Science & Justice
Volume 51, Issue 1 , Pages 28-37, March 2011

Digital imaging and image analysis applied to numerical applications in forensic hair examination

  • Elizabeth Brooks

      Affiliations

    • Forensic and Data Centres, Australian Federal Police, Canberra, Australia
    • Corresponding Author InformationCorresponding author. Australian Federal Police, GPO Box 401, Canberra ACT, Australia 2601. Tel.: +61 2 62233765; fax: +61 2 62233270.
    • ,
  • Bruce Comber

      Affiliations

    • Forensic and Data Centres, Australian Federal Police, Canberra, Australia
    • ,
  • Ian McNaught

      Affiliations

    • Petroleum Institute, Abu Dhabi, United Arab Emirates
    • ,
  • James Robertson

      Affiliations

    • Forensic and Data Centres, Australian Federal Police, Canberra, Australia

Received 27 April 2010; received in revised form 28 June 2010; accepted 29 June 2010. published online 02 August 2010.

Article Outline

Abstract 

A method that provides objective data to complement the hair analysts' microscopic observations, which is non-destructive, would be of obvious benefit in the forensic examination of hairs. This paper reports on the use of objective colour measurement and image analysis techniques of auto-montaged images. Brown Caucasian telogen scalp hairs were chosen as a stern test of the utility of these approaches. The results show the value of using auto-montaged images and the potential for the use of objective numerical measures of colour and pigmentation to complement microscopic observations.

Keywords: Objective numerical analysis, Forensic hair examination, Colour models, Pattern matching, Digital imaging

 

Back to Article Outline

1. Introduction 

The human scalp hair as a ‘morphological’ unit does not easily divide into component parts nor do these parts lend themselves to a number value [1]. Forensic hair examiners have long acknowledged this difficulty, even inability, of assigning numerical values to the features that discriminate one hair as being different from other hairs. Features of hair currently measured ‘objectively’ include hair length and shaft diameter, medullary index and mean scale count index, however, Robertson [2] suggests such features add little discriminatory value to hair evidence. Conversely, the features with good discriminatory value such as hair colour, pigmentation characteristics and medulla type have, to date, been relatively numerically resistant. While comparison microscopy allows a questioned hair to be assessed as not significantly different from known source hairs, previous attempts to measure colour and hair pigmentation features have generally fallen short of true objectivity. Other attempts to attribute an individual as the source of the hair evidence by means other than microscopy have also met with limited success [3], [4], [5], [6], [7], [8], [9] the exception being that of nuclear DNA (nDNA) analysis obtained from an anagen hair root, [10], [11], [12], [13] or one that bears sheath or tissue material, and to some extent, mitochondrial DNA analysis [14], [15], [16], [17]. Even if it is accepted that the gold standard for hair identification is some form of nDNA testing, microscopic examination will remain an important part of any serious hair examination protocol. The reason for this is simple. In real case work hair examination most often involves the examiner having to look at many hairs. The cost factor in subjecting all the hairs recovered from a typical forensic case scenario would be prohibitive. Hence, microscopical examination has an important and necessary role to play in the screening or triage of recovered or questioned hairs. In our view, microscopic examination up to and including examination at ×40 objective magnification enables an experienced (and competent) examiner to achieve significant levels of discrimination and elimination of hairs based on observed/significant differences [18]. However, we accept that in the absence of numerically assessable features such discrimination relies on the professional judgement of the examiner, which should be informed by experience. In an ideal world it would be highly desirable to be able to support expert assessment with objective and measurable evidence. It is generally accepted that colour and microscopically observable contributing pigment components are the features most relied upon by forensic hair experts. Hence, this study was designed to assess the potential for these features to be numerically captured and how these might then contribute to a more effective and efficient examination protocol.

Back to Article Outline

2. Materials and methods 

The techniques developed in this study used high quality digital images where the image pixel values enabled the assignation of numerical values to such features as colour and pigmentation characteristics. Telogen growth phase scalp hairs [10], [11], [12], [13], [18] obtained from the hairbrushes of 10 nominally brown haired male and female Caucasians were selected because telogen hairs comprise the vast majority (up to 95%) of hairs recovered in forensic casework. Furthermore, telogen hairs are generally less suitable for routine nDNA testing [19], [20]. Both male and female adult donors provided hair samples, many of whom, including the author, worked in the same forensic area. All samples were collected with appropriate ethics approval. Scalp hairs from donors of Caucasian origin were selected, because differentiation of hairs of this type is typically considered the most informative type of hair examination [11]. The focus of the research was twofold:

∘compare colour analysis of hair images from brown haired Caucasians within three standard, internationally recognised colour models, namely Red–Green–Blue (RGB) colour model; CIE XYZ Tristimulus (1931) colour model; and CIE L*a*b* (1976) colour model and

∘using the same sets of digital images, undertake pigment pattern recognition analysis looking at both intra- and inter-individual variations.

2.1. Preparation of hair samples 

Telogen hairs from ‘nominally’ brown haired individuals – at least 20 hairs – were collected from the donor's hair brush, and placed in a folded A4 piece of paper and sealed in a snap seal plastic bag. Hair donors were provided with a new hairbrush with which to brush their hair vigorously and collect the hairs retained in the brush. Each sample was labelled with a donor number (1 to 10) from which 10 hairs were randomly chosen from the 10 bulk samples, measured and mounted in Hystomount™, a semi-permanent medium, on glass slides — one hair per slide. Hairs were placed on a slide in a straight configuration or in figure of eight configurations until the full hair length was taken up. Each slide was labelled A to J for each set of samples (10 sets of samples were mounted). A #1 cover slip was carefully placed over the hair, aiming to avoid air bubbles forming in the Hystomount™. Slides were then placed in a 30°C oven to dry slowly for approximately a week, and stored in wooden slide boxes ready for imaging. Images collected from mounted hairs were stored on CD-ROMs for analysis.

2.2. Equipment and software 

The equipment for this project included a compound transmitted light microscope, a digital camera designed for light microscopy and a computer. The software required involved an image analysis package, image management system and image combining software. The Olympus DP70 camera [21] and its associated image management system were used to take all images that were then combined into z-plane montages using Auto-montage Pro software [22]. Finally, the image analysis package selected was V++ Version 4.0 [23].

Images collected from mounted hair samples used a Leitz Diaplan compound light microscope in bright field mode, which was chosen by default, because the Olympus DP70, the camera/computer system of choice, was attached to this microscope. This equipment, that belonged to the Commonwealth Scientific and Industrial Research Organisation, Division of Entomology, Canberra, Australia, was hired on a daily basis for image collection only.

2.3. Sampling procedure 

From the 10 sets of mounted samples, five hair samples were randomly selected for imaging. Each of the five hairs per person was imaged in five separate regions along the hair shaft, beginning one centimetre from the root end and then one set of five images was taken every second centimetre for five regions along the hair shaft, towards the hair tip, see Fig. 1. In one instance, where the hair was shorter, only four sets of images were collected for the selected hair. The numbering was consistent throughout for each of the 10 samples and comprised the number of the hair, followed by the number of the region being imaged and then the image number; for example — sample 1A2-3 indicates hair A of person 1, image region 2, image 3 of the set of images, see Fig. 2. In this manner each person acquired 25 sets of images from which montages were formed (one montage for each of the five regions imaged along the shaft) and further analysis undertaken. All images were stored on CD-ROMs and formed the basis of the materials for analysis.

2.4. Image acquisition 

Images saved as Tiff files (Tagged Image File Format) were collected under the same camera/light and microscope conditions. A ×40 Fluorotar oil immersion lens was used; the depth of the focal unit was 2μm; tungsten lamp illumination voltage was set at 10.5V. The camera setup included the following parameters: automatic exposure, image size 1360×1024pixels (experimentation indicated 1.3megapixel images allowed for rapid acquisition and analysis while providing suitable image quality); image acquisition sensitivity set at — ISO200; spot size — 0.1%; and white and black balance was corrected before capture of each set of images. Image collection proceeded in the following manner. The number of focal steps through the hair – from the uppermost ‘in focus’ plane to the lowermost ‘in focus’ plane – was determined using the fine focus knob on the microscope. This number was divided by 3, (for example, if the focal range was 21 focal steps, this number was divided by 3 which meant that only the middle 7 images were collected) the fine focus knob on the microscope was moved to what corresponded to the first of the mid-third steps of the total number of steps through the hair. Manually selected focal planes at 2μm steps (in the z-plane) were photographed through the mid-third of the hair shaft. Images captured for each subsample were saved in separate folders ready for the image montaging step, see Fig. 1.

2.5. Image combining software 

Auto-montage Pro [22] software enabled the acquisition of fully focused images from an optical microscope. The software is designed to overcome the problems traditionally associated with the inadequate depth of field for imaging 3-D specimens, by allowing a composite, focused image to be produced. This is achieved by combining the focused parts from a series of images focused at different heights or optical sections (often called multi- or extended focus), through the specimen to form an in-focus composite. If the focal planes are too few, or too far apart, the resulting final image will not be totally in focus. Auto-montage Pro [22] software uses a wide range of algorithms (known as Montage Methods (details of the algorithms are commercial-in-confidence) for calculating the way in which the focused image is produced. The Montage Methods allow the microscopist to view the results from the most suitable algorithm for the particular sample before processing takes place. This software is compatible with any digital camera system attached to an optical microscope from which source images can be imported into Auto-montage Pro [22], where they are automatically analysed and all high detail regions are combined into a single, focused, image — see Fig. 2.

2.6. Image analysis software 

V++ Version 4.0 [23] is a precision scientific imaging software package that supports advanced data types that range from simple binary and byte images to 96-bit floating point colour and 128-bit complex images. V++ Version 4.0 is fully compatible and integrated with Microsoft® Windows 95, 98, NT and XP; the imaging functions support monochrome and colour image processing operations and include a variety of data types for representing images. Data types with extended precision — such as the word or short-integer types (16bits/pixel) were used in this project.

2.7. Generating numerical values from digital images 

Digital images are generally square or rectangular, arranged in a typical Cartesian coordinate system, allowing an x,y coordinate to represent each pixel that results from the digitised image [25]. The x coordinate specifies the horizontal position or column location of the pixel, while the y coordinate indicates the row number or vertical position. A digital image is thus composed of a series of intensity values (pixels) and ordered through an organised (x,y) coordinate system. Image montage formation software, Auto-montage Pro [22] (Syncroscopy) and image analysis software, V++, Version 4.0 [23] (Digital Optics) combined with the Cartesian coordinate pixel system in the digital image enabled the generation of numerical values in an objective manner from information contained within pixel intensities of the digital image.

2.8. Generating numerical values from hair colour using colour models 

Determination of the value of component colours for the three colour models – RGB, [26], [27] CIE XYZ [28], [30] and CIE L*a*b* [29], [31] – was performed within V++ Version 4.0 [23] by viewing the normalised montaged image of the hair that excluded all background colour (see Fig. 3).

Each image was initially captured after performing white balancing because a white reference point is required for both the CIE XYZ and the CIE L*a*b* models [30], [31], [32]. No modifications were made to any of the images in terms of brightness and contrast, gamma or colour channels. For each person, 25 image sets (five sets of montaged images by five hairs per person) were analysed within the three colour models. Direct comparison of values obtained within each colour model assessed (RGB, CIE XYZ and CIE L*a*b* [30], [31], [32]) was achieved by conversion of all stimulus values to the range of 0–255. This was necessary as the CIE L*a*b* model is nonlinear and device independent, therefore, when the RGB colour model is changed to the CIE L*a*b* colour model a conversion program [31], [32] is required within V++. During conversion from RGB to CIE L*a*b* the image is transformed into three intensity scales that represent the three axes — L*, a* and b*. L* measures the lightness of the image on a scale with values from 0 to 100, where 0 is black and 100 is white. a* measures the colour saturation from red to green with a scale from +100 to −100, where positive values indicate varying intensities of red; and b* measures the colour saturation from yellow to blue where the scale is also +100 to −100, with positive values indicating varying intensities of yellow. However, because the script function [31], [32] does not accept any data that are negative and to maintain intermodel comparison, the values of L*, a* and b* were transformed to the range of 0 to 255 in the following manner:

Detection of the component colours was based on colour values for all the pixels making up the montaged image and averaging these mean values. The resulting data provided a unique set of coordinates for the five regions along the five hairs investigated for each of the 10 brown haired persons within each of the three colour models. Discriminant Analysis in SPSS, Version 12.0.1 [33] assessed the ability of the three colour models to assign the investigated hairs to their rightful owner. Discriminant analysis is useful for detecting the variables that allow discrimination between different (naturally occurring) groups and for classifying cases into different groups with a better than chance probability.

2.9. Generating numerical values from hair pigmentation characteristics using pattern matching 

From a single digital image (one optical section) of each set of images (see Fig. 4, Fig. 5) a small section was ‘extracted’ to create a new exact image (achievable within the Edit tab in V++ Version 4.0 [23]) and used as a pattern comparison tool, see Fig. 6.

For each person a ‘pattern extract’ was created and compared with a complete single image from another imaged region within the same group of samples, e.g., Sample 3D3 (see Fig. 5). The pattern extract was then compared with a single image from each of the other four sample sets for person 3 (3D1, 3D2, 3D4, and 3D5) and finally with single images from each of the other 20 image sets for person 3 (3E1–3E5; 3G1–3G5; 3H1–3H5 and 3I1–3I5, etc). The mechanics of pattern matching and the generation of ‘pattern matching values’ was based on a script written by Comber [34], [35], for operation within V++ Version 4.0 [23], and is both described and illustrated in Fig. 6. Pigment pattern values were based on subtraction of the extracted pattern image from the sample image at each pixel location along the x-axis of the sample image. Single images for pattern matching were used in the first instance, however, pattern matching using the montaged images was equally successful. Because each image was digital the pixel number on the x-axis (0 to ~≥1000) provided a data point, to which the subtracted value of the pattern match was assigned. If this value was zero then the match was ‘perfect’. However, if the subtracted average value was greater than zero, then this indicated the degree of numerical mismatch, see Fig. 6. In total, each pattern extract was matched to approximately 25 images within a person sample set and one pattern extract (from person 10) was matched to five randomly selected single images from all person sample sets (person 1 to person 10).

Back to Article Outline

3. Results 

3.1. Generating numerical values from hair colour using colour models 

The protocol developed here used digital images as the primary material and three internationally recognised colour models (RGB, CIE XYZ and CIE L*a*b*colour models) to discriminate between 10 nominally brown haired Caucasians. The three colour models operated within the image analysis software — V++ Version 4.0 [23]. This software supports the RGB true colour model and facilitates conversions between linear RGB and CIE XYZ (1931) coordinates using a 3×3 colour transformation matrix. The transformation matrix required to convert RGB to CIE L*a*b* (1976) colour was written in V++ Version 4.0 script [23], [30], [31], [32]. Fig. 7 presents examples of raw data generated for hair colour within the three colour models studied. The complete sets of raw and analysed data can be viewed at the following reference [24].

The raw values were averaged by colour model integer sets, see Table 1, and discriminant analysis undertaken where hairs were attributed to their owners based on colour alone, with varying degrees of validity (SPSS [24], [33]).

Table 1. Example of mean values by person and hair sample for the RGB colour model.
Source: SPSS analysis [24], [33].
PersonHairR meansG meansB means
11178.43166.40133.28
12186.44165.63118.77
13188.21162.88115.65
14189.14176.88136.28
15192.37173.70127.35
21186.58172.39130.10
22189.99173.78133.51
23183.14175.81139.72
24190.15175.48133.04
25179.82173.44137.04
31179.67160.95139.95
32185.11170.18153.67
33197.25405.55106.79
34181.94163.37145.27
35177.74125.33139.57

Table 1 lists some of the mean numerical values obtained from the analysis of the digital images of brown hair samples in the RGB colour space.

Two discriminant variables were used to develop the three colour model canonical discriminant plots and determine the statistical validity of the analysis. For each model, two functions describe almost all of the variations in the sample (RGB model — 99.9%; CIE XYZ model — 99.0%; and CIE L*a*b* model — 98.1%) [33]. Fig. 8 indicates the grouping of persons by hair colour as analysed in the RGB colour model. The RGB model readily separates the persons' hair colours into two distinct groups of five, where persons 1 to 5 are clustered on the right hand region of the function 1 axis and persons 6 to 10 on the left hand region of the function 1 axis. This separation generally corresponds to the nominal hair colour observed visually, persons 1 to 5 having ‘lighter’ brown hair and persons 6 to 10 having ‘darker’ brown hair.

The individual groupings have two obvious features. First, persons 1 to 4 indicate a reasonable degree of overlap, as do persons 6, 7 and 8. Second, there does not appear to be any clear allocation of a person with hair colour that is entirely separate from any other. The CIE XYZ and the CIE L*a*b* models also separated the hair colours into two distinct groups of five in the canonical discriminant plots (not shown here [24]). The CIE XYZ plot separation was similar in most aspects to that of the RGB model plot, with the differences being that person 3, person 9 and to some extent person 6 appeared relatively separated from the other allocations of samples. The CIE L*a*b* plot appears to represent the inverse of the previous two colour models in that persons 1 to 5 are clustered on the left hand region of the function 1 axis and persons 6 to 10 on the right hand region of the function 1 axis. This separation also generally corresponds to the nominal hair colour observed visually, where persons 1, 2, 4 and 5 are closely associated and appear to overlap with each other, as do persons 6, 7, 8 and 10, while persons 3 and 9 appear relatively separated from the other groupings of samples [24]. A technique that is often used in discriminant analysis is to draw an ellipse around each centroid to produce a ‘95% confidence region’ for the true centroid (the centroid that would be produced 95% of the time from studies of a large number of samples). The centroid ellipses are not drawn on the plot (Fig. 8) because this study was more interested in the location of the single colour values. Closest centroid analysis was applied to the data pertaining to the three colour models and indicates the ability of each model to correctly assign a hair to an individual based only on numerical colour values. The results of correctly assigned hairs are summarised in Table 2.

Table 2. Total number of hairs allocated to the person with the closest centroid for the three colour model analyses — case-wise analyses.
Correct allocation of hair to person by colour modelPerson 1Person 2Person 3Person 4Person 5Person 6Person 7Person 8Person 9Person 10
RGB1243444154
CIE XYZ1353444244
CIE L*a*b*3243442052

Number indicates number of times (for a maximum value of five) the hair for a person is correctly assigned to that person, using case-wise analysis for each of the three colour models [24], [33].

Based on these figures (Table 2) the accuracy of canonical discriminant plots expressed as a percentage of correctly allocating by colour only a hair to the individual is as follows — 64% for the RGB model; 68% for the CIE XYZ model and 58% for the CIE L*a*b* model. Another way of separating individuals by hair colour using discriminant analysis is to assume a null hypothesis where the centroid for the first person in the pair is the same as the centroid for the second person in the pair. An F-test of this null hypothesis is then carried out [24], [33]. If the P value is >0.05 then the null hypothesis is accepted (i.e., the pair-wise centroid is the same). Where the P value is <0.05 then the hypothesis is rejected (i.e., the pair-wise centroid is different) [24], [33]. The results from the pair-wise analysis within the RGB colour model (not shown [24]) indicates that the null hypothesis was accepted for 7 of the 45 pairs of persons (i.e., the centroid for 7 of the 45 pair-wise comparisons could not be separated). It appears, from these tests, that the data for 38 of the centroid pairs are different enough to allow a significant separation. However, for persons 1, 2, 3 and 4 who are known to be different, the data acquired for their centroids did not establish a significant difference between their centroids and the centroid of at least one other person. The same analysis criteria were applied to the other two colour models studied and Table 3 summarises the number of colour allocations for each colour model where the closest centroid was significantly different, for the 45 pairs compared.

Table 3. Pair-wise allocation of hairs to the person where the closest centroid was significantly different in the RGB, CIE XYZ and CIE L*a*b* colour model analyses.
Source: SPSS analysis [24], [33].
Colour ModelPair-wise centroid significantly differenta
RGB38
CIE XYZ40
CIE L*a*b*39

Table 3 summarises the ability of each of the colour models to allocate a hair sample to the correct owner for 45 pairs of analyses.

aNull hypothesis accepted; Probability is<0.05 (45 pairs).

Analyses of digital images of scalp hair colour using internationally recognised and standardised colour models removed much of the usual subjectivity associated with human colour perception. This study has not attempted to ‘define’ the hair colours imaged in terms of ‘light brown’ or ‘dark brown’, instead it has, through the assignation of specific colour model integer sets, attempted to allocate hair samples to the rightful owner. Further, the study compared colour models to identify which, if any, achieves the best allocation of hair sample to its owner. Table 4 presents two measures that show how well the data from each colour model are able to demonstrate that the colour characteristics of one person's hair are different from those of another person's hair.

Table 4. Summary of percentage allocation of hair samples by colour model and analysis type.
Source: SPSS analysis [24], [33].
Colour ModelCase-wise analysis % individual samples correctly allocatedPair-wise analysis % paired samples for which centroids were distinct
RGB6484.45
CIE XYZ6688.9
CIE L*a*b*5886.7

Table 4 indicates the percentage of correctly allocated hair samples to their individual donor.

As indicated in Table 4, it appears that by a small percentage, the CIE XYZ colour model best allocates hair samples to the correct person within the case-wise analysis; and by a rather greater difference, distinguishes between pair-wise centroids. The differences between the three models in the case-wise analyses are greater between the two CIE colour models (58% and 66% for CIE L*a*b* and CIE XYZ, respectively) than between either of these models and the RGB model (64% and 66%; 64% and 58% respectively).

3.2. Generating numerical values from hair pigmentation using pattern analysis 

Pattern matching [34], [36], [37] raw data were generated from ~25 single images for each of the 10 persons, with values collected for at least 1000pixel positions for each image [34] (see [24] for complete data sets). The means and maximum values for each of ~25 data sets per person were calculated, (SPSS [24], [33]) and used to investigate the distributions of the means and the maximum values of the pigment pattern matching data. The null hypothesis in this context was that ‘the distribution from which these values came was a normal distribution’. Kolmogorov–Smirnov [33] tests were used to confirm the validity of a data set as a normal distribution where, once the data were log-transformed, the null hypothesis was tested. Logarithms are to base 10, and it was necessary to add 1 for each transformation because some values were zero, or even slightly negative. Fig. 9, Fig. 10 show the distribution of log(mean +1) and log(maximum +1), respectively, for the pigment pattern matching data.

The distributions (Fig. 9, Fig. 10) for log(mean +1) and log(maximum +1) are for pigment patterns of persons matched with themselves. These distributions may be used to investigate whether the values obtained when matched with unknown hair samples are of a similar magnitude to those shown in these distributions, or if they are generally greater in magnitude. To validate this assumption, the pattern extract from a hair region of person 10 was matched to five randomly selected images from each of the hairs that constituted a person sample for every other person. The results for the mean values of this pattern matching exercise, for each person, are shown in Table 5.

Table 5. Mean, log(mean +1) and percentage values where a pattern region from person 10 was matched with a random selection of hairs from all other persons.
Person 1 (mean)Person 2 (mean)Person 3 (mean)Person 4 (mean)Person 5 (mean)Person 6 (mean)Person 7 (mean)Person 8 (mean)Person 9 (mean)Person 10 (mean)
Hair samplea
137.71948.19555.72749.07529.46015.15814.38515.47411.0474.270
240.52051.78359.64762.20154.81815.5039.16813.5994.8411.985
349.11556.68452.31755.10043.87817.64810.93219.9655.9374.296
459.42551.33343.95042.03247.64416.60220.35112.17615.6373.042
559.68963.90019.42139.99139.92412.22410.76015.2898.5933.293

Log (mean +1)b
11.5881.6921.7541.7001.4841.2081.1871.2171.0810.722
21.6181.7221.7831.8011.7471.3121.0071.1640.7660.475
31.7001.7611.7271.7491.6521.2711.0771.3210.8410.724
41.7811.7191.6531.6341.6871.2461.3291.1201.2210.607
51.7831.8121.3101.6131.6121.1211.0701.2120.9830.633

Percentagec
10.30.10.00.10.98.79.98.218.365.0
20.20.10.00.00.04.126.111.458.989.9
30.10.00.10.00.25.618.73.848.364.8
40.00.10.20.20.16.73.614.88.078.8
50.00.04.20.20.214.719.38.529.276.0

aRandom selection of one image from each of five hair samples per person. Means of 1000 values calculated for each hair and mean of mean values calculated.

bLog(mean +1) transformation of overall mean values (mean of five data sets per person).

cPercentage of the distribution obtained from persons matched with themselves that exceeds the value of log(mean +1).

For example, the mean for the first hair of person 1 was 37.719, giving a log-transformed value of 1.588. When this value was referred to the normal distribution in Fig. 9 it was found that the probability of 1.588 or a more extreme value was 0.3%. It may therefore be inferred that this hair was unlikely to be a hair matched with another from the same person (as was known to be the case). By way of contrast, the second hair from person 10 gave a log-transformed value of 0.475 and a probability of 89.9%. The conclusion that could be drawn from this inference is that there was a good chance that this was an example of a hair matched with another from the same person. In this instance, this was also the correct conclusion. While a similar set of inferences may be made with regard to the maximum values (not shown [24]) they can also be used as a ‘quick check’. For example if a ‘quick check’ of the maximum value for say person 1 has a high value of mismatch comparative to a hair from person 10 with a low maximum value of mismatch this may indicate that the two hairs are most likely to be from different sources, as is known to be the case. For complete data sets please refer to the SPSS analysis [24], [33].

The mean and maximum data for the 10 persons matched with themselves (Fig. 9, Fig. 10; Table 5 ) were used to estimate the components of variance. The three sources of variability throughout the pigment pattern matching were:

1.The individual donor of the hair sample, i.e., person 1 to person 10,

2.Hair samples within a person, i.e., five individual hairs from the same known source,

3.Sites within each of the five individual hairs, i.e., five separate regions along an individual hair shaft were imaged.

To estimate the components of variance a random model ANOVA was carried out using SPSS [24], [33] (2003) (for SPSS generated data see [24] showing the estimates of covariance parameters). The results for log(mean +1) and log(maximum +1) are shown in Table 6.

Table 6. Percentage of the total variance due to three sources of variability for pigment matching data, for both log(mean +1) and log(maximum +1).
Source of variance% log(mean +1)% log(maximum +1)
Person13.211.1
Hair within person69.070.5
Site within hair within person17.818.5

Back to Article Outline

4. Discussion 

In this study, it has only been possible to examine a very small population of individuals and it is recognised that it is now necessary to examine many more individuals to properly assess the full utility of the colour and pattern recognition approaches.

4.1. Colour measurement of hair 

Colour measurement is not a new concept in forensic science. Microspectrophotometry (MSP) applications to forensic examination of fibres, paints and inks have been used for around 40years [38]. Based on the chemistry of the dye used to colour the material, the absorption characteristics of the visible spectra can be used to determine the colour, which is itself determined by colorimetry. Colorimetry was developed in order to objectify human colour perception, thus allowing precise colour recording and effective communication about it [30], [38]. MSP colour measurement uses the CIE XYZ [30] (1931) tristimulus values and the complementary chromaticity coordinates symbolised as x′, y′, which operate in absorbance rather than transmission mode. The recent development of extensive and validated fibre databases [39], [40], [41], [42] has seen instrument measured colour as both objective and acceptable in forensic science.

In 2003 Bednarek [27] selected 50 blond and 50 brown hairs graded according to Ogle and Fox [43] classifications of D01–D03 and D21–D25, respectively. Bednarek [27] correctly identified 91 of the 100 hairs using an RGB colour model to determine unique colour coordinates from digital images. He compared these results with microscopic analysis based on criteria proposed by Ogle and Fox [43] and found for the same 100 hairs, 74 achieved correct identification and 26 were incorrectly identified for colour. Ogle and Fox [43] attempted to provide criteria for the definition and description of colours based on establishing and publishing photographic standards of colours and their shades. Verma et al. [7] reported a partially automated system (Hair-MAP) for forensic hair comparison and analysis also based on digital image capture and image analysis software. Verma et al. [7] used Khoros (image processing software) to convert coloured hair images into matrices of red, green and blue bands, as grey scale images. Colour classification values were created from histogram comparisons between the three colour bands, producing a value for each band. While classification by colour was just one aspect of Hair-MAP, it still relied on descriptive colour subtypes such as colourless/translucent, yellow brown, yellow red, reddish brown, etc., which are primarily colour perception descriptors. However, human perception of colour is a significantly complex field within colorimetry, thus, development of objective criteria for colour definition of hairs, or any other coloured item biological or otherwise, must necessarily be based within the colorimetry concepts and the internationally accepted, standardised colour models.

The methodology developed in this study compared hairs scored for colour within the measurement parameters of three standard colour models (RGB, CIE XYZ and CIE L*a*b* [30], [31]). It was found after discriminant analysis that pair-wise separation (closeness of paired centroids [33]) gave the best results for separating brown hair samples from each other. Furthermore, each model differed in degree of separation accuracy by two percentage points only. Based on the pair-wise analysis, the percentage of the 10 brown haired persons who were correctly separated was 84.45% by the RGB model; 86.7% for the CIE L*a*b* model; and 88.9% for the CIE XYZ model. Bednarek [27] concluded that hair colours could be described in a digital manner within the RGB model and provide the hair examiner with a set of component values (for red, green and blue) as opposed to a perceptual evaluation. Verma et al. [7] proposed the development of a database using Hair-MAP as the prototype, but it appears that colour would not play an important characterising role. Vaughn et al. [44], [45] have investigated and compared hair colour measurement within the CIE L*a*b* colour model using techniques including digital imaging and reflective spectrophotometry. Vaughn et al. [44], [45] concluded that using the CIE L*a*b* colour model, reflective spectrophotometry more accurately assigned macro images of hair colour to the individual. Clinicians treating skin complaints [46], [47], [48], [49], [50] have established the link between digital imaging, colour measurement, descriptions and colour models. The appeal to the clinician is clear: use of a method that is quantitative with principles that are internationally recognised and the colour space parameters that allow unambiguous communication of skin colour. The efficacy of the same methods, principles and unambiguous communication of colour for hair analysis also has much to recommend it to the forensic hair examiner. Measured colour values based on the colour models did remove the subjective evaluation of the ‘human observer’ and allowed meaningful objective measures of colour. Another advantage of describing colour via digital imaging and colour model analysis is the ease of communicating these results with the verifying examiner. As specific integer sets there can be no misunderstanding of what colour constitutes, for example, ‘light brown’, it will be a set of three coordinate values as described by the colour model of choice.

The technique described here has the ability to unambiguously demarcate ‘light’ brown hair from ‘dark’ brown hair for all three colour models. The separation of individuals within these two distinct groups was less comprehensive. In each group at least one of the five donors could be distinguished from the remainder (person 3 in the ‘light’ brown group and person 9 in the ‘dark ‘brown group). Finally, despite the apparent objectivity of hair allocations to rightful owners by colour model analysis, the forensic hair examiner is really interested in:

•how well does this type of analysis distinguish hairs from persons with ‘similar’ hair colouring?

Fig. 8 is an example of the canonical plots from the discriminant analyses. In general, all of the samples tended to separate into two main groups — persons 1–5 with the ‘lighter’ brown hair and persons 6–10 with the ‘darker’ brown hair. In both the CIE XYZ and CIE L*a*b* colour models, persons 3 and 9, and to a lesser extent, person 6, were relatively separated from any other person. The apparent complexity within the two major groupings did not readily allow individual separation of sets of hair samples, but it is important to remember that each of the colour models did separate individual hair donors within a group of brown haired people. The separation, while not perfect, was achieved by assigning hairs to an individual based only on the ‘shade’ of the brown hair. However, the best analysis for separating a similar hair colour from another was the pair-wise analysis, because this looked at how close the pairs' centroids were to each other. As shown in Table 4, the CIE XYZ model separates one hair sample from another, based on the hair colour, 88.9% of the time.

4.2. Pigment pattern matching 

The pattern matching of one hair with another did not attempt to isolate pigment features from each other, and identify them numerically. The efficacy of the pattern matching here is its simplicity and its ability to separate or associate patterns within or between hairs from 10 brown haired Caucasians. It also allows for the identification of two hairs from different people that appear to be difficult to separate. For example, the maximum values (raw data values) can indicate the degree of pattern mismatch between the extracted pattern and the questioned image. A large maximum value is a reasonable indication that the pattern extract is relatively dissimilar to the questioned hair pattern. On the other hand, where maximum values are numerically close the inference is that the small numerical difference between these two hair samples requires further investigation. Further investigation would imply careful comparison microscopy and possibly mitochondrial DNA analysis to exclude or include this person's hair sample as coming from the known source. Pattern values statistically represented from a questioned hair sample and compared with the data distributions in Table 6, (data defined by the log(mean +1) and log(maximum +1) and Fig. 9, Fig. 10) do allow an inference of association or exclusion from a known source. This is achieved by comparing the confidence indices (sample mean +1×standard deviation) from the questioned hair sample with those from the distribution described in Table 6 and Fig. 9, Fig. 10.

Verma et al. [7] have with Hair-MAP, provided a prototype database that claims to “provide [ing] more precise measures of the degree of intra- and inter-subject variation along the different microscopic features [of hair]”. Verma et al. [7] have not attempted to develop a database for pigment features because to do so (where every pigment feature had a numerical value) would be a significant computing challenge, even if agreement could be reached as to what each feature actually represented. However, the notion of databasing or data handling of hair and particularly pigment characteristics has much appeal. Hair examiners also wish to avoid Type II errors [51], [52], [53], [54] and acknowledge the merit of an objective means of confirming a microscopic hair comparison that does not involve mtDNA testing, which is ultimately destructive.

Once generated, the pattern matching values for persons matched with themselves were transformed into log(mean +1) and log(maximum +1) values, and their distributions determined [24], [33]. As for colour analysis the pigment pattern matching data showed clear separation between ‘light’ and ‘dark’ brown hair donors, see Table 6. Further, when the pattern extract from one individual (person 10) was compared with the others in the group (persons 1–9) there was no overlap between the log(mean +1) values, see Table 5, or log(maximum +1) (data not shown see [24]). In fact, the average scores for the other donors were between two and greater than three standard deviations from the values obtained for person 10 (five donors had values greater than three standard deviations from the mean). Two donors had standard deviations within two standard deviations of the mean; and two donors were located within one standard deviation of the mean, as calculated from the average values of the log(mean +1) data in Table 5 and Fig. 9. This not only gives an indication of the different pigmentations seen in the hairs but also allows for the extent of that differentiation to be quantified. Further, the probability of a pattern match can be inferred or denied from the calculated distributions and the log(mean +1) and log(maximum +1) data, for example as seen Table 5 for log(mean +1).

The mean and maximum data for the 10 persons matched with themselves (Fig. 9, Fig. 10) were used to estimate the components of variance. By ANOVA [33] analysis the variance between hairs within one person appears to be significantly greater than the variation between people with brown hair. Such a variance is counter-intuitive as the variation between the ‘lighter brown haired persons’ and those with ‘darker brown hair’ is visually obvious. This finding may indicate that more hairs per person need to be sampled and possibly imaging fewer and more widely spaced regions along the hair may decrease the variance of the intra-person data. To address the small apparent variation between brown haired people – approximately 12% – a selection of visually defined groups of Caucasians could be sampled. Comparing inter-person variation between groups of, for example, visually blond and red haired persons may indicate that the lack of inter-person variation is to be expected or otherwise. Brown hairs are particularly challenging for the hair examiner to discriminate. Sampling a group of brown haired Caucasians significantly challenged the techniques developed, however, this would, no doubt, be similar for other hair colour groups. Coordinates obtained from the montaged images and objectively compared across three colour models distinguished colour here. Analysis, based entirely on the shade of brown, determined groups of brown haired donors and within these shades it was possible to identify trends where all hairs were associated with the correct donor. Pigmentation discrimination using numerical values obtained from pattern recognition within digital images also divided the donors into light and dark brown haired groups. When the pattern from one individual was matched with the other nine donors this technique allowed for complete separation of the individual from the remainder.

Back to Article Outline

5. Conclusions 

The outcomes of this project include the novel application of auto montage techniques to produce superior in-focus images of hairs. This optical sectioning, and the images produced, will be of real value in a number of ways. These include images that can be used to demonstrate hair microscopic features for training and court purposes and images that can be subjected to numerical analysis and hence form the basis of future databasing of character frequencies. A second project outcome is the new method of applying colour and image analysis techniques to hair examination. The results of this study show the clear potential of both approaches as a tool to hair examiners that would give them methods to assess and classify colour and pigment characteristics (in the form of repeating patterns) in an objective way. It is not envisaged that such an approach would replace the human observer but rather that these approaches would provide an objective backup to what is seen by that observer.

This study only examined a very small population of individuals and recognises the necessity of expanding the techniques developed to encompass many more individuals to properly assess the full utility of the colour and pattern recognition approaches. However, the choice of brown hair donors in this study was seen as providing a stern test of the potential for these applications.

Back to Article Outline

References 

  1. Verma MS. The concept of total morphological pattern and human hair comparison. In Proceedings of the 33rd Annual Meeting of the American Academy of Forensic Sciences. 1981;Los Angeles, CA
  2. Robertson J. An appraisal of the use of microscopic data in the examination of human head hair. Journal of the Forensic Science Society. 1982;22:390–395
  3. Lee HC, Gaensslen RE, et al. Simultaneous identification and determination of species origin. ABH antigens and isoenzyme markers in the same blood stain. Forensic Science International. 1985;29:191–198
  4. Marshall RC. Hair comparison by protein analysis. Proceedings of the International Symposium on Forensic Hair Comparisons. Quantico, Virginia, USA: FBI Academy; 1985;
  5. Roe GM, McArdle W, Pole K. The detection of cosmetic treatments on human scalp hair. Screening of forensic casework samples. Proceedings of the International Symposium on Forensic Hair Comparisons. Quantico, Virginia, USA: FBI Academy; 1985;
  6. Smith JR. A quantitative method for analysing AFM images of the outer surfaces of human hair. Journal of Microscopy. 1998;191:223–228
  7. Verma MS, Pratt L, et al. Hair-MAP: A prototype automated systems for forensic hair comparison and analysis. Forensic Science International. 2002;129:168–186
  8. Ball TB, Griggs M, et al. Image analysis of Egyptian mummy hair. Microscopy and Microanalysis. 2002;8:92–93
  9. Sato H. Preliminary study of hair form of Japanese head hairs using image analysis. Forensic Science International. 2003;131:202–208
  10. Chase HB. Growth of hair. Physiology Review. 1954;34:113–126
  11. Robertson J. Forensic and microscopic examination of human hair. In:  Robertson J editors. Forensic Examination of Hair. London, U.K.: Taylor and Francis; 1999;
  12. Harding H, Rogers G. Physiology and growth of human hair. In:  Robertson J editors. Forensic Examination of Hair. London, U.K.: Taylor and Francis; 1999;
  13. Bisbing RE. Forensic hair comparisons. In:  Saferstein R editors. Forensic Science Handbook.Vol 1, 2nd edition. Englewood Cliffs, NJ: Prentice-Hall; 2002;
  14. Houck MM, Budowle B. Correlation of microscopic and mitochondrial DNA hair comparisons. Journal of Forensic Science. 2002;47:964–967
  15. Budowle B, Adams DE, et al. Mitochondrial DNA: a possible genetic material suitable for forensic analysis. In:  Lee HC,  Gaensslen RE editor. Advances in Forensic Science. Chicago: Year Book Medical Publishers; 1990;p. 76–79
  16. Linch CA, Whiting DA, Holland MM. Human hair histogenesis for the mitochondrial DNA forensic scientist. Journal of Forensic Science. 2001;46:844–853
  17. Wilson MR, Polanskey D, et al. A family exhibiting heteroplasmy in the human mitochondrial DNA control region reveals both somatic mosaicism and pronounced segregation of mitotypes. Human Genetics. 1997;100:167–171
  18. Chase HB, Eaton GJ. The growth of hair follicles in waves. Annals of the New York Academy of Science. 1959;83:365–368
  19. Houck MM, Bisbing RE. Forensic human hair examination and comparison in the 21st century. Forensic Science Review. 2005;17:51–66
  20. Kolowski JC, Petraco N, et al. A comparison study of hair examination methodologies. Journal of Forensic Science. 2004;49:430–432
  21. Olympus Optical Co Ltd. Olympus Australia, Technical Data — DP70 Digital Camera, 2003.
  22. Syncroscopy™, www.syncroscopy.com www.auto-montage.com.
  23. V++, Digital Optics Ltd., www.digitaloptics.co.nz.
  24. E.M. Brooks, An appraisal of the use of numerical features in the forensic examination of hair. http://erl.canberra.edu.au./public/adt-AUC20080624.144159; less http://erl.canberra.edu.au./public/adt-AUC20080624.144159; http://research.nla.gov.au/main/results?sort=primary_title&keyword=A+Brooks+publication.
  25. Bovik AC, Desai MA. Basic image processing techniques. In:  Bovik A editors. Handbook of Image and Video Processing. USA: Academic Press; 2000;
  26. Seta S, Sato H, Miyake B. Forensic hair investigation. In:  Maehly A,  Williams RL editor. Forensic Science Progress, Volume 2. Berlin: Springer-Verlag; 1988;
  27. Bednarek J. An attempt to establish objective criteria for morphological examinations of hairs using the image analysis system. Problems of Forensic Science. 2003;(LVI):65–77
  28. Shriver MD, Parra EJ. Comparison of narrow-band reflectance spectroscopy and tristimulus colorimetry for measurements of skin and hair color in persons of different biological ancestry. American Journal of Physical Anthropology. 2000;112:17–27
  29. Setaro M, Sparavigna A. Quantification of erythema using digital camera and computer-based colour image analysis: a multicentre study. Skin Research and Technology. 2002;8:84–88
  30. CIE . Commission Internationale de l'Eclairage Proceedings. Cambridge: Cambridge University Press; 1931;
  31. CIE . The CIE system is defined in Publication CIE No. 15.2, Colorimetry, Second Edition. 1986;
  32. Comber B. V++, Version 4.0 VPascal program for conversion of RGB images to L*a*b* images. Canberra, ACT: Australian Federal Police; 2006;
  33. SPSS 12.0.1 for Windows: Release 12.0.1 (11 Nov, 2003), SPSS Inc., 1998–2003. www.spss.com.
  34. Comber B. V++, Version 4.0 VPascal programmes for pattern matching in images. Canberra, ACT: Australian Federal Police; 2004;
  35. Excel Microsoft® Office XP Professional. www.microsoft.com/office 2002.
  36. Duda RO, Hart PE, et al. Pattern Classification. 2nd edition. New York: Wiley; 2001;
  37. Theodoridis S, Koutroumbas K. Pattern Recognition. 3rd edition. The Netherlands: Elsevier; 2006;
  38. Adolf FP, Dunlop J. Microspectrophotometry/colour measurement. In:  Robertson J,  Grieve M editor. Forensic Examination of Fibres. London: Taylor and Francis; 1999;
  39. Wiggins KG, Clayson NJ. The use of the Zeiss UMSP50 microspectrophotometer in textile fibre analysis. MPFSL Report. 1994;91:1–14
  40. Christensen P, Adams J, et al. A brief evaluation of the Zeiss MCS/16 photo-diode array with corresponding J & M Spectralys software. FSS Report. 2001;TN854:1–22
  41. Drummond P, Adams J, et al. Evaluation and validation of the J & M MSP800 microspectrophotometer for use in forensic textile fibre examination. Part I: visual Range. FSS Report. 2002;TN866:1–32
  42. FSS: Report, TN895. Valuation and evaluation of the J & M MSP800 microspectrophotometer for use in forensic textile fibre examination. Part II: UV Wavelength. FSS Report TN866. (2005) 1–30.
  43. Ogle RR, Fox MJ. Atlas of human hair. Microscopic characteristics. Florida, USA: CRC Press LLC; 1999;
  44. Vaughn M, van Oorschot R, Baindur-Hudson S. Measurement of classification of hair colour. Proceedings of 17th ANZFSS International Forensic Symposium. April 2006;Perth, WA
  45. Vaughn M, Brooks EM, van Oorschot R, Baindur-Hudson S. A comparison of macroscopic and microscopic hair color measurements and a quantification of the relationship between hair color and thickness. Microscopy and Microanalysis. 2009;15:189–193
  46. Laing DK, Hartshorne AW, Harword EJ. Colour measurements on single textiles. Forensic Science International. 1986;30:65–77
  47. Gardner BB, MacAdam DL. Colourimetric analysis of hair colour. American Journal of Physical Anthropology. 1934;19:187
  48. Sunderland E. Hair colour variation in the United Kingdom. Annals of Human Genetics. 1956;20:312–330
  49. Findlay GH. An optical study of human hair colour in normal and abnormal conditions. The British Journal of Dermatology. 1982;107:517–527
  50. Hanna BL. Colourimetric estimation of the pigment concentration on hair of various colour grades. American Journal of Physical Anthropology. 1956;14:153
  51. Houck MM, Bisbing RE, et al. Locard exchange: the science of forensic hair comparisons and the admissibility of hair comparison evidence: Frye and Daubert considered. Modern Microscopy Journal. 2004;1–10March
  52. U.S. v. Frye, 293 F.2d (D.C. Cir. 1923).
  53. Daubert v Merrell Dow Pharmaceuticals, Inc., 509 U.S. 579, 113 S.Ct. 2786, 125 L.Ed.2d 469 (1993).
  54. Driskell Inquiry . Commission of inquiry into certain aspects of the trial and conviction of James Driskell. Transcript of Proceedings. 2006;24:2269–5624

PII: S1355-0306(10)00075-4

doi:10.1016/j.scijus.2010.06.008

Science & Justice
Volume 51, Issue 1 , Pages 28-37, March 2011

Article Tools

* Image Usage & Resolution