Abstract
Integral Fast Fourier Color Constancy (IFFCC) is a high-efficiency algorithm designed for multi-illuminant white balance. By introducing an "integral histogram" technique, the method enables the extraction of local color distributions in constant time. This allows IFFCC to achieve accuracy on par with heavy deep-learning models while being 400x smaller (12k parameters) and up to 100x faster (5.8ms on CPU). These characteristics make it exceptionally well-suited for real-time deployment on mobile hardware.
Introduction
Most modern cameras rely on Auto White Balance (AWB) to ensure that colors appear natural regardless of the scene's lighting. However, traditional AWB algorithms typically assume a single, global light source. Real-world environments are frequently "multi-illuminant," featuring complex combinations of light, such as a room illuminated by both warm indoor lamps and cool natural light from a window.
While recent deep learning architectures can address these complex scenes by predicting pixel-level illumination maps, they are often computationally prohibitive for mobile devices. These models typically require powerful GPUs and possess millions of parameters, making them unsuitable for real-time video or mobile preview screens. IFFCC bridges this gap by combining the mathematical rigor of classical statistical methods with the spatial awareness of modern neural networks.
Methodology
IFFCC is an evolution of the Fast Fourier Color Constancy (FFCC) algorithm. FFCC treats color constancy as a localization problem in log-chromaticity space, solved efficiently via the Fast Fourier Transform (FFT). IFFCC extends this framework to handle spatial variation through several key innovations.
Integral UV Histogram
The defining innovation of IFFCC is the Integral UV Histogram. Drawing inspiration from "integral images" (summed-area tables), the authors precompute a 3D structure representing the image's color distribution. This structure allows the algorithm to extract a color histogram for any rectangular sub-region of the image in constant time ($O(1)$), independent of the region's size. This efficiency is what enables the analysis of hundreds of local windows without a significant performance penalty.
Parallelized Prediction and Smoothing
Instead of a single global estimate, IFFCC utilizes an overlapping grid to calculate local illuminants across the frame. These calculations are performed simultaneously using parallelized Fourier-based convolutions. To prevent the final output from appearing "patchy," the local estimates are refined through a spatial smoothing process. This involves linear interpolation and Guided Filtering, which ensures smooth transitions between lighting zones while maintaining sharp edges at physical boundaries.
Experimental Results
The authors evaluated IFFCC on two major multi-illuminant benchmarks: the LSMI Dataset (7,400+ images) and the Shadow Dataset (2,500 images focused on indoor/outdoor transitions).
Performance and Efficiency
The results highlight a significant leap in efficiency. IFFCC utilizes only 12,000 parameters, a reduction of over 400x compared to leading pixel-level neural networks like AID or Pix2Pix. Despite this compactness, its accuracy remains competitive. On the Shadow dataset, IFFCC achieved a mean angular error of 2.19, outperforming several significantly larger models.
In terms of speed, IFFCC can process a 64x48 preview image in just 5.8ms on a standard CPU. This 20–100x speedup over network-based approaches is largely due to the pre-processing efficiency of the integral histogram and the use of optimized FFT operations common in mobile Image Signal Processors (ISPs).
Conclusion
IFFCC demonstrates that mathematically optimized classical frameworks can still outperform "black box" neural networks in specific computer vision tasks. By prioritizing efficiency through the integral histogram and parallelized Fourier transforms, the authors have created a solution that is both highly accurate and hardware-friendly. While future work may further refine adaptive window sizing to handle extremely sharp lighting transitions, IFFCC currently stands as a state-of-the-art solution for real-time, multi-illuminant color constancy on mobile devices.
Summary updated