Most of the machine learning tasks are largely dependent on labeled data. Additionally, obtaining accurate and high-quality annotated data represents a significant train-ing process bottleneck. If we consider image clustering al-gorithms for creating labels, they are usually focused on working with data of small dimensions. Images data, they usually have a high dimension (Height x weight x No of Channels). It would be computationally intensive to clus-ter images. We have explored two techniques which would extract the latent representations of images in a low dimensional space to which clustering algorithms can be effectively applied. In the first approach we have built autoencoder network for an auxiliary task through which the latent representations of images are extracted which were further reduced using t-SNE and later clustered using K-Means. In the second technique we built a neural net-work and trained it on a arbitrary task to predict the rota-tion of input images. The latent representations or encod-ings from the hidden layer are extracted which were fur-ther reduced using t-SNE and are clustered using K-Means. Both the techniques were validated successfully on MNIST dataset and extended it on a real world COVID-19 CT scan images dataset.