Fast Color Matching Using Weighted Subspace on Medicine Package Recognition Kenjiro Sugimoto and Sei-ichiro Kamata Graduate School of Information, Production and Systems, Waseda University 2-7 Hibikino Wakamatsu-ku Kitakyushu-shi Fukuoka, 808-0135 JAPAN sugimoto@asagi.waseda.jp, kam@waseda.jp Abstract This paper presents a color matching technique us- ing weighted subspace on medicine package recogni- tion. The proposed method is more compact and lower- complex than scalable color descriptor and dominant color descriptor, which are employed by MPEG-7. Our method is based on subspace matching: A color object is treated as a subspace derived from its color distri- bution. Unlike mutual subspace method, it is specially designed for color matching. Specifically, weighted sub- space and a distance-based dissimilarity are employed instead of normalized subspace and similarity based on canonical angles of MSM. Experiments show that the proposed method outperforms the conventional methods in terms of description size, building/matching speed, and recognition rate. 1 Introduction Medicine package recognition is a significant tech- nique for preventing dispensing errors, namely incor- rect prescription of a medicine or its dosage. Fast and accurate recognition algorithms are required in or- der to actualize smooth and reliable dispensing oper- ation. This research targets Press-Through Package (simply called a package in the paper), which is one of the most popular packages for pills/tablets/capsules, as shown in Figure 1. Color information is an im- portant features of the packages. A package con- tains several pills/tablets/capsules and some charac- ters/symbols are regularly printed on its surface. Ev- ery package is uniquely color-designed, normally show- ing in a few distinct colors. Thus, color matching is a reasonable solution for medicine package recog- nition. Although symbol recognition is also effective, color matching is nevertheless useful to fast prune can- didates prior to symbol recognition. This paper focuses on low-level color descriptors. Many descriptors including MPEG-7[1] have been proposed in the past. The conventional descriptors can be classified into two groups: description based on histogram[2, 3] and dominant color[4, 5]. MPEG- 7 employs scalable color descriptor (SCD) as the for- mer and dominant color descriptor (DCD) as the lat- ter. Histogram description is easy to build and match, whereas the size is relatively large, e.g. RGB space di- vided into 8×8×8 requires 512 bins. Dominant color description is much more compact, whereas clustering process to find dominant color consumes much com- puter resources. In order to resolve the drawbacks, a color descriptor using eigenvectors and eigenvalues was proposed in [6]. The method treats a color distri- bution as eigenvectors weighted by their correspond- ing contribution ratio. However, three unconvincing points remain: (1) the reason why the weighted eigen- vectors can efficiently characterize color distributions, (2) concrete difference from mutual subspace method (MSM)[7], a successful technique based on subspace matching on pattern recognition, (3) effect of dimen- sion reduction. Therefore this paper specifies the three points and presents a modified color matching method based on the analysis. The rest of the paper is organized as follows: Sec- tion 2 analyzes subspace description of color distribu- tions and outlines drawbacks of MSM for color match- ing. Section 3 proposes a modified subspace method specialized for color matching. Section 4 evaluates its performance by comparing with MSM, SCD and DCD. Finally, Section 5 provides our conclusions. 2 Color Matching Using Subspace Method 2.1 Analysis on subspace description Subspace description requires the following condi- tion: A color distribution in K-dimensional color space consists of C color clusters where C is unknown but satisfies C ≤ K + 1, e.g. K = 3 for RGB space and K = 4 for CMYK space. Each cluster is assumed to be a point. Figure 2 illustrates some examples of dis- tributions with C =1, 2, 3, 4 under K = 2. If C =1 (see Figure 2a), the cluster locates on a point. If C =2 (see Figure 2b), the two clusters can produce a line. If C = 3 (see Figure 2c), the three clusters can produce a plane. The point, line, and plane are zero-, one-, and two-dimensional subspaces respectively. By contrast, if C = 4 (see Figure 2d), the four clusters still pro- duce a plane because of limitation of K = 2. Thus, as long as C ≤ K + 1, eigenvalues are effective to repre- sent C and power relationship between clusters. Next, another helpful information is the direction of eigen- vectors. Even if C is the same, eigenvectors can distin- guish distributions with distinct clusters. Accordingly, a subspace in K-dimensional space can characterize up to (K + 1) clusters. Another positive effect of subspace description is noise reduction for more practical cases. Practically, each cluster is a dense manifold, not a point. Con- sider an ideal distribution (see Figure 2b) and its more practical one (see Figure 2e) composed of two clusters. The two clusters in the ideal case are exactly located on a one-dimensional subspace and its orthocomplement contains no energy. By contrast, the orthocomplement in the practical case will contain some energy. Projec- tion to the subspace can moves the practical distribu- tion closer to the ideal one (see Figure 2f). Therefore, subspace description behaves also as noise reduction if C ≤ K. MVA2011 IAPR Conference on Machine Vision Applications, June 13-15, 2011, Nara, JAPAN 9-15 287