Aviation security services have significant value for aviation safety. Aviation security services include personnel and technical equipment for dangerous and prohibited object detection. The X-ray screening system is the main equipment for baggage contain determining. The X-ray security devices give the possibility to detect weapons, including handguns, knives, and others. The high value of the probability of false detection of X-ray screening systems requires the development of new methods of image processing. Therefore, this paper concentrates on the synthesis and analysis of methods for handgun recognition while operating the X-ray screening system. The synthesis is based on a special image processing technique using a comparison of verified images with etalon images. During this, we use the correlation coefficient to determine verified object similarity with etalons inside the database and define their mutual rotation and resizing. The analysis is associated with computer modeling for estimating probabilistic characteristics of method efficiency.
One of the main problems in civil aviation is to
increase the safety and regularity of aircraft flights
[
The deliberate behavior of terrorists and criminals can significantly reduce the level of aviation safety. To counter such events, the aviation security service operates at airports. Each passenger and their baggage must be checked before they enter the aircraft.
The aviation security service includes
personnel and technical equipment to implement
the function of dangerous and prohibited object
detection. Personnel must be appropriately
trained. Technical equipment creates a system of
levels, on each of which different threats must be
detected and eliminated [
A new challenge in the era of digital
information technology development and
utilization is protection against cyberterrorism.
This threat can be realized using different types of
cyberattacks, spreading false information aimed
to impair airport operations [
Following international requirements and
recommendations, screening inspection measures
should be organized at the airports. Screening
usually contains three or five levels. To analyze
the internal structure of the baggage, the personnel
of the aviation security service uses X-ray devices
[
To improve the efficiency of X-ray screening
systems, designers use two approaches. The first
approach is related to the modernization of
equipment operation processes [
The literature considers many different
methods of image processing for X-ray systems.
The research [
Effective data processing methods can
significantly simplify security control processes
based on automating the detection of dangerous
objects [
Research [
A new trend in image processing is the usage of artificial intelligence techniques and Neural Networks (NNs). In the branch of X-ray image processing, the most convenient type of NN is Convolutional NN (CNN). A literature search gives many examples of CNN utilization for the tasks of dangerous object detection.
Publication [
Paper [
The literature considers different approaches
for data processing and detection in many fields
of study, for example, those given in publications
[
Mentioned analysis of the literature shows insufficient probabilistic characteristics of detection while X-ray security system image processing. Therefore, the aim of this paper is a synthesis of a new method for dangerous object recognition, an analysis of the efficiency of the proposed method, and the formation of recommendations for this method's improvement.
The X-ray security system consists of a transmitter and receiver of X-ray radiation. The transmitter unit contains a radiation source (X-ray tube), a power supply unit, and collimators (to form the scan beam). The receiver unit contains a detector line, optical-to-electrical and electricalto-optical signal convertors, a unit for image processing, and a device for information display. Baggage is placed on a conveyor and moved through the scanning area between the transmitter and receiver.
The detector line of the receiver records one of the parameters of attenuated radiation that propagates through the baggage. Received parameter values are encoded in gamut luminosity or grayscale. In this situation, the received parameter is a three-dimensional function of the coordinates (x, y, z). However, since the monitor of the X-ray security system displays a twodimensional image, we will consider the brightness function to be two-dimensional for a further solution to our problem. Let the brightness function is ( x, y) . In this case, the abscissa and ordinate will indicate a specific image pixel on the monitor of the X-ray security system.
The problem of improving the unit for image processing is not new. However, this problem is still relevant today due to several reasons: 1. A constant increase in the variety of prohibited and dangerous items. 2. The possibility of noises that distort the quality of the image. 3. An increase in the complexity of airport structure and the increasing level of airport passenger traffic. 4. Limitation for time to make correct decisions on aviation security measures.
There are two ways to the improvement of the unit for image processing associated with the hardware and software. However, both approaches require new efficient methods of data processing. In these conditions, statistical and filtering techniques are very relevant.
To examine the proposed method, this research uses computer modeling and statistical simulation. In addition, it should be noted that our approach is at the first stage of development, so some limitations will be used.
The flowchart of data processing procedures while recognizing dangerous objects on the X-ray security system image is shown in Fig. 1. s t c e jb e o s su aab o t reg ad n a D
probabilistic characteristics for detected object
The first procedure is image obtaining. This procedure is implemented in an X-ray security system based on encoding the level of attenuated radiation into the brightness. The image has the form of a two-dimensional array with the corresponding value of brightness. The metal objects completely absorb radiation, and then the value in the array for this case will be equal to zero. The maximum possible value in the array corresponds to receiving radiation without attenuation.
The second procedure is signal preprocessing. This procedure concentrates on the initial preparation of the image for subsequent operations. Preprocessing can contain image noise filtration, image dimensions change, selection of the objects with given attenuation of radiation, and others. For example, if we want to recognize handguns, we will select only metal objects with complete attenuation of the signal. However, such selection decreases the quality of dangerous object detection, because weapons can be produced using 3D printers and other materials.
The next procedure is possible to position determination for the searched object. This procedure can be implemented using different transformations to decision space, various clustering techniques, spatial spectra calculation, and others.
The detection procedure assumes decisive statistics calculation and comparison of their values with a threshold. The threshold can be computed using some priori information, for example about the probability of false alarm for given statistical characteristics of noise.
To synthesize the efficient algorithm for different object detection, the reference database of searched object masks can be used. In the general case, the masks give the possibility to train the detector and improve the quality of detection. The X-ray security system must contain filter banks for different dangerous objects. This study uses seven types of etalons for handguns that need to be detected. The etalons description is shown in Fig. 2.
The last procedure of data processing is an estimation of probabilistic characteristics for detected objects. This procedure assumes calculation receiver operating characteristics for different noise situations, calculation of error matrix, and others.
In this paper, we concentrate on only techniques for handgun detection. This technique is based on the correlation coefficient calculation of analyzed images and etalons.
Before explaining the step-by-step procedure of detection, let's introduce limitations: 1. The object of the search is the handgun. 2. The preprocessing procedure filter all noises in the image. 3. The object of search has an arbitrary angle of rotation and arbitrary scale factor. 4. The area of possible location of a handgun is selected in the image. 5. The handgun can be produced using 3D printers.
The calculations of all steps for the detection procedure were carried out in the MathCAD program.
Consider the step-by-step procedure of detection.
Step 1. Reading images of selected areas of analyzed items and etalons. It is possible to perform simultaneous studies for all etalons. For simplicity, this methodology presents only one of them. The example of the selected area and etalon 6 images are shown in Fig. 3.
a b
In the MathCAD program, images were obtained using the built-in operator read_image(). In addition, at this step, it is possible to distort the image by adding noise to it. Two noise generators have been implemented: Gaussian and Rayleigh. Noise parameters can be adjusted, and it is possible to estimate the signal-to-noise ratio.
Step 2. Image binarization. To speed up the computing process, it may be useful to filter out pixels of small amplitudes—insignificant ones.
pixels can be considered as significant. To implement this procedure, the thresholds for etalons and analyzed images were chosen. The results of the calculation are matrices B(x, y) that contain zeroes and ones.
The example of
binarization for images presented in Fig. 3 is shown in Fig. 4.
a b 1 0 1 0 0 0 c = where Nx and Ny are image dimensions.
Obtained estimate of the center is weighed center with taking into account pixel intensity.
The results of the calculation generate centers for analyzed images (xc; yc) and etalons (xcEi; ycEi).
The image movement is realized using the matrix method of graphics processing [38]. The translation matrix in this case will be E = (0 0 c − cE 1 c − cE ).
1 1
The coordinates of each pixel of the etalon are transformed according to the equation Em the sums of all distances in the radial matrices for all possible sectors. The example of obtained dependencies is shown in Fig. 6. a b images: a) tested image; b) image of etalon 6
The movement of the etalon image to the position of the analyzed one may require not only rotation but also reflection. Therefore, we created a reversible
matrix of accumulated sums of distances in the discrete sectors for the etalon image to further use it in sliding processing for step-by-step evaluation of the correlation with a reflected copy and to make a decision about the need to perform reflection transformation.
The next stage of calculation is an estimation of the correlation coefficient. For this purpose, direct and reverse matrices were extended to two complete rotations. After that, we used the estimation
windows that moved along extended matrices. correlation coefficient allows making decisions about image reflection. In this numerical case, reverse sliding provides the greater value of the global maximum, so the etalon image should be reflected.
To perform
reflection transformation, we used the matrix equation with reflection matrix about the x-axis of the following type = (0 1 0 0 0 −1 0 0).
1 for (a) direct sliding and (b) reverse sliding
Then we can estimate the RA α. This angle corresponds to the argument of a maximum of global maximums for the correlation coefficient for direct and reverse sliding, i.e. α = − 2
argmax( dir( ), rev( )), s where j is several discrete scan sectors, dir( ) and rev( ) are correlation coefficients for direct and reverse sliding, respectively.
For considered numerical example, the estimate of RA is equal to—6.075 radians.
To perform rotation of etalons, we used the matrix equation with rotation matrix about the center of the coordinate system of the following type
cosα − sinα 0 E = ( sinα cosα 0).
0 0 1
It should be noted that to implement this procedure, etalon images were moved to the coordinate system origin, reflected in case of necessity, rotated, and moved back to the initial point. Therefore, this operation requires two translation matrices of the following types 1 0 − c 1 0 c 1 = (0 1 − c) , 2 = (0 1 c).
0 0 1 0 0 1
The complete transformation matrix for Step 4 will take the following form α = { 2
E 1, if max( dir)< max( rev),
2 E 1, otherwis e.
After performing the matrix calculation, we can obtain the following result a) if max( dir)< max( rev),
2 E 1 = cosα − sinα c − c cosα + csinα = (− sinα − cosα c + c sinα + c cosα) ; 0 0 1 b) if max( dir)≥ max( rev),
1 E 2 = cosα − sinα c − c cosα + csinα = ( sinα cosα c − c sinα − c cosα).
0 0 1
The example of the transformation result is shown in Fig. 8. Visual analysis of obtained radial matrices shows the approximately same shape of both dependencies.
Step 5. Estimation of the scale factor. This step is associated with solving the optimization task for the correlation coefficient.
For this purpose, we perform direct and reverse linear movement of images of radial transformations of the etalon image relative to the tested one and vice versa with sliding correlation coefficient estimation. a
a b Figure 8: Radial matrices for images: a) tested image; b) image of etalon 6
In this case, we use the property of radial matrices, which is connected with the fact that the number of distance rings between any two arbitrary points of the same image with different scaling factors is the same, which allows estimating the correlation coefficient using a sliding distance window. The success of the described procedure requires sliding the radial matrix of larger objects relative to the radial matrix of smaller objects. However, since the size of each of the objects is unknown, it is necessary to check both sliding options.
Fig. 9 presents the dependence of the correlation coefficient estimates on the scale factor for both sliding options. Visual analysis of dependencies shows the existence of maximum in one of two cases. The maximum values are equal to 0.69 and 0.108 for direct and reverse sliding, respectively. b Figure 9: Estimates of the correlation coefficient for (a) direct sliding and (b) reverse sliding 0 0 0 0 1 0). correlation coefficient allows estimating the scale factor s. This factor corresponds to the argument of the maximum correlation coefficient for direct and reverses sliding, i.e.
= 2 arg max( dir( ), rev( )) .
numerical examples the estimate of the scale factor is equal to—0.479.
To perform the scaling procedure, we used the matrix equation with a scaling matrix of the following type
E = (0
It should be noted that to implement this procedure, etalon images were moved to the coordinate system origin, scaled, and moved back to the initial point. The complete transformation matrix for Step 5 will take the following form = 1 E 2 = (0 0 0 c − c 0 c − c ).
1
An example of the transformation result is shown in Fig. 10. a) tested image; b) image of etalon 6
Step 6. Final decision-making. This procedure consists of two operations. The first operation is a calculation of the correlation coefficient between the analyzed image and the transformed image of the etalon. The second operation is a comparison of the correlation coefficient with the threshold. If the calculated correlation is greater than the threshold, the decision on dangerous object presence will be made.
The main problem is the threshold value determination. In this research, the threshold value was calculated based on computer modeling. The value of the threshold corresponds required value of the probability of a false alarm. For our study, we computed the approximate value of the threshold for a decision about handgun presence for a 0.01 probability of a false alarm. This threshold is approximately equal to 0.71. In future research, we will try to find the value of the threshold taking into account noise influence based on statistical simulation.
For the presented numerical example, the final correlation coefficient is equal to 0.799. This value exceeds the threshold, so we have the correct detection of handguns.
The flowchart of the proposed method for handgun detection is shown in Fig. 11.
Start Tested image,
etalons Data arrays for images forming Binarization of
images Combination of images centers Estimation of RA and etalons rotation max(rdir)< max(rrev)
Yes
Etalon reflection Estimation of scale factor and etalons
scaling Comparing correlation coefficient with
threshold
Decision on handgun presence
Finish
To estimate the efficiency of the proposed method of detection, it is necessary to perform an analysis. In classical interpretation, the analysis of the detector assumes the calculation of receiver operating characteristics. However, according to the introduced limitations we process the image without noise.
In this case, the analysis concentrates on the calculation of a matrix of correct and erroneous decisions while handgun recognition.
The analysis was carried out using the designed program in MathCAD. To test the possible decision-making, two types of tested images were generated. The first type is one of the etalon images with an arbitrary angle of rotation and arbitrary scale factor. The second type is one of the tested images, including other dangerous objects (weapons) and non-dangerous objects. The basic types of tested images are shown in Fig. 12.
Image 1 Image 4 Image 7
Image 2
Image 3 Image 5
Image 6 Image 8
Image 9 Image 10
Image 11
Image 12 Image 13
Image 14
Image 15
Image 16 Image 17 Figure 12: The images for detector testing Image 18
The results of correlation coefficient calculation in the case of etalon (E) and tested image (I) recognition are given in Table 1 and Table 2, respectively.
The data in Table 1 was obtained without rotation and scaling. The diagonal elements in Table 1 correspond to the maximum value of the correlation coefficient. Rotating and scaling reduce the correlation coefficient estimate to 0.95. Computer simulation made it possible to estimate the probability of correct detection of the handgun, which was 0.882 for false alarm probability equal to 0.01. The non-detected weapons (and correspondingly low correlation coefficient) in Table 2 are due to the absence of etalons in the filter bank for tested images 1–7. The data in Table 2 was obtained for random values of rotation and scaling of tested images.
The paper concentrates on the synthesis and analysis of methods for handgun recognition while X-ray security system operation. The proposed method is based on a special image processing technique using a comparison of verified images with etalon images. The modeling gives the possibility to determine the probabilities of correct detection and false alarm (0.882 and 0.01, respectively). The future scope is associated with increasing the efficiency of detection by introducing new etalons and combining the correlation approach with the spectral technique of detection.