Image restoration in frequency space using complex-valued CNNs.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Real-valued convolutional neural networks (RV-CNNs) excel in spatial domain image restoration but struggle with full frequency spectrum processing.
• Limitations in spectral information processing by RV-CNNs can lead to loss of textural and structural details.

Purpose of the Study:

• To explore complex-valued convolutional neural networks (CV-CNNs) for frequency domain image restoration.
• To address the limitations of RV-CNNs in preserving spectral information for tasks like denoising and super-resolution.

Main Methods:

• Proposed novel CV-CNN models incorporating complex-valued attention gates for frequency domain image denoising and super-resolution.
• Evaluated models on structured illumination microscopy (SR-SIM) and conventional image datasets.

Main Results:

• CV-CNN models demonstrated superior performance compared to their RV-CNN counterparts in denoising and super-resolution tasks.
• Experimental results confirmed that CV-CNNs better preserve the frequency spectrum during denoising compared to RV-CNNs.

Conclusions:

• CV-CNN-based methods offer a promising deep learning approach for frequency domain image restoration.
• The proposed CV-CNN models effectively address spectral information limitations, improving image restoration quality.