Honeycomb Artifact Removal Using Convolutional Neural Network for Fiber Bundle Imaging.

Area of Science:

• Optical imaging
• Deep learning
• Image processing

Background:

• Fiber bundle imaging can suffer from honeycomb artifacts due to optical path blocking.
• These artifacts degrade image quality and hinder accurate analysis.
• Existing methods for artifact removal are often insufficient or require complex setups.

Purpose of the Study:

• To introduce a novel deep learning framework, HAR-CNN, for efficient honeycomb artifact removal.
• To enable end-to-end artifact correction from raw fiber bundle images.
• To validate the framework's performance without extensive ground truth data collection.

Main Methods:

• A convolution neural network (CNN) forms the core of the HAR-CNN framework.
• Honeycomb patterns were synthesized onto ordinary images for network training and validation.
• The framework was tested using a 1961 USAF chart sample for performance evaluation.

Main Results:

• HAR-CNN demonstrated significant improvement in removing honeycomb artifacts.
• The framework effectively preserved fine details in the images, as shown with the USAF chart.
• Compared to conventional methods, HAR-CNN offered superior performance.

Conclusions:

• HAR-CNN provides an effective, end-to-end solution for honeycomb artifact removal in fiber bundle imaging.
• The synthesis-based training approach simplifies network validation and reduces hardware requirements.
• The GPU-accelerated framework supports real-time processing and enhances image mosaicking.