In this paper we propose an approach for learning low dimensional optimized\nfeature space with minimum intra-class variance and maximum inter-class\nvariance. We address the problem of high-dimensionality of feature vectors\nextracted from neural networks by taking care of the global statistics of\nfeature space. Classical approach of Linear Discriminant Analysis (LDA) is\ngenerally used for generating an optimized low dimensional feature space for\nsingle-labeled images. Since, image retrieval involves both multi-labeled and\nsingle-labeled images, we utilize the equivalence between LDA and Canonical\nCorrelation Analysis (CCA) to generate an optimized feature space for\nsingle-labeled images and use CCA to generate an optimized feature space for\nmulti-labeled images. Our approach correlates the projections of feature\nvectors with label vectors in our CCA based network architecture. The neural\nnetwork minimize a loss function which maximizes the correlation coefficients.\nWe binarize our generated feature vectors with the popular Iterative\nQuantization (ITQ) approach and also propose an ensemble network to generate\nbinary codes of desired bit length for image retrieval. Our measurement of mean\naverage precision shows competitive results on other state-of-the-art\nsingle-labeled and multi-labeled image retrieval datasets.\n