FR2623927A1 - METHOD AND DEVICE FOR DETECTING AND LOCATING AN IMAGE PORTION - Google Patents

Abstract

A method and a device are described which make it possible to detect and locate a portion in an overall image by means of an adaptive preprocessing and a normalized correlation, in which are processed image point signals from opto-electronic converters. Each image (overall image or image portion) is associated with an image memory and a variable discrete filter connected to a computer which performs adaptive preprocessing. Each variable filter can be combined with a fixed discrete filter and it is possible to provide switching. A setup for determining the normalized values of the disturbance autocorrelation is shown.

Description

- 1.- Method and device for the detection and localization of a

  Portion d'imaoe The invention relates to a method and a device for detecting and locating an image portion in an overall image using an image processing device, in which are processed image point signals divided into lines and columns from opto-electronic converters, and

comprises at least one filter.

  Detecting the position of a given image portion in an overall image is an important problem of image processing. This designated spot can be solved in different ways. The most important methods are based on cross-correlation and on the determination of error quantities. For both processes, the literature provides many modifications and improvements. However, under certain conditions, these processes can be transformed into

  others and lead to similar results.

  The following is first of all discrete images that can be described by pixels. In order to simplify, it will be assumed that the offsets between the images also correspond to pixels

(picture elements) whole.

  Two rectangular B and R images are given which are represented by their gray values B (i, j) and R (i, j) at the discrete points with the coordinates i and J, with B (i, j) = 0 if i <0 or I> nbi or j <0 or i> nbj (1) 2 5 R (ij) = if i <0 or i> nri or j <0 or j> nrj The image R represents a smaller portion of image B, on the same scale. So nbi> nri (2) - 2 - nbj> nri It will be admitted that the gray values R (i, j) of the image R are distinguished from the corresponding gray values B (i, j) of the image B by an additive disturbance N (l, j). With the actual offset position ivt and ivt in the direction of columns and lines, we obtain: R (i, j) = B (i + ivt, j + jvt) + N (i + ivt, j + jvt) ( 3) An optimal location strategy provides the Civo shift vector, jivo. which is most likely among all the vectors

  offset [iV, jv] ackdmissibles.

  For the conditional probabilities p (a / b) therefore has an optimal strategy: p (ivoJivo] = [ivt, jvtj / B, R)> p (liv, jv) = tivt, Jivt] / B, R) 1 5 with [ivj [ivo, Jvo] (4) For this purpose, we must determine all the conditional probabilities p (liv, jv) = civt, Jvt] / B, R) which express with which probability the shift livJv] coincides with the real offset

  [ivt, jvt], assuming that the images B and R exist.

  Bayes' theorem allows us to transform this equation: - p (B, R / [ivoivo = [ivtJvt]) * P ([ivo, Jvc = tivt, Jvt]) / p (BR)> p (B, R / [ ivJv3 = [ivtiv) * p ([ivjv = [ivtJvtI) / p (B, R) with [iViv] J [ivoJvol (5) 2 5 In the case where all shift positions have the same probability a priori or none information is available concerning these probabilities, and these are therefore assumed to be identical, it follows: p (B, R / [ivoJv = [ivt, Jvtl)> p (B, R / [VivJvl = [ivtjvt]) withCivJv] l [ivo'ivol (6) - 3- By designating the difference between the image R (i-iv, j-jv), shifted from [iv, jvl, and the image B (i, j) by N- (i, j) tiv, jvl = R (i-iv, i-jv) - B (i, j) (7) we obtain for the necessary probabilities in equation 6: p (B, R / Livjv ] = CivtJvt]) = p (N (i, j) = Nd (i, j) Liv, jv]) (8) Calculating these probabilities is usually possible only under certain conditions. In the case where the perturbations N- (i, j), defined according to equation 7, have no local correlation and have a probability density function of Gauss, it is possible to indicate the optimal location strategy. according to equation 6 using equation 8. After a logarithmic calculation and the omission of identical terms from both sides of the inequation, we get: 1 ((N (I,) [voJvo])) <X ( (N '(l, j) [tviv]) zJ) (9) ij: 1, j If we introduce in a similar way to equation 7 an image R (I, j) iv, jv, with / 0 if i <O or i> nri or j <O or i> nrj R, (i, j) [ivJvj = <\ B (i + iv, j + jv) in other cases (10) we get, by some transformations and by omission of identical terms from both sides of the equation from equation 9: R (t, j) .R- (iJ) [1voJvo]> 1 R (IJ) * R '(1 , j) [ivJv] (11) ij ij Now, these two known equations, namely the determination of error quantities according to equation 9 and correlation equation 11, provide an optimal location. only in the

  stringent conditions given above.

  4 - A discrete filter, with which the image B (i, j) and the partial image R (i, j) are filtered, makes it possible to modify the correlation of the gray values of the disturbances. Such a filter, which completely eliminates existing correlations, is called a decorrelation filter. However, the determination of the filter coefficients is only possible if the statistical characteristic quantities of the perturbation N. (i, j) are known, which is not technically possible for reasons of causality. In the known processes with fixed discrete filters (pre-treatment filters), one simply strives with a statistical hypothesis to reduce the local correlations of the disturbances. The result depends

  nevertheless to a large extent images.

  The object of the invention is to indicate a method and a device for locating a portion of an image in a larger image (overall image), making it possible to obtain a better localization frequency, even when the disturbances , defined

  according to equation 7, are in correlation.

  In accordance with the present invention, there is provided a method for detecting and locating an image portion in an overall image using an image processing device, in which signals are processed. image dots divided into rows and columns from opto-electronic converters, said device comprising at least one filter, characterized in that for the prior adaptive image processing, a variable discrete filter (fv) is associated to each of the image memories (Memo 1, 2) for - the overall image (B (i, j)), - the partial image or image portion (R (i, j)), and the two filters (fv) are connected to a calculator (C), which performs a calculation of the filter coefficients for each image from the statistical characteristic variables of the images

filtered or unfiltered.

  The invention also proposes a device for implementing an image localization method (part of an overall image), in which image point signals from optoelectronic converters are transformed. in image data and filtered, characterized in that each variable discrete filter (fv) is combined with a fixed discrete filter (ff) and is selected between two filters and / or modes, or is optionally loaded with at least two groups of filter coefficients determined differently. According to a particular, nonlimiting embodiment of the device according to the invention, is connected to the computer (C) an input stage (E) for determining the autocorrelation which - serially extracted, by lines (point by point in the columns) or by columns (point by point in the lines), the image data for each image from the memory (Memo 1, 2) associated, - control of the multiplication members (M) for the multiplication of these image data and - control, for the summation, of the addition members (A) whose

  Outputs are each connected to a division stage (D).

  Some or all of the functions of the input stage (E)

  above can also be done in the calculator (C).

  The method and the device according to the invention also make it possible to obtain a better decorrelation of the disturbances in the case

different images.

  In contrast to known methods and devices, image localization with normalized correlation is performed by means of adaptive pretraining, where the filter coefficients

  are adapted to the content of the image and calculated in a calculator.

  The computer is associated with a preliminary stage for the approximate determination of the autocorrelation function of the perturbation. For the determination of the autocorrelation as well as for the determination of the filter coefficients, one applies a

calculation model (algorithm).

  The calculation model, applied in the preliminary stage E of calculator C - according to FIGS. 1b and 2 "determination of

Z623927

  The autocorrelation "- and, for a general understanding, the normalized correlation according to FIG. 3 are presented below by way of explanation: When, as is the most frequent, the two images (overall image and partial image) are generated by different units of registration, they are further distinguished by a multiplicative (positive) factor which is taken into account below.In addition, it is advisable to separate a constant from the additive disturbances, in order to reduce to a minimum the energy of the

disturbance N (i, j).

  Using the real offset position ivt and iJvt, we obtain analogously to equation 3: R (l, j) = km * B (i + ivt, j + jvt) + ka + N (i + ivt, j + jvt) (12) 1 5 km> 0 The substitution B '(i + ivt, j + jvt) = km * B (i + ivt, j + Jvt) + ka (13) makes it possible to carry out the localization essentially with the location algorithms described in the preceding paragraph according to equation 9 or 11, if the perturbations N (i, j) have a probability density function of Gauss and are uncorrelated. In the preceding equations, the image B (i, J) must be replaced by an estimate B "(i, j) and the image R (i, j) [iv, jv] 2 5 by / O if i < O or l> nri or j <O or j> nri RI '- (i, ji) [ivjv] = <(14) \ B "(i + iv, j + jv in other cases For an estimation of the image B "(l, j), the constants km and ka are not always of interest and are not treated here in detail.As shown below, it is possible to establish an estimate also from the images B '(i, j) and R (i, j) available.

26239-27

  The expression N '(I1i) tIViJv]) 2 (R (t-iv, j-Jv) - B "(, j)) 2 (15) 1, ji: used in equation 9 constitutes a measure for the energy of the disturbing parts The equation B "(i, j) = ci * (B '(ij) - c2) (16) makes it possible to determine, by partial differentiation and by subsequent zeroing, the constants c1 and c2 so that the energy according to equation I5 takes a minimum value. In this case, the summation must only be performed in the image areas,

in which the image R exists.

  In this model, the linear average value of the gray values does not influence the location. It is therefore acceptable to pre-normalize the image R (i, j) so that the linear average value is equal to zero. In this case, the equations are substantially simpler. It is possible to find an approach solution, if the partial images R "'- (i, j) Civ, jv] defined in equation 10 are normalized in the same way as the image

R (i, j).

  The combination of two successive normalization stages is suitable for the standardization of the chosen model. During the first step, the identical elements of the gray values of the images are eliminated. Then, the two images are each multiplied by one

  This factor makes the two images have the same energy.

  Figure 3 shows the block diagram of an image location operating according to this principle. In this block diagram, it is assumed that the determination of the individual values of the correlation function or the error function is done in successive stages. We consider as ivoivo shift the one with which is associated the highest value of the correlation function or

  the smallest value of the error function.

  The known method of normalized correlation so far can be improved by pre-treatment filters, which increases

the location frequency.

  However, for the determination of the coefficients for a fixed filter it is disadvantageous that the filter coefficients can be optimized only for an individual case. In the case where the statistical characteristic variables of the images change in an unfavorable manner, the use of a preprocessing filter may be worse than the probability of location errors, that is to say the accuracy of location. Therefore, it is necessary for the optimization of the filter to use a very large number of representative images. The correlation of the gray values between them can be described by a two-dimensional normalized autocorrelation function Ln (i, j): X (N (iijj) * N (ii.1, jj + j) Ln (ij In this case, the perturbations N (i, j) are uncorrelated, if: Ln (i, j) = <(18) \ 0 in other cases In the case where the values of the normalized autocorrelation function are different from zero for i * O or jiO, a filter should be used of decorrelation, in order to obtain a better localization.

  In the description below, we start from the assumption that

  Two-dimensional pretreatment filter with the quantities nfi and nfj in the direction of the columns and these lines is used. The filter coefficients are denoted by F (i, j), for which: F (lj> = O sI I <0 or 1> nfi or j <0 or j> nfj (19) -9- The decorrelation of the perturbation N (i, j) is equivalent to a minimization of the perturbation energy Enf of the filtered perturbation, for which: ## EQU1 ## Allocation: 1i, Jj The partial differentiation and the zeroing make it possible to determine the filter coefficients F (i, j) of the pre-treatment filter, a filter coefficient having to remain fixed, because for this method only the relations between the coefficients The coefficients are determined by a linear equation system that can be solved

relatively simple.

  For the calculation of a decorrelation filter it is necessary to know the statistical characteristic variables of the perturbation N (i, j). It is possible to indicate this disturbance, which must be determined according to equation 3, only if the correct location has been carried out. In order to be able to determine the statistical characteristic quantities, one must

  very large number of locations.

  The invention particularly relates to an adaptive method for image preprocessing. For this, estimations concerning the autocorrelation of the disturbances are made according to the

  content of the image and the search area concerned. -

  For a large search area, it is assumed that the disturbance according to equation 3 consists largely of the non-Image. Therefore, the autocorrelation of the perturbation will substantially coincide with the autocorrelation of the image. For a large search area, the autocorrelation according to FIG.

  to be determined approximately from the image B (i, j).

  Determining the exact location requires additional correlation around the most likely location point found previously. The assumption on disturbances,

- 10 -

  previously used, being in this case no longer valid, a switch to a fixed discrete filter is then necessary, resulting in a determination of the fixed filter coefficients from a statistical assumption on disturbances. This principle is represented in FIG. 2. Another possibility consists in using a single variable filter instead of two filters and a switching device, and in this case we load and use, as desired, fixed or adaptive filter coefficients. (calculated according to

Equation 20).

  Embodiments of the invention are illustrated in FIGS. 1 to 4, of which FIG. 1a represents a sought portion in an overall image, FIG. 1b represents an image location with adaptive preprocessing, FIG. an image location with switchable adaptive pretreatment, Fig. 3 shows a block diagram of an image location, and Fig. 4 shows a determination of the values

  normalized autocorrelation function.

  The invention relates in particular to a pre-translation of an image

  adaptive for partial decorrelation.

  The use of decorrelation makes it possible to obtain extreme values, defined locally very clearly, in the error or correlation function. The determination of extreme values can therefore often be made only after the use of an interpolation method. When it is desired to avoid using an expensive interpolation method, the predefinition of an autocorrelation function of the filtered images makes it possible to adapt the shape of the extreme values of the error or correlation function. In this case, the calculation of the filter coefficients becomes considerably more expensive. The convolution of the coefficients, calculated from equation 20, with a function

- 11 -

  model, leading to the desired correlation function, nevertheless makes it possible here also to obtain a simple determination of the

final filter coefficients.

  For an embodiment of the invention for video real-time image localization (<1 s), structures as simple as possible should be used. In the concept presented here, a fast hardware, working in parallel allows to carry out a localization by means of the pretreatment and the correlation in time

real video.

  With the use of a pretreatment, in this example, a filter with the magnitude nfi = 1 (values in the direction of the columns) has been optimized nij = 5 (values in the direction of the lines) i5 Nevertheless, with adaptive pretreatment alone three coefficients have been determined according to equation 20. By designating these coefficients by F '(i, j), we obtain the following system of equation: Enf = Ln (O, -2) * (F' (1,1 ) * F '(1,3)) + Ln (O, -1) * (F' (1, 1) * F '(1,2) + F' (1,2) * F '(1,3) )) + Ln (O, O) * (F '(1,1) 2 + F' (1,2) 2+ F '(1,3) 2) (21) + Ln (0,1) * ( F '(1,1) * F' (1,2) + F '(1,2) * F' (1,3)) + Ln (0,2) * (F '(1,1) * F '(1,3)) One of the three filter coefficients to be determined can be chosen freely, for example:

2 5 F (1,1) = 1

  If we consider the symmetry qualities of the autocorrelation function Ln (i, j) = Ln (i, -j) and its value at position (0,0) according to equation 17 Ln (O, O ) = 1,

- 12 -

  we obtain from Equation 21: Enf = 2 * Ln (0,2) * F '(1,3) + 2 * Ln (0,1) * (F' (1,2) + F '( 1,2) * E (1,3))

+ (1 + F '(1,2) 2 + F' (1,3) 2)

  For the determination of F '(1,2), it is necessary to partially differentiate this equation according to F' (1,2) and to set it to zero: 0 = 2 * Ln (0,1) + 2 * Ln (0,1 ) * F '(1,3) + 2 * F' (1,2) For the quantity F '(1,3), we obtain in the same way: 0 = 2 * Ln (0,2) + 2 * Ln (0,1) * F '(1,2) + 2 * F' (1,3) From the system of linear equation it follows: F '(1,1) = 1 (according to the hypothesis) F' (1,2) = Ln (0,1) * (Ln (0,2) - 1) / (1 - Ln (0,1) 2) F '(1,3) = (Ln (0,1) 2 - Ln (0,2)) / (1 - Ln (0,1) 2) For an easier determination of maximum correlation value, it is possible, if desired, to carry out moreover

  a convolution of this result vector with a shape vector.

  In this case, for the coefficients of the filter used:

F (1,1) = F '(1,1)

F (1,2) = 2 * F '(1,1) + F' (1,2)

  F (1,3) = F '(1,1) + 2 * F' (1,2) + F '(1,3)

F (1,4) = F '(1,2) + 2 * F' (1,3)

F (1.5) = F '(1.3)

  The determination of these coefficients, which must be determined again for each image, requires only the correlation values Ln (O, 1) and Ln (0.2). These can be easily determined in real-time video, as shown in the block diagram in Figure 4. In this case, it is assumed that the image data is transmitted serially by lines. A

- 13'-

  command makes it possible to obtain that the summation is carried out only on

  the points of a line. The order is included in Figure 4.

  In FIG. 4, both the image portion R (i, j) and the larger image or residual image B (i, j) (overall image portion = residual image) can be extracted from the memories The multiplication elements are designated by M, the addition members by A and the division stages by D. The addition members are controlled by the drive circuits.

their are associated.

- 14 -

Claims (3)

claims   A method for detecting and locating an image portion in an overall image using an image processing device, in which image point signals divided into lines are processed and in columns derived from optoelectronic converters, said device comprising at least one filter, characterized in that for the prior adaptive image processing, a variable discrete filter (fv) is associated with each of the image memories (Memo 1 , 2) for - the overall image (B (i, j)), - the partial image or image portion (R (i, j)), and the two filters (fv) are connected to a calculator (C), which performs a calculation of the filter coefficients for each image from the statistical characteristic quantities of the images filtered or unfiltered.   2. Device for implementing an image localization method (part of an overall image), in which image point signals from opto-electronic converters are transformed into image data and filtered, characterized in that each variable discrete filter (fv) is combined with a fixed discrete filter (ff) and is selected between two filters and / or modes, or is optionally loaded with at least two   groups of filter coefficients determined differently.   3. Device for carrying out a method according to claim 1 or 2, characterized in that the computer (C) is connected to an input stage (E) for determining the autocorrelation which - extracted Serially, by lines (point by point in the columns) or by columns (point by point in the lines), the image data for each Image & from the memory (Memo 1, 2) associated, - command of multiplication units (M) for the multiplication of these image data and - 15 -   control, for the summation, of the addition members (A) whose outputs are each connected to a division stage (D), some or all the functions of this input stage (E)   can also be made in the calculator (C).