Abstract: Manifold-based subspace feature extraction methods have recently been deeply studied in data dimensionality reduction. Inspired by PCA Hashing (PCAH), if the Locality Preserving Projection (LPP) is directly used in the hash image retrieval, it is prone to shortcomings such as being inefficient and time-consuming. In order to address these deficiencies, this paper mainly combines Principal Component Analysis (PCA) and manifold subspace feature extraction method LPP, and we present a RLPH framework using random rotation. Among them, PCA processing solves the eigenvalue problem encountered in the calculation of LPP, thereby improving the recognition effect of the algorithm. The PCA projection needs to ensure that the variance of the sample points after projection is as large as possible. However, projections of small variance may produce unnecessary redundancy and noise. Therefore, in the subspace after the PCA projection, we only extract the eigenvectors that contain most of the information at the top of the PCA projections. Then, we utilize a random orthogonal matrix to randomly rotate and shifts the eigenvectors and the reduced-dimensional sample obtained after the top eigenvectors of the PCA projection is subjected to LPP mapping. Random rotation produces many thin projection matrices blocks that are then concatenated into one final projection matrix. Random rotation is a key step in this paper that minimizes the quantization error for codes. The proposed method greatly improves the retrieval efficiency, and extensive experiments demonstrate its effectiveness.

Key words: manifold, data reduction, hashing, PCA, LPP, random rotation