MWDNs: reconstruction in multi-scale feature spaces for lensless imaging.

Area of Science:

• Computational imaging
• Optical systems engineering
• Machine learning for image reconstruction

Background:

• Lensless cameras offer miniaturization and flexibility for constrained applications.
• Current reconstruction methods often suffer from model mismatches, limiting image quality.
• Existing algorithms typically combine iterative physical deconvolution with deep learning perception.

Purpose of the Study:

• To develop an improved reconstruction algorithm for lensless camera systems.
• To address the limitations of model mismatch in current lensless imaging reconstruction.
• To enhance image fidelity and computational efficiency in mask-based imaging.

Main Methods:

• Introduced a novel Multi-channel Wiener Deconvolution Network (MWDN).
• MWDN operates in multi-scale feature spaces, employing Wiener filters for deconvolution.
• The network corrects input data to improve model accuracy, reducing information loss.

Main Results:

• The proposed MWDN significantly outperforms state-of-the-art reconstruction algorithms.
• Achieved superior image quality in both simulated and real-world lensless imaging scenarios.
• Demonstrated improved computational efficiency by eliminating iterative processes.

Conclusions:

• MWDN offers a robust and efficient solution for lensless camera image reconstruction.
• The method effectively mitigates model mismatch issues inherent in physical imaging models.
• This approach advances the practical application of lensless imaging technologies.