A residual-based deep learning approach for ghost imaging.

Area of Science:

• Quantum Imaging
• Computational Imaging
• Deep Learning Applications

Background:

• Ghost imaging using deep learning (GIDL) aims to improve imaging efficiency.
• Existing GIDL methods suffer from poor generalization due to using identical training and testing patterns, limiting detail reconstruction.
• This leads to GIDL only reconstructing object profiles, losing crucial image details.

Purpose of the Study:

• To enhance the generalization ability and imaging quality of GIDL.
• To optimize the simulation algorithm for ghost imaging (GI) by introducing batch processing.
• To develop an advanced deep learning framework for improved GI.

Main Methods:

• Introduced "batch" concept into the pre-processing stage of GI simulation to reduce data acquisition time and generate reliable simulation data.
• Developed a novel residual-based framework for GI, termed the double residual U-Net (DRU-Net).
• Enhanced GIDL by improving simulation data reliability and network generalization.

Main Results:

• Significantly reduced data acquisition time and created reliable simulation data through batch processing.
• Appreciably enhanced the generalization ability of GIDL.
• Achieved a threefold improvement in GI imaging quality, as evaluated by the structural similarity index, using the proposed DRU-Net.

Conclusions:

• The optimized simulation algorithm and DRU-Net framework significantly improve GIDL performance.
• The enhanced GIDL approach overcomes previous limitations in generalization and detail reconstruction.
• This work presents a substantial advancement in computational quantum imaging quality and efficiency.