MULTI-LAYER SEMANTIC REPRESENTATION LEARNING FOR IMAGE RETRIEVAL (TP-P6)

Author(s) : Wei Jiang (Department of Automation, Tsinghua University, China) Guihua Er (Department of Automation, Tsinghua University, China) Qionghai Dai (Department of Automation, Tsinghua University, China)

Abstract : Long-term relevance feedback learning is an important learning mechanism in content-based image retrieval. In this paper, our work has two contributions: (1) A Multi-layer Semantic Representation (MSR) is proposed, and an algorithm is implemented to automatically build the MSR for image database through long-term relevance feedback process. (2) The accumulated MSR is incorporated with the short-term feedback learning to aid subsequent users' retrieval. The MSR memorizes the multi-correlation among images, and integrates these memories to build hidden semantics, which are distributed in multiple layers, for images. In experiment, an MSR are built based on real retrieval from 10 different users, which can precisely describe the hidden semantics underlying images, and help to bridge the gap between high-level semantics and low-level features, thus improves the retrieval performance significantly.

Menu