Optimized feature space learning for generating efficient binary codes for image retrieval

In this paper, a novel approach for learning low-dimensional optimized feature space image retrieval with minimum intra-class variance and maximum inter-class is proposed. The classical of Linear Discriminant Analysis (LDA) generally used generating an single-labeled images. Since involves images multiple objects, LDA cannot be directly dimensionality reduction optimization. This problem addressed by utilizing the relationship between Canonical Correlation (CCA) eigenvalues to generate both multi-labeled A CCA-based network architecture which correlates vectors label We design loss function such that correlation coefficients CCA are maximized. Our experiments prove we could train neural reach theoretical lower bound corresponding negative sum coefficients. Once generated, binarized Iterative Quantization (ITQ) approach. Finally, propose ensemble binary codes desired bit length retrieval. measurement mean average precision shows proposed outperforms results other benchmarks at same numbers in considerable number cases.