C#与OpenCV实战：基于FCM算法的模糊图像分割入门指南

1次阅读

共计 2476 个字符，预计需要花费 7 分钟才能阅读完成。

图像分割 (Image Segmentation) 是计算机视觉中的基础任务，目的是将图像划分为具有相似特性的区域。传统方法如阈值分割 (Thresholding) 和边缘检测 (Edge Detection) 在处理模糊图像时往往效果不佳，因为：

难以应对光照不均
对噪声敏感
无法处理渐变边界

FCM(Fuzzy C-Means, 模糊 C 均值)算法通过引入模糊隶属度概念，能更好地处理这类不确定性问题。

核心思想：每个像素可以同时属于多个类别，隶属度表示归属程度。目标函数为：

$$
J = \sum_{i=1}^N \sum_{j=1}^C u_{ij}^m |x_i – c_j|^2
$$

其中：
– $u_{ij}$：第 i 个像素对第 j 类的隶属度
– $m$：模糊因子(通常取 2)
– $c_j$：第 j 类的聚类中心

通过交替优化隶属度和聚类中心使目标函数最小化。

安装 NuGet 包：

Install-Package OpenCvSharp4
Install-Package OpenCvSharp4.runtime.win

验证安装：

using OpenCvSharp;
Mat testImage = new Mat();
Console.WriteLine("OpenCV 加载成功");

Mat PreprocessImage(string path)
{
    // 读取图像并转为灰度
    Mat src = Cv2.ImRead(path, ImreadModes.Color);
    Mat gray = new Mat();
    Cv2.CvtColor(src, gray, ColorConversionCodes.BGR2GRAY);

    // 归一化到 [0,1] 范围
    Mat normalized = new Mat();
    gray.ConvertTo(normalized, MatType.CV_32F, 1.0/255);

    return normalized;
}

public class FCMCluster
{public int ClusterCount { get; set;} = 2;
    public float FuzzyFactor {get; set;} = 2.0f;
    public int MaxIterations {get; set;} = 100;
    public float Epsilon {get; set;} = 0.01f;

    public Mat Run(Mat data)
    {
        // 初始化隶属度矩阵
        Mat membership = InitMembership(data.Rows);

        for (int iter = 0; iter < MaxIterations; iter++)
        {Mat centers = UpdateCenters(data, membership);
            Mat newMembership = UpdateMembership(data, centers);

            // 检查收敛条件
            if (CheckConvergence(membership, newMembership))
                break;

            membership = newMembership;
        }

        return membership;
    }

    private Mat InitMembership(int sampleCount)
    {
        // 随机初始化隶属度矩阵
        Random rand = new Random();
        Mat membership = new Mat(sampleCount, ClusterCount, MatType.CV_32F);

        for (int i = 0; i < sampleCount; i++)
        {
            float sum = 0;
            for (int j = 0; j < ClusterCount; j++)
            {float val = (float)rand.NextDouble();
                membership.Set(i, j, val);
                sum += val;
            }

            // 归一化
            for (int j = 0; j < ClusterCount; j++)
            {float normalized = membership.At<float>(i, j) / sum;
                membership.Set(i, j, normalized);
            }
        }

        return membership;
    }

    // 其他关键方法省略...
}

void VisualizeResult(Mat membership, Mat original)
{
    // 获取硬分类结果
    Mat labels = new Mat(membership.Rows, 1, MatType.CV_8U);
    for (int i = 0; i < membership.Rows; i++)
    {
        int maxIdx = 0;
        float maxVal = 0;
        for (int j = 0; j < membership.Cols; j++)
        {float val = membership.At<float>(i, j);
            if (val > maxVal)
            {
                maxVal = val;
                maxIdx = j;
            }
        }
        labels.Set(i, 0, maxIdx);
    }

    // 上色显示
    Mat colored = new Mat();
    Cv2.ApplyColorMap(labels.Reshape(1, original.Rows), colored, ColormapTypes.Jet);

    Cv2.ImShow("Result", colored);
    Cv2.WaitKey();}