Image Clustering Algorithm Based on Predefined Evenly-Distributed Class Centroids and Composite Cosine Distance
The clustering algorithm based on deep neural network clustersby obtaining the optimal feature representation,but the clustering performance is still not ideal in the face of the complexity of the image.This paperpresents an image clustering algorithm based on predefined evenly-distributed class centroids(PEDCC) and composite cosine distance.Compared with current popular auto-encoder structure, we design an encoder only network structure with normalized latent features, and two effective loss functions in latent feature space by replacing the Euclidean distance with cosine distance.We find that (1) contrastive learning plays a key role in clustering algorithm and greatly improves the quality of learning latent features; (2)Compared with Euclidean distance, composite cosine distance can be more suitable for the normalized latent features and PEDCC-based Maximum Mean Discrepancy(MMD) loss function.
You must have the following already installed on your system.
1、Pytorch 1.0
2、sklearn
3、python 3.6
For reproducing the experiments on MNIST、Fashion-Mnist、COIL20 datasets and cifar-10, run the following codes.
1、python PEDCC.py : to Initialize the PEDCC, You need to set the cluster number, and every kind of dimension. We suggest that the MNIST every picture extract 60 dimension feature vector.
2、Modify data_transform.py: you should choose datasets.
3、python main.py for training.