Abstract
Reconstructing high-quality RAW images from standard sRGB files is a critical task for preserving the linear radiance and high dynamic range necessary for advanced computer vision applications. This paper presents a novel approach using Implicit Neural Functions (INF) to bridge the gap between compressed sRGB outputs and their original RAW counterparts. By embedding a minimal amount of metadata (1.5% uniform sampling) and utilizing the SIREN architecture, the proposed method achieves a significant performance leap, outperforming existing state-of-the-art techniques by over 10dB PSNR without requiring hardware-specific pre-training.
Introduction
The Image Signal Processor (ISP) in modern cameras performs a series of non-linear operations—including tone mapping, gamma correction, and white balancing—to convert RAW sensor data into the sRGB format. While sRGB is efficient for storage and display, the ISP process irreversibly discards scene information. RAW reconstruction aims to reverse these transformations. Traditional methods often struggle with the complexity of ISP mappings and fail to generalize across different camera hardware, often leaving significant artifacts in saturated regions where pixels have been "clipped."
Methodology: Implicit Neural Functions
The core of this research is the transition from traditional mapping functions to coordinate-based Implicit Neural Functions. The authors employ a Multi-Layer Perceptron (MLP) to represent the RAW image as a continuous function of 2D coordinates $(x, y)$, conditioned on local sRGB values.
A pivotal component of this architecture is the use of SIREN (Sinusoidal Representation Networks). Unlike standard ReLU-based networks, which tend to behave like smooth interpolators and lose high-frequency detail, SIREN uses periodic sine activation functions. These are uniquely suited for modeling high-frequency textures and the intricate non-linearities of the ISP curve.
The network is structured into a two-branch architecture: a spatial branch for coordinate processing and a color branch for sRGB guidance. This decoupling allows for independent regularization, optimizing the network's ability to prioritize color information—which is more highly correlated with the target RAW values—while maintaining spatial consistency. The model is trained on-the-fly for each image using a tiny amount of embedded metadata (1.5% uniform sampling) that acts as "anchors" for the interpolation process.
Experimental Evaluation
The framework was rigorously tested on the NUS dataset, which includes images from Samsung NX2000, Olympus E-PL6, and Sony SLT-A57 cameras. This diversity ensures the model's effectiveness across various ISP implementations. Additionally, the framework's versatility was demonstrated on the Middlebury dataset for guided depth super-resolution tasks.
Results and Discussion
Quantitative analysis reveals that the proposed INF method achieves a PSNR of 61.48 dB on the Samsung dataset, a substantial improvement over the 49.51 dB achieved by the previous state-of-the-art (CAM). As illustrated in the qualitative comparison above, the INF method significantly reduces "ghosting" and reconstruction errors. This is particularly evident in bright, saturated areas like the sky, where other methods typically fail to recover detail.
The success of this approach stems from treating the problem as a signal reconstruction task rather than a simple mathematical mapping. By optimizing the network on-the-fly using sampled RAW pixels, the INF effectively learns a custom "reverse ISP" for every image, bypassing the need for extensive pre-training on specific camera sensors.
Conclusion and Future Work
This paper demonstrates that Implicit Neural Functions, specifically when paired with periodic activations like SIREN, offer a powerful new paradigm for RAW image reconstruction. By leveraging minimal metadata, the proposed framework achieves unprecedented accuracy and visual quality. While the current per-image optimization takes approximately two hours on high-end hardware, future research will focus on optimizing computational efficiency for real-time applications and exploring adaptive sampling strategies to further enhance reconstruction in complex image regions.
Summary updated