With the rapid development of network communication technologies and the advent of the 5G era, breakthroughs have been achieved in high-definition video encoding/decoding and real-time mobile video transmission. As a result, global online interactions have witnessed exponential growth. Against this backdrop, the video conferencing market has experienced explosive expansion, with major international platforms such as Zoom, Microsoft Teams, Tencent Meeting and Feishu, emerging and evolving rapidly. Meanwhile, driven by advances in deep learning algorithms, facial recognition-based biometric authentication has become a mainstream identity verification method in key domains such as government and finance services, owing to its convenience and security. However, in scenarios such as video conferencing and facial recognition, eyeglass reflection often leads to significant occlusion of facial features. In video conferencing, such reflection degrades visual perception and distracts online participants. In facial recognition, reflection regions interfere with facial localization, increasing the recognition errors.

Eyeglass reflection typically involves either specular highlights or ambient scene reflections, or a combination of both. The captured image with reflection can be regarded as a superposition of a transmission layer and a reflection layer. The objective of reflection removal is to effectively separate these two components. Currently, reflection removal methods can be broadly categorized into single-image-based and multi-image-based approaches, depending on the input modality. Multi-image-based methods utilize either flash image sequences or polarization image sequences to facilitate reflection removal. Single-image-based methods lack sufficient cues, such as temporal, angular, or polarization information, to distinguish between the reflection and transmission layers, and thus typically rely on modeling the statistical characteristics. For instance, some approaches have exploited the higher blur level and weaker gradient magnitude of the reflection layer to identify reflection regions and reconstruct the transmission content using a non-local self-similarity sparse prior and a gradient sparsity constraint^1,2. Similarly, Yang Y et al.³ proposed to leverage the blurry nature of reflection, removing reflection edges through gradient thresholding and solving the reflection-free image using convex optimization⁴. Several studies have approached the problem from the perspective of unwanted artifact removal and image restoration, aiming to achieve more general image reconstruction and quality enhancement, such as MB-TaylorFormer v2⁵, GDP⁶, MC-Blur⁷, and DBLRNet⁸. Sandhan T et al.⁹ focused on eyeglass reflection and integrated facial symmetry prior into a non-convex optimization scheme to achieve reflection removal by exploiting the sparsity of reflection gradient. Subsequent work has gradually shifted toward deep neural network-based solutions. Perceptual loss and exclusion loss were employed to train end-to-end networks¹⁰, while other losses based on the characteristics of eyeglass reflection, such as eye-symmetry loss, were designed to avoid visual artifacts¹¹. Li C et al.¹² integrated the iterative convex optimization process into deep networks via cascaded architecture to improve the reliability of blind separation. More recently, Huang JJ et al.¹³ designed the Deep Exclusion Unfolding Network, focusing on minimizing the shared features between the transmission and reflection layers to enhance separation. However, these methods typically rely on paired reflection/reflection-free data for supervised training, which significantly hinder their scalability and applicability in real-world scenarios. To enable unsupervised single-image reflection removal, Deep Image Prior (DIP) was introduced^14,15. Due to the semantic ambiguity between reflection and transmission layers, DIP-based methods often fail to achieve accurate separation. In the past three years, several studies have explored diffusion model, leveraging the generative diversity to reconstruct the missing transmission content in regions heavily occluded by reflection¹⁶⁻¹⁸. Yi ZL et al.¹⁸ focused on eyeglass reflection scenes and proposed DL2G, which employs a pretrained degradation estimation module to detect reflective regions, followed by a local structure-aware diffusion model to recover fine-grained textures hidden under reflection.

Nonetheless, single-image reflection removal methods remain heavily reliant on strong scene assumptions, making it difficult to achieve both generality and fidelity under diverse real-world conditions. In contrast, multi-image-based approaches can capture diverse reflection characteristics under different illumination, angular, or polarization conditions, thereby addressing the ill-posed problem of separating transmission and reflection layers. One common strategy is to leverage illumination control. By capturing paired images under ambient lighting with and without the flash, and subtracting the two, one obtained a “flash-only” image that suppressed reflection^19,20. Recently, Shi WQ et al.²¹ proposed a deep-learning-based control system that dynamically adjusted a ring-shaped lighting system in four polarized directions to eliminate specular highlights for the Fourier light field microscopy system. However, such methods require integration of high-power illumination modules and the sensors, resulting in bulky systems with high energy consumption. In contrast, many thin, compact, and portable snapshot polarization imaging systems already exist, such as the commercialized Sony division-of-focal-plane polarization sensor and the emerging research focus on metasurface-based polarimetric imaging sensors^22,23. Metasurfaces, due to their flexibility in multi-parameter modulation, are strong candidates for polarization manipulation²⁴. In particular, dielectric metasurfaces can realize highly efficient polarization-dependent wavefront manipulation²⁵ and polarization multiplexing²⁶. Favorable conditions are created for acquiring polarization information in real-world scenarios. Since both transmitted and reflected light can be partially polarized, polarization filtering has been explored to suppress reflection. Nevertheless, complete suppression only occurs under ideal conditions, such as when the incident angle matches the Brewster angle. As a result, physical polarizing filters cannot achieve complete extinction.

Consequently, data-driven polarization imaging technologies have gradually found applications in reflection removal²⁷. Wen SJ et al.²⁸ constructed polarization chromaticity images from four polarization-direction intensity images to cluster pixels with similar diffuse colors, using the cluster result as a prior in iterative optimization for separating specular reflection. Wang X et al.²⁹ further optimized the network structure for polarization-guided reflection removal by proposing a dual-stream architecture with feature-attention guidance to estimate both the transmission and reflection layers. To construct paired polarization datasets, the synthetic pipeline was proposed to generate realistic reflection polarized data from curved surfaces and dynamic environments^30,31. This work demonstrates the potential of integrating traditional polarization optics with differentiable computational imaging, offering an efficient pathway for reflection removal. However, there still exists the domain gap between synthetic data and real-world data distributions. A field-of-view-aligned polarization dataset covering over 100 types of glass materials was introduced, including reflection-free ground truth and corresponding images captured from four polarization angles, to improve real-world performance³². Most recently, PolaRGB³³ was released, a large-scale dataset of indoor and outdoor RGB and polarization images under diverse lighting conditions. In this work³³, the PolarFree leverages the diffusion process to generate reflection-free cues. However, currently available public polarization datasets do not include eyeglass reflection scenarios, limiting their effectiveness for eyeglass reflection removal.

In summary, existing optimization-based polarization reflection removal methods only work under strong assumptions and tend to degrade the details of the transmission scene. Deep learning-based approaches typically depend on paired polarization reflection images and reflection-free ground truth, which poses significant challenges in data acquisition and generalization. To address these challenges, we propose an untrained method: PDPrior (polarization-guided diffusion prior for eyeglass reflection removal). Figure 1 illustrates the results of PDPrior and highlights its advantages and uniqueness compared with existing approaches. The PDPrior operates solely on polarized observations acquired at test time, free of pretrained models, training data, and ground-truth reflection-free images. The DIP, as an untrained method^14,34, requires thousands of iterations to update network parameters. Unlike DIP, the proposed PDPrior keeps the network parameters frozen during the optimization process and only updates the network input. This design significantly reduces the number of iterations while maintaining performance, leading to approximately two orders of magnitude improvement in inference speed. The PDPrior leverages the generative prior of a diffusion model and introduces polarization information as guidance to control the generation process. During each sampling step, the reflection and transmission variables are alternately and iteratively updated via gradient descent. Specifically, a division-of-focal-plane polarization camera is used to capture intensity images at 0°, 45°, 90°, and 135° polarization angles. The reflection diffusion model takes the unpolarized intensity image as input. The degree of linear polarization (DoLP) is then used to preliminarily identify reflection regions and the diffusion prior of progressively darkening facial content during generation process, which enables the model to focus on reflective regions. The optimal polarization-direction image with minimal reflection is adaptively selected as input to the transmission diffusion model based on intensity, gradient, and entropy constraints. The transmission diffusion model uses the DoLP to compute the transmission coefficient. The generation of transmission image is guided by the forward model of reflection formation, ensuring physical interpretability. To our knowledge, this is the first work to apply untrained diffusion model to the task of polarized reflection removal, marking a novel and promising direction in the field. We validate PDPrior on our polarization eyeglass reflection datasets under various indoor and outdoor lighting conditions. Experimental results show that PDPrior achieves artifact-free, high-fidelity reflection removal, and outperforming state-of-the-art methods. Moreover, the PDPrior consistently obtains higher and more robust scores in face image quality assessments for recognition performance.

Eyeglass reflection typically exhibits image characteristics distinct from background transmission, and these differences can be exploited for reflection removal. However, such characteristics are sensitive to illumination and difficult to unify into a general representation. To address this issue, we leverage the flexible generative capability of generative model to reconstruct reflection-free transmission images. Nevertheless, without proper constraints, generative model may produce results that deviate from physical reality, making effective guidance a critical challenge. As one of the three fundamental properties of light, along with color and intensity³⁵, polarization can effectively distinguish reflection regions. By incorporating polarization information, the PDPrior achieves reflection and transmission image generation without requiring training data, ensuring the physical reliability of the generated results.

To tackle the challenge of eyeglass reflection obscuring facial features, we propose an untrained polarization reflection removal method, PDPrior. The PDPrior requires no training data, and no ground-truth reflection-free images. The PDPrior leverages the generative prior of a diffusion model and introduces polarization information as guidance to control the generation process. The reflection diffusion model utilizes the DoLP to preliminarily identify reflection regions and the diffusion prior of progressively darkening facial content during the generation process, enabling the model to focus on reflective regions. Using the DoLP to compute the transmission coefficient, the generation of transmission image is guided by the physical forward model of reflection formation. During each sampling step, the reflection and transmission variables are alternately and iteratively updated via gradient descent based solely on the test sample, making PDPrior more adaptable to complex lighting conditions and diverse scenes. We collected real-world eyeglasses reflection images using a division-of-focal-plane polarization camera under various indoor and outdoor lighting environments. Experimental results demonstrate that PDPrior effectively removes eyeglass reflection, producing high-fidelity face reconstructions with no visible artifacts and achieving more robust face image quality assessment scores for recognition performance. The PDPrior provides a general framework for reflection removal, which can be extended to diverse real-world scenarios such as window photography, showcase displays, and driver monitoring, paving the way for broader applications in consumer photography and intelligent vision systems.