使用K-means和高斯混合模型對圖像進行聚類

導入圖片

%matplotlib inline
import numpy as np  
import skimage.io as SKimg  
import matplotlib.pyplot as plt  
from sklearn.cluster import KMeans
from sklearn.mixture import GaussianMixture
from skimage import io,data,color
from scipy.ndimage import gaussian_filter
from mpl_toolkits.mplot3d import Axes3D  # noqa: F401 unused import
import time
import matplotlib.cm as cm

imgpath=r"H:\02Course\pics\35049.jpg"
img =SKimg.imread(imgpath)
io.imshow(img)
print(type(img))  #顯示類型
print(img.shape)  #顯示尺寸
print('圖片高度:',img.shape[0])  #圖片高度
print('圖片寬度:',img.shape[1])  #圖片寬度
print('圖片通道數:',img.shape[2])  #圖片通道數
print('總像素個數:',img.size)   #顯示總像素個數
print('最大像素值:',img.max())  #最大像素值
print('最小像素值:',img.min())  #最小像素值
print('像素平均值:',img.mean()) #像素平均值
print('圖像的像素值:',img[0][0])#圖像的像素值

image_height, image_width, image_channels = img.shape
image_pixels = np.reshape(img, (-1, image_channels))#將三維矩陣轉為通道特征矩陣  每列代表R G B值
_mean = np.mean(image_pixels,axis=0,keepdims=True)
_std = np.std(image_pixels,axis=0,keepdims=True)
image_pixels_ = (image_pixels - _mean) / _std # 歸一化
X=image_pixels_ #特征矩陣

#繪制每個通道的直方圖
Kwars=dict(histtype='stepfilled',alpha=0.4,bins=np.arange(0,255))#,normed=True
intervals=plt.hist(image_pixels,color=[ "red","green","blue"],**Kwars)

 

#聚類個數從3遞增至8
for ncomp in range(3,9):
    fig =plt.figure(figsize=(12,10))
    plt.axis('off')
    t0 = time.time()

    gmm = GaussianMixture(n_components=ncomp,covariance_type='full').fit(X)#調用GMM算法，設置聚類的目標類別數
    Clus_Centers=gmm.means_#每個類別的聚類中心 
    Labels=gmm.predict(X)

    elapsed_time = time.time() - t0
    Labels_show=np.reshape(Labels,(image_height, image_width))#把分類標簽展開為二維數組
    dst=color.label2rgb(Labels_show)#將標簽轉為RGB
    plt.title('n_cluster=%i,(time%.2fs)'%(ncomp,elapsed_time), size=25)
    plt.imshow(dst)
    fig.savefig('pics/gmm_cluster_%i.jpg'%(ncomp), format='jpg', dpi=300,bbox_inches = 'tight',pad_inches = 0)#保存聚類結果圖
    
    Labels_arr=np.reshape(Labels_show, (-1, 1))
    combine=np.append(image_pixels, Labels_arr, 1)
    fig = plt.figure(figsize=(15, 10))
    ax = plt.axes(projection='3d')
    xs = combine[:,0]
    ys = combine[:,1]
    zs = combine[:,2]
    ax.scatter(xs, ys, zs,c=combine[:,3],edgecolors='none',s=1)
    ax.set_xlabel('R channel')
    ax.set_ylabel('G channel')
    ax.set_zlabel('B channel')
    plt.show()
    fig.savefig('pics/gmm_cluster3d_%i.jpg'%(ncomp), format='jpg', dpi=300,bbox_inches = 'tight',pad_inches = 0)#保存RGB通道三維特征空間圖
    
    fig = plt.figure(figsize=(6,5))
    plt.xlabel('R channel')
    plt.ylabel('G channel')
    plt.axis('equal')
    plt.xlim(0, 255)
    plt.ylim(0, 255)
    plt.scatter(combine[:, 0], combine[:, 1], c=combine[:,3],s=.2)
    plt.colorbar()
    plt.show()
    fig.savefig('pics2/gmm_clusterR&G_%i.jpg'%(ncomp), format='jpg', dpi=300,bbox_inches = 'tight',pad_inches = 0)#保存RG通道二維特征空間圖
    
    fig = plt.figure(figsize=(6,5))
    plt.xlabel('R channel')
    plt.ylabel('B channel')
    plt.axis('equal')
    plt.xlim(0, 255)
    plt.ylim(0, 255)
    plt.scatter(combine[:, 0], combine[:, 2], c=combine[:,3],s=.2)
    plt.colorbar()
    plt.show()
    fig.savefig('pics/gmm_clusterR&B_%i.jpg'%(ncomp), format='jpg', dpi=300,bbox_inches = 'tight',pad_inches = 0)#保存RB通道二維特征空間圖
    
    fig = plt.figure(figsize=(6,5))
    plt.xlabel('G channel')
    plt.ylabel('B channel')
    plt.axis('equal')
    plt.xlim(0, 255)
    plt.ylim(0, 255)
    plt.scatter(combine[:, 1], combine[:, 2], c=combine[:,3],s=.2)
    plt.colorbar()
    plt.show()
    fig.savefig('pics/gmm_clusterG&B_%i.jpg'%(ncomp), format='jpg', dpi=300,bbox_inches = 'tight',pad_inches = 0)#保存GB通道二維特征空間圖

 

聚類算法換位K-means，只需將聚類部分代碼換為：

kmeans = KMeans(n_clusters=ncomp,random_state=0)
Labels = kmeans.fit_predict(X)

　　

 

#高斯濾波 可平滑噪聲 #替換即可
img =SKimg.imread(imgpath)
img[:,:,0] = gaussian_filter(img[:,:,0], sigma=1)
img[:,:,1] = gaussian_filter(img[:,:,1], sigma=1)
img[:,:,2] = gaussian_filter(img[:,:,2], sigma=1)
plt.axis('off')
plt.imshow(img)
plt.savefig('pics/gausfilter.jpg', format='jpg', dpi=300,bbox_inches = 'tight',pad_inches = 0)

　　

 

 

 

　　

　　

 

 

參考：

https://jakevdp.github.io/PythonDataScienceHandbook/05.12-gaussian-mixtures.html

https://scikit-learn.org/0.15/modules/generated/sklearn.mixture.GMM.html