Fast Multi-Scale Structural Patch Decomposition for Multi-Exposure Image Fusion

Area of Science:

• Computer Vision
• Image Processing
• Computational Photography

Background:

• High dynamic range (HDR) imaging is essential for capturing scenes with extreme lighting variations.
• Existing multi-exposure fusion (MEF) methods, like structural patch decomposition for MEF (SPD-MEF), face challenges with ghosting and halo artifacts, and computational cost.
• The relationship between SPD-MEF and other MEF techniques remains unclear.

Purpose of the Study:

• To develop a more efficient and effective multi-exposure fusion (MEF) method.
• To address the limitations of existing SPD-MEF techniques, specifically halo artifacts and computational expense.
• To clarify the connection between SPD-MEF and other MEF approaches.

Main Methods:

• Developed an unnormalized version of SPD-MEF by omitting explicit structural patch decomposition, achieving a 30x speed-up.
• Introduced a fast multi-scale SPD-MEF algorithm to mitigate halo artifacts.
• Related the proposed method to pixel-level MEF and standard two-layer decomposition techniques.

Main Results:

• The unnormalized SPD-MEF variant demonstrated a significant speed improvement (30x) over the original SPD-MEF.
• The fast multi-scale SPD-MEF effectively reduced visible halo artifacts.
• Experimental results confirmed the proposed method's superior performance in both speed and image quality.

Conclusions:

• The proposed unnormalized and multi-scale SPD-MEF methods offer a substantial advancement in HDR imaging.
• The new techniques provide a computationally efficient and high-quality solution for multi-exposure fusion.
• This work clarifies theoretical connections and offers practical improvements for MEF.