AutoLumNet: Monotone Optimal Transport for Single-Shot Exposure Correction

Technical Summary: AutoLumNet

Problem Statement
Single-shot exposure correction (SEC) aims to map a single degraded image—whether under-exposed, over-exposed, or a spatial mixture of both—to a well-exposed output. Existing approaches face significant structural limitations. Low-light image enhancement (LLIE) methods assume a monotone darkening degradation and apply monotone brightening; when applied to over-exposed inputs, they exacerbate saturation. Conversely, multi-exposure fusion (MEF) methods require multiple aligned captures, which are rarely available in dynamic or resource-constrained scenarios. Furthermore, most existing SEC methods are evaluated purely empirically, lacking formal guarantees that the enhanced image preserves the relative ordering of scene luminances. This violation of luminance ordering often manifests as halo artifacts, tonal reversals, and detail loss.

Methodology: AutoLumNet
The authors propose AutoLumNet, a framework that decomposes exposure correction into a global monotone tone curve and a bounded local residual. The architecture is built on three core principles:

1. Honest Degradation Model: The method distinguishes between mathematically recoverable regions (where the tone map is invertible) and clipped regions (where recovery relies on prior-based estimation rather than inversion).
2. Global Monotone Tone Map (Structural Guarantee):
   • The global luminance correction $T_\theta$ is parameterized as the normalized cumulative integral of a strictly positive density function $m_\theta$.
   • Specifically, $T_\theta(y) = \frac{\int_0^y m_\theta(t) dt}{\int_0^1 m_\theta(t) dt}$, where $m_\theta$ is generated via a neural network output passed through softplus with a positive floor $\epsilon$.
   • This construction ensures strict monotonicity by design, rather than enforcing it via a penalty term. Consequently, the map preserves the pairwise luminance ordering of all pixels and all spatial extrema unconditionally.
   • The tone curve is derived from a global image descriptor (global average pooling of the deepest encoder feature), ensuring exactly one curve per image to maintain spatial order preservation.
3. Optimal Transport Alignment:
   • The parameterized family of tone curves is proven to be dense in the space of valid tone corrections and contains the one-dimensional optimal-transport (OT) map from the input luminance distribution to a target well-exposed distribution.
   • Training utilizes a differentiable sorted-sample Wasserstein-2 objective ($L_{align}$) to drive the learned curve toward the OT optimum.
4. Bounded Local Residual:
   • To handle local shading, chrominance shifts, and clipped-region restoration (which the global map cannot address), a bounded spatial residual $r_\theta$ is added.
   • The residual is constrained by $|r_\theta| \leq \rho$ using a tanh activation, ensuring the correction remains within a defined budget.
   • The decoder employs a dual-branch convex fusion mechanism. It maintains separate sub-networks for shadow recovery ($U^i$) and highlight restoration ($O^i$), fusing them via a softmax gate. This ensures the fused features lie within the convex hull of the two branches, preventing extrapolation.

Key Contributions

• Formal Monotonicity: The paper introduces a "monotone-by-construction" parameterization that guarantees strict monotonicity and the preservation of global luminance ordering and spatial extrema without relying on soft penalties.
• Optimal Transport Connection: It proves that the proposed parameterized family is dense in the space of valid tone curves and contains the 1-D OT map, justifying the use of an OT-based alignment objective.
• Conditional Local Order Preservation: The authors provide an explicit sufficient condition for local order preservation when the bounded residual is applied, clarifying the limits of local adaptivity.
• Unified Architecture: AutoLumNet is the first method to unite structural monotonicity, optimal-transport optimality, and bounded local adaptivity within a single trainable architecture for single-shot correction.

Experimental Results
Experiments were conducted on five benchmarks: MSEC, SICE, LCDP, LOL-v1, and LOL-v2-real.

• Performance: AutoLumNet achieves state-of-the-art PSNR and SSIM across both under-exposure and over-exposure regimes on MSEC, SICE, and LCDP. For instance, on MSEC, it achieves 23.75 dB PSNR and 0.907 SSIM, outperforming prior SOTA methods like Exposure-Slot and CTAS.
• Efficiency: The model processes frames at 11.2 ms (approx. 8.39 ms in specific hardware tests mentioned in Table IV), significantly faster than many competing multi-exposure methods.
• Zero-Shot Generalization: A model trained exclusively on the multi-exposure MSEC dataset generalizes zero-shot to pure low-light benchmarks (LOL-v1 and LOL-v2-real) without retraining, achieving competitive results against methods specifically trained on those datasets.
• Qualitative Analysis: Visual comparisons demonstrate that AutoLumNet avoids the halo artifacts, color shifts, and detail loss common in other methods, particularly in mixed-exposure scenarios.

Significance
The paper claims that AutoLumNet represents a significant step forward by moving exposure correction from purely empirical optimization to a framework with formal theoretical guarantees. By ensuring that the global tone map is strictly monotone by construction, the method eliminates a primary source of failure in enhancement (tonal reversals and halos). The integration of optimal transport provides a principled objective for distribution matching, while the bounded residual ensures local adaptivity without sacrificing global structural integrity. The authors position this as the first method to successfully combine these three properties—structural monotonicity, OT optimality, and bounded local adaptivity—within a single architecture.