Research On Image Clustering Based On Auto Encoder

Clustering is a fundamental technology of unsupervised learning,aiming to segment unlabeled data into several clusters or groups.It is widely used in computer vision,data mining and other fields.The traditional clustering algorithms can achieve good results for low-dimensional data,but it is ineffective and inefficient for high-dimensional complicated and large-scale data such as images,due to highly redundancy or noise inherent in high dimensional data,and huge computational cost.Auto Encoder network(AE)is the mainstream dimensionality reduction method for clustering.Using well-designed AE and appropriate clustering loss function,some existing methods have shown effectiveness in extracting low-dimensional clustering-friendly features,reducing the computation cost,and improving the clustering effect.In this paper,a new structure of AE network is designed,and a new clustering model and algorithm is presented.The main contributions are as follow.First,a single input and multi output Auto Encoder(SIMOAE)network is proposed to improve the existing single input and single Output Auto Encoder(SISOAE)networks.Image features extracted by SISOAE network have single semantic meaning,while those extracted by SIMOAE have multiple semantic meaning.Experimental results on multiple datasets show that SIMOAE is significantly superior to SISOAE in two cluster evaluation indicators.Secondly,inspired by the idea of contrast learning,a Contrastive Clustering Loss(CCL)function is proposed.The purpose of the CCL is to make the extracted features of subjects from the same cluster gather compactly,thus facilitate data segmentation.Experimental results show that,with the CCL,our SIMOAE network can significantly improve the clustering effect,and the proposed algorithm significantly outperforms the benchmark algorithms on all test data sets we used.