On selection of centroids of fuzzy clusters for color classification

Dae-Won Kim,Kwang H. Lee

2024-07-09

Abstract:A novel initialization method in the fuzzy c-means (FCM) algorithm is proposed for the color clustering problem. Given a set of color points, the proposed initialization extracts dominant colors that are the most vivid and distinguishable colors. Color points closest to the dominant colors are selected as initial centroids in the FCM. To obtain the dominant colors and their closest color points, we introduce reference colors and define a fuzzy membership model between a color point and a reference color.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

The problem that this paper attempts to solve is **the selection of initial center points in the Fuzzy C - Means (FCM) algorithm for the color clustering problem**. Specifically, the author proposes a new initialization method to extract the most vivid and distinguishable colors as initial center points, thereby improving the effect of color clustering.

### Detailed Explanation

1. **Problem Background**:
- The goal of the color clustering task is to divide a set of colors into multiple homogeneous color clusters.
- Color clustering is essentially a fuzzy task because color boundaries are often fuzzy, especially in applications such as image segmentation.
- Fuzzy set theory and fuzzy logic are very suitable for dealing with this uncertainty, so fuzzy clustering methods are more applicable in this case.

2. **Existing Challenges**:
- **Determining the Number of Clusters**: Most algorithms require the user to specify the number of clusters.
- **Selection of Initial Center Points**: Most algorithms use the method of randomly selecting initial center points, but this method may lead to local optimal solutions.
- **Handling the Diversity of Data**: A method is needed to handle data points with different shapes, densities, and quantities.

3. **The Solution Proposed in the Paper**:
- The author proposes a new initialization method based on reference colors and membership functions.
- **Steps**:
1. **Define Reference Colors**: 14 common reference colors (such as red, green, blue, etc.) are selected, which represent the main colors that may be encountered in natural scenes.
2. **Calculate Membership**: By defining a membership function, calculate the similarity between each color point and the reference color. The membership function takes into account the distance from the color point to the reference color and uses a fuzzy parameter \(\lambda\) to control the fuzziness of the membership.
3. **Select Dominant Colors**: According to the maximum value of the membership, select the most vivid and distinguishable colors as the dominant colors.
4. **Determine Initial Center Points**: Select the color point closest to the dominant color as the initial center point.

4. **Mathematical Formulas**:
- **Distance Formula**:
\[
\delta(x, R_i)=\sqrt{(x_L - r_{iL})^2+(x_a - r_{ia})^2+(x_b - r_{ib})^2}
\]
- **Membership Function**:
\[
\mu_{R_i}(x)=\begin{cases}
1.0 & \text{if } \delta(x, R_i) = 0\\
0.0 & \text{if } \delta(x, R_j) = 0\\
\left(\sum_{j = 1}^{k}\left(\frac{\delta(x, R_i)}{\delta(x, R_j)}\right)^\lambda\right)^{-1} & \text{otherwise}
\end{cases}
\]

5. **Experimental Verification**:
- The author shows how to use this method for initialization through a simple example and finally determines the initial center points.

### Conclusion
This paper provides a new initialization method by introducing reference colors and membership models, which can select initial center points more effectively, thereby improving the performance of the Fuzzy C - Means algorithm in the color clustering problem.