Deep Multiple Instance Hashing for Fast Multi-Object Image Search.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Multi-keyword search is common in text retrieval, but multi-object queries are underdeveloped in image retrieval.
• Existing object-based image retrieval methods are often multi-step and lack seamless integration.

Purpose of the Study:

• To propose a weakly-supervised Deep Multiple Instance Hashing (DMIH) approach for efficient multi-object image retrieval.
• To integrate object detection and hashing learning into a unified framework.

Main Methods:

• Developed a Deep Multiple Instance Hashing (DMIH) approach using a CNN model for end-to-end image-to-object hash code learning.
• Treated object detection as a binary multiple instance learning (MIL) problem, extracting instances from multi-scale convolutional features.
• Incorporated a conditional random field (CRF) module to model semantic and spatial relationships between class labels.
• Employed a multi-task learning scheme for hashing training, sampling image pairs to learn semantic relationships via object predictors.
• Introduced a two-level inverted index for accelerated multi-object query retrieval.

Main Results:

• The DMIH approach effectively supports multi-object queries by integrating object detection and hashing.
• Demonstrated superior performance compared to state-of-the-art methods on public benchmarks for object-based image retrieval.
• Achieved promising results specifically for the challenging task of multi-object queries.

Conclusions:

• The proposed DMIH method offers a unified and effective solution for multi-object image retrieval.
• Weakly-supervised learning and integrated object detection significantly advance the field.
• The approach shows strong potential for real-world applications requiring complex image search capabilities.