Attribute-Aware Deep Hashing With Self-Consistency for Large-Scale Fine-Grained Image Retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Large-scale fine-grained image retrieval faces challenges due to subtle inter-class variations and vast datasets.
• Existing methods struggle with efficiency and establishing clear links between hash codes and visual attributes.

Purpose of the Study:

• To develop an efficient and accurate method for large-scale fine-grained image retrieval.
• To generate attribute-aware hash codes that correspond to visual attributes without requiring manual annotations.

Main Methods:

• Proposes attribute-aware hashing networks with self-consistency.
• Utilizes an encoder-decoder structure for unsupervised distillation of attribute-specific vectors via attention.
• Incorporates feature decorrelation and an image reconstruction path to enhance self-consistency and combat simplicity bias.

Main Results:

• Achieved superior performance over competing methods on six fine-grained and two generic retrieval datasets.
• Demonstrated that generated hash codes strongly correlate with crucial fine-grained object properties.
• Showcased the effectiveness of self-consistency designs in overcoming simplicity bias.

Conclusions:

• The proposed attribute-aware hashing networks offer an efficient and effective solution for fine-grained image retrieval.
• The method successfully generates meaningful, attribute-aware hash codes, improving retrieval accuracy and interpretability.
• Self-consistency principles are crucial for enhancing deep hashing models in fine-grained retrieval tasks.