Global Structure-Aware Diffusion Process for Low-light Image Enhancement

Authors: Jinhui HOU, Zhiyu Zhu, Junhui Hou, Hui LIU, Huanqiang Zeng, Hui Yuan

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental evaluations reveal that the proposed diffusion-based framework, complemented by rank-informed regularization, attains distinguished performance in low-light enhancement. The outcomes indicate substantial advancements in image quality, noise suppression, and contrast amplification in comparison with state-of-the-art methods. We believe this innovative approach will stimulate further exploration and advancement in low-light image processing, with potential implications for other applications of diffusion models. The code is publicly available at https://github.com/jinnh/GSAD. 4 Experiments 4.1 Experiment Settings 4.2 Comparison with State-of-the-Art Methods 4.3 Ablation Study
Researcher Affiliation Academia Jinhui Hou1, Zhiyu Zhu1, Junhui Hou1 , Hui Liu2, Huanqiang Zeng3, and Hui Yuan4 1City University of Hong Kong, 2Caritas Institute of Higher Education 3Huaqiao University, 4Shandong University
Pseudocode Yes Algorithm 1 Forward process 1: Repeat 2: X0 q(X0) 3: t Uniform({1, 2, ..., T}) 4: ϵ N(0, I) 5: Take gradient descent step on 6: θ ϵ ϵθ( αt X0 + 1 αtϵ, t) 2 7: Until converged Algorithm 2 Reverse process 1: XT N(0, I) 2: For t = T, ..., 1 do 3: z N(0, I) if t > 1, else z = 0 4: Xt 1 = 1 αt (Xt 1 αt 1 αt ϵθ(Xt, t))+σtz 5: End for 6: Return X0 Algorithm 3 Training a diffusion-based low-light enhancement model Require: Normal and low-light image pairs (X0, Y), and a pre-trained models µθ1 and ϵˆθ. 1: Initialize parameters of µθ1( ) (resp. ϵθ( )) from pretrained ˆθ1 (resp. ˆθ), and freeze ˆθ1. 2: Repeat 3: Sample an image pair (X0, Y), αt p( α) and ϵ N(0, I) 4: Xt = αt X0 + 1 αt ϵ 5: Pt µθ1(Y, Xt, αt) 6: Construct the learnable closed-form sample Xt 1 via Eq. (2) 7: Perform a gradient descent step on 8: θ{λ Pt ϵ Pt ϵθ(Y, Xt, αt) 1 + Lt s }, where λ is a balancing factor, and Lt s refers to Eq. (4) 9: Until converged 10: Return the resulting diffusion model ϵ θ
Open Source Code Yes The code is publicly available at https://github.com/jinnh/GSAD.
Open Datasets Yes We employed seven commonly-used LLIE benchmark datasets for evaluation, including LOLv1 [45], LOLv2 [54], DICM [11], LIME [6], MEF [14], NPE [39], and VV2. Specifically, LOLv1 contains 485 low/normal-light image pairs for training and 15 pairs for testing, captured at various exposure times from the real scene. LOLv2 is split into two subsets: LOLv2-real and LOLv2-synthetic. LOLv2-real comprises 689 pairs of low-/normal-light images for training and 100 pairs for testing, collected by adjusting the exposure time and ISO. LOLv2-synthetic was generated by analyzing the illumination distribution of low-light images, consisting of 900 paired images for training and 100 pairs for testing.
Dataset Splits No The paper specifies training and testing splits for datasets but does not explicitly mention a dedicated validation dataset or split for reproducibility.
Hardware Specification No The paper does not provide specific details about the hardware used, such as GPU models, CPU specifications, or memory.
Software Dependencies No The paper mentions "Py Torch" and "Adam optimizer" but does not specify their version numbers or any other software dependencies with versions.
Experiment Setup Yes We trained our models for 2M iteration steps with Py Torch. We employed an Adam optimizer [10] with a fixed learning rate of 1 10 4 without weight decay. We applied an exponential moving average with a weight of 0.9999 during parameter updates. During training, we utilized a fine-grained diffusion process with T = 500 steps and a linear noise schedule with the two endpoints of 1 10 4 and 2 10 2. The patch size and batch size were set to 96 and 8, respectively. The hyper-parameter λ was empirically set to 10.