WO2006131701A1

WO2006131701A1 - Comparison of patterns

Info

Publication number: WO2006131701A1
Application number: PCT/GB2006/002001
Authority: WO
Inventors: Frederick Warwick Micheal Stentiford
Original assignee: British Telecommunications Public Limited Company
Priority date: 2005-06-10
Filing date: 2006-06-01
Publication date: 2006-12-14
Also published as: EP1732030A1; EP1889208B1; EP1889208A1; US7574051B2; US20080205760A1

Abstract

A first visual image or other pattern (A) is represented by a first ordered set of elements (xi) each having a value and a second pattern (B) is a represented by a second ordered set of element (ui) each having a value. Iteratively, a comparison is made by: (i) selecting a plurality of elements from the first ordered set ; (ii) for each selected element of the first ordered set, selecting an element from the second ordered set, such that the selected elements of the second ordered set have, within the second ordered set, a set of positional relationships relative to each other that is the same as, or a transformation of, the set of positional relationships that the selected plurality of elements of the first ordered set have relative to each other; (iii) selecting a random shift value ( ); and (iv) determining whether the selected elements meet the match criterion that the value of each of the selected elements of the first ordered set differs from the value of the correspondingly positioned selected element of the second ordered set by an amount that is equal to, or within a set margin of, and in the same sense as, said shift value. In the event of a match, a similarity score (SAB) is updated. Alternatively, or in addition, the shift amount in the event of a match is recorded and, at the conclusion of the iteration, a single representative shift value is estimated from said recorded shift value.

Description

Comparison of Patterns

The present invention is concerned with comparing one pattern with another, and is of particular interest in the comparison of two-dimensional patterns such as visual images, although applicable also to one-dimensional patterns and patterns having three or more dimensions. Some techniques for analysis of images or other patterns where the pattern is compared with other parts of the same pattern are described in our earlier patent applications as follows.

European patent application 00301262.2 (publication No. 1126411) ; International patent application PCT/GB01/00504 (publication No. WO 01/61648); International patent application PCT/GB01/03802 (publication.No. WO02/21446)); U.S patent application 977,263/09 filed 16 October 2001 (publication No. 20020081033); - as well as the following papers published by the inventor:

Stentiford F W M, "An estimator for visual attention through competitive novelty with application to image compression", Proc. Picture Coding Symposium 2001, Seoul, 25 - 27 April, pp 101 - 104, 2001. Stentiford F W M, "An evolutionary programming approach to the simulation of visual attention", Proc. Congress on Evolutionary Computation 2001, Seoul, pp 851 - 858, 27 - 30 May, 2001.

Methods and apparatus for comparing patterns are described in our earlier international patent application WO03/081532 and our co-pending international patent application WO2006/030173. Aspects of the present invention are set out in the claims.

Some embodiments of the present invention will now be described, with reference to the accompanying drawings.

Figure 1 shows an apparatus consisting of a general purpose computer programmed to perform image analysis according to a first embodiment of the invention. It has a bus 1, to which are connected a central processing unit 2, a visual display 3, a keyboard 4, a scanner 5 (or other device, not shown) for input of images, and a memory 6.

In the memory 6 are stored an operating system 601, a program 602 for performing the image analysis, and storage areas 603, 604 for storing two images, referred to as image A and image B. Each image is stored as a two-dimensional array of values, each value representing the brightness and/or colour components of a picture element within the array. It will be understood, however that the apparatus might more generally be arranged to analyse data sets other than images, in which case the storage areas 603, 604 would each contain a data set A, B, each being an ordered set of values, ordered in any number of dimensions (the same number of dimensions in each set). The image arrays are shown schematically in Figures 2a and 2b. Image A consists of an array of picture elements X| =( X₁, yO where X₁ and y; are the horizontal and vertical positions of the elements within the image. A 20x20 array is shown, for the purposes of illustration. In the general case, Xj =( Xii, Xi2,Xi3, ... Xi_n) where Xy are the n coordinates of x, in n dimensions. Each element has a respective value a = a (x,). This may be a scalar value a or vector (multidimensional) value a. In the case of an image these would be the brightness a, or alternatively a set of colour components such as (in r,g,b representation) a = (a_n a_g, a_b) Similarly, the image B consists of an array of picture elements Ui = (u;, v,) having brightness values b(%) or colour vectors bfe). In the general case, U_| =(

Mil, Mj₂, «ϋ, .-.. Wj_n).

Figure 3 is a flowchart explaining the application of the method as applied to two- dimensional images A, B. In Step 100, a score C_AB is set to zero, and a counter t set to zero.

The first task of the method is to make a random selection of¹ elements in image A. This may be visualised as stabbing the image with an irregular fork having m tines. The actual number m of elements may be fixed, or if preferred could be selected at random within a range m_mJn≤ffz≤/W_max. The fork size may be limited by a diameter D, which could be fixed or selected as a random number D_mi_n≤D≤D_nωx . Other approaches to varying the distribution of fork size are discussed in our above- mentioned U.K. patent application 0420727.0. Note that, throughout this specification, references to a random selection also envisage the possibility of selection by means of a pseudo-random process. Also, a reference to random selection envisages the possibility not only of a random selection for which the selection of any value (within a range specified) is equally probable, but also of a selection in accordance with a non-uniform probability distribution.

At Step 101 a set Nx of m elements at random positions x; is selected in image A, subject to the constraint that all elements lie within the image. If desired, the elements may be constrained to lie within a certain distance of one another, though this is not essential. One possible method of doing this is as follows: a) Choose a random element x₀ in A. b) Choose a set of random displacements dj = (d_xi, dyi), where 0 ≤ d^ < D/2 and 0 < d_yi < D/2. c) If the Euclidean distance |di| < D/2 for all i, then continue; otherwise repeat the previous step for any / needing it until this condition is satisfied. Where |dj| = V{ d,/ + d_yi²} . d) Compute positions Xj = xo +di e) If any x_; lies outside the image area, repeat steps (b) to (e) in respect of the offending di. Note that this process permits the occurrence of one or more pairs x; = Xj i≠j i.e. the Xi are not constrained to be distinct. This is not objectionable in practice but if desired may be eliminated by testing for it and reselecting one member of the matching pair in the same way as for elements falling outside the image area..

Step 102 is optional and will be described later. At 103, a counter n is initialised to zero. The next stage (Step 104) is to make a selection N_u of m elements u_; within image B, at a random overall position, but having the same positions relative to one another as the selected elements of image A. That is, x; - X_j = u; - U_j for all ij. This may be visualised as stabbing image B with the same fork as used on image A.

Assuming that a set of displacements dj has been calculated as above, this can be done by: Choose a random element u₀ in B.

Compute positions u; = U₀ +di If any u, lies outside the image area, repeat these steps in respect of the offending dj.

It is to be determined whether each of the selected elements in image A matches the element having the same relative position in image B. First, however, a random colour shift γ is chosen. In the case of monochrome image this would simply be a brightness shift but in the colour case it requires selection of a random shift value for each colour axis, e.g., in an r, g, b system,

where each shift γ_r, γ_g, γ_b is selected to lie in the range -Y_MAX to +Y_MAX (different ranges could be defined for the different colour components if desired - for example in y, u, v systems one range might be chosen for the luminance and a different range for the chrominance values). Supposing that the colour components take values in the range 0 to 255, then a typical value for J_MAX might be 64, or perhaps 128.

The test for a match occurs at Step. 105. In the r, g, b case a match occurs if Dist[a(x;) - b(u;) -+γ] < δ for all / = 1 ... m Where Dist is some distance operator (e.g. Euclidean or city-block), or the vector components may be thresholded separately, e.g. for colour components the match condition would be Ia₁(X₁) - b_r(u_;) + γ_r I < δ_r and Ia₈(Xi) - b_g(u_;) + γ_g I < δ_g and |a_b(xi) - bb(ui) + γ_b I < δ_b for all / = 1 ... TO where δ_r, δ_g and δ_b are some small threshold values which will commonly be equal.

If a match occurs, then at Step 106 the Score C_AB is incremented. It could simply be incremented by 1 (or other fixed value).

Once the score has been updated, then the iteration counter t is incremented at 109 and tested at 110 to see if it has reached a maximum iteration count T, and if so, the process terminates. Otherwise the process returns to Step 101 for a further iteration. If no match occurs, then the fork is retained and the process repeats from 104. However this repetition with the same x,- is subject to a limit of N iterations: the counter n is incremented at 107 and if at 108 it reaches Ntlien control passes to 109 for a fresh fork. The result of this process is the final value of C_AB, which is a score indicating the degree of similarity between image A and image B. As compared with the method described in our international patent application WO03/081523, this method has the advantage that a high score can be obtained despite the fact that there may be differences in colour between the two images, due for example to differences in lighting. In the case of images A, B both containing large plain areas of similar brightness (or colour), a large number of matches and hence a high score might be obtained even though there may be little similarity between the parts of the image that contain detail. In this case, the process need not necessarily be carried out for the whole image. For example, if regions of images A and B have been identified as being of special interest - perhaps using the method described in one of our earlier patent applications referred to above - then the picture elements x, u dealt with may be just those lying in the vicinity of the identified region. This could be implemented by constraining the selection of x and u such that at least one element x, is an element of a special interest region of image A and at least one element u; is an element of a special interest region of image B. This means that each match establishes a similar relationship in each image between the foreground and background, or foreground and foreground (if all tines lie on high interest pixels). Small diameter forks will tend to obtain similarities between localised regions in A and B, and larger forks will determine global relationships between foreground and background. Alternatively, we find that improved results can be obtained if the process is biased towards making comparisons involving the more detailed parts of the image, and with this in mind, the operation represented in Step 102 serves to test each new fork to determine whether, when it is in its initial position in image A, each of its tines points at an element that differs by at least a threshold amount from those pointed at by at least h of its other tines. This is shown in detail in Figure 4. Counters q, p are initialised at 200, 201 so each distinct combination of p and q represents a pair of tines of the fork. The process then enters a loop where at 202 the picture elements at the corresponding two points in image A are compared to see whether they differ by more than a threshold amount. If not, then/? is incremented (205) until it exceeds m (206) and the comparison repeated. If the values match, then a match counter Z is incremented at 203 . As long as (204) the counter Z has not reached h, then the comparison is again repeated for a new value of p. Once Z has reached h, it has been established that picture element x₉ differs significantly from at least h others in the fork and at 208, 209 Z is reset and q advanced for a further iteration. If all values of q result in a value of z that reaches h, then the test is passed (211). But is any value of q fails to result in Z reaching h them the tests fails 207.

The above process is effective but rejects many candidate forks and is, therefore, slow. A faster but cruder alternative is to select elements at random until m-h have been found that match each other, and then looking for h random pixels that mismatch all of the first group ofh pixels.

In this context a mismatch between two pixels means that they differ by ε or more in at least . one colour component.

Figure 5 is a flowchart showing the operation of a second embodiment of the invention. Steps that are identical to those of Figure 3 are given the same reference numerals. Here, the object is not to generate a similarity score, but rather to estimate the colour difference between the two images, though the score C_AB is still calculated, but for a different purpose. At the start, a set of counters C_r(z), C₆(O, Cb(O is initialised. The number of counters corresponds to the number of possible values of γ_r, γ_B, γι_» so there are 3(Y_MAX + l)counters, corresponding to i = 0 ... Y_MAX-

Operation is then as before up to the point where a match is found. Then the value of γ that resulted in the match is recorded by incrementing the corresponding counters. Thus, at step 200 the counters C_r (γ_r), C_g (γg), C_b (γ_b) are incremented. The remained of the iteration process is as before, until the limit T is reached at 110. At this point the counters C_r(/), C₅(J), Cb(Z) contain, in effect, three histograms indicating how successful each of the values of γ_r, γ_g, γbhave been in enabling matches to be obtained. Figure 6 shows actual histograms plotted to show the results of a test in which image A was a photograph of a bowl of fruit under normal illumination and image B was the same scene but with a green cast. These results were obtained with T= 500, m = 6, N= 500.

The next step is 201, where the aim is to determine a single, representative colour shift for each colour component. This is achieved by taking the average.of the colour shifts that resulted in matches, that is to say the shift [X _AB = (jJ_r , ju_g,ju_b) given by

YMAX

∑iC_b(i) μ_b = ^

C AB

Note that C_AB was counted separately, but could alternatively be calculated from

/MAX C_AB = ∑ C_b(i) .

!=0

Of course, rather than counting the number of times each value occurs, the same result could be obtained by storing in a memory each value that resulted in a match, and then taking the average of the stored values.

For the example shown in Figure 6, a value of μ_AB = (-28, 42, 2) was obtained. If desired, this result can be subtracted from image B by way of a correction, that is to say, the original pixel values b(u,) are replaced by b(u,) - μ_AB- Or if image B is considered to be correct the correction could be added to image A.

As an alternative to taking the average, the values of μ_r, μg, μ_b could be obtained by estimating the position of the peak of the histogram. This may be done by looking for the maximum value, the median, or the position of a window of width w in which the sum of w values was maximised.

In practice it is found that good colour correction can be obtained in this manner. In tests, it has become clear that it is not necessary for colour correction purposes that the two images should be identical in all respects except brightness/colour. For example, good results have been observed in a case where image A covered only part of the scene shown in image B. More significantly, the images do not have to be of the same scene: it is only necessary to have a range of matching colour adjacencies that allow forks to match in both images in order to obtain a good estimate, though similar structures present in both images will allow a greater proportion of colour shifted forks to match and therefore produce a more statistically reliable colour, shift estimate. This offers the opportunity for automatically tailoring images to individual preferences by simply supplying examples of acceptable images and allowing the system to adjust the colour balance accordingly. If desired, the system could be incorporated in a camera for setting the exposure or aperture, and/or (in a digital camera) for adjusting the colour balance at the time a photograph is taken, or immediately afterwards.

The introduction of random colour variations into the scoring process, as described above with reference to Figure 3 can also be combined with the system described in our aforementioned international patent application WO2006/030173 which suggested the introduction of random variations in the number of tines m, and/or in the maximum fork diameter D and/or the fork shape to be applied to the second image (reflection, rotation, scaling). However, the more variables that are perturbed, the greater is the number of iterations need to produce a good result, so for fast performance one prefers to vary only two or three parameters at once.

Thus, in such an implementation, Step 105 of Figure 3 would be preceded by one or more of the following steps for transformation of the displacements d;.:

A. a random reflection is applied to the fork:

Choose a random integer R as 0 or 1;

If R = 1, set djd = - d_xi for all i = 1 ... m. dy, are unchanged.

This is a reflection about a single vertical axis. Reflections about other axes are encompassed by this reflection plus the rotation step that comes next.

B. choose a random rotation angle β in the range β_max to P_1nJ_n. Then the fork represented by d; is rotated by this angle by performing, for each i: d_xi' = d_xi.cosβ - dyi.sinβ dyi'

+ dyi.cosβ. C. a random scale factor α is chosen in the range (X_nUn to O_m3x. The fork is then scaled by this factor:

The scaling and rotation, are relative to (0,0), though scaling or rotation about any other centre would do just as well.

These processes has been described in terms of values representing the colour or brightness of elements of a two dimensional visual image. This would include a pattern that is viewable as a visual image even if it was not originally generated by detection of visible light, such as an x-ray image or tomography. It may be applied also to three-dimensional visual images (as in tomography), or to other patterns, especially where the values represent measurements where it may be of interest to identify a global shift which is hard to spot using conventional techniques; examples include measurements of magnetic fields, or of temperature.

Claims

1. A method of comparing a first pattern (A) represented by a first ordered set of elements (X_J) each having a value with a second pattern (B) represented by a second ordered set of element (u,) each having a value, comprising iteratively performing the steps of: (i) selecting a plurality of elements from the first ordered set ;

(ii) for each selected element of the first ordered set, selecting an element from the second ordered set, such that the selected elements of the second ordered set have, within the second ordered set, a set of positional relationships relative to each other that is the same as, or a transformation of, the set of positional relationships that the selected plurality of elements of the first ordered set have . relative to each other;

(iii) determining whether the selected elements meet the match criterion; and (iv) in the event of a match, updating at least one similarity score (S_AB); characterised by, on each iteration, generating a random shift value (γ), and in that the match criterion is that the value of each of the selected elements of the first ordered set differs from the value of the correspondingly positioned selected element of the second ordered set by an amount that is equal to, or within a set margin of, and in the same sense as, said shift value.

2. A method of comparing a first pattern (A) represented by a first ordered set of elements (xθ . each having a value with a second pattern (B) represented by a second ordered set of element (uj) each having a value, comprising iteratively performing the steps of:

(i) selecting a plurality of elements from the first ordered set;

(ii) for each selected element of the first ordered set, selecting an element from the second ordered set, such that the selected elements of the second ordered set have, within the second ordered set, a set of positional relationships relative to each other that is the same as, or a transformation of, the set of positional relationships that the selected plurality of elements of the first ordered set have relative to each other; and

(iv) determining whether the selected elements meet a match criterion; characterised in that each iteration includes the step of generating a random shift value (γ); the match criterion is that the value of each of the selected elements of the first ordered set differs from the value of the correspondingly positioned selected element of the second ordered set by an amount that is equal to or within a set margin of, and in the same sense as, said shift value; and by, in the event of a match, recording the shift value; and, at the conclusion of said iteration, estimating from said recorded shift values a single representative shift value.

3. A method according to claim 2 including applying to one image a correction based on said single representative shift value.

4. A method according to any one of the preceding claims in which the selection of the parameter(s) is random or pseudo-random.

5. A method according to any one of the preceding claims, in which each value (a) comprises a plurality of components, and the shift value has a corresponding number of components .

6. A method according to claim 6 in which the patterns are visual images and the components are brightness and/or colour values.

7. A method according to any one of the preceding claims, comprising firstly processing the first and second ordered set to identify regions thereof containing significant detail, and in which the selection of elements is constrained such that at least one of the selected elements of the first ordered set shall lie in the or an identified region of the first ordered set and/or that at least one of the selected elements of the first ordered set shall lie in the or an identified region of the first ordered set.