CN106295668A - Robust gun detection method - Google Patents

Abstract

The invention provides a robust gun detection method, which carries out effective pretreatment and target significance detection, and searches a candidate area of a target detection object through color segmentation and screening of specific size characteristics; the method is to design a gun classifier based on the risk of classification errors of guns and non-guns; then, a plurality of multilayer deep learning neural networks are cascaded, and gun possibility is finally judged through gun statistics of candidate areas, so that a complete gun detector is finally formed. The gun detection method provided by the invention has the advantages of good anti-interference performance, strong real-time performance and high identification precision.

Description

A Robust Gun Detection Method

technical field

The invention belongs to the technical fields of computer vision, image processing and pattern recognition, and in particular relates to a robust gun detection method.

Background technique

With the increasing rampant of terrorist attacks, drug trafficking, smuggling and other criminal activities around the world, the governments of various countries have continuously strengthened the monitoring of various public places (such as airports, stations, shipping terminals, convention centers, government agencies, large stadiums and port, etc.) security checks. The security inspection system in the prior art generates X-rays to penetrate the luggage to obtain an X-ray image of the luggage, and displays the X-ray image on the screen, and then the staff judges whether the luggage contains dangerous items based on experience. This method requires the staff to have a certain experience in image judgment, and the work intensity is high, and the judgment must be wrong. For some dangerous items such as guns, the danger of missing inspection is immeasurable. With the mature development of computer vision, image processing and pattern recognition technology, the security inspection system of the prior art begins to adopt the automatic image recognition technology to automatically determine whether dangerous articles are contained in the luggage.

Contents of the invention

Aiming at the problems existing in the prior art, the object of the present invention is to propose a firearm detection method with a low rate of missed detection and false positive rate.

In order to achieve the above object, the present invention proposes a robust firearm detection method, which performs effective preprocessing and target saliency detection, and searches for candidate areas of target detection objects through color segmentation and screening of unique size features; The method is to design a gun classifier based on the risk of misclassification of guns and non-guns; then cascade multiple multi-layer deep learning neural networks, and finally judge the possibility of guns through the candidate area statistics, and finally form a complete gun detection device;

In a system composed of an X-ray machine image acquisition device and a computer, the detection method includes a training phase and a detection phase;

1. The training phase includes the following steps:

1.1. Collection of samples;

Use the X-ray machine to go through the luggage package, and manually calibrate and cut the guns in the collected pictures, and randomly cut non-gun pictures from the X-ray pictures that do not contain guns; establish a gun sample database;

1.2. Normalization processing;

Contains linear normalization of sample illumination and size, that is, normalizes the gun and non-gun images obtained in step 1.1 to a specified size, and then grayscales the size-normalized image;

1.3. Extraction of sample feature library;

1.3.1. Calculate the integral map of each sample

1.3.2. Extraction of microstructure feature library

Three microstructural features using harr features: up-down type, left-right type, oblique symmetry type, plus an edge description value as a microstructural feature, plus a mean value as a microstructural feature, a total of five microstructural feature templates to extract The high-dimensional microstructural features of gun samples, for the five types of microstructural feature vectors, are expressed as follows:

The up-and-down Harr feature is to subtract the total gray value of the upper half of the field of view from the total gray value of the lower half of the field of view in the N*N field of view, and then divide it by the number of pixels in the field of view N , to obtain a microstructure feature, assuming that the upper and lower parts are symmetrical up and down and have the same area, w represents the width of each part, h represents the height of each part, and N is a natural number greater than zero;

The left and right Harr feature is to subtract the total gray value of the left half of the field of view from the total gray value of the right half of the field of view in the N*N field of view, and then divide it by the number of pixels in the field of view N , to obtain a microstructural feature, assuming that the left half and the right half are symmetrical and have equal areas, and the definitions of w and h are the same as the upper and lower Harr features;

The oblique symmetric Harr feature is that in the N*N field of view, the total gray value of the upper left half of the field of view and the total gray value of the lower right half of the field of view are added to obtain the total gray value of the oblique direction, and the field of view Add the total gray value of the upper right part and the total gray value of the lower left part to obtain the total gray value in the reverse oblique direction, and then subtract the total gray value in the oblique direction from the total gray value in the reverse oblique direction, Then divide by the number of pixels N in the field of view , to obtain a microstructure feature, assuming that the areas of the upper left half, the lower left half, the upper right half, and the lower right half are equal, and the definitions of w and h are the same as the upper and lower Harr features;

The maximum value of the absolute number of the above three microstructural features is used as the edge description value, and this value is also used as a microstructural feature;

The mean feature is to add the gray values in all the fields of view in the N*N field of view, and then divide it by the number of pixels in the field of view N , get a microstructural feature;

1.3.3. Extraction method of microstructure features

For a sample image of 72*72 pixels, an 8*8 resolution field of view is read every 4 pixels in the horizontal direction, there can be 17 fields of view in the horizontal direction, and 17 fields of view in the vertical direction, so the resolution is 72*72 There are a total of 289 fields of view, and 5 microstructural features are calculated for each field of view, thus forming 1445 features to describe a sample;

1.4, classifier design

Use the features designed above and the DBN algorithm in deep learning to train multiple gun classifiers, and combine these multiple classifiers into a complete gun detector in a hierarchical cascade, including the following steps:

1.4.1. Initialize i=1; initialize and define the training target of each layer classifier, the target is to train one with a missed detection rate of less than 1% on the gun training set and a false positive rate on the non-gun training set Less than 20% of the classifiers, and then train a classifier with a missed detection rate of less than 20% on the gun training set and a false positive rate of less than 1% on the non-gun training set, and then combine these two classifiers into one The classifier is used as the classifier of this layer; define the target of the whole gun detector, the false negative rate on the gun training set is less than 5%, and the false positive rate on the non-gun training set is less than 1%; each DBN classifier, Two hidden layers are used, the input layer is 1445 neurons, the first hidden layer is 578 neurons, the second hidden layer is 300 neurons, and the output layer is 2 classifications. Link;

1.4.2. Train the i-th layer classifier;

1.4.3. Use the trained i-level classifier to detect the sample set, and calculate the missed detection rate and false positive rate;

Among them, the missed detection rate = the number of non-gun samples identified as guns / the total number of non-gun samples * 100%,

False positive rate = number of gun samples identified as non-guns/total number of gun samples*100%;

1.4.4. If the missed detection rate and false alarm rate do not reach the predetermined value set in step 1.4.1, then , return to step 1.4.2 to continue training, otherwise stop training;

In the detection phase, the following steps are used to determine whether the input image contains a gun:

2.1. Load the trained parameters and initialize the classifier;

2.2. Input the picture to be detected into the firearm detector obtained in step 1.4;

2.3. Perform image preprocessing on the image to be detected, using white balance and color equalization;

2.4. Scaling of the input image;

2.5. For gun target detection, use color segmentation and limit the size of connected domains to screen candidate areas;

2.6. Calculation of integral image;

2.7. Through the integral map, extract the features of the candidate area,

Use a 72*72 window sliding in the candidate area, use the integral map in the window to calculate the characteristics of the sliding window, the feature of a window is used as the feature of a sample to be predicted, and the number of windows in the candidate area is determined by the target saliency detection;

2.8. Use the trained classifier to predict the possibility of guns in the candidate area,

Specifically, in each candidate area, use a classifier to calculate whether each sliding window contains targets such as guns;

2.9. The possibility of accumulating guns in the candidate area,

In each candidate area, if the number of windows containing guns exceeds a certain threshold, it is determined that there are guns. This threshold is the number of windows in the candidate area multiplied by a coefficient, and the coefficient range is greater than 0.3 and less than 0.9.

The robust gun detection method of the present invention can also be used to detect whether a battery is contained in an image.

Description of drawings

Fig. 1 is the flowchart of the training stage in the firearm detection method in a specific embodiment of the present invention;

Fig. 2 is a flow chart of the detection stage in the firearm detection method in a specific embodiment of the present invention.

detailed description

The robust gun detection method proposed by the present invention will be further described below in conjunction with the accompanying drawings and embodiments. The following examples are only used to illustrate the present invention, but not to limit the scope of the present invention.

As shown in Figure 1 and Figure 2, the robust gun detection method proposed by the present invention, the method has carried out effective preprocessing and target saliency detection, through color segmentation and screening of unique size features, to find the target detection object Candidate area; this method is to design a gun classifier based on the risk of misclassification of guns and non-guns; then cascade multiple multi-layer deep learning neural networks, and finally judge the possibility of guns through the candidate area statistics, and finally form a complete firearm detectors;

In a system composed of an X-ray machine image acquisition device and a computer, the detection method includes a training phase and a detection phase;

1. The training phase includes the following steps:

1.1. Collection of samples;

Use the X-ray machine to go through the luggage package, and manually calibrate and cut out the guns in the collected pictures to establish a gun sample database; randomly cut non-gun pictures from the X-ray pictures that do not contain guns; after screening, a total of 17,800 pictures were obtained Gun samples and 90,000 non-gun samples are used as training sample sets;

1.2. Normalization processing;

Contains linear normalization of sample illumination and size, that is, normalizes the gun and non-gun images obtained in step 1 to a specified size, and then grayscales the size-normalized image;

In the sample library, the original target width is 180 pixels, and the height is 180 pixels. After zooming, the target width is 72 pixels, and the height is 72 pixels. After zooming, it will be grayed out;

1.3. Extraction of sample feature library;

1.3.1. Calculate the integral map of each sample

use by definition Calculate the integral map corresponding to each sample , and have , ;

1.3.2. Extraction of microstructure feature library

Three microstructural features using harr features: up-down type, left-right type, oblique symmetry type, plus an edge description value as a microstructural feature, plus a mean value as a microstructural feature, a total of five microstructural feature templates to extract The high-dimensional microstructural features of gun samples, the microstructural features are used Expression, for the five types of microstructure eigenvectors, respectively expressed as follows:

The up-and-down Harr feature is to subtract the total gray value of the upper half of the field of view from the total gray value of the lower half in the field of view of N*N, and then divide it by the number of pixels in the field of view N , to obtain a microstructure feature, assuming that the upper and lower parts are symmetrical up and down and have the same area, w represents the width of each part, h represents the height of each part, and N is a natural number greater than zero;

=

The left and right Harr feature is to subtract the total gray value of the left half of the field of view from the total gray value of the right half of the field of view in the N*N field of view, and then divide it by the number of pixels in the field of view N , to obtain a microstructural feature, assuming that the left half and the right half are symmetrical and have equal areas, and the definitions of w and h are the same as the upper and lower Harr features;

=

The oblique symmetric Harr feature is that in the N*N field of view, the total gray value of the upper left half of the field of view and the total gray value of the lower right half of the field of view are added to obtain the total gray value of the oblique direction, and the field of view Add the total gray value of the upper right part and the total gray value of the lower left part to obtain the total gray value in the reverse oblique direction, and then subtract the total gray value in the oblique direction from the total gray value in the reverse oblique direction, Then divide by the number of pixels N in the field of view , to obtain a microstructure feature, assuming that the areas of the upper left half, the lower left half, the upper right half, and the lower right half are equal, and the definitions of w and h are the same as the upper and lower Harr features;

=

The maximum value of the absolute number of the above three microstructural features is used as the edge description value, and this value is also used as a microstructural feature;

The mean feature is to add the gray values in all the fields of view in the N*N field of view, and then divide it by the number of pixels in the field of view N , get a microstructural feature;

=

1.3.3. Extraction method of microstructure features

For a sample image of 72*72 pixels, an 8*8 resolution field of view is read every 4 pixels in the horizontal direction, there can be 17 fields of view in the horizontal direction, and 17 fields of view in the vertical direction, so the resolution is 72*72 There are a total of 289 fields of view, and 5 microstructural features are calculated for each field of view, thus forming 1445 features to describe a sample;

1.4, classifier design

Use the features designed above and the DBN algorithm in deep learning to train multiple gun classifiers, and combine these multiple classifiers into a complete gun detector in a hierarchical cascade, including the following steps:

1.4.1. Initialize i=1; initialize and define the training target of each layer classifier, the target is to train one with a missed detection rate of less than 1% on the gun training set and a false positive rate on the non-gun training set Less than 20% of the classifiers, and then train a classifier with a missed detection rate of less than 20% on the gun training set and a false positive rate of less than 1% on the non-gun training set, and then combine these two classifiers into one The classifier is used as the classifier of this layer; define the target of the whole gun detector, the false negative rate on the gun training set is less than 5%, and the false positive rate on the non-gun training set is less than 1%; each DBN classifier, Two hidden layers are used, the input layer is 1445 neurons, the first hidden layer is 578 neurons, the second hidden layer is 300 neurons, and the output layer is 2 classifications. Link.

1.4.2. Train the i-th layer classifier;

1.4.3. Use the trained i-level classifier to detect the sample set, and calculate the missed detection rate and false positive rate;

Among them, the missed detection rate = the number of non-gun samples identified as guns / the total number of non-gun samples * 100%,

False positive rate = number of gun samples identified as non-guns/total number of gun samples*100%;

1.4.4. If the missed detection rate and false alarm rate do not reach the predetermined value set in step 1.4.1, then , return to step 1.4.2 to continue training, otherwise stop training;

In the detection phase, the following steps are used to determine whether the input image contains a gun:

2.1. Load the trained parameters and initialize the classifier;

2.2. Input the picture to be detected into the firearm detector obtained in step 1.4;

2.3. Perform image preprocessing on the image to be detected, mainly including white balance and color equalization;

2.4. Scaling operation on the input image,

After zooming, the target width is 72 pixels, and the height is 72 pixels. After zooming, it will be grayscaled;

2.5. Gun target detection,

Specific steps are as follows:

Step 1: Color Conversion

Convert an image in RGB color space to an image in HSV color space;

Step 2: Color Analysis

Segment according to the color type of the target object. When the target object is specified as an inorganic substance, if it appears blue and green, then segment according to the H channel in HSV, and the image part with H value > 140 and < 281 will be retained. The rest is replaced by pure white; when the specified target is organic, it appears yellow and green, then it is segmented according to the H channel in HSV, that is, the image part with H value > 0 and < 180 is retained, and the rest is replaced by pure white ;

Step 3: Brightness Analysis

Step 3.1: Histogram Statistical Segmentation

Perform histogram statistics on the brightness of the image after color analysis, select the area with a brightness of 0 to 10% as the black mask area, and the rest as the white area;

Step 3.2: Density Analysis

In the image that has been statistically segmented by histogram, calculate the integral map of the image, use the integral map to retain the area with a resolution of 25*25 with more than 70% of black pixels, and replace the rest of the area with white;

Step 4: Saturation Analysis

Step 4.1: Histogram Statistical Segmentation

Perform histogram statistics on the saturation of the image after color analysis, select the area with a brightness of 0 to 10% as the black mask area, and the rest as the white area;

Step 4.2: Density Analysis

In the image that has been statistically segmented by histogram, calculate the integral map of the image, use the integral map to retain the area with a resolution of 25*25 with more than 70% of black pixels, and replace the rest of the area with white;

Step 5: Color Density Analysis

Step 5.1: Colored Segmentation

Set all non-pure white areas in the image after color analysis to black;

Step 5.2: Density Analysis

For the colored segmented image, calculate the integral map, use the integral map to retain the colored area with a density exceeding 40%, and replace the rest of the area with white;

Step 6: Get the mask area

Merge the results of brightness analysis, saturation analysis and color density analysis to get darker areas;

Step 7: Connected Domain Analysis

For the image after the color density analysis, perform connected domain analysis, select a part with an appropriate area, and replace the connected domain with an area larger than twice the size of the target object and smaller than half the size of the target object with the background white;

Step 8: Get the final result

Combine the results of steps 6 and 7.

2.6. Calculate the integral map of the input image,

use Calculate the integral map corresponding to each sample , and have , ;

Generally, after this preprocessing, 1 to 30 candidate regions can be obtained, most of which are 3-5 candidate regions.

2.7. Through the integral map, extract the features of the candidate area,

Use a 72*72 window sliding in the candidate area, use the integral map in the window to calculate the characteristics of the sliding window, the feature of a window is used as the feature of a sample to be predicted, and the number of windows in the candidate area is determined by the target saliency detection;

2.8. Use the trained classifier to predict the possibility of guns in the candidate area,

Specifically, in each candidate area, use a classifier to calculate whether each sliding window contains targets such as guns;

2.9. The possibility of accumulating guns in the candidate area,

In each candidate area, if the number of windows containing guns exceeds a certain threshold, it is determined that there are guns. This threshold is the number of windows in the candidate area multiplied by a coefficient, and the coefficient range is greater than 0.3 and less than 0.9.

The gun detection method of the present invention can also be used to detect whether an image contains a battery.

The above usage methods are only used to illustrate the present invention, rather than to limit the invention. Those of ordinary skill in the relevant technical field can also make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all Equivalent technical solutions also belong to the protection category of the present invention.

Claims (3)

1. A robust firearm detection method, characterized in that the method has carried out effective preprocessing and target saliency detection, and searches for candidate regions of target detection objects by color segmentation and screening of unique size features; the method A gun classifier is designed based on the risk of misclassification of guns and non-guns; then multiple multi-layer deep learning neural networks are cascaded, and finally the probability of guns is discriminated through the candidate area statistics, and finally a complete gun detector is formed; In a system composed of an X-ray machine image acquisition device and a computer, the detection method includes a training phase and a detection phase; 1. The training phase includes the following steps: 1.1. Collection of samples; Use the X-ray machine to go through the luggage package, and cut out the guns in the collected images with manual calibration, and randomly cut non-gun images from the X-ray images that do not contain guns; establish a gun sample database; 1.2. Normalization processing; Contains linear normalization of sample illumination and size, that is, normalizes the gun and non-gun images obtained in step 1.1 to a specified size, and then grayscales the size-normalized image; 1.3. Extraction of sample feature library; 1.3.1. Calculate the integral map of each sample 1.3.2. Extraction of microstructure feature library Three microstructural features using harr features: up-down type, left-right type, oblique symmetry type, plus an edge description value as a microstructural feature, plus a mean value as a microstructural feature, using five microstructural feature templates to extract The high-dimensional microstructural features of gun samples, for the five types of microstructural feature vectors, are expressed as follows: The up-and-down Harr feature is to subtract the total gray value of the upper half of the field of view from the total gray value of the lower half in the field of view of N*N, and then divide it by the number of pixels in the field of view N , get the first microstructure feature; The left-right Harr feature is to subtract the total gray value of the left half of the field of view from the total gray value of the right half of the field of view in the N*N field of view, and then divide it by the number of pixels in the field of view N , to obtain the second microstructure feature; The oblique symmetric Harr feature is that in the N*N field of view, the total gray value of the upper left half of the field of view and the total gray value of the lower right half of the field of view are added to obtain the total gray value of the oblique direction, and the field of view Add the total gray value of the upper right part and the total gray value of the lower left part to obtain the total gray value in the reverse oblique direction, and then subtract the total gray value in the oblique direction from the total gray value in the reverse oblique direction, Then divide by the number of pixels N in the field of view , to obtain the third microstructure feature; The maximum value of the absolute numbers of the first microstructural feature, the second microstructural feature and the third microstructural feature is used as an edge description value, and this value is used as a fourth microstructural feature; The mean feature is to add the gray values in all the fields of view in the N*N field of view, and then divide it by the number of pixels in the field of view N , to obtain the fifth microstructure feature; 1.3.3. Extraction method of microstructure features For a sample image of 72*72 pixels, an 8*8 resolution field of view is read every 4 pixels in the horizontal direction, there can be 17 fields of view in the horizontal direction, and 17 fields of view in the vertical direction, so the resolution is 72*72 There are 289 visual fields in total, and each visual field calculates the five microstructural features described in step 1.3.2, thus forming 1445 features to describe a sample; 1.4, classifier design Using the microstructure features designed above and the DBN algorithm in deep learning, train multiple gun classifiers, and combine these multiple classifiers into a complete gun detector in a hierarchical cascade, including the following steps: 1.4.1. Initialize i=1; initialize and define the training target of each layer classifier, the target is to train one with a missed detection rate of less than 1% on the gun training set and a false positive rate on the non-gun training set Less than 20% of the classifiers, and then train a classifier with a missed detection rate of less than 20% on the gun training set and a false positive rate of less than 1% on the non-gun training set, and then combine these two classifiers into one The classifier is used as the classifier of this layer; define the target of the entire gun detector, the false positive rate on the gun training set is less than 5%, and the false positive rate on the non-gun training set is less than 1%; 1.4.2. Train the i-th layer classifier; 1.4.3. Use the trained i-level classifier to detect the sample set, and calculate the missed detection rate and false positive rate; Among them, the missed detection rate = the number of non-gun samples identified as guns / the total number of non-gun samples * 100%, False positive rate = number of gun samples identified as non-guns/total number of gun samples*100%; 1.4.4. If the missed detection rate and false alarm rate do not reach the predetermined value set in step 1.4.1, then , return to step 1.4.2 to continue training, otherwise stop training; In the detection phase, the following steps are taken to determine whether an input image contains a gun: 2.1. Load the trained parameters and initialize the classifier; 2.2. Input the image to be detected into the firearm detector obtained in step 1.4; 2.3. Perform normalization processing on the image to be detected, including preprocessing the image to be detected and normalizing the image to be detected to a specified size, and then grayscale the normalized image; 2.5. For gun target detection, use color segmentation and limit the size of connected domains to screen candidate areas; 2.6. Calculate the integral map of the image to be detected; 2.7. Through the integral map, extract the features of the candidate area, Use a 72*72 window sliding in the candidate area, use the integral map in the window to calculate the characteristics of the sliding window, the feature of a window is used as the feature of a sample to be predicted, and the number of windows in the candidate area is determined by the target saliency detection; 2.8. Use the trained classifier to predict the possibility of guns in the candidate area, Specifically, in each candidate area, use a classifier to calculate whether each sliding window contains targets such as guns; 2.9. The possibility of accumulating guns in the candidate area, In each candidate area, if the number of windows containing guns exceeds a certain threshold, it is determined that there are guns. This threshold is the number of windows in the candidate area multiplied by a coefficient, and the coefficient range is greater than 0.3 and less than 0.9. 2. the robust firearm detection method as claimed in claim 1, is characterized in that, in described step 1.4.1, each DBN classifier adopts two hidden layers, and the input layer is 1445 neurons, the first The first hidden layer is 578 neurons, the second hidden layer is 300 neurons, the output layer is 2 classifications, and all layers of the classifier are fully connected. 3. The robust firearm detection method according to claims 1 to 2, wherein the method can also be used to detect whether an image contains a battery.