Feature Extraction of Bisindo Alphabets Using Chain Code Contour

— This paper proposes a new method of features extraction of BISINDO (Indonesian Sign Language) alphabet forms. The extracted features play a very important role in the process of hand pattern recognition which expresses BISINDO alphabets and recognize letters when communicating. The proposed method consisted of 5 processes; segmentation, morphology closing, edge detection, contour following, and chain codes formation. 26 hand patterns representing the 26 BISINDO alphabets from A to Z are used during the experiments. The extracted features of each hand pattern (each BISINDO alphabet) are presented in terms of the probability value of occurrence of each chain code direction. The analysis result shows that every hand pattern expressing BISINDO letters has different features. This indicates that the extracted feature by the proposed method is very accurate to be used for BISINDO alphabets recognizing based on the hand pattern feature when communicating.

. Schema of the proposed system for the shape feature extraction of BISINDO alphabets

A. Images of Bisindo Alphabets
BISINDO alphabet images used are in RGB format (Red, Green, Blue). Handshape of the images representing BISINDO alphabets has different color to the background. The background used in this study is white. Fig. 2 shows 4 handshapes i.e.: A, B, C, D respectively. A, B, and D use two hands while C uses one hand. Based on Fig. 2, it can be seen that the hand color is used as the parameter to differentiate between the hand and background.

B. Segmentation
Image segmentation process aims to extract objects of interest (OoI) from others which makes it easier to conduct the following analysis [9], [10], [11]. In the process of image segmentation, the most commonly used technique is the threshold value which is used to select the optimal threshold value to separate the desired object area from its background in the image based on gray-level. Thresholding has a very important role in the process of forming a binary image [12]. The segmentation process uses color distance measurement algorithm and threshold values to determine the similarity of skin tone with the colors of each pixel in the image. Referring to the images used is RGB color images, the formula distance used is city block distance or Manhattan distance which can be seen in the equation (1). ∆ Where Rs, Gs, Bs, are red, green, blue color component of skin color whereas Rp, Gp, Bp, are red, green, blue color component of each pixel P color in the image. If the value ofΔE>Th, then the pixel P is the part of the object area of interest.
In Fig. 3.(b)shows the result of the desired object segmentation that represents the shape of the Alphabet A ( Fig. 3.(a)). Fig. 3.(c) is a gray-level image from an image of Fig. 3.(b). The result of segmentation using city block distance is not perfect yet due to some black holes in the fingers. This indicates that it needs another further action to cover the holes.

C. Mathematical Morphology
Mathematical morphology is one part of digital image processing used to analyze image [13]. Mathematical morphology uses techniques for digital image processing based on the shape of the object [14]. The mathematical morphology is commonly used the image processing both the binary image and gray-level image. The basic operation of mathematical morphology is dilation and erosion. Dilation is a morphological process to expand the area or size of an object. Dilation image of f by the structure of element B is shown in Equation 2. Erosion is a morphological process to erode or reduce the width of the surface size of an area or object. The erosion image of f by the element structure of matrix B shown in the equation 3.
(3) From these operations, then two other processes were developed, closing and opening. Closing and opening are used as morphological filtering in digital image processing [15]. Closing was the combination of the dilation process and erosion process conducted successively. The dilation result was as the input for the erosion operation as shown in the equation 4. Closing functioned to cover the holes throughout dilation operation and to prevent the change of the object size through erosion operation.

D. Edge Detection
Edge detection is an image processing technique used to identify and discover sharp discontinuities or to identify object boundaries in an image [16], [17]. The boundaries can be seen due to the intensity differences or the different color between one pixel with the closest neighbor pixels.
In this study, we use the Roberts operator as edge detection. The operator consists of a pair of 2x2 convolution mask is shown in Fig. 5. The results of the edge detection process of the image in the Fig. 4.(b) with use Roberts operator are shown in Fig. 6. Fig. 7 shows the result of edge detection process of alphabet B, C, and D. -

E. Contour Following
Contour is a segment that has a width of one pixel and has a length of one or more pixels while the boundary is an interconnected or uninterrupted contour [18]. Identification of contour or contour extraction is a very important part to analyze the contents of an image.
Contour following is a method to find out the start point of the object edge and then follow it forward to the end of the edge. The following forward technique to determine the movement direction of the contour conducted by using eight neighbor pixels. In our case, the entire contour objects were represented by white pixels where the thickness of the contour was one pixel. The process of the contour following as below: 1.Check each pixel in each row for each column. Either it was the part of the contour or not. If the start point of the contour was found then the second stage was taken, however if it was not found then the fourth stage was taken. 2. Do the following forward trace by using eight neighbor pixels to determine the next contour pixel. Do it continuously to the last pixel from this contour. 3. Return to the first stage to find out the other contour. 4. The following is completed.

F. Chain Code
Chain code is used to represent the shape of the contour of the object by using the direction code of the contour movement. The chain code gives the boundary of the characteristics of an image which representing the direction in which pixel is located and has a connection to the starting point [19]. The method for finding contours is also called the Freeman chain code [20]. In this study, we used the eight-direction chain code.
As mentioned above that the contour movement is searched by the approximate direction of the eight neighboring pixels as shown in Fig. 8. The center point is the contour pixel which is tracked by the coordinates (x = 0, y = 0), the next pixel shows the direction of the contour motion encoded with 0 -7 as shown in Table I.  In this study, we modify the chain code because BISINDO alphabets have different contours i.e. if there is more than one contour then each contour will be added number 9 as a boundary between the contour one and the other contour.
The chain code contours of the BISINDO alphabets A, B, C, and D in Fig. 6 and 7 are shown in Fig. 9 to 12. Alphabets A and D have two contours so there is only one '9' as the boundary, the alphabet B has three contours so there are two '9' as boundaries, while the alphabet C has only one contour no boundary.

III. RESULTS AND ANALYSIS
The suggested method and algorithm resulted a clear edge in one-pixel width. This eases to extract the chain code contour of the handshape that represents BISINDO alphabets. Fig. 13 shows the edge result of the BISINDO alphabets from A to Z. The BISINDO alphabets can be classified according to the number of contours; C, E, F, G, H, I, J, K, L, R, S, T, U, V, W, X, Y, and Z have one closed contour, A, D, N, O, P, and Q have two closed contours, whereas B and M have three closed contours. The contours shaping each alphabet have different chain codes. Therefore, the feature number of the contour of each alphabet and the chain code in each contour can be used as the feature in the process of recognizing the BISINDO alphabets.
The problems in the object size or hand image size were sometimes smaller or bigger during the image acquisition. The size change caused the different length of the chain code even though the shape of the alphabets were still similar. Therefore, in this study, the probability feature of each code in each contour is used to make the chain code independent from the object size change. The probability of each chain code and the contour number (Table II) as a feature in each alphabet is saved in the database which would be used in the process of recognizing BISINDO alphabets.

IV. CONCLUSION
The suggested method in this study consists of the algorithm for segmentation, closing operation, edge detection, contour following and chain code contour. The method is very effective and produce the edge with a thickness of one pixel. It facilitate the chain code extraction to get the probability which would be used as feature of each BISINDO alphabets. These features were saved in the database to become the feature references in the recognizing process of BISINDO alphabets.