Recognition of Oracle Bone Inscriptions by Extracting Line Features

on Image Processing

Lin Meng

College of Science and Engineering, Ritsumeikan University, 1-1-1 Noji-higashi, Kusatsu, Shiga 525-8577, Japan

menglin@fc.ritsumei.ac.jp

Keywords:

Recognition of Oracle Bone Inscriptions, Hough Transform, Clustering.

Abstract:

Oracle bone inscriptions is a kind of characters, which are inscribed on cattle bone or turtle shells with sharp

objects about 3000 years ago. Understanding these inscriptions can give us a lot of insight into world history,

character evaluations, global weather shifts, etc. However, for some political reasons the inscriptions remained

buried in ruins until their discovery about 120 years ago. The aging process has caused the inscriptions

to become less legible. In this work, we design a system and proposal a recognition method for recognizing

oracle bone inscriptions as a template image from an oracle bone inscription database, by using the line feature

of the inscriptions. First we use Gaussian ﬁltering and labeling to reduce noise and use afﬁne transformation

and thinning to extract the skeleton. Then we use Hough transform to extracting the line feature points by

proposing a method of clustering. Finally, we calculate the minimum distance of the line feature points

between the original image and the template images to perform the recognition. Experimental results shows

that almost 80% of inscriptions are recognized as the most minimum distance and the second-most minimum-

distance. And the proposal can recognized well, even if the noise and tilt happened in original images.

1 INTRODUCTION

Oracle bone inscriptions (OBIs), which ﬁrst came into

being about 3000 years ago in China, are some of

the oldest characters in the world. OBIs were in-

scribed on cattle bone or turtle shells with sharp ob-

jects (Ochiai, 2008),(Pu and Xie, 2009) and are a kind

of early literature used to record the history, weather,

political activity, etc. taking place in China at that

time. Understanding these OBIs can give us a lot

of insight into world history, notable births, charac-

ter evaluations, global weather shifts, etc. However,

for political reasons OBIs remained buried in ruins

until their discovery about 120 years ago. The aging

process has caused these inscriptions to become less

legible, and due to a lack of early research on the sub-

ject, it is now increasingly difﬁcult to understanding

what it is the OBIs have to say.

Because few people can read OBIs, the major

OBIs recognition is that the experts of historian rec-

ognize OBIs by their experience. Currently, some re-

searchers have suggested using image processing to

recognize OBIs automatically. However, the recog-

nition rate with this approach is not high enough and

needs to be improved.

The most common OBI recognition method is

Figure 1: An example of OBIs.

rubbing, where the OBI surface is reproduced by plac-

ing a piece of paper over the subject and then rubbing

the paper with rolled ink. Figure 1 shows an example

of an oracle bone rubbing with the middle part show-

ing an enlarged view. Several characters visible on the

left side of the rubbing refer to a divination predicting

that it will rain that day from 11 p.m. to 1 a.m.

A lot of these characters are made up of lines,

which makes sense since they were inscribed by sharp

objects. With this feature in mind, we have designed

an OBI recognition system that uses Hough transform

606

Meng, L.

Recognition of Oracle Bone Inscriptions by Extracting Line Features on Image Processing.

DOI: 10.5220/0006225706060611

In Proceedings of the 6th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2017), pages 606-611

ISBN: 978-989-758-222-6

image processing. The system recognizes inscriptions

from an inscription database that contains images of

normalized inscriptions similar to a dictionary. The

normalized inscriptions are generated using character

font software to make the characters smooth, clear,

and straight, with uniformly thick strokes. These

characters have been examined by historians, and the

database is created by the researchers who belongs

the letters college of Ritsumeikan University. More

than 2000 normalized inscriptions are stored in the

database(Ochiai, 2014).

The recognition system is comprised of four steps

for recognition. The ﬁrst step is noise reduction pro-

cessing, where Gaussian ﬁltering and labeling are

applied to reduce noise. The second step is fea-

ture extraction pre-processing, which includes afﬁne

transformation(Schneider and Eberly, 2003) and thin-

ning(L. Lam and Suen, 1992) for extracting the skele-

ton of OBIs. The third is line feature processing,

which extracts the line feature points by Hough trans-

form (Ballard, 1981). The fourth is recognition by

calculating the minimum distance between the ex-

tracted line feature points of original and template

OBI images.

The contributions of this paper are as follows:

1. Design of an OBI recognition system from

noise reduction to recognition.

2. Proposal of a method for OBI recognition by

Hough transform and clustering.

Section 2 of this paper discusses related work and

section 3 describes the recognition method. Experi-

ments and results are reported in section 4. We con-

clude in section 5 with a brief summary.

2 RELATED WORK

As technologies evolve, various researchers have at-

tempted to recognize OBIs by image processing.

However, few English papers have reported on OBI.

We do know that the recognition rate needs to be im-

proved.

(Li and Woo, 2008) and (Q. Li, 2011) presented a

recognition method that treats OBIs as a non-directed

graph for recording the features of end-points, three-

cross-points, ﬁve-cross-points, blocks, net-holes, etc.

However, due to the age of OBIs, some of the holes

and cross-points that occur are not actually a part

of the OBIs themselves, which increases the difﬁ-

culty of the recognition. (Li and Woo, 2000) pro-

posed a DNA method for recognizing OBIs. How-

ever, neither (Li and Woo, 2000) nor (Q. Li, 2011)

provided details on any experiments.We have previ-

ously proposed several methods for recognizing OBIs

Figure 2: Flow of OBI recognition.

by template matching and by using Hough transform

(L. Meng, 2016),(L. Meng and Oyanagi, 2015). How-

ever, the template matching was weak when the orig-

inal character tilt, and (L. Meng and Oyanagi, 2015)

did not properly process the tilt, either.

In the present work, we propose a complete recog-

nition system from noise reduction to recognition, and

consider the tilt.

3 RECOGNITION PROCESSING

Figure 2 shows the OBIs recognition ﬂow. The main

processing includes noise reduction processing, fea-

ture extraction pre-processing, line feature extraction

processing and recognition processing.

3.1 Noise Reduction Processing

Due to aging, many noises both big and small exist

on OBI rubbings. Noise reduction processing is there-

fore an important part of the recognition process. Fig-

ure 3a) is the original image, the character means the

period of ”zi”, which is a rubbing image cut from (Pu

and Xie, 2009). As shown, both smaller noises such

as fog and some bigger noises exist in the image.

We use Gaussian ﬁltering and binarization for re-

ducing the smaller noises. Formula (1) shows the

Gaussian ﬁlter used for blurred images. Figure 3b)

shows the Gaussian ﬁltering results and Fig. 3e)

shows the histogram of the Gaussian ﬁltering results

divided into two peaks. The Otsu method(Sezgin and

Sankur, 2004) was used to decide the threshold for

binarization and reduce the smaller noise. Figure 3c)

shows the results of binarization, where the smaller

noises (such as fog) are reduced successfully and the

Recognition of Oracle Bone Inscriptions by Extracting Line Features on Image Processing

607

Figure 3: Noise reduction result.

character becomes clear. However, the bigger noises

remain.

f(x, y) =

2πσ

exp



−

+ y

2σ



(1)

To reduce the bigger noise, labeling (L.F. He,

2008) is used. Labeling is a method that last scans

the binarization image and counts the pixel numbers

of each connected object. We know that bigger noises

have only a few pixels while the connected characters

have a lot of pixels, so pixel number represents a big

change between noisy objects and character objects.

We use a histogram method to detect big changes in

the histogram of objects for detecting the threshold.

If an objectfs pixel number is more than a threshold,

the object will be left alone, and otherwise, the object

is treated for noise reduction.

Figure 3d) shows the result of labeling. As shown,

the bigger noises are reduced successfully and the

character becomes more clear.

In Fig. 3d), there are some noises that we were

not able to reduce due to the noise being closely

connected with the characters. However, the experi-

mental results discussed in section 5 demonstrate that

these few remaining noises do not have a serious ef-

fect on the recognition.

Figure 4: Feature extraction pre-processing.

3.2 Feature Extraction Pre-processing

Feature extraction pre-processing includes Afﬁne

transformation (Schneider and Eberly, 2003) and

Thinning thinning (L. Lam and Suen, 1992).

For comparison with the template image in the

database, the original image needs to be normalized

into the database image space. Afﬁne transforma-

tion is a map that transforms points and vectors in

the original image space into points and vectors in the

database image space. In other words, for the exam-

ple of ”zi”, the labeling results space of Fig. 3d) need

to be changed into to the template space of Fig. 4a).

Formula (2) show the changing method, where (x

, y

)

is the pixel axis in the original image, (x

, y

) is the

character center axis of the template image, θ is the

angle to which the original image will be changed as

a result of the rubbing, and M is the size of the exten-

sion from the original image size and to the template

image size.

Figure 4b) shows the afﬁne transformation results,

namely, the correct changing of the labeling result

space of Fig. 3d) into the template space of Fig. 4

a).



∗





cosθ − sinθ

sinθ cosθ



M × x

M × y







(2)

After the afﬁne transformation, we extract the skele-

ton from the original image using Hilditch’s algo-

rithm. The method considers each of the eight neigh-

borhoods (p2,p3...P9) of the target pixel as one pixel

(p1) and decides whether to peel it off or keep it a

skeleton. Figure 4c) shows the thinning results of Fig.

4b).

3.3 Line Feature Extraction by Hough

Transform

Hough transform is wildly widely used for extracting

lines from images, by transforming the (x, y) space to

the (r, θ) space.

This transformation is shown in Formula 3, where

(x, y) is the size of the original image, r is the distance

from the origin to the closest point on the straight line,

ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods

608

Figure 5: Hough transform.

and θ is the angle between the x axis and the line con-

necting the origin with that closest point.

r = xcosθ + ysinθ (3)

Every point in the (x, y) space will be transformed

into a curve in the (r, θ) space by changing the θ from

0 to 2π. The method records the time the curve is

passed in every pixel of the (r, θ) space. Finally, the

recording time will be used for deciding the line.

Figure 5 e) and f) shows the three dimensions of

Hough transform results on the (r, θ) space for the

thinning result of the original ( Fig. 5 a) ) and tem-

plate ( Fig. 5 c). We found there are eight largest

points that make up the feature line point in Fig. 5

e),f), especially in the template of 5 f). The three

points on the left and right are the same line. This

is the case of (θ) being 0

◦

and 360

◦

. Therefore, the

feature line points are ﬁve.

If we catch the eight largest points correctly and

transform the (r, θ) space into (x, y) space again, it

will be possible to generate the results of the Hough

transform in Fig. 5 b) d). The times at which we ﬁnd

the largest points of the template and original are the

same in Fig. 5 e),f). Hence, deciding on the largest

points and calculating the distance between the origi-

nal image and the temp image is helpful for recogniz-

ing the OBIs.

Below are deﬁnitions pertaining to the line feature

point decision. Algorithm 1 shows the line feature

decision.

(r, θ)

is all points of the (r, θ) space.

(r, θ)

is the largest point in (r, θ) space that still

does not be checked.

LPs

(r, θ)

is a set of larger points that is decided as

the line feature point. It keeps the area of every point

by radius.

SDis(LP, LC) are the distances between LC

(r, θ)

and all of the LPs

(r, θ)

MinDis

(r, θ)

is the smallest distance of

SDis(LP, LC)LP

(r, θ)

SLP

(r, θ)

is a point of LP

(r, θ)

that generates the

MinDis with LC.

Algorithm 1: Line feature point decision.

C ⇐ SortC

while C 6= NULL do

Search LC

Generate SDis by using LC and LPs, Search

MinDis

if MinDis is lower than the radius of SLP then

do nothing

else {MinDis is lower than (the radius of

SLP+30)}

record (the radius of SLP ⇐ MinDis ) into LPs

else

input the LC into LP and keep the axis

end if

end while

Recognition of Oracle Bone Inscriptions by Extracting Line Features on Image Processing

609

Figure 6: Template image.

Figure 7: Original image.

3.4 Matching by Distance Calculation

After the line feature extraction processing, the line

points in the (r, θ) space will be decided. Then the

system calculates the minimum distance of the line

points for the templates and original image. How-

ever, the line points of the template and the original

are often different. Hence, The minimums are nor-

malized by line feature point number. We deﬁned the

line points of the templates and original image into

LargerNumber and SmallerNumber by comparing the

numbers. We use the ﬂown ﬂume to normalize the

minimum distance.

NormilzedDistance =

Distance∗ LargerNumber

SmallerNumber

(4)

4 EXPERIMENTAL RESULTS

We used 24 templates as the dictionary and used 10

kinds original OBIs characters to do the experiment.

Every kind original OBIs characters has 10 pieces

OBIs, which are cutting from rubbing. Figure 6 shows

the template image, with the character number shown

below the images. Figure 7 shows a part of an original

OBI image.

4.1 Recognition Results

Figure 8 shows the recognition rate of our method.

We extract the most minimum distance time, second-

most minimum distance time, and the others which

and comparing them with the template. From the av-

erage, we found that almost 70% of ROIs are rec-

ognized as the most minimum distance, and 10%

Figure 8: Recognition Rate.

are recognized as the second-most minimum distance

time. The horizontal axis shows the character number

of the original which that can be found in Fig.6.

4.2 Recognition Analysis

Figure 9 shows our analysis after the experi-

ment,Figure 9 a) shows the original image. The three

rankings of the character along with the template and

the thinning are shown. The ranks 1 means the tem-

plate which has the most most minimum distance with

the original image. The original thinning is the result

of the afﬁne transformation by the top of the template

in Fig. 9. About the line number, left is the template

and right is the original. From the minimum distance

ranking, we found the results to be correct. As for the

thinning results, although noise was found, it did not

have a negative effect on the recognition.

Figure 10 shows our analysis after the experiment,

ﬁgure 10 a) shows the original image which have a

large tilt. From the results, we found the tilt are re-

new which is shown in 10 b) of original. From the

minimum distance ranking, we found the results to be

correct. Although tilt happened in original image, it

can be recognized correctly.

5 CONCLUSION

In this paper, we discussed our design of an OBIs

recognition system and proposal proposed a recogni-

tion method by using Hough transform. The recogni-

tion includes noise reduction processing which to re-

duces the noise by Guaussian ﬁltering and Labeling,

feature extraction pre-processing, (which including

afﬁnetransformation and Thinning ) for extracting the

skeleton of OBIs, and line feature processing, which

extracts the line feature points by Hough transform.

Then do the recognition is performed by calculating

the minimum distance between the extracted line fea-

ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods

610

Figure 9: Recognition analysis.

Figure 10: Recognition analysis about tilt.

ture points of the original OBI and the its correspond-

ing template image of OBI. The method results shows

showed that almost 80% of ROIs are recognized as the

most minimum distance and the second-most mini-

mum distance time. As for the thinning results, al-

though noise was found, it did not have a negative ef-

fect on the recognition and although the tilt happened

in original image, it can be recognized correctly.

ACKNOWLEDGEMENTS

This work was supported in part by a grant-in-aid for

scientiﬁc research (60615938) from JSPS.

REFERENCES

Ballard, D. H. (1981). Generalizing the hough transform to

detect arbitrary shapes. In Pattern Recognition. Else-

vier.

L. Lam, S. W. L. and Suen, C. Y. (1992). Thinning method-

ologies: A comprehensive survey. In IEEE Trans. on

Pattern and Machine Intelligence. IEEE.

L. Meng, e. a. (2016). Recognition of inscriptions on bones

or tortoise shells based on graph isomorphism. In Int.

J. of Computers Theory and Engineering. IJCTE.

L. Meng, T. I. and Oyanagi, S. (2015). Recognition of orac-

ular bone inscriptions by clustering and matching on

the hough space. In J. of the Institute of Image Elec-

tronics Engineers of Japan. IIEEJ.

L.F. He, e. a. (2008). A run-based two-scan labeling algo-

rithm. In IEEE Tras. on Image Processing. IEEE.

Li, F. and Woo, P. Y. (2000). The coding principle and

method for automatic recognition of jia gu wen char-

acters. In Int. J. of Human-Computer Studies. Sci-

enceDirect.

Li, F. and Woo, P. Y. (2008). Sticker dna algorithm of

oracle-bone inscriptions retrieving, computer engi-

neering and applications. In Computer Engineering

and Applications. CEA.

Ochiai, A. (2008). Reading History from Oracular Bone

Inscriptions. Chikuma Shobo, Japan, 1st edition.

Ochiai, A. (2014). Oracle bone inscriptions database. In

http://koukotsu.sakura.ne.jp/top.html. No.

Pu, M. Z. and Xie, H. Y. (2009). Shanghai Bo Wu Guan

Cang Jia Gu Wen Zi. Shanghai Bo Wu Guan, China,

1st edition.

Q. Li, e. a. (2011). Recognition of inscriptions on bones or

tortoise shells based on graph isomorphism. In Com-

puter Engineering and Applications. CEA.

Schneider, P. K. and Eberly, D. H. (2003). Geometric tools

for computer graphics. Morgan Kaufmann Publishers,

USA, 1st edition.

Sezgin, A. and Sankur, B. (2004). Survey over image

thresholding techniques and quantitative performance

evaluation. In J. of Electronic Imaging. SPIE.

Recognition of Oracle Bone Inscriptions by Extracting Line Features on Image Processing

611