Research On Image Clustering Method With Semantic Features Fusion

Nowadays,a large number of image data are produced from both real life and cyber world.Due to the huge number and variety of images,it is a hard work to label and classify all images.Therefore,in the research field of machine learning,clustering unlabeled image has been a research hotspot for a long time.Usually,the feature dimension of image is very high,for the image clustering tasks,feature extraction and dimensionality reduction have import influence for clustering performance.Although many mature feature extraction methods have been proposed in the past decades,these methods mainly focus on the low-level features of image in clustering tasks while neglect the middle level features.In fact,the middle-level features of images often have discriminative information for clustering research,they can enrich image feature expression and improve clustering performance.As the middle-level features,the semantic features plays a crucial role in image retrieval,natural language processing.Therefore,this thesis attempts to combine the low-level features with the middle-level features and applying them to image clustering.As the dimensionality of low-level features of image is very high,the combination of low-level features and semantic features may cause dimension disaster.In order to achieve better clustering performance,we should not only effectively fuse the two kinds of features,but also eliminate redundant features,this thesis proposes a Deep Semantic Embedding Clustering Algorithm,firstly,the low-level features and semantic features of the image are simply cascaded together,then,fusing the two features by the depth autoencoder meanwhile,to reduce the dimensionality of image features.By this way,we finally achieve effective image clustering by using the low-dimensional fused features.A Weighted Superposition Fusion Algorithm is proposed to improve the simple feature fusion.This method can clearly measure the contribution of two features to clustering and distribute different weights.Experiments on several image datasets show that the clustering performance with the fused features is improved obviously,which is about 3 percentage points higher than that of the previous sole feature clustering.This fully proves the effectiveness of the proposed clustering approaches.