Learning Hierarchically Consistent Disentanglement with Multi-Channel Augmentation for Public Security-Oriented...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Sketch re-identification (Re-ID) is vital for public security tasks like criminal investigations and missing person searches.
• A significant challenge in Re-ID is the modality gap between sketch images and real photographs, especially due to color information differences.
• Existing methods struggle to extract robust, modality-invariant features essential for accurate cross-modal matching.

Purpose of the Study:

• To propose a novel network architecture for sketch re-identification (Re-ID) that overcomes the modality gap.
• To mitigate the interference of color bias in cross-modal matching between sketches and photographs.
• To develop a method for decomposing pedestrian representations into modality-invariant and modality-specific features.

Main Methods:

• A novel network architecture integrating multi-channel augmentation and hierarchically consistent disentanglement learning.
• A multi-channel augmentation module to reduce color bias interference.
• A modality-disentangled prototype (MDP) module for feature-level decomposition and a cross-layer decoupling consistency constraint.

Main Results:

• The proposed approach significantly improves performance on sketch re-identification tasks.
• Experimental results on two public datasets demonstrate the superiority over state-of-the-art methods.
• The network effectively bridges the modality gap, extracting discriminative modality-invariant features.

Conclusions:

• The novel network architecture effectively addresses the challenges of sketch re-identification.
• The integration of multi-channel augmentation and disentanglement learning enhances cross-modal matching accuracy.
• The proposed method offers a promising solution for real-world applications in public security.