Deep constrained multi-frame blind deconvolution with learned regularizations

Area of Science:

• Optics and Photonics
• Computer Vision
• Astrophysics

Background:

• Multi-frame blind deconvolution (MFBD) traditionally corrects atmospheric turbulence in images.
• Existing MFBD methods rely on handcrafted regularizations, leading to instability and slow processing.

Purpose of the Study:

• To develop an efficient and effective end-to-end trainable network for space target image restoration.
• To overcome limitations of traditional MFBD techniques in terms of stability and inference time.

Main Methods:

• An end-to-end network performing alternating estimations on deep image features using a constrained least-squares filter.
• Adaptive learning of smoothing filters directly from image features.
• Two specialized convolutional neural network modules for point spread function (PSF) estimation and deconvolution, cycled iteratively.
• Training on a large-scale dataset simulating real turbulence scenarios.

Main Results:

• The proposed network demonstrates robust and effective performance in high-resolution space target image restoration.
• Achieved significant improvements in image clarity and stability compared to traditional methods.
• Validated through extensive experiments under simulated turbulence conditions.

Conclusions:

• The developed deep learning approach offers a more efficient and stable solution for space target image restoration.
• This method holds promise for applications requiring high-quality imaging in turbulent environments.
• The adaptive feature learning and iterative deconvolution process are key to its success.