Optical correlation-based cross-domain image retrieval system

Area of Science:

• Optics and Photonics
• Computer Vision
• Artificial Intelligence

Background:

• Cross-domain image retrieval is challenging with conventional methods.
• Optical correlators offer high-speed processing capabilities.
• Integrating deep learning with optical systems requires novel data conversion techniques.

Purpose of the Study:

• To present a novel cross-domain image retrieval system utilizing a high-speed optical correlator.
• To introduce a new conversion module enabling data transformation for optical correlators.
• To demonstrate a sketch-based image retrieval application using the developed system.

Main Methods:

• A coaxial holographic optical correlator was employed.
• A novel conversion module using an autoencoder was designed to transform diverse data into uniform optical intensity pagedata.
• Existing deep learning models were integrated as feature extractors via the conversion module.

Main Results:

• The conversion module successfully processed various data types into uniform optical intensity pagedata.
• The system enabled the utilization of deep learning models for feature extraction within the optical correlator.
• A sketch-based cross-domain image retrieval system was experimentally demonstrated, successfully discovering similar images from sketch queries.

Conclusions:

• The proposed optical correlation-based system offers a viable solution for cross-domain image retrieval.
• The novel conversion module enhances the versatility of optical correlators by integrating deep learning capabilities.
• This research expands the potential applications of optical correlators in areas like image retrieval and content-based searching.