Abstract:The existing diffusion models (DMs) have shown impressive performance in face super-resolution tasks under normal illumination conditions. However, when applied to low-resolution (LR) facial images captured under non-uniform illumination (NI) conditions, the performance of DMs significantly deteriorates due to the lack of sufficient illumination information constraints. To address this challenge, we present ILSR-Diff, a novel illumination-constrained DM, designed to restore authentic high-resolution (HR) facial images while compensating for NI conditions. ILSR-Diff comprises three key components: (1) The illumination conditional constraint module extracts global and spatial illumination constraints from conditional facial images and seamlessly embeds them into the denoising network, which effectively addresses facial distortion caused by complex facial illumination variations in LR faces. (2) The noise guider harnesses facial priors by blending them with Gaussian noise to guide the initial noise sampling, which provides superior sampling initialization and prior knowledge for the denoising process to enhance the efficiency and precision of reconstruction. (3) The contrastive diffusion loss function utilizes high-level semantic features of high-quality faces to supervise the training of the denoising network, thereby enhancing the recovery details of facial features. Extensive experiments demonstrate that ILSR-Diff achieves authentic HR facial images under NI conditions and outperforms other state-of-the-art methods both qualitatively and quantitatively. Compared with SR3, our method achieves improvements of 2.2 dB and 2.8 dB in PSNR on the Multi-PIE and Extended-YaleB datasets, respectively.

Non-ideal Class Non-Point Light Source Quotient Image for Face Relighting

An Accurate Algorithm to Estiamte Illumination Parameter for Facial Relighting Synthesis

Adaptive Illumination Normalization Via Adaptive Illumination Preprocessing and Modified Weber-Face

Illumination Normalization For Robust Face Recognition Against Varying Lighting Conditions

Illumination transition image: Parameter-based illumination estimation and re-rendering

No-Reference Image Quality Assessment Using Independent Component Analysis and Convolutional Neural Network

Face Recognition Under Generic Illumination Based On Harmonic Relighting

An efficient illumination normalization method for face recognition

Nonsubsampled Contourlet Transform-Based Algorithm for No-Reference Image Quality Assessment

Towards High Fidelity Face Relighting with Realistic Shadows

Lighting Coefficients Transfer Based Face Illumination Normalization.

Designing An Illumination-Aware Network for Deep Image Relighting

Normalization of face illumination using basic knowledge and information extracted from a single image

Homomorphic Filtering Based Illumination Normalization Method for Face Recognition.

Face Illumination Normalization Based on Generative Adversarial Network

Non-Uniform Illumination Image Enhancement via a Retinal Mechanism Inspired Decomposition

Low-Light Image Enhancement via Implicit Priors Regularized Illumination Optimization

DifFRelight: Diffusion-Based Facial Performance Relighting

Illumination normalization based on correction of large-scale components for face recognition.

Normalization of Face Illumination Based on Large-and Small-Scale Features

ILSR-Diff: joint face illumination normalization and super-resolution via diffusion models