Author Name(s): Morarjee Kolla, Dr. T. VenuGopal
Author Email:


Deep Clustering learns feature representations in embedded space suitable for clustering. In Deep Convolutional Embedded Clustering (DCEC) algorithm, the last convolution layer feature map of encoder is used to build the embedded space. This considers spatial information retains in the last convolution layer of encoder, which unable to identify the discriminative parts of the image. To address this issue, we propose a solution using Global Average Pooling (GAP) of the last convolution layer feature maps in the encoder.  This will encourage the network to identify all discriminative regions and an extent of an object to formulate semantic image clusters (SIC). Our experimental results prove the efficiency of proposed Global Average Pooled Deep Convolutional Embedded Clustering (GAPDCEC) for simultaneous feature learning and clustering.


Clustering is an Unsupervised data analysis and visualization methodology for grouping similar images in Content based image retrieval (CBIR). Most of the clustering methodologies deal with different distance functions to measure the similarity between the raw pixels on data space. These distance driven approaches on handcrafted images are still in an infancy stage to extract meaningful clusters. Mapping raw pixel representations with human understandable representations is a quite challenging task. Recently data-driven approaches are more successful in building cluster friendly feature space. Clustering is a complex process in large data space. To avoid this, recent approaches [1,2,3] are using shallow embedded space.

Semantic Image Clustering (SIC) is the unsupervised machine learning approach using knowledge representation concepts to group unlabelled images meaning fully. It leans the efficient discriminative features to reduce semantic gap in retrieval systems. Most popular methodology in this area uses relevance feedback(RF) [4]. Some other approaches are using knowledge representation mechanisms like Ontology [5] in their contributions. RF deals with low level representation based traditional approach with human judgement. Ontology deals with concepts, constraints and domain knowledge. However, these approaches are still facing difficulties with complex structure of unlabelled images.

Recent success in deep learning approaches are using a shallow embedded space through representation learning [1] for feature learning and clustering individually. Some other approaches [3] are jointly learning features and clustering in embedded space with high performance.

Some of the interesting questions in deep learning are still unclear. For example, which framework CNN features are more suitable for clustering? How is the memory space to store these deep features? What type of autoencoders are more suitable for effective clustering? Which methodology will be useful in creating cluster friendly embedded space? Which loss functions are effective in reducing both clustering and reconstruction loss of deep autoencoders?

Our main idea is to reduce the space for valid deep features suitable for clustering. This embedded latent space captures the structure of data and used to group similar data points


This paper proposes Semantic Image Clustering with GAPDCEC algorithm, which identifies all discriminative regions of image. This framework avoids an overfitting problem and concentrates on user query-based cluster target regions effectively. It learns discriminative features with localization, which can improve the cluster accuracy. The experiments conducted on five different comparative methods on three benchmarked data sets demonstrated the efficacy of GAPDCEC in terms of accuracy, NMI of clusters. In future, experiments will be conducted on complex and real time datasets to explore novel deep learning frameworks.

482 total views, 2 views today

Download PDF File

About the author: admin