Neural Method for Explicit Mapping of Weighted Locally Linear Embedding in Image Retrieval

A new explicit nonlinear dimensionality reduction(DR) method, on account of neural networks, is presented for image retrieval tasks. We first propose a Weighted Locally Linear Embedding (WLLE) for training set, based on which linear relations in neighborhood of each sample are guaranteed. Then, a neural method (NM) is proposed to solve the out-of sample problem. As a combination of WLLE and NM, we provide an explicit nonlinear DR approach for efficient image retrieval. The experimental results in three benchmark datasets illustrate that our algorithm could get outstanding performance than other state-of-the-art out-of-sample methods.