Multiview High Dynamic Range Image Synthesis Using Fuzzy Broad Learning System.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• High dynamic range (HDR) images offer superior detail compared to standard low dynamic range (LDR) images.
• Existing HDR synthesis methods struggle with dynamic scenes, producing artifacts.
• Current learning-based HDR techniques often lack interpretability and have lengthy training times.

Purpose of the Study:

• To propose a novel approach for synthesizing multiview HDR images using a fuzzy broad learning system (FBLS).
• To address the limitations of existing methods, specifically artifact generation in dynamic scenes and long training durations.
• To enhance the interpretability and efficiency of HDR image synthesis.

Main Methods:

• Utilized a set of multiview LDR images with varying exposures as input.
• Employed Takagi-Sugeno (TS) fuzzy subsystems and expanded structures in 'enhancement groups' with nonlinear transformations.
• Integrated fuzzy subsystems and enhancement groups using trained weights for HDR image generation.
• Applied incremental learning and pseudoinverse methods for efficient weight computation.

Main Results:

• Successfully synthesized high-quality multiview HDR images.
• Demonstrated significantly faster training speeds compared to other learning-based methods.
• The fuzzy system's IF-THEN rules effectively detected and rejected artifacts.
• Developed a new dataset of multiview LDR images with HDR ground truth for training.

Conclusions:

• The proposed FBLS method offers an effective and efficient solution for multiview HDR image synthesis.
• FBLS overcomes key challenges in HDR generation, including artifact reduction and training time.
• The approach provides better interpretability, making it a valuable alternative to complex deep learning models.