LMT-GP: Combined Latent Mean-Teacher and Gaussian Process for Semi-supervised Low-light Image Enhancement

TL;DR

This paper tackles the challenge of semi-supervised low-light image enhancement (LLIE) under domain shift and limited labeled data. It introduces LMT-GP, a framework that combines a latent mean-teacher network with Gaussian-process regression to jointly leverage labeled and unlabeled data through latent representations, guided by an assisted GPR loss and a pseudo-label adaptation module (PAM). By modeling latent vectors with GP priors and employing PCA-based basis selection, the method mitigates domain shift and improves both perceptual quality and downstream task performance. Empirical results across LOL, VE-LOL, Darkface, and ACDC demonstrate superior LLIE quality and generalization, with notable gains in detection and segmentation tasks. The approach offers a principled, generalizable path for semi-supervised LLIE and related low-level vision problems.

Abstract

While recent low-light image enhancement (LLIE) methods have made significant advancements, they still face challenges in terms of low visual quality and weak generalization ability when applied to complex scenarios. To address these issues, we propose a semi-supervised method based on latent mean-teacher and Gaussian process, named LMT-GP. We first design a latent mean-teacher framework that integrates both labeled and unlabeled data, as well as their latent vectors, into model training. Meanwhile, we use a mean-teacher-assisted Gaussian process learning strategy to establish a connection between the latent and pseudo-latent vectors obtained from the labeled and unlabeled data. To guide the learning process, we utilize an assisted Gaussian process regression (GPR) loss function. Furthermore, we design a pseudo-label adaptation module (PAM) to ensure the reliability of the network learning. To demonstrate our method's generalization ability and effectiveness, we apply it to multiple LLIE datasets and high-level vision tasks. Experiment results demonstrate that our method achieves high generalization performance and image quality. The code is available at https://github.com/HFUT-CV/LMT-GP.

Figures (8)

• Figure 1: (a) A comparison on the real-world night images shows that existing LLIE methods have issues such as under/over-enhanced predictions and noisy ouputs. (b) Performance of the compared methods on three tasks: enhancement (measured by PSNR, SSIM, and LPIPS), detection (measured by mAP), and segmentation (measured by mIoU).
• Figure 2: (a) Basic mean-teacher framework. (b) Latent mean-teacher framework. Compared with basic mean-teacher framework, there are three differences: 1) the labeled data needs to be input into both the teacher and student networks; 2) the latent vectors and pseudo-latent vectors are used for network learning; 3) and the mean-teacher-assisted GP learning is performed on latent vectors and pseudo-latent vectors.
• Figure 3: Overview of LMT-GP. LMT-GP is based on the latent mean-teacher framework. Labeled and unlabeled data are simultaneously input into the student network and the teacher network. The student network generates latent vectors $z_l^{stu}$ and $z_u^{stu}$ for labeled and unlabeled data, respectively. The teacher network generates pseudo-latent vectors $z_l^{pseudo}$ and $z_u^{pseudo}$ for labeled and unlabeled data, respectively. After being processed by PAM, we perform mean-teacher-assisted GP learning on the latent vectors, pseudo-latent vectors and the generated vector sequence $F_l$.
• Figure 4: Visual comparison of state-of-the-art LLIE methods on the LOL dataset. More results can be found in the Supplementary Materials.
• Figure 5: Visual comparison of state-of-the-art LLIE methods on the VE-LOL dataset. More results can be found in the Supplementary Materials.