature point recognition method, which is described in detail below in conjunction with an embodiment:

[0047] (1) The specific process of the bullet mark recognition method of this method is as follows Figure 1As shown, in this example, the Fadul, Weller and other dimple mark datasets are used as a joint dataset, which has a total of 158 marks, with 2-5 marks for each type; the center needle impact mark of the above bullet mark image is cut out in Matlab, and the characteristic concentration area of ​​the bullet bottom dimple mark is retained; then the dimple mark image is processed by two-dimensional Gaussian regression filtering to highlight the roughness characteristics of the printing pressure; 120 samples are taken as the training set and 38 samples are taken as the verification set, among which the traces from the same gun are taken as the true value pairs for training.

[0048] (3) The images of the ImageNet ILSVRC 2012 dataset are affine transformed to generate positive image pairs, a total of 12,000 pairs of size 192×192, of which 9,000 pairs are used as training data and 3,000 pairs are used as verification data to guide the loss function calculation to update the network weights.

[0049] (4) Using Python and TensorFlow framework, we build a feature point extraction network model Key.Net, which includes manual feature point filter, learning feature point rate filter, and multi-scale spatial index suggestion layer. Figure 3 As shown in the figure; the Key.Net model is initialized with He weights and L2 kernel regularization; the ImageNet dataset is used as the model input to two Key.Net instances to share weights and iteratively update the weights simultaneously until the optimal model parameters are obtained.

[0050] (5) Import the bullet mark training set into the optimal parameter model in step (4) to continue iterative training, and freeze the Key.Net model parameters with the highest matching accuracy of the validation set as the final key point detection model;

[0051] (6) There are two additional dimple samples I1 and I2, both of size 1000 × 1000. Key point detection is performed using the trained Key.Net model, with the scale factor set to θ i ={0.4, 0.6, 0.8, 1, 1.2, 1.4}, i = 1, ..., 6, that is, N = 6, Key.Net detects all pixel-level feature points on multiple scale planes, and the number of feature points of the two images on each scale plane is Images I1 and I2 are then described by HardNet, and the matching algorithm uses the RANSAC filtering method. As shown in Figure 5(a) and Figure 5(b), the number of correctly matched key points is much higher than the number of matching pairs of the SIFT method.

[0052] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A trace recognition method combining manual features and learning features, characterized by The specific steps are as follows: Step S1: cropping and filtering the bullet trace image collected by a three-dimensional confocal microscope to highlight its roughness characteristics; The preprocessed bullet scar image samples are constructed into a total bullet scar data set, 80% of the sample set is taken as the algorithm training set, and 20% is taken as the validation set; Step S2: Establish an improved feature point extraction network model Key.Net, which includes a manual feature point filter, a learning feature point filter, and a multi-scale spatial index suggestion layer, and improves the original Key.Net learning feature point filter to adapt to the bullet mark task; Step S3: Perform image affine transformation on the ImageNet large-scale dataset to generate an image pair dataset, which is input into the Key.Net twin model established in step S2 to obtain the optimal model parameters; Step S4: Import the bullet mark training set in step S1 into the optimal parameter model in S3 to continue iterative training, and use the validation set in step S1 to test the model accuracy, freeze the Key.Net model parameters with the highest matching accuracy in the validation set, and use them as the final key point detection model; Step S5: Collect the bullet mark image to be registered, input it into the Key.Net feature point detection model of step S4, obtain the key point response map, and then use HardNet to construct feature descriptions for the key points. By calculating the Euclidean distance of the feature descriptions between the key points, registration with the reference bullet mark image can be achieved.

2. The trace recognition method combining manual features and learning features according to claim 1 is characterized in that: In step S1, the sampling and filtering process includes: removing irrelevant areas around the dimple marks at the bottom of the bullet by cutting, retaining the middle feature concentration area, and then performing two-dimensional Gaussian regression filtering on the dimple marks at the bottom of the bullet to obtain a dimple image highlighting the roughness characteristics; wherein the two-dimensional Gaussian regression filtering formula is: Among them, t(ξ,η) is the input surface, ξ,η are the horizontal and vertical coordinates under the t surface respectively; s(x,y) is the output surface after filtering, x,y are the horizontal and vertical coordinates under the s surface respectively; ρ(r) is the square term.

3. The trace recognition method combining manual features and learning features according to claim 1 is characterized in that: In step S2, the Key.Net model contains a multi-layer structure, and the CNN module uses He weight initialization and L2 kernel regularization: Step S2.1: To construct a three-level scale space, the input image, with a size of W×H, is downsampled at three scale levels, and the layer-by-layer sampling coefficient and Gaussian blur coefficient are 1.2; Step S2.2: The manual feature point filter provides an anchor point structure for the CNN filter to locate, score and sort the feature points with strong repetitiveness. The corner point detection algorithm is based on Harris and Hession: the corner point is in the area where the grayscale changes rapidly, and its gradient and gradient change rate are large. The prominent corner features are obtained by calculating the first and second order derivatives of the image at each scale, and the Soble operator is used to calculate the corner point. d x ×d y There are 10 parameters in total, constituting an output tensor of 10 channels; Step S2.3: The learned feature filter consists of three basic learning blocks and one separate convolution block. Improvements are made based on the original Key.Net learning module. 1×1, 3×3, and 5×5 convolution kernels are used to build the Inception module to replace the original convolution block so that the model can adaptively select the convolution kernel size. The activation layer is changed to the Sigmoid function, whose function is: The basic learning block consists of an Inception module, a batch normalization layer, and an activation function layer. The input channel of the first learning block is 10, and the input and output channels of the remaining learning blocks are set to 8. The individual convolutional blocks consist of a convolutional layer with a kernel size of 3×3 and a ReLU function. Its output channel is 1, that is, the output is the response feature map of the bullet mark image. The network scores each pixel. Step S2.4: The output images obtained after the images at the three scale levels are processed by the manual feature layer and the learning feature layer are respectively upsampled, and each layer obtains a different scale feature response map with a size of W×H, which is cascaded into a 3-channel scale space image; Step S2.5: The image truth value is obtained based on the affine transformation, which is used to train the entire model. The single index proposal layer divides the image feature response map into grids multiple times with a grid size of N×N, and calculates a key point of each grid based on the true value of the affine transformation. The loss function on a single scale is: in and is an image With image The characteristic point response map, H ab Representing images and There is an affine transformation matrix between them, with parameter α i Determine the value of the feature point response map when calculating the loss based on the position of each feature point; The multi-scale loss function is expressed as: Where s is the scale factor, N s ∈[8, 16, 24, 32, 40] is the window size, λ s ∈[256, 64, 16, 4, 1] is the control parameter when the scale factor is s, N s and λ s Determined by performing a hyperparameter search on the validation set.

4. The trace recognition method combining manual features and learning features according to claim 1 is characterized in that: In the step S4, the registration between the image to be registered and the reference image is performed by detecting and describing the key points using Key.Net+HardNet, obtaining the key point coordinates and the 128-dimensional feature vectors of the key points, and calculating the Euclidean distance of the feature vectors of each key point between the two images is less than 0.7, that is, it is determined to be a matching key point. If the number of matching key points of the two images exceeds a certain value, it can be determined as a correctly registered bullet mark image; The Euclidean distance formula is: min(E w (x1,x2))=||G W (x1)-G w (x2)|| Among them, x1 is the reference bullet trace image, x2 is the image to be registered, G W is the model mapping function, E w is the Euclidean distance.

Images