ImageCluster

The ImageCluster project is aimed to cluster unlabeled images based on the SOTA models.
It designed for engineers and students to fast feature maps and cluster the image according to your cluster-algo hyperparameter.
This flow contains two main steps:

1. Use SOTA pre-trained models(egs:VGG16,VGG19,ResNet50 ) on imagenet to extract feature maps.
   You can choose the bone model and resize the images(default size is (224,224)) as you like.
2. Choose a clustering algorithm (eg: kmeans ) to label these feature maps. Therefore the images can be labeled to K classes.
   So,the data flow is shown as below
   Image->array->feature map->labeling(cluster algorithm)

Getting Started

Installing

• system: win10/Ubuntu
• python:3.5+
• Clone or download the repository
• Commands to install requirements via Conda or pip wheels
  pip install -r requirements.txt

Usage

You can import this model script to your local the same-level python scripts.
egs:

from model import ImageCluster
m=ImageCluster(
    base_model='vgg16',#your feature map extractor model
    resorted_img_folder='resorted_data',#the folder for clustered images
    cluster_algo='kmeans',#cluster algorithm
    maxK=30,#the max k num is 30, which means ImageCluster calculates every k in range(2,30+1)
)
#calculate the feature maps
m.get_feature_map(
    resize_shape=(224,224) # (w,h)  a tuple for resizing the input images to the same shape images
)
#clustering for feature maps
m.imagecluster()

Then, you can see the cluster sse value map, which draws a line shows each sse value for the k in range [2,maxK].  
Elbow criterion is applied to choose the best K for your cluster

As we can see, 21 may be the best cluster number for this dataset.  
So,we can call the resorted_img function to label the images under different folders

m.resorted_img(
    selected_k_num=21# a int number in range[2,maxK]
)

In the end, the neatly arranged label folders and the image under these folders are shown below.  

Authors

• Chris Li - all work

License

This project is licensed under the Apache License - see the LICENSE file for details