Double enhanced residual network for biological image denoising

Area of Science:

• Deep learning
• Image processing
• Computational biology

Background:

• Deep convolutional neural networks (CNNs) are widely used for image denoising.
• Existing CNN methods face limitations in GPU memory, restricting network depth and complexity.
• Biological imaging often requires high-fidelity denoising to preserve subtle details.

Purpose of the Study:

• To propose an efficient deep learning model for biological image denoising.
• To address GPU memory constraints in deep neural networks for image restoration.
• To enhance the extraction of deep features and restoration of image details.

Main Methods:

• Development of the Double Enhanced Residual Network (DERNet), a multi-level network architecture.
• Utilizing a U-Net inspired, two-sub-network structure with encoder-decoder hierarchies.
• Incorporating residual blocks for feature extraction and a novel Residual Dense Instance Normalization (RDIN) block for detail restoration.

Main Results:

• DERNet demonstrates efficient utilization of GPU memory for complex network structures.
• The proposed RDIN block effectively extracts deep features and restores image details.
• Qualitative experiments and visual assessments confirm the algorithm's effectiveness in biological image denoising.

Conclusions:

• DERNet offers an effective solution for biological image denoising, overcoming GPU memory limitations.
• The novel architecture and RDIN block contribute to superior feature extraction and image restoration.
• This method holds promise for improving the quality and interpretability of biological imaging data.