As one of the most in uential professors in the eld of Chinese language research in the last century, Professor Shirakawa had left many precious hand-notated research documents that have not been organized and made public. This paper aims to automatically organize hand-notated Oracle Bone Inscriptions research documents by image processing. In detail, Area-Mutation-Segmentation is applied for separating Oracle Bone Fragments from Oracle Bone Inscriptions document. Three-Stage-Denoise and Spider-Web-Extinction are proposed to reduce outside noise and border noise respectively. The experimental results show that the accuracy of Noise Reduction achieves 97.8%. The border noises of Oracle Bone Fragment images have been reduced perfectly without errors, and the number of valid Oracle Bone Inscriptions extracted has increased by 32.5%, comparing with previous research. In summary, the experimental results demonstrate the e ectiveness of our proposed method.
are treated as outside noise. Three-Stage-Denoise is proposed to reduce the outside noise. Spider-Web-Extinction is raised to reduce the border noise that refers to the border of each OBF.
This method not only provides us a new Noise Reduction method but also generates a relatively su cient OBIs dataset for enhancing the robustness of OBIs recognition. Therefore, organizing the hand-notated OBIs research documents by Professor Shirakawa is of great signi cance for the research and protection of Chinese culture.
In section 2, some methods of Noise Reduction of images, recognition and extraction of OBIs are introduced. In section 3, the processing of the segmentation of OBFs, Noise Reduction of outside noise and border noise are depicted in detail. We report and discuss the evaluation results in section 4, and Section 5 proposes a conclusion and mentions the future work. 2
Thanks to their importance in potentially unlocking the history of ancient China and
helping with the evaluation of early Chinese characters, OBIs have recently been attracting more
research attention. Anyang Normal University has created an OBI database that stores
photographs and rubbings of OBI [10]. Many researchers have extracted features for recognition
from OBIs images by various methods. For example, [
During the preprocessing of segmentation, Fig.2 shows intermediate results of segmentation
processing. Figure 2(a) shows the original image, and as Fig.2(b) shows, the original image
is converted into grayscale for reducing the amount of calculation. In binarization, the OTSU
method [
In OBFs segmentation, Area-Mutation-Segmentation is proposed to identify and locate OBFs on OBIs images. In detail, there is a large gap between the sizes of a single OBI character feature and a single OBF feature. In this algorithm, we get the threshold by catching the mutation point where the gap is located to decide whether it is OBF or not, and take a suitable range to lter out OBF features. The processing is divided into organizing the size of bounding boxes, catching mutation points, and ltering OBFs.
1. Organizing size of bounding boxes. Perform feature detection on OBIs images by cv2.Findcontours in OpenCV and its return value can be used to draw a bounding box. The sizes of the bounding boxes are sorted in order by quick sort and saved in the list. The sudden change between the sizes of a single OBI and a single OBF is detected by getting the max rate of change(gap). Avoiding being disturbed by the sudden changes in the size between the OBFs and the entire OBIs image, we delete the max data, which is the size of the OBIs image, from the list. 2. Catching mutation point. Starting with the second element in the list, we calculated the di erence between every element and the previous element, as shown in Equation 1. S(i) means the size of the ith bounding box, and D(i) refers to di erence between S(i) and S(i + 1). The mutation point of size can be caught by nding the max D(i). Figure 3 shows a part of the size of bounding boxes ordered including mutation point. Assuming that D(k) is the max rate of change, the larger size S(k + 1) of the two bounding boxes means the smallest single OBF's size, and the smaller size S(k) refers to the size of the 3. Filtering OBFs. All the bounding boxes with a size between D(k) and the OBIs image size border the OBF feature. Ensuring that every bounding box includes an OBF feature completely, we expand the range of 100 around the bounding boxes before cutting them out. Although the expanded bounding box may border outside noise by mistakes, these can be reduced clearly by the next processing. 3.2
Three-Stage-Denoise is proposed to perform Noise Reduction. We take Fig.4(a) as an example
to show the speci c processing of Noise Reduction. Originally, the image is converted into
grayscale, as shown in Fig.4(b). Meanwhile, in the median ltering [
1. Eroding processing . Some noise like the OBF number that is generally composed of three digits, multiple iterations eroding operation needs to be performed to make the three digital features connected. However, too many iterations will make the noise feature connected with the border of the OBF by mistake, so that the outside noise feature can not be identi ed as one object. In order to deal with such problems, at the rst stage, multiple iterations are proposed to make some noise that is far away from the border connected as much as possible. In the second stage, we slightly reduce the iterations of the eroding operation so that the noise near the OBF border can be connected without contacting the OBF border. In the third stage, a light iteration of the eroding operation is utilized to target some noises that have not yet been reduced clearly during the rst and second Noise Reduction. 2. Organizing each OBF image by list. Scan each row of the OBF image from right to left and record how many pixels we have traversed until encounter the rst black pixel. If there are no black pixels in a row, then record the distance we traverse at the row Wi, which is the width of the picture. In this way, from the rst row to the last row, we save Hi (Hi refers to the height of the picture) data in the list. The same processing should be carried out again from right to left for organizing the image by another list. 3. Getting the point where noise may be appear . After morphological processing, the black pixels of the OBF's border should be continuous relatively. Hence, the rate of change between two adjacent data in the two lists should be smooth besides some corner points and noise. We get this point to capture where outside noise may appear, and try to set a threshold as a criterion to decide whether the rate of change in the list is smooth or not. If the rate of change exceeds this threshold, the positions are marked as places where the noises may appear by a circle, as shown in Fig.5. 4. Filtering noise out. Drawing bounding boxes and judging whether the positions where noise may appear are inside a small bounding or not. The gap between the sizes of OBF and the outside noise is large, and the threshold can be set easily. Further, we treat the marked positions bordered by a small bounding box as outside noise and reduce it by turning it into white. 3.3
Spider-Web-Extinction is used to reduce border noise. During this algorithm, the OBF image was scanned from top to bottom, bottom to top, left to right, and right to left respectively. The case of left to right is chosen as an example for a detailed explanation. Firstly, start scanning from left to right every T th row (T determines how big the border noise can be decomposed into pieces). Secondly, stop scanning when black pixels are encountered. Thirdly, turn the N pixels into white (N determines the degree of decomposition). Because all of the outside noise has been reduced in the second step, the black pixel encountered by scanning is judged as a part of border noise. The conversion of border pixels to white is a decomposition of the border noise.
To ensure that the border noise can be completely decomposed, the maximum value of the border width is set as N during Spider-Web-Extinction. Although this will a ect part of the OBI features, the nal erosion process can retain most of the character features. When the four aspects of processing are over, we can nd that the border of the OBF is broken down into small squares and strips, which look like a spider web. The border noise is decomposed into many tiny features smaller than T N . By turning all the features with a size below T N into white, the Noise Reduction is completed. It is noteworthy that avoid reducing the small OBI features inside the border due to mistakes, we try not to choose too big T . 4
In the experiment, we choose 79 OBIs images from Shirakawa's hand-notated document and process these OBF images for the organization in CPU(i7). OBF segmentation, Noise Reduction of outside noise and border noise are the three key components of the organization. In this section, we show and evaluate the experimental results of these three parts. 4.1
In segmentation, we chose to use the Area-Mutation-Segmentation to lter the threshold to help us cut each OBF out. The result is that we cut 364 OBF images from 79 OBIs images from Shirakawa's hand-notated documents, such as Fig.6.
During Three-Stage-Denoise, outside noise has been reduced perfectly. As shown in Fig.7, it is the control group where we processed the two OBF images respectively. Fig.7(a) and Fig.7(d) are the original image of OBF a and OBF b respectively. And so on, Fig.7(b) and Fig.7(e) are the a and b that after the rst stage of Noise Reduction. Similarly, Fig.7(c) and Fig.7(f) are the results of the second Noise Reduction. The outside noise in b is far away from the OBF's border, so it was reduced in the rst stage. Relatively, the position of noise is close to the border of the OBF so that outside noise can be connected with the border after the rst stage Noise Reduction. As a result, the noise has to be reduced in the second stage. The third time result is not shown here, because all the outside noise had been cleared mostly. The third stage is utilized to target some noises that have not been reduced clearly during the rst and second stages. We performed Noise Reduction on the 364 OBF images totally, and on 356 pieces of them, outside noises have been reduced perfectly, and the accuracy of Noise Reduction is 97.8%.
In Noise Reduction of border noise, Spider-Web-Extinction has been used to decompose the border into strips and block features like spider webs, as shown in Fig.8(a). Fig.8(b) is a partially enlarged image of Fig.8(a). These features are reduced by ltering the size of the bounding box, and the result is shown in Fig.8(c). In a summary, the border noise of 356 OBF images has been reduced perfectly without errors. In terms of segmentation, the OBF features di erentiated from OBI features by catching a large rate of di erence between two bounding boxes' size at the rst time. It is intuitive to help us distinguish OBF features accurately. However, problems still exist in using AreaMutation-Segmentation to cut OBFs out. If the rate of change between the size of OBFs is particularly large, the position we locate by Area-Mutation-Segmentation may not be the dividing line between the OBIs and OBFs, but a dividing line between large OBFs and small OBFs. Fortunately, Professor Shirakawa sorted out and put OBFs with similar sizes in the same OBIs image as much as possible. However, facing these special cases, the algorithm should be optimized in the future.
During Noise Reduction of outside noises, we masterly utilize the characteristic that the border of OBFs and the outside noise are almost not smoothly connected together to reduce the noises by three stages. About the reduction of border noise, we propose a creative method to e ectively eliminate the OBF border, and it can almost deal with the reduction of any border noise during the organization perfectly. 5
In light of the recent discovery of Professor Shirakawa's hand-notated OBIs research documents,
this paper aims to organize hand-notated Oracle Bone Inscriptions research documents by image
processing. It is helpful for us to extract more OBI features from the limited OBIs images [
http://jgw.aynu.edu.cn/. [11] Rohit Verma and Jahid Ali. A comparative study of various types of image noise and e cient noise removal techniques. International Journal of advanced research in computer science and software engineering, 3(10), 2013. [12] Yikang Zhang, Heng Zhang, Yongge Liu, Qing Yang, and Chenglin Liu. Oracle character recognition by nearest neighbor classi cation with deep metric learning. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 309{314. IEEE, 2019.