Tattoo Image Search at Scale: Joint Detection and Compact Representation Learning.

Area of Science:

• Computer Vision
• Biometrics
• Forensic Science

Background:

• The increasing volume of digital images necessitates advanced person identification methods beyond traditional biometrics.
• Tattoos serve as valuable soft biometric traits for enhancing person identification in forensic applications.
• Existing tattoo search methods often fail in real-world scenarios due to limitations in matching cropped images.

Purpose of the Study:

• To develop an efficient tattoo search system capable of handling large, unconstrained image collections.
• To integrate tattoo detection and compact representation learning within a unified framework.
• To address the challenges of real-world tattoo search applications.

Main Methods:

• A multi-task learning approach using a single convolutional neural network (CNN) for joint tattoo detection and compact representation.
• Shared backbone network features optimized for both detection and representation learning tasks.
• Techniques including random image stitching and feature buffering to overcome small batch size limitations.

Main Results:

• The proposed system demonstrates superior performance in both tattoo detection and large-scale tattoo search compared to existing methods.
• Effective evaluation on public tattoo benchmarks and a large gallery set of 300K images.
• Successful introduction of a tattoo sketch dataset for sketch-based tattoo search.

Conclusions:

• The joint learning approach offers an efficient and accurate solution for tattoo search in forensic and surveillance applications.
• The method effectively bridges the gap between current tattoo search techniques and real-world requirements.
• The developed system shows significant potential for improving person identification accuracy in large-scale image retrieval.