Fully Convolutional Network-Based Multifocus Image Fusion.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Optical lenses have limited depth of field, resulting in images with varying focus.
• Multifocus image fusion aims to create a single, all-in-focus image from multiple partially focused images.
• Existing fusion methods often struggle with defining effective fusion rules.

Purpose of the Study:

• To propose a novel multifocus image fusion method based on focus region detection using a fully convolutional network (FCN).
• To improve the accuracy and quality of fused images by learning pixel-wise focus regions.
• To overcome limitations of previous fusion methods by developing a data-driven approach.

Main Methods:

• Synthesized 4500 multifocus image pairs using Gaussian filtering on PASCAL VOC 2012 dataset to train the FCN.
• Developed an FCN to generate pixel-wise focus score maps from input images.
• Utilized a fully connected conditional random field (CRF) for refining decision maps and a weighted strategy for final image fusion.

Main Results:

• The proposed FCN-based method accurately detects pixel-wise focus regions.
• The fusion process effectively combines focus information from multiple images.
• Experimental results demonstrate superior fusion performance compared to state-of-the-art methods on both grayscale and color datasets.

Conclusions:

• The novel FCN-based multifocus image fusion method achieves superior performance.
• The approach offers improved human visual quality and objective assessment metrics.
• This learning-based strategy provides a robust solution for multifocus image fusion challenges.