TL;DR
This research presents a Fuzzy Broad Learning System (FBLS) for synthesizing High Dynamic Range (HDR) images from multiview Low Dynamic Range (LDR) inputs. The approach achieves state-of-the-art visual quality while drastically reducing training time from days to minutes and offering superior interpretability compared to traditional deep learning models.
Introduction: The Challenge of Multiview HDR
HDR imaging typically involves merging multiple LDR photos with varying exposures. In static scenes, this is straightforward. However, in dynamic scenarios—where objects move or the camera viewpoint shifts (multiview)—traditional methods often produce "ghosting" artifacts, which appear as blurry or translucent overlaps.
While Convolutional Neural Networks (CNNs) have improved deghosting, they remain "black boxes" that are computationally expensive to train. This paper addresses these limitations by proposing a system that is both fast to train and logically interpretable through fuzzy logic.
Methodology: The Fuzzy Broad Learning System
The proposed pipeline moves away from deep architectures in favor of a "broad" approach that expands horizontally rather than vertically.
- Alignment: Three LDR images (low, medium, and high exposure) are pre-aligned using optical flow, with the medium exposure acting as the reference.
- FBLS Synthesis:
- Fuzzy Subsystems: Aligned images pass through Takagi-Sugeno (TS) fuzzy subsystems. These use "IF-THEN" rules to extract features, allowing the model to logically identify and reject artifacts.
- Enhancement Groups: These groups apply nonlinear transformations to preserve input characteristics as the network expands broad-wise.
- Pseudoinverse Weight Calculation: Instead of time-consuming backpropagation, FBLS uses ridge regression (pseudoinverse) to calculate weights. This direct mathematical approach is significantly faster than iterative training.
- Incremental Learning: The system can dynamically add fuzzy subsystems or enhancement nodes if the output quality (PSNR) is insufficient, updating only the new weights without a full retrain.
Datasets and Evaluation
To validate the model, the authors developed a new Multiview HDR Dataset consisting of 100 training and 100 test scenes across various environments (indoor, outdoor, sunny, and cloudy).
The model was also benchmarked against standard datasets like Kalantari’s and Gallo’s to ensure robustness across architectural scenes and high-motion subjects.
Key Results
The FBLS method demonstrates a clear advantage in both efficiency and fidelity:
- Performance: It achieved a PSNR-T (tonemapped) of 43.24, surpassing the 42.72 achieved by leading CNN-based methods.
- Training Speed: Training was completed in approximately 10 minutes, a massive improvement over the 3 days required by traditional deep learning models on the same hardware.
- Visual Quality: Qualitative comparisons show the system excels at recovering details in saturated areas (like bright skies) and deep shadows without the blur typical of patch-based methods.
Conclusion
By replacing deep "black-box" networks with a Fuzzy Broad Learning System, the authors have made multiview HDR synthesis significantly more practical for real-world applications. While the current model is limited to three LDR inputs and has high memory requirements for matrix operations, its speed and interpretability mark a major step forward for the field.
Summary updated