Evalution of hair and scalp condition based on

microscopy image analysis

Hyungjoon Kim, Woogeol Kim, Jeheok Rew, Seungmin Rho
Jisoo Park and Eenjun Hwang* Department of Media software
Sungkyul University
School of Electrical Engineering
Anyang, Republic of Korea
Korea University, Seoul, Republic of Korea
[email protected]
{hyungjun89, woogulzzang, rjh1026, jisoo_park,
ehwang04}@korea.ac.kr

Abstract— Due to the rapid deployment of IT technology, features from the hair image, which are hair thickness, hair
health care service has entered a new era. Some services such as density and scalp blotch. By performing chronological analysis
cardiac monitoring are critical for life and have contributed to on the accumulated feature data, we can evaluate the user’s hair
saving lives. On the other hand, monitoring hair loss is another and scalp condition. The remainder of this paper is organized as
interesting health care service. Even though it is not critical for follows. In Section 2, we introduce several related works for hair
life, people tend to pay much attention to their hair condition. Hair and scalp image analysis. Detailed techniques for hair/scalp
loss is one of the major issues related to the hair condition since image analysis are described in Section 3 and then we explain
excessive and uncared hair loss might lead to bald head. Hair care our prototype system and experiment in Section 4. We conclude
can be done professionally at the hair care shop but it requires
this paper in Section 5.
much time and cost. Recently, due to inexpensive smart devices,
self-diagnosis on the hair condition has become possible. Still, few
applications have been developed to evaluate hair condition. In
this paper, we propose a new scheme to evaluate the condition of II. RELATED WORKS
hair and scalp by extracting diverse features from their
In order to analyze the scalp image, hair and scalp should be
microscopy image. The features include hair thickness, hair
separated from each other. The most noticeable difference
density and scalp blotch. We show the effectiveness of our scheme
by extensive experiments on the prototype system.
between the two is their color; the scalp is relatively bright, and
the hair is relatively dark. Hence, in many studies, hair and scalp
Keywords—scalp analysis; microscopy image analysis; feature areas were classified based on color, and image analysis was
extraction; done based on this separation.

I. INTRODUCTION In [2], Huang addressed three critical issues of hair

segmentation and counting and then proposed a new hair-
Hair loss can occur due to diverse factors, which include counting algorithm. He showed the algorithm is substantially
genetics, stress, abuse of hair products and poor nutrition. Even more accurate than the Hough-based one and robust to curls, oily
environmental changes are reported to cause hair loss [1]. Since scalp, noise corruption, and overlapping hairs, under various
excessive and uncared hair loss could lead to bald head, many white balance. In [3], Hayashi et al. developed a quantitative
efforts have been made to remedy or prevent the hair loss. Such method for measuring hair growth using optical microscopy and
efforts include visiting professional hair care shop or using image analysis. The hairs were cut from an area 7-8 mm in
special hair or scalp care products. Since they are usually diameter and 24 h and 72 h later, images of the areas were
expensive and inconvenient to the users, there has been an obtained using an optical microscope and were recorded on a
increasing demand for inexpensive and convenient way to video disc. Then, they measured regrowing hairs using the
monitor hair and scalp condition. image analyzer. The length of each hair was measured on the
On the other hand, with the popularity of smartphones with monitor by designating two points with the digitizer. In [4],
decent computational power, various smartphone-based Hoffmann Rolf presented the TrichoScan as a method that
applications have been developed for health care purpose. Their combines epiluminescence microscopy with automatic digital
focus was fast diagnosis using cheap equipment. Even though image analysis for the measurement of human being, and
their performance was not that satisfactory compared to the potentially animal hair, in situ. The software was developed to
professional services, still they are gaining more popularity due analyze hair density, hair diameter, hair growth rate, and
to the convenience and cost effectiveness. anagen/telogen ratio. In [5], Abbas et al. proposed hair detection
and repair techniques. The proposed hair removal algorithm
Hence, in this paper, we propose an efficient scheme to consists of hair detection, refinement by morphological
automatically analyze hair and scalp condition using cheap techniques and hair repair by fast marching image inpainting
microscopy camera. To do that, we first extract three critical technique.

*
Corresponding author; Tel.: +82-02-3290-3256

978-1-5090-5140-3/17/$31.00 ©2017 IEEE

 Fig. 1. Overall steps for feature extraction

III. FEATURE EXTRACTION SCHEME Binarization: After aforementioned pre-processing, the

In this section, we describe how to extract hair and scalp resulting image is converted into gray-scale image, and then into
features from a microscopy image. To improve the accuracy of binary image based on Otsu threshold [6]. In the binary image,
feature extraction, we first perform some pre-processing, and ‘0’ and ‘1’ denote hair pixel and scalp pixel, respectively.
then analyze the resulting image for feature extraction. Fig.1
shows the overall steps for feature extraction.
B. hair/scalp image analysis
Hair detection: Hair contour can be acquired from binary
A. Pre-processing image by using Canny edge detection algorithm [7]. Fig.2 shows
Since RGB-based color images are easily influenced by light all the steps for detecting and modeling hairs. For the original
source, it is not good to perform feature extraction for the scalp microscopy image in Fig.2 (a), we can calculate hair contour and
image directly. If so, the image could suffer from diverse skeleton. In particular, hair skeleton is calculated by using the
problems such as vignetting effect, which could drop the quality thinning operation on the binary image. Thinning [8] is a
of image analysis dramatically. To solve this, original image morphological operation that removes foreground in overall
should be pre-processed appropriately before further processing. binary image. Fig.2 (d) shows the final image by superimposing
For the hair and scalp image, we perform the following pre- hair contour and skeleton.
processing:
Cropping image: One of the typical problems caused by
camera hardware limitation is vignetting effect. Due to the
effect, the outer edges of an image is getting darker due to the
reduced light in the periphery of camera lens. This problem can
be solved by concentrating on the center of the image and
cropping the boundary of the image out.
Contrast stretching: Hair area and scalp area can be (a) Original image (b) Hair contour
effectively separately by utilizing their color contrast. However,
in many cases, color contrast of original image leads to some
ambiguity due to diverse noises such as moles, red blotches and
wrinkles. This problem can be solved effectively by using
contrast stretching, which expands the dynamic range of the
intensity levels allowing it to span the color distance between
hair and scalp.
Morphological opening: The oily and moist hair makes
bright spots in its microscopy image. This is due to the light (c) Hair skeleton (d) Contour + skeleton
reflection of the microscopy camera. However, those bright Fig. 2. Example of hair detection
spots would give incorrect result during the feature extraction.
This problem can be solved by using morphological opening Hair thickness calculation: Hair thickness can be defined
approach, which is derived from the fundamental operations of by the length of perpendicular line to the hair. To get the
erosion and dilation. perpendicular line, we first need to calculate the hair direction.
We calculate the direction of each pixel by considering its
neighboring pixels and applying PCA (Principal Component
Analysis) algorithm [9].
When the direction is calculated for all the points on the hair
skeleton, the perpendicular line for each point can be calculated.
Then, the distance between the intersection points of the
perpendicular line and the hair boundary is the thickness of the
hair and can be calculated by using Euclidean distance as shown (a) Normal scalp (b) Scalp blotch
in Eq. (1). Based on the distance, we can calculate average hair
thickness using Eq. (2). Fig. 4. RGB values vs red blotch

∑ (1) First, a blotch region is always on the scalp, thus we only

consider scalp in the original image for blotch detection. To do
that, we consider coordinates of the original image and binary
Here, the unit of Thicknessavg is pixel thus it needs to be image together. For instance, if a point in the binary image is 0,
changed into the meter unit using Eq. (2) as shown below. the point is considered to correspond to a hair. If the point is 1,
/ then it belongs to a scalp. Based on this, we can extract scalp
(2) area from the original image. Secondly, since each pixel of the
original color image has R, G, B values, we can calculate the
average of R values, Ravg. Third, for all the points in the image,
Where is the magnification of camera and is the unit if the R value of a pixel (i, j) is bigger than Ravg and the point is
length which indicates the micrometer length of one pixel. not in the hair area, then the point is considered as a blotch
Hair counting: Typical portable microscope camera candidate. Lastly, for the candidate points, we calculate the ratio
generally covers a rectangle of 5mm x 5mm. Thus, if a hair is of R and G and test if G/R < 0.85. Likewise, we calculate the
longer than 5mm, it must go beyond the rectangle. Based on this ratio of R and B and test if B/R < 0.8. If the point satisfies these
observation, we made two assumptions for hair counting. two conditions, then the point is considered to be in the blotch
area. Two threshold values 0.85 and 0.8 are decided
1. Each hair is represented by one start point and one end point. experimentally. For instance, in Fig.4 (a), R and G, B values of
2. The start point is positioned inside the rectangle, and the marked points have no significant difference, and their ratios are
end point is positioned on the boundary of the image. bigger than 0.9. Whereas, as in the Fig.4 (b), R values of pixels
in the blotch area are bigger than their G or B values.
For hair counting, we use the skeleton image that was
obtained by thinning operation. Based on the assumptions, we
count the hair if one point of its skeleton line is inside the IV. EXPERIMENT
rectangle and the other point is on the boundary. Fig.3 shows an
example of hair counting. In the figure, the counted pixels are In order to evaluate the performance of our proposed
marked with red dots and those dots are superposed onto the scheme, we built a prototype system based on the scheme. The
original image. Then, hair density can be calculated by dividing system is implemented using Matlab R2013a. The test images
the number of counted pixels by the image dimension. were made using x50 microscopy camera compatible with
smartphone as shown in Fig.5 (a), Fig.5 (b) shows the electron
microscope for measuring actual hair thickness.

(a) Counted hairs (b) Superposed onto image

Fig. 3. Example of hair counting
(a) Portable camera (b) Electronic microscope
Scalp blotch detection: Blotches on the scalp are usually Fig. 5. Equipments for experiment
reddish. In the dermoscopy image, their R values are larger than
those of G or B. Therefore, if points in a specific region have We constructed a dataset of scalp images from 50 subjects
largest R value, the region might represent a blotch region. Fig.4 aging from 20s to 50s. To measure the accuracy of our hair
(b) shows an example of scalp blotch. However, the fact that R thickness evaluation method, we pulled out 4 hairs per persons
value is largest doesn’t necessarily represent a blotch. For and took 3 microscopy partial images of start, middle and end
instance, Asians have a scalp with a relatively large R value in portion of each hair. So, thickness was calculated for a total of
the microscopy image even though they don’t have blotches as 600 hair portions. For hair counting, 4 scalp images were
shown in Fig.4 (a). Therefore, it is necessary to establish a captured per person, so total 200 samples were tested.
guideline by which the reddish area can be counted as a blotch.
A. Thickness V. CONCLUSION
To evaluate the accuracy of our hair thickness evaluation In this paper, we proposed an efficient scheme for extracting
method, we measured actual hair thickness using electronic hair/scalp features from microscopy images. The features
microscope. The accuracy of hair thickness measurement is include average hair thickness, hair density, and scalp blotches.
shown in Table 1. The average hair thickness of Koreans is In order to improve the quality of hair/scalp image for feature
reported to be 84.9μm [10]. Compared with this, our result of extraction, extra works were carried out during the pre-
90.29μm is reasonably good. There is small difference between processing.
actual (calculated by electron microscope) and estimated values.
One possible reason is the shadow effect. Another possible To measure the performance of our proposed scheme, we
reason is the camera lens distortion. The image has 640*480 implemented a prototype system and carried out several
size. However, the real shape of the area covered by the camera experiments for the images collected from subjects using cheap
is almost an ellipse. Thus, camera magnification has difference microscopy camera attachable to smartphone. We measured the
between rows and columns. Depending on the shot angle, performance of our feature extraction scheme by calculating
magnification could be different. error rate and precision/recall. Overall, our scheme showed very
good performance.
TABLE I. AVERAGE HAIR THICKNESS AND ERROR RATE
ACKNOWLEDGMENT
Actual Estimated Error rate Accuracy This work was supported by Institute for Information &
Average 85.91 μm 90.29 μm 5.10% 94.90% communications Technology Promotion (IITP) grant funded by
the Korea government (MSIP) (No. R0190-16-2012, High
Performance Big Data Analytics Platform Performance
B. Hair Counting Acceleration Technologies Development).
In order to evaluate our hair counting algorithm, we
performed an experiment for the dataset of 200 scalp images.
Table 2 shows the result of precision/recall between actual REFERENCES
number of hairs and estimated number of hairs calculated by our [1] “Alopecia Market Analysis By Treatment (Oral, Topical, Injectable), By
algorithm. The average precision and recall were 91.35% and Gender (Men, Women) And Segment Forecasts To 2024,” Market
92.01%, respectively. Even though the performance is quite Research Report, 6. 2016.
good, the precision/recall can be improved further. One critical [2] H. Chia, “An Unsupervised Hair Segmentation and Counting System in
reason for the error is preprocessing and the quality of original Microscopy Images,” IEEE Sensors Journal, Vol. 15, No. 6, pp. 3565-
3572. 2015.
image. In our experiment, once the image has a blurred spot in
the image, it is not possible to remove all the noises in the [3] S. Hayashi et al. "Measurement of human hair growth by optical
microscopy and image analysis," British Journal of Dermatology, Vol.
preprocessing step. Therefore, considering the camera noise, the 125, No. 2, pp. 123-129, 1991.
accuracy is very good. Similarly, we tested our hair pore [4] H. Rolf, “TrichoScan: a novel tool for the analysis of hair growth in vivo,”
counting algorithm, and the result shows about 90% of average Journal of Investigative Dermatology Symposium, Vol. 8, No. 1, pp. 109-
accuracy rate. 115, 2003.
[5] A. Qaisar et al. "Hair removal methods: a comparative study for
TABLE II. PRECISION AND RECALL OF HAIR RECOGNITION dermoscopy images," Biomedical Signal Processing and Control, Vol. 6,
No. 4, pp. 395-404, 2011.
Precision Recall [6] N. Otsu, "A threshold selection method from gray-level histograms,"
Automatica, Vol. 11, No. 23-27, pp. 285-296, 1975.
91.35% 92.01%
[7] J. Canny, “A computational approach to edge detection,” IEEE
Transactions on Pattern Analysis and Machine Intelligence, Vol. 8, No. 6,
pp. 679-698, 1986.
C. Scalp blotch detection [8] L. Lam, “Thinning Methodologies-A Comprehensive Survey,” IEEE
To measure the accuracy of detecting reddish blotches on the Transactions on Pattern Analysis and Machine Intelligence, Vol. 14, No.
scalp, we performed an experiment for 150 blotch images from 9, pp. 869-885, 1992.
30 subjects and 200 normal images from 40 subjects. As shown [9] Y. S. Lee et al. “A straight line detection using principal component
analysis,” Pattern Recognition Letters, Vol. 27, No. 14, pp. 1744-1754,
in the Table 3, most of scalp blotches in the images were 2006.
detected. Relatively low recall score was caused by a few
[10] G. S. Choi, “Hair characteristics and androgenetic alopecia in Koreans,”
subjects with reddish scalp even though they didn’t have Journal of the Korean Medical Association/Taehan Uisa Hyophoe Chi,
blotches. This phenomenon happens frequently for reddish scalp Vol. 57, No. 1, 2014
image. Since our method is based on R, G, B values, subjects
with comparatively red skin color tend to have higher R value.

TABLE III. PRECISION AND RECALL OF DETECTING SCALP BLOTCH

Precision Recall

93% 87.5%