A Novel Feature Point Detection Algorithm Based on Strips

Feature point detection plays a crucial role in shape representation and object recognition. A feature point set that can represent the shape honestly and consistently under different scale and environment is desired. The method used should be able to cater to these requirements as much as possible. Regretfully, no method has done completely well. Dynamic two-strip algorithm (Dyn2S) used the strip to extract features. Digitization noise can be tolerated because the strip has width and it can enclose points that can be approximated as a straight line. Unfortunately, its performance seems not very satisfactory on curves. In this paper, further investigation has been carried out along this direction. The proposed method, based on the long and narrow strips that are prominent and reliable, has been applied on logos with encouraging results. This approach is more capable of extracting consistent feature points fiom one instance of a model shape to another. 1 Dynamic two-strip algorithm [I] Dyn2S is a curve representation method using feature points. The whole procedure has two stages. In the first stage, important properties of the data points are derived. Two strips are fitted to the left and right sides of each point on the curve respectively, and the points inside each strip are approximated as the left or right straight line. The best fitted strips are derived by adjusting the width of the strip and rotating it up or down dynamically using the starting point as the pivot. Longer and narrower strips are preferred. The ratio, which is called the elongate value of a strip, is defined as E=UW of its length L over its width W to represent the merit of fit. The merit f for a point can be computed as f = E'@s'E"~~' , where El@ and Engh' are the elongate values of the left and right strips, and so =I 180° 8 1 is the acuteness of the angle 8 between the two strips. In the second stage, a subset of the most representational points is selected based on the local maximum of point merits. Address: Singapore Technologies Dynamics Pte Ltd, 249 Jalan Boon Lay, Singapore 6 19523. Email: [email protected] Maylor K. H. ~ e u n ~ ~ School of Computer Engneering Nanyang Technological University One example is illustrated in Figure 1 where the minimum and maximum strip widths can be easily pre-selected as 1 and 0.1 x D pixels, where 1 is considered small enough while D is the length of the image diagonal. Let Po be the current point to be processed, L, (i=l, 2, ..., 6) be points on Po's left side and R j be points on Po P right side. An initial strip of minimal width would first extend fiom Po to encompass as many points as possible to its left side (i.e. the dashed-line box). Its width is then increased and its orientation is adjusted to include as many points as possible, until the maximum strip width is reached (i.e. the left solid-line box). The measure of fit (i.e. elongate value) is recorded for each strip as a function of strip width but only the strip with the highest value is marked and used. The same procedure is applied to the other side of the point. Figure 1. The illustration of the dynamic two-strip algorithm. There are advantages using strips to extract local features. First, digitization noise can be tolerated because the strip has width and it can enclose points that can be approximated as a straight line. Then the support region of one point can be reflected honestly by the strips. However, the Dyn2S does not work well on curves with no obvious sharp points or comers. This gives rise to inconsistent point sets being generated on different instances of the same model. Here, the robust shape feature (i.e. the long and narrow strip) is retained and further developed into a new approach to extract feature points. Address: BLK N4, School of Computer Engineer, NTU, Singapore 639798. 2 The proposed method Teh and Chin [2] indicated that feature point detection relies more on the determination of a local support region rather than on the estimation of discrete curvature. A novel feature point extraction technique based on the local support region (i.e. strip from dyn2S) is proposed here. Instead of starting from a point and looking for support regions from its left and right sides, the proposed method looks first for the supports (i.e. the strips) and then computes the location of the feature point. Strips have the good property of tolerating noises as described in Dyn2S. In addition, one can sort all the strips according to their elongate values and start searching for feature points based on strips of the largest local elongate values i.e. the long and narrow strips that are prominent and reliable. This approach is more capable of extracting consistent feature points from one instance of a model shape to another since it relies on the local most reliable strips. It also allows labeling feature points according to their important attributes, i.e. the strip elongate value. Some characteristics of strips can be observed from Figure 2 (a) where the strip lengths and elongate values become smaller and smaller when one goes up-hill. The values start to become larger and larger when the strips are heading down the hill. In addition, the rotation directions of the strips also change from clockwise to counterclockwise. This direction change can be detected as shown in Figure 2 (b) where the intersecting angles (a and P) from strip S1 to strips S 2 , S, ,..., Sk are changing in a monotonous manner. In this case, it is decreasing monotonously. The angle starts to change in different direction (see angle y) when one considers S,,, which rotates in counter-clockwise manner. The feature point (e.g., point D in Figure 3) can then be located by two nearest maximal strips (e.g., S1 and St in Figure 2 (b)) whose elongate values are local maxima. At the same time, the rotation direction after Sk should turn to the other direction. Hence, one has two indicators for manipulation. The next step in computing a feature point can be illustrated as shown in Figure 3. First, one needs to build a triangle AAl3C by extending the center lines going through the strips Sl and Sk . The base, BC, is then moved up until it touches the curve at only one point at D. Point D would be declared as the feature point. Here, the feature point detected from the above process is declared as major feature point. (a) Changing pattern of the ship lengths (b) Changing panern of the strip angular difference. Figure 2. The illustration of the proposed concept for detecting feature point. Figure 3. The illustration of feature point location. The main concern of feature point detection is whether the detected points are good enough to represent the shape of a curve. For example, in Figure 4, point p , and p, are detected from curve to represent the curve segment f romp, to p2 as line p 1 p 2 (the dashed line). Obviously, deviations from points on the curve to p l p z are large. In this case, it might be appropriate to add supplementary point, such as ql and q , . In this study, iterative polygonal approximation method [3] is used to add supplementary feature points. Since major feature points have been detected in the previous process, supplementary feature points can be properly computed from these good starting points. PI, p2 : major feature points. qi, q 2 : supplementary f e w e points. Figure 4. Sample of major and supplementary feature points.