The article deals with the task of information synthesis of on-board recognition system of ground vehicles. Machine training of the system is carried out within the framework of information-extreme intellectual technology of data analysis, which is based on maximizing the information ability of the recognition system. The modified Kulbak information measure is used as a criterion for optimizing the parameters of machine learning. The proposed algorithm of machine learning is realized on the example of recognition of monochrome coloring cars on approximately identical chassis.
The widespread use of unmanned aerial vehicles to monitor the earth's surface makes it urgent to create autonomous on-board recognition systems (ORSs) for ground objects, including vehicles of various uses [1, 2]. The main way of solving this problem is to apply methods of machine learning and the theory of pattern recognition [3 – 5]. In this case, the functional efficiency of the ORS essentially depends on the method of processing images of the object being recognized and the method of machine learning. In addition, when creating the ORS it is necessary to ensure the invariance of the constructed in the process of machine learning decisive rules to the arbitrary position of the vehicle in the frame of the area of interest. One of the ways to solve this problem is to process images of recognition objects in a polar coordinate system [6, 7]. But within the framework of this approach necessarily there is a need to solve the problem of determining the polar coordinate system on the ground vehicle. The overwhelming majority of the methods of the information synthesis of the ORS, which are taught, is based on the application of neural networks [8, 9]. In this case, there are complications of scientific and methodological nature, associated with arbitrary initial conditions for the formation of the input mathematical description, the intersection in the space of signs of recognition classes, a large measure of the space of signs and the complexity of retraining.
One of the promising ways of information synthesis of highly effective ORS is the application of ideas and methods of the so-called information-extreme intellectual technology (IEI-technology) of data analysis, which is based on maximizing the information ability of the recognition system in the process of its machine learning [10 – 12]. The main idea of the methods in the framework of IEI-technology as in neural networks is to adapt the input mathematical description of the recognition system to the maximum functional efficiency of machine learning. But in contrast to the neural networks built on the results of machine learning in the framework of the geometric approach, the decisive rules are practically invariant to many of the dimensionality of the recognition signspace.
The article discusses, in the framework of the IEI-technology, the formulation of the problem and the algorithms for the operation of the onboard vehicle recognition system in machine learning modes of the exam. 2
Consider the formalized formulation of the task of information-extreme machine learning on-board recognition of a land vehicle. Let the alphabet {X mo | m 1, M } of the recognition classes characterizing the terrestrial vehicles and the input brightness || ym(j,)i || of the pixels of the receptor field of the image objects of the recognized objects be given. In this case, the line {ym(j,i) | i 1, N} of the matrix, where N is the number of signs of recognition, is a structured vector-realization (hereinafter simply realization) of the image, and the matrix column is a random educational sample of {ym(j,)i | j 1, J} with a volume of J .
In the process of machine learning it is necessary: in accordance with the concept of IEI-technology, convert the input training matrix into a binary working matrix || xm(j,)i || , which, through admissible transformations, can be adapted to the maximum full probability of making the correct classification decisions;
optimize according to the information criterion the parameters of the machine
learning ORS, which for each class of recognition X mo are given by structured vector
gm xm , dm , ,
(
At the same time, the following restrictions are set for the machine learning parameters:
– the range of values of the radius of the hyperspherical container of the recognition class d m , is given by the inequality dm d (xm xc ) , where d ( xm xc ) is the inter-center coding distance between the reference implementation of class xm of class X mo and the reference implementation of xc neighboring class X co nearest to it; – symbol of the logical operation of adding by module 2;
– the region of values of parameter is given by inequality H / 2 , where H is the normalized tolerance field on the recognition signs.
In the functioning of the on-board system in the exam mode, it is necessary to confirm the high functional efficiency of the machine learning onboard recognition system. 3
The categorical model of on-board computer training includes an input mathematical description of the on-board vehicle recognition system in terrain, which has the form B T , G, , Z , K ,Y , X ; 1, 2 , where T is the set of moments of the time of obtaining information; G is the space of the functional states of the recognized object; Ω the space of signs of recognition; Z – space of functional states of the recognition system; K – the set of frames of the electronic map of the area; Y is a sample set that forms the input training matrix; X – working binary training matrix; 1 :G T Z Y the operator of the formation of matrix Y ; 2 :Y X – s an operator for transforming the input training matrix Y into a binary matrix X .
Fig. 1 shows a categorical model of information-extreme learning of recognition system with optimization of geometric parameters of containers of recognition classes and system of control tolerances on recognition signs.
V
U G K
D
Y 2 Ф1 Ф2 1
X r
E ~ M φ γ |q | I| S| ognition classes on the fuzzy partition of the |M | binary traits of spaces, and the classification operator checks the basic statistical hypothesis of the validity of the class X mo mplementation and thus forms the set of hypotheses I |l| , where l – is the number of statistical hypotheses. The operator γ, by evaluating the hypotheses received, forms the set of exact characteristics |q| |, where q l 2 , and operator calculates the set of values of the information criterion E , which is a function of the exact characteristics. The contour of the model, which is closed by the operator r , restores, at each step of the machine learning, the recognition class containers that are built in the radial basis of the feature space. In this case, the iterative process of optimizing the geometric parameters of the partition |M | is carried out by finding the global maximum of the information criterion in the working (admissible) region of its function definition. In fig. 1 contour of optimization of control tolerances for recognition signs is closed through set D – the system of control tolerances for recognition signs and allows in the training process to change the value of the working binary training matrix X , adapting it to the maximum functional efficiency of the classifier. Shown in Fig. 1 categorical model implies, in accordance with the principle of deferred deci sions, the transition to other types of radial-basic decision rules. For this purpose, its outer contour contains a set of V types of decisive rules, which are built using more complex radial-basic separation functions. The training process is regulated by the operator U : V G T Z K .
Fig. 2 shows a categorical model of the functioning of the on-board recognition system in the mode of examination, which tests the functional efficiency of machinebased learning. f2
U E
x f1 G K y Р * 1
2 F
I | M + 1 | Fig. 2. Category model of functioning of on-board recognition system in the exam mode In the categorical model (Fig. 2), the operator f1 forms an incoming implementation of the recognizable object. Operator f2 , obtained on the learning stage, obtains optimal control tolerances for recognition attributes, converts incoming readmissions into binary realization x , and operator reflects the implementation of the recognizable object on the optimal breakdown of * recognition classes built at the stage of machine learning. The operator 1 for each vector-realization calculates the values of constructed at the learning stage of the deciding rules and forms the term-set F , and the operator 2 by the maximum value of the deciding rule, assigns implementation to one of the classes of a given alphabet { X mo } . The purpose of the operator UЕ is the regulation of the exam. 4
(
Optimization of the parameters of the vector (
E* 1 M
max Em (d) ,
M m1 GE Gd where Em (d ) – information criterion for optimizing the parameters of the training system to recognize the implementation of class X mo ; d – distance measure of radiuses of hyperspherical containers of recognition classes; GE working (admissible) area of determining the function of information criterion optimization of machine learning parameters; Gd is the permissible range for changing the radiuses of the recognition class containers.
As criteria of optimization of machine learning parameters in the methods of IEItechnology, modifications of the Kullback information measure and Shannon's entropy measure are mainly used. For example, the modified Kullback information measure for two alternatives to the a priori equivalence hypothesis has the form 2 (m (d ) m (d )) Em (d ) 1 (m (d ) m (d )) log2 , m (d ) m (d ) where m (d) is a mistake of the first kind of acceptance of a classification decision at every step of machine learning; m (d ) is a mistake of the second kind.
On practice, when calculating the information criterion (
K1,m (d ) ; (mk ) (d )
K2, m (d )
, nmin nmin where K1,m (d ) is the number of events that indicate the inappropriateness of "their" implementations of the recognition class X mo ; K2,m (d ) the number of events that indicate the belonging of "alien" implementations of class X mo , nmin the minimum amount of representative sample of study.
The working modification of the Kullback criterion after the corresponding
substitution of the estimates (
1 2n 10r K1(k ) (d ) K2(k) (d ) , Em(k ) (d ) {n [K1(d ) K2 (d )]}log2 n K1(k ) (d ) K2(k ) (d ) 10r where 10r is a sufficiently small number which is entered to avoid division into zero and in practice it is chosen in the interval 1 r 3 .
According to the categorical model (Fig. 1), the algorithm of ORS informational
and emergency machine learning with optimization of the system of control
tolerances can be represented as a two-cycle procedure for finding the global
maximum of the information criterion (
G GE Gd
Let's consider the main stages of implementing the algorithm (
AKH ,i [k ] y1,i [k ] ; AKB,i [k ] y1,i [k] ;
(
1, if AKH ,i ym(j,)i AKB,i ; xm(j,)i
0, if else. 9. calculation for class X mo binary averaging vector хm according to the rule 1, if xm,i 1 n
xm(j,i) m ;
n j1
0, if else ,
where m is the level of selection of the coordinates of the averaged binary vector
of the recognition class X mo , which by default is equal to m 0, 5 ;
10. pairwise decomposition of the set of averaged vectors of recognition classes by
the method of closest neighbors;
11. formation for the breakdown of |m2| xm , xc educational matrix;
12. calculation of the information criterion for optimizing the parameters of machine
learning GIS, for example, in the form of modification of the information measure
of Kullback (2.5);
13. if d d (xm xc ) , then paragraph 6 is executed, otherwise, paragraph 14;
14. if k H / 2 , then paragraph 4, is executed, otherwise, paragraph 15;
15. if m M , then paragraph 2 is executed, otherwise, paragraph 16;
16. calculation of averaged alphabet classes recognition information criterion (
Based on the results obtained in the process of machine learning, the optimal geometric parameters of the classes of recognition containers are based on decisive rules, which, when the recognition system operates directly in the operating mode, verifies the functional efficiency of machine learning. For hyperspherical containers of recognition classes, decisive rules have the form
(X mo |M | )(x( j) |M | )[if (m 0 ) & (m max{m}) then x( j) X mo ], (
m
In expression (
d (x( j) xm ) ; m 1 d m* (9) where d m* – obtained in the process of machine learning the optimal radius of the 1. {xm* | m 1, M } is an array of reference binary vector-image implementations that determine the geometric centers of the corresponding optimal containers of recognition classes constructed at the stage of machine learning; 2. { d m* } is an array of optimal radiuses constructed at the stage of training of the corresponding containers; 3. {xs( j) | s 1, SMAX ; j 1, n} is an array of binary vectors-implementations of identified frames, where SMAX is the number of frames of a reconstructed terrain; 4. {*k,i | i 1, N} is the optimal system of control tolerances for recognition signs, determined at the stage of training.
According to the categorical model (Fig. 2), the algorithm of the exam within the
framework of the IEI-technology is based on the analysis of the values of the decisive
rules formed at the stage of learning (
The analysis of decisive rules (
Approbation of the proposed algorithm for information-extreme machine learning was carried out on the example of the recognition of three cars that were moving along the highway (Fig. 3). At the same time, in order to check the functional efficiency of machine learning ORS cars specially selected monotonous and with approximately the same contours.
а b c
The formation of the input training matrix was carried out by processing images of cars in the polar coordinate system, which allowed to ensure the invariance of the decision of the rules to the arbitrary position of the object of recognition in the frame of the zone of interest. As a sign of recognition, the average value of the brightness of the pixels of the reading range, built around the center of the polar coordinate system, was taken. At the same time, the definition of the center of the polar coordinate system on the car was carried out by the class SelectedObject, which, moreover, handles the image of the object in the polar coordinate system and forms the input training matrix ORS.
Fig. 4 shows a screenshot of the result of the Class SelectedObject program, which shows the center of the polar system on a class X 3o (Fig. 3c) at different levels of quantization of the brightness of pixels of the area of interest frame.
To form the implementation of the input training matrix, all pixels of the frame of the interest zone, which was accepted as the first quadrant of the Cartesian coordinate system, was numbered. This allowed us to determine polarization centers of cars as average arithmetic numbers of pixels whose brightness exceeded the corresponding quantization level. Then the center of the Cartesian coordinate system was transferred to the found center of the polar system, around which the area of the given radius was asked. In the given region, the coordinates of the pixels were converted to polar and formed arrays of pixels with the same radiuses. For each RGB-component image of the car for recognition, the average brightness value of the pixels of the corresponding array was taken.
Fig. 5 shows the graph of the dependence of the information criterion recognition
(
The results of the physical simulation of the on-board recognition system in the exam mode showed that the full probability of correct recognition of the vectorrealization of class X1 is Pt 0,84 , class X 2 – Pt 0,80 and class X 3 – Pt 0, 78 . 6
The method of information-extreme machine learning of the ORS ground vehicle with
optimization of the system of control tolerances on recognition signs is proposed. The
formation of the input mathematical description was carried out on the basis of the
results of processing images of vehicles in the polar coordinate system, which ensured
the invariance of the deciding rules to the arbitrary position of the object of
recognition in the frame of the interest zone. The results of machine learning were not
allowed to construct non-error-based educational matrix deciding rules due to the high
degree of intersection of classes of recognition in the space of signs. Therefore, in
order to increase the functional efficiency of the ORS, it is necessary to increase the
depth of machine learning by optimizing other learning parameters, including the
parameters of image processing of terrestrial objects.
9. Ciresan, D., Meier, U. and Schmidhuber, J., Multi-column deep neural networks for image
classification, Computer Vision and Pattern Recognition (CVPR), 2012 IEEE Conference
on, pp. 3642-3649 (2012)
10. Dovbysh, A.S., Rudenko, M. S. Information-extreme learning algorithm for a system of
recognition of morphological images in diagnosing oncological pathologies. Cybernetics
and Systems Analysis, vol. 50, pp. 157-163 (2014) DOI:10.15587/1729-4061.2016.71930.
11. Dovbysh, A. S., Moskalenko, V. V., Rizhova, A. S. Information-Extreme Method for
Classification of Observations with Categorical Attributes. Cybernetics and Systems
Analysis, vol. 52(