LCR-Net: a progressive soft correction network for controlled low-light image enhancement
Abstract
Low-light image enhancement (LLIE) requires a practical balance among visibility recovery, reconstruction fidelity, chromatic stability, exposure safety, and computational efficiency. Aggressive enhancement can reveal dark content but may introduce highlight saturation and color distortion, whereas conservative enhancement may leave substantial residual darkness. To address this trade-off, this paper presents LCR-Net, a lightweight controlled LLIE framework based on functionally decoupled progressive correction. LCR-Net consists of a stable base enhancement backbone, a lightweight global–local brightness calibration module, and a deterministic luminance-only refinement stage with fixed soft-gating parameters. The three stages respectively establish a stable enhancement basis, compensate for residual brightness inconsistency, and selectively refine dark and highlight regions while bypassing chrominance components. Experiments on three paired benchmarks and three no-reference datasets evaluate reconstruction fidelity, perceptual quality, exposure behavior, computational efficiency, post-processing alternatives, and Stage III parameter sensitivity. Under the final unified native-resolution evaluation protocol, LCR-Net achieves an equal-weight average PSNR of 18.05 dB across LOL-v1, LOL-v2 Real, and LOL-v2 Synthetic. Although URetinex-Net and PairLIE obtain higher reconstruction fidelity, LCR-Net achieves the lowest mean underexposure ratio of 0.0000 among the compared methods while maintaining a mean overexposure ratio of 0.0265. LCR-Net also reaches 154.22 FPS at 256 × 256 resolution. Parameter-sensitivity experiments further show that moderate variations in the Stage III thresholds, gate slopes, and correction strengths do not cause abrupt performance degradation. These results position LCR-Net as a lightweight controlled-enhancement alternative that emphasizes residual-darkness reduction, restrained exposure behavior, and computational efficiency rather than unrestricted reconstruction accuracy.