背景

业务当中需要给图片风格进行标签化，关于图片格化处理，这里想的是根据图像分类的算法来做，因为分类属于有监督学习，需要整理相关的训练样本，如果上来就通过人工的方式去整理，成本太高，再加上前期大家认知的不同意，样本质量也难得到保障；所以需要先对已有商品进行聚类分析，在聚类的结果集上进行相关样本的整理。

实现流程

整个流程并不复杂，首先采用深度学习模型进行图片特征抽取，然后采用K-means对特征进行聚类。

特征抽取

CNN模型是图片特征提取常用网络，这里将使用经典图像分类模型Vgg16的卷积网络进行相关图像处理

from keras.applications.vgg16 import VGG16
from keras.preprocessing import image
from keras.applications.vgg16 import preprocess_input
 
import numpy as np
import pylab
import io
from PIL import Image
from numpy import linalg as LA

base_model = VGG16(weights='imagenet', pooling='max',include_top=False)  #这里也可以使用自己的数据集进行训练
    
def get_image_feature(path):
    
    model = base_model
#     print(model.summary())
     
    img = Image.open(path)
    img = img.convert('RGB')
    img = img.resize((224,224), Image.NEAREST)
    x = image.img_to_array(img)
    x = np.expand_dims(x, axis=0)
    x = preprocess_input(x)
    features = model.predict(x)
    return features[0]
 
features = get_image_feature()

vec = features/LA.norm(features)
 
print(vec.shape, vec.tolist())

图片聚类

采用k-means对图片特征进行聚类分析，特征长度512

from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.cluster import KMeans
import pandas as pd
import numpy as np


features = [
    [],
    [] 
]
names = ['', '']

 
df = pd.DataFrame( features )
name_df = pd.DataFrame( names )
samples=df.values

#未进行标准化
kmeans=KMeans(n_clusters=3)
kmeans.fit(samples)#训练模型
labels=kmeans.predict(samples)#预测
name_df[labels==1]

完整代码

网上整理了凯旋门和埃菲尔铁塔的图片，简单测试下效果

from keras.applications.vgg16 import VGG16
from keras.preprocessing import image
from keras.applications.vgg16 import preprocess_input
import os 
import numpy as np
import pylab
import io
from PIL import Image
from numpy import linalg as LA
from sklearn.cluster import KMeans

base_model = VGG16(weights='imagenet', pooling='max',include_top=False)  #这里也可以使用自己的数据集进行训练
    
def get_image_feature(path):
    
    model = base_model
#     print(model.summary())
     
    img = Image.open(path)
    img = img.convert('RGB')
    img = img.resize((224,224), Image.NEAREST)
    x = image.img_to_array(img)
    x = np.expand_dims(x, axis=0)
    x = preprocess_input(x)
    features = model.predict(x)
    return features[0]


feature_ls = []
names = []
for name in os.listdir("d:/tmp/天池/jd"):
    
    features = get_image_feature(os.path.join("d:/tmp/天池/jd", name))
    #特征标准化
    vec = features/LA.norm(features)

    #print(name,"==", vec.tolist())
    names.append(name)
    feature_ls.append(vec.tolist())
    

df = pd.DataFrame(feature_ls)
names_df = pd.DataFrame(names)
samples=df.values 
kmeans=KMeans(n_clusters=3)
kmeans.fit(samples)#训练模型
labels=kmeans.predict(samples)#预测
print(labels, names_df[labels==0] )

效果

结果打印

 # 将标签值为0的数据进行输出
[1 1 1 1 1 0 0 0 0 0 0 0]             0
5    test.jpg
6     tt1.jpg
7     tt2.jpg
8     tt3.jpg
9     tt4.jpg
10    tt5.jpg
11  ttkxm.jpg

可以看出所有的铁塔被划分到一个类了