Bionic Vision Descriptor for Image Retrieval

摘要

Human visual system gets remarkable performance by processing low-level features. In the last decade, many descriptors have been proposed for feature extraction. However, fewer of them get satisfying performance with low-level features. Compared to high-level ones, low-level features make use of natural underlying elements like texture and they are extracted directly, which makes low-level features more efficient in image retrieval domains. In this paper, a new descriptor named Bionic Vision Descriptor (BVD), which is based on the principle of human visual system, is proposed. The descriptor fuses uniform low-level features extracted from color, texture and gradient elements. Moreover, matrix calculation and feature selection are utilized to accelerate the calculation of BVD. Experimental results show that our method outperforms other state-of-the-art traditional descriptors with less runtime and fewer initial dimensions on benchmark datasets.