Abstract
The illumination of improperly exposed photographs has been widely corrected using deep convolutional neural networks or Transformers. Despite with promising performance, these methods usually suffer from large parameter amounts and heavy computational FLOPs on high-resolution photographs. In this paper, we propose extremely light-weight (with only ~8K parameters) Multi-Scale Linear Transformation (MSLT) networks under the multi-layer perception architecture, which can process 4K-resolution sRGB images at 125 Frame-Per-Second (FPS) by a Titan RTX GPU. Specifically, the proposed MSLT networks first decompose an input image into high and low frequency layers by Laplacian pyramid techniques, and then sequentially correct different layers by pixel-adaptive linear transformation, which is implemented by efficient bilateral grid learning or 1x1 convolutions. Experiments on two benchmark datasets demonstrate the efficiency of our MSLTs against the state-of-the-arts on photo exposure correction. Extensive ablation studies validate the effectiveness of our contributions.
Introduction
Capturing perfectly exposed photographs is a persistent challenge in digital photography. Inaccurate camera settings often result in images that are either under-exposed (too dark) or over-exposed (too bright). While deep learning has significantly advanced automated exposure correction, state-of-the-art methods based on Convolutional Neural Networks (CNNs) or Transformers typically require millions of parameters and substantial computational resources (FLOPs). This high complexity makes them impractical for processing 4K-resolution content in real-time or for deployment on resource-constrained edge devices like smartphones.
To address these limitations, we introduce the Multi-Scale Linear Transformation (MSLT) network. MSLT is an ultra-lightweight model (~8K parameters) designed to maintain high visual quality while achieving unprecedented processing speeds, reaching 125 FPS for 4K resolution content.
Methodology
The MSLT framework adopts a "divide and conquer" strategy based on image frequency components, allowing for targeted and efficient correction.
Multi-Scale Decomposition
The network utilizes a Laplacian pyramid to decompose an input image into a single low-frequency layer, which captures global illumination, and several high-frequency layers that contain fine spatial details. This separation is key to efficiency: global exposure issues primarily reside in low frequencies, while details are concentrated in high frequencies.
Low-Frequency Correction via Bilateral Grids
Exposure correction is fundamentally a global illumination problem. MSLT focuses its primary computational effort on the low-frequency layer using a Bilateral Grid Network (BGN). The BGN learns a 3D grid of affine transformation coefficients, enabling pixel-adaptive color and brightness correction that remains smooth and edge-preserving. This approach avoids the heavy cost of large-kernel spatial convolutions.
Architectural Components
Beyond frequency separation, the network incorporates several novel modules:
- Context-aware Feature Decomposition (CFD): A parameter-free module that extracts global context (channel-wise mean and standard deviation) to guide the correction process without adding learnable parameters.
- Hierarchical Feature Decomposition (HFD): A cascaded structure that utilizes context features to estimate the bilateral grid.
- High-Frequency Correction: Fine details are corrected using extremely simple 1x1 convolutions that predict pixel-wise masks, ensuring detail preservation with minimal overhead.
The variant MSLT++ further optimizes performance by using learnable 3x3 convolutions for decomposition and skipping the correction of the largest high-frequency layer to maximize throughput.
Experiments and Results
The proposed method was evaluated on two primary benchmarks: the Multi-Exposure (ME) dataset (derived from MIT-Adobe FiveK) and the Single-Image Contrast Enhancement (SICE) dataset.
Quantitative Performance
MSLT++ demonstrates a dramatic improvement in the performance-efficiency trade-off. As shown in Figure 1, it occupies the optimal top-left region of the comparison charts, indicating high quality at low computational cost.
- Efficiency: MSLT++ contains only ~8,098 parameters, orders of magnitude fewer than competing models like MSEC (7M+ parameters). It processes 4K images at 125 FPS on a Titan RTX GPU.
- Quality: Despite its minimal size, MSLT++ achieves superior PSNR (Peak Signal-to-Noise Ratio). On the ME dataset, it reached a PSNR of 21.12 dB, outperforming larger models such as IAT (20.17 dB) and MSEC (19.38 dB).
Qualitative Analysis
Visual results indicate that MSLT effectively recovers details in over-exposed regions (e.g., clouds) and enhances textures in deep shadows without the "over-processed" artifacts common in traditional HDR methods. Its pixel-adaptive nature allows it to simultaneously brighten dark foregrounds and dim bright backgrounds within the same frame.
Conclusion
The MSLT network provides a highly efficient solution for high-resolution photo exposure correction. By leveraging frequency separation and bilateral grid learning, it achieves state-of-the-art quality with a fraction of the parameters and computational cost of previous methods. Future research will focus on enhancing robustness in extreme noise conditions and ensuring temporal consistency for real-time video applications.
Summary updated