Large scale near-duplicate celebrity web images retrieval using visual and textual features

Area of Science:

• Computer Vision
• Information Retrieval
• Machine Learning

Background:

• Near-duplicate image retrieval is crucial for applications like image annotation and content-based retrieval.
• Web image searches, especially for celebrities and historical figures, frequently involve near-duplicate images.
• Current methods, such as bag-of-visual-words (BoVW), primarily rely on visual features, limiting their effectiveness.

Purpose of the Study:

• To propose a novel text-based, data-driven reranking framework for near-duplicate image retrieval.
• To enhance retrieval accuracy by integrating textual features with existing visual methods.
• To address the limitations of purely visual feature-based retrieval systems.

Main Methods:

• Developed a reranking framework that utilizes textual features alongside state-of-the-art bag-of-visual-words (BoVW) schemes.
• Constructed a large-scale dataset comprising 2 million images of 1089 celebrities with associated text.
• Analyzed various categories of near-duplication present in the constructed dataset.

Main Results:

• The proposed text-based framework achieved a 21% average improvement in mean average precision (mAP) compared to visual-feature-only methods.
• The framework successfully integrates textual information to refine near-duplicate image retrieval results.
• Retrieval time was not significantly increased by the addition of the text-based reranking component.

Conclusions:

• The novel text-based reranking framework offers a significant advancement in near-duplicate image retrieval accuracy.
• Integrating textual data provides a powerful complement to visual features for improving retrieval performance, particularly for celebrity images.
• The approach is efficient, offering substantial accuracy gains without compromising retrieval speed.