Abstract

Translate

The abrupt light-to-dark or dark-to-light transitions at tunnel entrances and exits cause short-term, large-scale illumination changes, leading traditional RGB perception to suffer from exposure mutations, glare, and noise accumulation at critical moments, thereby triggering perception failures and blind zones. Addressing this typical failure scenario, this paper proposes a closed-loop enhancement solution centered on polarization imaging as a core physical prior, comprising a real-world polarimetric road dataset, a polarimetric physics-enhanced algorithm, and a beyond-fusion network, while satisfying both perception enhancement and real-time constraints. First, we construct the POLAR-GLV dataset, which is captured using a four-angle polarization camera under real highway tunnel conditions, covering the entire process of entering tunnels, inside tunnels, and exiting tunnels, systematically collecting data on adverse illumination and failure distributions in day–night traffic scenes. Second, we propose the Polarimetric Physical Enhancement with Adaptive Modulation (PPEAM) method, which uses Stokes parameters, DoLP, and AoLP as constraints. Leveraging the glare sensitivity of DoLP and richer texture information, it adaptively performs dark region enhancement and glare suppression according to scene brightness and dark region ratio, providing real-time polarization-based image enhancement. Finally, we design the Polar-PENet beyond-fusion network, which introduces Polarization-Aware Gates (PAG) and CBAM on top of physical priors, coupled with detection-driven perception-oriented loss and a beyond mechanism to explicitly fuse physics and deep semantics to surpass physical limitations. Experimental results show that compared to original images, Polar-PENet (beyond-fusion network) achieves PSNR and SSIM scores of 19.37 and 0.5487, respectively, on image quality metrics, surpassing the performance of PPEAM (polarimetric physics-enhanced algorithm) which scores 18.89 and 0.5257. In terms of downstream object detection performance, Polar-PENet performs exceptionally well in areas with drastic illumination changes such as tunnel entrances and exits, achieving a mAP of 63.7%, representing a 99.7% improvement over original images and a 12.1% performance boost over PPEAM’s 56.8%. In terms of processing speed, Polar-PENet is 2.85 times faster than the physics-enhanced algorithm PPEAM, with an inference speed of 183.45 frames per second, meeting the real-time requirements of autonomous driving and laying a solid foundation for practical deployment in edge computing environments. The research validates the effective paradigm of using polarimetric physics as a prior and surpassing physics through learning methods.

Full text

Turn on search term navigation

Translate

1. Introduction

With the rapid development of autonomous driving technology, visual perception systems need to maintain highly reliable performance in various complex environments [1,2]. However, traditional visible light imaging technology performs poorly in environments with drastic light-to-dark transitions such as tunnel entrances and exits. Instantaneous dramatic changes in light intensity cause image sensors to be overexposed or underexposed, severely affecting the performance of object detection and recognition [3,4]. In these high dynamic range (HDR) illumination scenarios, the imaging quality of traditional RGB cameras degrades significantly, leading to sharp deterioration in the performance of perception tasks (such as object detection), posing serious threats to driving safety [5,6]. Particularly, in scenarios such as tunnel entrances and underground parking exits, traditional vision sensors require considerable adaptation time to adjust exposure parameters, during which transient “perception blind zones” are created, bringing enormous safety hazards to high-speed vehicles [3]. Although deep learning-based image enhancement methods have made significant progress, these methods primarily rely on adjusting the brightness and color of RGB images through probabilistic “guessing” based on learned patterns. When facing traffic scenarios where illumination intensity changes by several orders of magnitude, they often expose problems such as color distortion, detail loss, and noise amplification [7,8], as they lack stable physical priors to guide enhancement decisions.

Polarization imaging, as an emerging imaging technology, provides a fundamentally different approach by offering physical “measurement” rather than probabilistic inference [9]. It can capture information about the vibration direction of light waves, thereby providing additional physical dimension information such as material, shape, and surface properties that traditional intensity images cannot obtain [10,11]. These unique physical characteristics enable polarization imaging to demonstrate distinctive advantages in challenging scenarios such as adverse weather, underwater environments, and low-light conditions [12,13,14,15]. Recent advances have demonstrated the effectiveness of polarimetric methods in scattering and turbid media, where traditional imaging fails due to light scattering and absorption [16]. Particularly, polarization information can effectively distinguish surface reflections of different materials, which is of significant value for enhancing object detection capabilities in environments with drastic illumination changes such as tunnels [17,18,19].

Although polarization imaging theoretically provides enormous potential for solving visual perception problems in light-to-dark transition environments, current research still faces three key challenges: first, there is a lack of specialized polarimetric datasets for such specific environments, which hinders algorithm development and validation; second, existing physical enhancement algorithms have insufficient adaptive capabilities for dynamically changing illumination conditions; and finally, how to effectively integrate polarization information with deep learning models to fully leverage their physical advantages remains an urgent problem to be solved.

To address these challenges, this paper proposes a closed-loop research solution encompassing “dataset construction–physical algorithm–learning beyond.” We first construct a polarimetric imaging traffic dataset covering light-to-dark transition tunnel scenes, subsequently develop an adaptive physical polarimetric enhancement algorithm, and ultimately design a polarimetric beyond-fusion network combining physical models with deep learning, focusing on its application in object detection tasks. The main contributions of this paper are as follows:

Construction of the first polarimetric imaging traffic dataset (POLAR-GLV) for light-to-dark transition tunnel road conditions: This dataset uses a four-angle polarization camera for real vehicle acquisition, completely covering the entire process from entering tunnels, inside tunnels, to exiting tunnels, providing critical data foundation for researching polarimetric image enhancement techniques in such scenarios.

Proposal of an adaptive Polarimetric Physical Enhancement with Adaptive Modulation algorithm (PPEAM): Based on Stokes parameters, this algorithm can dynamically balance dark region enhancement and glare suppression according to scene illumination conditions, providing an effective physical baseline for physics-prior-based enhancement.

Design of a beyond-physical polarimetric fusion network (Polar-PENet): This network achieves adaptive fusion of physical features and deep semantics through novel Polarization-Aware Gates (PAG) and attention mechanisms (CBAM), and driven by perception-oriented loss functions and a “beyond mechanism,” enables image enhancement effects to surpass the limits of pure physical methods in downstream detection tasks.

The remainder of this paper is organized as follows: Section 2 reviews related work. Section 3 provides a detailed introduction to the constructed dataset. Section 4 elaborates on the proposed methods. Section 5 presents experimental results and analysis. Finally, Section 6 concludes the paper.

2. Related Work

2.1. Traditional Low-Light Traffic Environment Enhancement Methods

Illumination changes in tunnel environments seriously affect the reliability and safety of traffic perception systems. The light-to-dark transition phenomenon at tunnel entrances produces “black hole effects” and “white hole effects,” causing cameras to be underexposed or overexposed [20,21]. Such drastic illumination changes require traditional cameras 3–5 s of adaptation time, forming perception “blind zones” that bring safety risks to high-speed vehicles [3,22]. Tunnels employ artificial lighting systems arranged at intervals, resulting in uneven light distribution [23], producing motion blur and signal noise that reduce the detection accuracy of perception systems [24].

To address these challenges, researchers have proposed various enhancement methods. Adaptive exposure control techniques achieve dynamic adjustment through gradient analysis [25] or region-of-interest switching [26]. HDR imaging technology can capture a wider brightness range [27], but dataset limitations and computational requirements still restrict its practical application [28]. Neural network-based low-light enhancement methods, such as RetinexNet [29], EnlightenGAN [30], and Zero-DCE [31], have improved low-light image quality to some extent, but still face problems such as color distortion, detail loss, and noise amplification in extreme illumination change environments [7].

Multi-sensor fusion approaches, such as the work of Wang et al. [6], combine LiDAR and visual data to enhance low-light perception capabilities. In terms of environmental optimization, improving tunnel lighting systems to moderate brightness transitions [32] reduces pressure on perception systems from an environmental level. However, coverage of tunnel scenarios in existing datasets remains limited, lacking comprehensive datasets with high-fidelity tunnel illumination conditions [27].

2.2. Polarization Imaging-Based Traffic Environment Enhancement Research

Polarization imaging technology, with its unique physical characteristics, demonstrates excellent potential in traffic scene perception. Research by Wang et al. [33] shows that polarization information can effectively enhance road scene perception under low visibility conditions, with average detection precision increasing by more than 15% through deep learning methods that fuse polarization features. The theoretical foundation of polarimetric imaging relies on the characterization of Mueller matrices and Stokes parameters, which provide comprehensive descriptions of light polarization states and their interactions with materials [34]. In the field of object tracking, polarization features such as mosaic gradient histograms [35] have proven effective for robust target tracking in DoFP infrared polarization imaging scenarios.

In high-light environment applications, Zhou et al. [36] developed a dual-discriminator generative adversarial network combining intensity and polarization information, which effectively enhances weak target detection while preserving rich background information. Huang et al. [18] proposed the YOLO-PRTD algorithm for road target detection under complex weather conditions, utilizing adaptive polarization encoding to enhance features, achieving 89.83% average precision and significantly reducing error rates. In tunnel infrastructure applications, Zhang et al. [19] demonstrated the effectiveness of polarization 3D imaging for crack detection in tunnel linings, achieving high precision in challenging structural environments through passive single-frame array polarimetric imaging.

In low-light environments, Zhou et al. [37] proposed a polarization-aware low-light enhancement algorithm that effectively fuses brightness and polarization information through a Stokes domain enhancement process, addressing the difficulty traditional RGB enhancement methods face in maintaining natural colors while improving contrast. Recent developments in deep learning-based polarimetric reconstruction have shown promising results; Lin et al. [38] developed an improved U-Net architecture for high-performance polarization image reconstruction in scattering systems under natural light conditions, demonstrating enhanced quality and robustness in challenging visibility scenarios. Multi-modal fusion technology further enhances the value of polarization imaging in harsh environments; for example, Wang et al. [39] fused polarization and infrared image features to construct a vehicle-road perception system suitable for low visibility environments.

For DoFP (division-of-focal-plane) polarization cameras, demosaicking is a critical preprocessing step to reconstruct full-resolution polarization images from the raw mosaic pattern. Several advanced demosaicking methods have been proposed, including polarization channel difference prior-based approaches [40], physics-based quality assessment methods [41], and frequency domain techniques [42]. While these methods offer sophisticated reconstruction capabilities, this work employs interpolation as a baseline demosaicking approach, which provides a reasonable trade-off between computational efficiency and reconstruction quality for real-time applications.

Although polarization imaging technology has enormous potential, it still faces challenges in traffic scene applications: (1) lack of polarimetric datasets for specific traffic scenarios; (2) the complexity of traffic scenes causes differences in polarization characteristics of different material surfaces, making processing complex; and (3) computationally intensive algorithms have difficulty meeting real-time processing requirements. The physics-guided polarimetric framework proposed in this paper fully leverages the physical characteristics of polarized light, enhancing image quality in traffic environments such as tunnels by combining intensity images and polarization properties, while meeting real-time application needs through efficient implementation.

3. POLAR-GLV Dataset

3.1. Polarization Imaging Principles

Polarization is a fundamental physical property of light waves, describing their vibration direction. Polarization imaging, by measuring polarization information at each point in a scene, can reveal features such as object surface material, roughness, and geometric shape that traditional intensity imaging cannot perceive. In polarization measurement, the Stokes Vector S is typically used to completely describe the polarization state of light:

(1) $S = {[S_{0}, S_{1}, S_{2}, S_{3}]}^{T}$

where

S_{0}

represents total light intensity,

S_{1}

and

S_{2}

describe linear polarization components, and

S_{3}

describes circular polarization components (which can typically be ignored in natural scenes). For linear polarization, Stokes parameters can be calculated by acquiring intensity images (

I_{0}

I_{45}

I_{90}

, and

I_{135}

) at four different polarization angles (0°, 45°, 90°, and 135°):

(2) $\begin{matrix} S_{0} & = \frac{I_{0} + I_{90}}{2} \end{matrix}$

(3) $\begin{matrix} S_{1} & = I_{0} - I_{90} \end{matrix}$

(4) $\begin{matrix} S_{2} & = I_{45} - I_{135} \end{matrix}$

Based on Stokes parameters, two key physical quantities can be further derived as follows: Degree of Linear Polarization (DoLP) and Angle of Linear Polarization (AoLP):

(5) $\begin{matrix} DoLP & = \frac{\sqrt{S_{1}^{2} + S_{2}^{2}}}{S_{0} + ϵ} \end{matrix}$

(6) $\begin{matrix} AoLP & = \frac{1}{2} arctan 2 (S_{2}, S_{1}) \end{matrix}$

where DoLP reflects the proportion of linear polarization components in light waves and is commonly used for glare suppression and material identification. AoLP represents the direction of linear polarization and is closely related to the normal direction of object surfaces. The dataset constructed in this study includes the above multi-modal polarization information.

3.2. Dataset Acquisition and Composition

To address the lack of specialized polarimetric data in light-to-dark transition traffic scenarios, we constructed the POLAR-GLV dataset. The data acquisition equipment is a DoFP (division-of-focal-plane) polarization camera (LUCID PHX050S-QC, LUCID Vision Labs, Richmond, BC, Canada), with an original resolution of 2448 × 2048 pixels, equipped with a Fujinon 25 mm lens (Fujifilm, Tokyo, Japan). The camera was rigidly mounted on the roof of a Volkswagen SUV at a height of approximately 1.8 m to simulate a vehicle-mounted perception perspective. The acquisition route is a typical highway tunnel section on China’s Shenhai Expressway, with vehicle speed maintained at 60–80 km/h. Data collection was conducted around 12:00 p.m. to cover the most drastic illumination change conditions. The acquisition process completely recorded the entire process from the vehicle entering the tunnel, traveling inside the tunnel, to exiting the tunnel. The camera was rigidly mounted on the roof of a Volkswagen SUV at a height of approximately 1.8 m to simulate a vehicle-mounted perception perspective. The acquisition route is a typical highway tunnel section on China’s Shenhai Expressway, with vehicle speed maintained at 60–80 km/h. Data collection was conducted around 12:00 p.m. to cover the most drastic illumination change conditions. The acquisition process completely recorded the entire process from the vehicle entering the tunnel, traveling inside the tunnel, to exiting the tunnel.

The POLAR-GLV dataset contains a total of 1487 pairs of multi-modal image samples. Each sample includes the original intensity image ( $S_{0}$ ) and corresponding polarization parameter maps (DoLP and AoLP). To support algorithm training and evaluation, we divided the dataset into training, validation, and test sets in a 7:1:2 ratio. Additionally, all images were annotated with detailed bounding boxes for vehicle targets to support performance evaluation of downstream object detection tasks. Figure 1 shows typical brightness distribution statistics in the dataset, clearly revealing the drastic illumination changes at tunnel entrances and exits, which directly lead to performance bottlenecks in traditional visual perception.

4. Methods: Physics Enhancement and Beyond-Fusion

4.1. Polarimetric Physical Enhancement with Adaptive Modulation (PPEAM)

We designed an adaptive Polarimetric Physical Enhancement with Adaptive Modulation algorithm (PPEAM) that can automatically adjust processing parameters according to scene illumination conditions, providing stable enhancement effects in different lighting environments. The algorithm flow is shown in Algorithm 1. This algorithm uses the Stokes parameters, DoLP, and AoLP defined in Section 3 as physical prior inputs. Its core innovation lies in adaptively calculating processing weights according to scene illumination conditions and polarization characteristics. The dark region enhancement weight is defined as

(7) $w_{d a r k} = \{\begin{matrix} 0.9 & if μ < 60 \\ 0.9 - \frac{0.8}{120} (μ - 60) & if 60 \leq μ \leq 180 \\ 0.1 & if μ > 180 \end{matrix}$

The glare suppression weight is defined as $w_{f l a r e} = 1 - w_{d a r k}$ .

Where $μ$ is the average brightness, $r_{d a r k}$ is the proportion of dark region pixels, and DOLP represents the degree of polarization.

Algorithm 1 Adaptive Polarimetric Physical Enhancement Algorithm Pseudocode. The algorithm processes polarimetric images through three main stages: the input and parameter calculation stage extracts physical features, the feature extraction and analysis stage calculates adaptive weights, and the processing and output stage implements dark region enhancement and glare suppression.

Input: Four images at different polarization angles

I_{0}

I_{45}

I_{90}

I_{135}

Output: Enhanced image

I_{f i n a l}

1:. function PPEAM( $I_{0}$ , $I_{45}$ , $I_{90}$ , $I_{135}$ )

2:. Stage 1: Input and Parameter Calculation

3:. Calculate Stokes parameters: $S_{0} = \frac{I_{0} + I_{90}}{2}$

4:. $S_{1} = I_{0} - I_{90}$

5:. $S_{2} = I_{45} - I_{135}$

6:. Calculate overall scene brightness: $μ = \frac{\sum_{i, j} I_{0} (i, j)}{H W}$

7:. Calculate scene brightness standard deviation: $σ = \sqrt{\frac{1}{H W} \sum_{i, j} {(I_{0} (i, j) - μ)}^{2}}$

8:. Calculate dark region pixel ratio: $r_{d a r k} = \sum_{i = 0}^{63} \frac{h i s t (i)}{N}$

9:. Stage 2: Feature Extraction and Analysis

10:. Calculate Degree of Linear Polarization: $DOLP = \frac{\sqrt{S_{1}^{2} + S_{2}^{2}}}{S_{0} + ϵ}$ { $ϵ$ is a small constant to prevent division by zero}

11:. if $μ < 60$ then

12:. $w_{d a r k} = 0.9$

13:. else if $60 \leq μ \leq 180$ then

14:. $w_{d a r k} = f (μ, r_{d a r k}, DOLP)$ {Adaptively calculate weight based on brightness, dark ratio, and polarization}

15:. else

16:. $w_{d a r k} = 0.1$

17:. end if

18:. $w_{f l a r e} = 1 - w_{d a r k}$ {Calculate glare suppression weight}

19:. Stage 3: Processing and Output

20:. $I_{b a s e} = B (S_{0})$ {Base layer decomposition using bilateral filter}

21:. $I_{d e t a i l} = S_{0} - I_{b a s e}$ {Calculate detail layer}

22:. $I_{l o w}^{e n h} = I_{l o w} (1 - {DOLP}^{e n h} \cdot β)$ {Dark region enhancement, $β = 0.7$ }

23:. $I_{f u s i o n} = w_{d a r k} \cdot I_{d e t a i l} + w_{f l a r e} \cdot I_{f l a r e}$ {Adaptive fusion}

24:. $I_{g u i d e d} = G (I_{f u s i o n}, I_{0})$ {Guided filtering, $γ = 0.6$ }

25:. $I_{l o c a l} = I_{f u s i o n} - I_{g u i d e d}$ {Local enhancement component}

26:. $I_{f i n a l} = I_{g u i d e d} + I_{l o c a l}^{e n h} \cdot (DOLP \cdot 0.5)$ {Final enhanced result}

return $I_{f i n a l}$

27:. end function

The PPEAM algorithm includes two main processing branches: 1. Dark region enhancement processing: Implements physics-guided multi-scale decomposition enhancement, including base layer and detail layer decomposition, gamma correction, and DOLP-guided detail enhancement. 2. Glare suppression processing: Targets high-light environments, particularly specular reflection regions, through DOLP-guided frequency domain decomposition and suppression.

The final enhanced image is obtained through adaptive fusion:

(8) $I_{f i n a l} = {\hat{w}}_{d a r k} I_{d a r k} + {\hat{w}}_{f l a r e} I_{f l a r e}$

Through adaptive weight calculation and multi-level processing, this algorithm achieves high-quality image enhancement in light-to-dark transition scenarios while maintaining physical consistency. It is particularly suitable for environments with drastic illumination changes such as tunnel entrances and exits in autonomous driving.

4.2. Polar-PENet: Beyond-Physical Polarimetric Fusion Network

Polar-PENet adopts the widely used U-Net encoder–decoder architecture, which is renowned for its effectiveness in image-to-image translation tasks by combining multi-level features, as shown in Figure 2. The network accepts a multi-channel input, typically concatenating standard RGB images with corresponding polarization information maps (AoLP, DoLP, and S0), and processes them through an encoder–decoder structure to generate enhanced image outputs. Key architectural components and features include

1.. Encoder Path: Composed of multiple convolutional blocks with downsampling operations (such as max pooling or strided convolution) to progressively extract hierarchical features, capturing low-level details and high-level semantic context from the combined RGB and polarization inputs.

2.. Decoder Path with Skip Connections: Composed of multiple convolutional blocks with upsampling operations (such as transposed convolution or upsampling). Crucially, skip connections concatenate feature maps from corresponding encoder stages with upsampled features in the decoder. This enables the network to reuse fine-grained details from the encoder when reconstructing images.

3.. Polarization-Aware Gate (PAG) Integration: The core PAG module is strategically integrated within decoder stages (e.g., after feature concatenation through skip connections). It adaptively fuses polarization-rich physical features from the encoder (via skip connections) and abstract deep features from the previous decoder layer.

4.. Attention Enhancement (CBAM): After the PAG module in decoder stages, Convolutional Block Attention Module (CBAM) is employed to further refine fused features by adaptively recalibrating feature responses in both channel and spatial dimensions, thereby focusing on more informative parts of feature maps.

5.. Output Layer: The final convolutional layer (e.g., 1 × 1 convolution) maps high-resolution features from the last decoder stage to the desired output format (e.g., 3-channel enhanced RGB image).

The Polarization-Aware Gate (PAG) mechanism is a core innovation designed to intelligently fuse information from two feature streams within the decoder. It is typically located after the concatenation step in skip connections, receiving two inputs:

1.. $F_{p o l a r}$ : Polarization-rich physical features passed from the corresponding encoder layer through skip connections.

2.. $F_{r a w}$ : Deep features carrying more abstract semantic information, passed from the previous upsampling layer in the decoder.

PAG computes an adaptive, spatially varying gating weight map $α$ (with values between 0 and 1) that is learned from the input features. This weight dynamically controls the contribution of each feature type at each location:

(9) $F_{f u s e d} = α ⊙ F_{p o l a r} + (1 - α) ⊙ F_{r a w}$

where ⊙ denotes element-wise multiplication. This adaptive fusion enables the network to prioritize physical polarization cues and deep contextual features according to specific image content and enhancement objectives. The resulting

F_{f u s e d}

is typically passed to the CBAM module.

Convolutional Block Attention Module (CBAM) consists of channel attention and spatial attention sub-modules, applied sequentially to input feature map F (the $F_{f u s e d}$ from PAG):

(10) $F^{'} = M_{s} (M_{c} (F) \otimes F) \otimes M_{c} (F) \otimes F$

where F represents input features,

M_{c}

and

M_{s}

represent channel attention and spatial attention maps, respectively, and ⊗ denotes element-wise multiplication. The CBAM module can effectively enhance the network’s focus on critical features, especially in complex scenarios and low-light conditions, accurately locating regions requiring enhancement.

4.3. Loss Function Design

To better balance image fidelity, physical consistency, and downstream detection performance, a two-stage multi-objective loss is adopted as follows:

Adversarial Loss (if adopting GAN framework):

(11) $\begin{matrix} L_{a d v} = E_{x, y} [log D (x, y)] + E_{x} [log (1 - D (x, G (x)))] \end{matrix}$

Reconstruction Loss (L1):

(12) $\begin{matrix} L_{r e c o n} = {∥ G (x) - y ∥}_{1} \end{matrix}$

Perceptual Loss (VGG features):

(13) $L_{p e r c e p} = \sum_{i} λ_{i} {∥ ϕ_{i} (G (x)) - ϕ_{i} (y) ∥}_{2}^{2}$

Physical Consistency Loss: Maintaining reasonableness of physical quantities such as DoLP/AoLP:

(14) $L_{p h y s} = λ_{1} L_{d o l p} + λ_{2} L_{a o p}$

YOLO-Guided Detection Loss: Integrating intermediate features and final predictions:

(15) $L_{d e t} = α L_{f e a t} + β L_{p r e d}$

Also, we introduce the “beyond mechanism”: when generated images surpass target images in detection performance on a batch (such as mean confidence or object count), gradually reduce the weight of imitating target images, increasing self-comparison and absolute metric reward terms to avoid being constrained by target upper limits:

(16) $L_{p r e d}^{s u r p a s s} \approx L_{p r e d} \cdot f (performance gap, epoch) .$

The total loss is

(17) $L_{t o t a l} = w_{a d v} L_{a d v} + w_{r e c o n} L_{r e c o n} + w_{p e r c e p} L_{p e r c e p} + w_{p h y s} L_{p h y s} + w_{d e t} L_{d e t} .$

5. Experiments and Results

5.1. Setup and Evaluation

Implementation is based on PyTorch (version 2.7.1, Meta AI, Menlo Park, CA, USA) with input resolution $512 \times 612$ , and training for 100 epochs in each of two stages. Detectors employ YOLOv8 series (n/s/m/l/x), with regional evaluation: entrance, inside, and exit. Image quality is measured by PSNR/SSIM/MSE/NRMSE, and detection by mAP/F1 as primary metrics.

Image size: Random crop to $512 \times 612$

Training epochs: Stage 1: 100 epochs; Stage 2: 100 epochs

Batch size: 4 (gradient accumulation to 16)

Optimizer: Adam ( $β_{1} = 0.5$ , $β_{2} = 0.999$ )

Learning rate: $2 \times 10^{- 4}$ , decayed to 0.5× at epochs 30/60/90

Data augmentation: Random crop to $512 \times 612$ , horizontal flip, photometric distortions ( $\pm 20 %$ ), brightness/contrast adjustment

Early Stopping: Patience = 10

L2 regularization: $5 \times 10^{- 4}$

Perceptual loss: VGG-19 conv3_3, conv4_3, conv5_3 layers

Detector: YOLOv8 series (input $640 \times 640$ , default pretrained; comparison includes YOLOv8x)

5.2. Image Quality Analysis

As shown in Table 1 and Table 2, both PPEAM and Polar-PENet significantly outperform original images.

Comparison with SOTA RGB-Based Enhancement Methods. To demonstrate the advantages of polarization-guided fusion over traditional RGB-based enhancement methods, we compare Polar-PENet with state-of-the-art RGB enhancement approaches. Table 3 presents quantitative comparisons using entropy, gradient, and mean intensity metrics.

Further detailed comparisons validate the proposed method’s superiority. A visual comparison of the enhancement results under different illumination conditions is presented in Figure 3. Specific image quality metrics are compared in Table 4, while illumination characteristics are analyzed in Table 5. Additionally, the computational efficiency is contrasted in Table 6.

5.3. Ablation Study

To analyze the contribution of each component in Polar-PENet, we conduct an ablation study by progressively adding key modules. Table 7 presents the results on the test set. Starting from a baseline U-Net (PSNR: 18.65 dB, SSIM: 0.5120, Avg Gradient: 13.50, mAP: 58.2%), adding CBAM attention mechanism improves mAP to 60.1%, demonstrating its effectiveness in focusing on critical features for detection tasks. The addition of PAG (Polarization-Aware Gate) brings significant improvements in gradient (16.80) and mAP (61.8%), highlighting PAG’s crucial role in adaptively fusing polarization physical features with deep semantic features, which preserves rich texture information. The full Polar-PENet configuration achieves the best performance across all metrics (PSNR: 19.37 dB, SSIM: 0.5487, Avg Gradient: 17.25, mAP: 63.7%), validating the synergistic effect of combining PAG and CBAM modules.

5.4. Overall Results: Original vs. Physics vs. Fusion

Table 1 and Table 2 show that: PPEAM and Polar-PENet significantly improve over original images; Polar-PENet outperforms PPEAM across all three illumination types, particularly notable at entrances and mixed lighting.

5.5. Regional Segment Analysis and Summary

As shown in Figure 4, entrance/mixed-light regions are most sensitive to enhancement strategies. The specific illumination metrics across these regions are further visualized in Figure 5. To provide a comprehensive view of the stability, the statistical distributions of these quality metrics are illustrated in the boxplots in Figure 6. Polar-PENet shows more significant improvements in F1/mAP at entrance segments; in low-light inside tunnels, both are effective, with PPEAM having slight advantages in maintaining contrast; and at exit regions, PPEAM scores slightly higher in PSNR/SSIM but the detection performance gap narrows.

To deeply analyze performance differences under different illumination conditions, we divided tunnel scenes into three regions: entrance region (0–290), tunnel middle section (291–1180), and exit region (1181–1486). As shown in Figure 4, the two methods demonstrate significant differences across different regions.

Table 8, Table 9 and Table 10 list detailed image quality comparisons for different regions. In entrance and middle sections, Polar-PENet outperforms PPEAM in all metrics; while at exit regions, PPEAM performs better. Below are detailed comparisons of illumination characteristics for each region, as shown in Table 11, Table 12 and Table 13.

5.6. Object Detection Performance Analysis

5.6.1. Model Scale and Overall Comparison

The optimal performance of different-scale YOLOv8 models on different enhanced images is shown in Table 14: small/medium models perform better on Polar-PENet, while large models can achieve high performance on either PPEAM or Polar-PENet.

Figure 7 shows the comparison of detection performance improvements by the two enhancement methods across different regions.

5.6.2. Relative Improvement and Parameter Sensitivity

The relative improvements of the proposed method compared to the baseline are analyzed. The detailed F1 score improvements relative to original images are listed in Table 15.

The impact of different IoU and confidence thresholds on detection performance is visualized in Figure 8.

Parameter Recommendations: Optimal confidence thresholds for lightweight models at entrance/inside/exit are approximately 0.25/0.20/0.30; extra-large models on original/PPEAM images suit higher thresholds (0.45–0.55), while on Polar-PENet images, approximately 0.30.

5.6.3. Regional Optima and Best Combinations

Table 16 summarizes the best image type and model for each region. Specifically for the challenging entrance region, Table 17 details the relative improvements achieved by different models. Finally, based on these analyses, Table 18 provides the recommended combination strategies for practical deployment.

6. Discussion and Limitations

Although the proposed Polar-PENet demonstrates superior performance on the collected POLAR-GLV tunnel dataset, directly applying it to mainstream public autonomous driving benchmarks such as KITTI or Cityscapes remains non-trivial due to the modality gap. Most existing datasets primarily provide standard RGB intensity images and do not contain raw polarization measurements at four specific angles ( $0^{\circ}, 45^{\circ}, 90^{\circ},$ and $135^{\circ}$ ). As a result, the key polarization state information required by our framework—namely the Stokes parameters $S_{0}, S_{1},$ and $S_{2}$ and the derived DoLP and AoLP maps—cannot be reconstructed, which limits a plug-and-play deployment of Polar-PENet on RGB-only data.

Nevertheless, the architecture of Polar-PENet is intrinsically scalable in two aspects. (1) Module generalizability: The proposed polarization-guided fusion strategy is designed to integrate physical priors with deep semantic features in a modular way. In principle, this “physics-aware” fusion mechanism can be extended to other multi-modal enhancement tasks by replacing the polarization branch with alternative modality encoders, such as RGB–thermal fusion for night-vision perception or RGB–depth fusion for structure-aware reconstruction, while reusing the same PAG and CBAM fusion blocks. (2) Methodological universality: Beyond specific datasets, this work validates a universal “physics-guided, beyond-physics” principle. Our strategy uses the polarization physical enhancement (PPEAM) as a stable prior, ensuring that the network generalizes better in complex lighting conditions than purely data-driven methods. This paradigm allows the network to learn the residuals between the physical model and the ground truth, effectively surpassing the performance limits of the physical model itself. It thus provides a scalable reference for other optical sensing tasks where paired data are scarce but reasonably accurate physical models are available.

Regarding implementation efficiency, deployment and real-time performance: the proposed model can run in real-time on both GPU and edge NPU platforms, while future work will further compress the network for deployment on more resource-constrained edge devices. Interpretability: visualization of PAG weight coordination with DoLP/AoLP demonstrates that the network makes physically consistent enhancement decisions in both glare and dark regions, which improves the transparency and trustworthiness of the enhancement process in safety-critical driving scenarios.

7. Conclusions

This paper proposes a complete technical pipeline of “dataset–physics enhancement–beyond-fusion network” with polarimetric physics as a prior: we (1) constructed the POLAR-GLV polarimetric dataset to characterize light-to-dark transition failures; (2) proposed adaptive PPEAM maintaining physical consistency; and (3) designed Polar-PENet to stably surpass physical enhancement in extreme scenarios such as entrances/mixed lighting. Experimental results show that Polar-PENet surpasses the physical enhancement method PPEAM in both image quality (PSNR 19.37, SSIM 0.5487) and downstream detection tasks, particularly achieving up to 105.3% detection performance improvement for lightweight models such as YOLOv8s in challenging regions like tunnel entrances. Additionally, its inference speed reaches 183.45 frames/second, 2.85 times faster than PPEAM, meeting real-time requirements. The combination of lightweight models with Polar-PENet achieves performance approaching large models, providing an efficient solution for edge deployment, validating that this research provides a feasible path for reliable vision in complex illumination traffic scenarios.

Author Contributions

Conceptualization, R.R.; methodology, R.R.; software, R.R.; validation, L.C.; formal analysis, R.R.; investigation, R.R., Z.O. and S.C.; resources, C.C.; data curation, R.R., Z.O. and S.C.; writing—original draft preparation, R.R.; writing—review and editing, R.R.; visualization, R.R.; supervision, C.C.; project administration, R.R.; funding acquisition, C.C. All authors have read and agreed to the published version of the manuscript.

Data Availability Statement

The data presented in this study are available on request from the corresponding author. The data are not publicly available due to privacy restrictions involving vehicular and pedestrian information.

Conflicts of Interest

Author Liang Chen was employed by the company XMZJY. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Footnotes

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Figures and Tables

Figure 1 Brightness statistics and trends of light-to-dark transitions in tunnel scenes. The blue dots represent the brightness values of individual frames, while the red lines and shaded areas indicate the statistical mean and standard deviation ranges.

Figure 2 Polar-PENet: Explicitly fusing physics and semantics through PAG at the decoder side, and recalibrating with CBAM. (Different colors represent distinct functional modules: green indicates the encoder and decoder backbone blocks, purple represents the fusion (PAG) and attention (CBAM) mechanisms, and blue indicates the input and output).

Figure 3 Visual comparison of enhancement under different illumination conditions: Original (S0)/PPEAM/Polar-PENet. The red bounding boxes indicate the ground truth vehicle targets.

Figure 4 Comparison of image quality metrics across different regions.

Figure 5 Illumination metrics across different regions (mean brightness, standard deviation, contrast, and brightness entropy).

Figure 6 PSNR/SSIM/MSE/NRMSE statistical boxplots: Original, PPEAM, Polar-PENet. The red lines indicate the median values, and the hollow circles represent outliers.

Figure 7 Comparison of detection performance improvements by two enhancement methods across different regions. (a) F1 Score comparison for different YOLOv8 models; (b) Performance gain in the Entrance Region; (c) Performance gain in the Tunnel Middle Section.

Figure 8 Impact of IoU and confidence thresholds on detection performance.

Table 1

Overall comparison of image quality and detection.

Method	PSNR	SSIM	mAP
Original Image	-	-	0.346
PPEAM	18.89	0.5257	0.571
Polar-PENet	19.37	0.5487	0.624

Table 2

Detection mAP under different illumination regions (entrance/mixed/outdoor).

Method	Inside Tunnel (Low Light)	Entrance/Exit (Mixed Light)	Outside Tunnel (Strong Light)
Original	0.278	0.319	0.425
PPEAM	0.532 (+91.4%)	0.568 (+78.1%)	0.603 (+41.9%)
Polar-PENet	0.589 (+111.9%)	0.637 (+99.7%)	0.638 (+50.1%)

Table 3

Quantitative comparison with SOTA RGB-based enhancement methods. Bold indicates the best performance.

Method	Type	Entropy	Gradient	Mean Intensity
S0 (Baseline)	Raw	3.7003	3.5446	6.9222
Physical Enhanced	Physics-based	5.3970	5.5451	35.5792
EnlightenGAN	RGB SOTA	1.4362	3.3261	40.2925
Zero-DCE	RGB SOTA	4.6441	11.7574	39.5884
Polar-PENet (Ours)	Fusion	5.4748	17.2531	38.3613

Table 4

Image quality metrics comparison (Polar-PENet vs PPEAM).

Quality Metric	Polar-PENet	PPEAM	Better
PSNR (dB)	19.37	18.89	Polar-PENet
SSIM	0.5487	0.5257	Polar-PENet
MSE	812.66	940.55	Polar-PENet
NRMSE	1.4632	1.5913	Polar-PENet

Table 5

Illumination characteristics comparison (Original/Polar-PENet/PPEAM).

Illumination Metric	Original	Polar-PENet	PPEAM
Mean Brightness	45.24	71.65 (+58.4%)	73.70 (+62.9%)
Brightness Std Dev	29.72	27.44 (−7.7%)	29.60 (−0.4%)
Brightness Entropy	3.36	3.75 (+11.6%)	3.97 (+18.2%)
Dark Region Ratio	0.76	0.63 (−17.1%)	0.57 (−25.0%)

Table 6

Computational efficiency comparison.

Metric	Polar-PENet	PPEAM
Avg. Processing Time (ms)	185.99	528.07
Avg. Frame Rate (FPS)	5.38	1.89 (−64.9%)
Inference Frame Rate (FPS)	183.45	N/A
Model Parameters (M)	8.73	N/A

Table 7

Ablation study of component contributions on the test set. Bold indicates the best performance.

Configuration	PSNR (dB)	SSIM	Avg Gradient	mAP (%)
Baseline (U-Net)	18.65	0.5120	13.50	58.2
Base + CBAM	18.90	0.5210	14.10	60.1
Base + PAG	19.15	0.5380	16.80	61.8
Polar-PENet (Full)	19.37	0.5487	17.25	63.7

Table 8

Entrance region image quality metrics.

Quality Metric	Polar-PENet	PPEAM	Better Method
PSNR	21.43	20.64	Polar-PENet
SSIM	0.9401	0.9055	Polar-PENet
MSE	509.57	876.21	Polar-PENet
NRMSE	0.2384	0.3029	Polar-PENet

Table 9

Tunnel middle section image quality metrics.

Quality Metric	Polar-PENet	PPEAM	Better Method
PSNR	18.76	17.78	Polar-PENet
SSIM	0.4212	0.3988	Polar-PENet
MSE	873.86	1088.95	Polar-PENet
NRMSE	1.8631	2.0218	Polar-PENet

Table 10

Exit region image quality metrics.

Quality Metric	Polar-PENet	PPEAM	Better Method
PSNR	19.73	21.01	PPEAM
SSIM	0.6984	0.7072	PPEAM
MSE	727.37	641.84	PPEAM
NRMSE	0.3781	0.3467	PPEAM

Table 11

Entrance region illumination characteristics.

Illumination Metric	Original	Polar-PENet	PPEAM
Mean Brightness	142.12	160.43 (+12.9%)	164.02 (+15.4%)
Brightness Std Dev	60.22	51.10 (−15.1%)	50.86 (−15.5%)
Contrast	0.43	0.32 (−25.6%)	0.31 (−27.9%)
Brightness Entropy	5.02	4.90 (−2.4%)	4.88 (−2.8%)
Dark Region Ratio	0.10	0.00 (−100.0%)	0.01 (−90.0%)
Bright Region Ratio	0.24	0.30 (+25.0%)	0.32 (+33.3%)

Table 12

Tunnel middle section illumination characteristics.

Illumination Metric	Original	Polar-PENet	PPEAM
Mean Brightness	9.84	39.01 (+296.4%)	40.63 (+312.9%)
Brightness Std Dev	12.06	13.30 (+10.3%)	16.06 (+33.2%)
Contrast	1.35	0.35 (−74.1%)	0.41 (−69.6%)
Brightness Entropy	2.72	3.30 (+21.3%)	3.60 (+32.4%)
Dark Region Ratio	0.99	0.86 (−13.1%)	0.79 (−20.2%)
Bright Region Ratio	0.00	0.00 (0.0%)	0.00 (0.0%)

Table 13

Exit region illumination characteristics.

Illumination Metric	Original	Polar-PENet	PPEAM
Mean Brightness	56.27	82.32 (+46.3%)	84.17 (+49.6%)
Brightness Std Dev	51.89	45.91 (−11.5%)	48.58 (−6.4%)
Contrast	1.58	0.63 (−60.1%)	0.64 (−59.5%)
Brightness Entropy	3.67	3.95 (+7.6%)	4.20 (+14.4%)
Dark Region Ratio	0.71	0.55 (−22.5%)	0.49 (−31.0%)
Bright Region Ratio	0.12	0.12 (0.0%)	0.12 (0.0%)

Table 14

Best detection performance of each YOLOv8 model.

Model	Best F1	Precision/Recall	Best Image Type
YOLOv8n	0.7534	0.7692/0.7383	Polar-PENet
YOLOv8s	0.8339	0.9328/0.7450	Polar-PENet
YOLOv8m	0.8069	0.8298/0.7852	Polar-PENet
YOLOv8l	0.8839	1.0000/0.7919	Original/Polar-PENet
YOLOv8x	0.8963	1.0000/0.8121	PPEAM

Table 15

F1 improvement relative to original images (%).

Model	Original F1	PPEAM	Polar-PENet
YOLOv8n	0.4242	+91.1%	+87.1%
YOLOv8s	0.3804	+129.0%	+123.3%
YOLOv8m	0.8593	+1.6%	+1.9%
YOLOv8l	0.8839	+0.5%	0.0%
YOLOv8x	0.8930	+0.4%	0.0%

Table 16

Best image type and average F1 for different regions.

Region	Best Image Type	Avg. F1	Best Model
Entrance	Polar-PENet	0.8234	YOLOv8l
Inside	PPEAM	0.8605	YOLOv8x
Exit	PPEAM	0.8741	YOLOv8x

Table 17

Entrance region relative improvement (%).

Model	PPEAM	Polar-PENet
YOLOv8n	+46.8%	+87.5%
YOLOv8s	+81.7%	+105.3%
YOLOv8m	+9.6%	+15.7%
YOLOv8l	+6.3%	+8.9%
YOLOv8x	+5.1%	+4.2%

Table 18

Best combination strategy (region/type/description).

Region	Best Image Type	F1	Description
Entrance	Polar-PENet	0.8234	Lightweight prioritized
Inside	PPEAM	0.8605	Resource-sufficient
Exit	PPEAM	0.8741	Resource-sufficient baseline

References

1. Zheng, Z.; Cheng, Y.; Xin, Z.; Yu, Z.; Zheng, B. Robust Perception Under Adverse Conditions for Autonomous Driving Based on Data Augmentation. IEEE Trans. Intell. Transp. Syst.; 2023; 24, pp. 13916-13929. [DOI: https://dx.doi.org/10.1109/TITS.2023.3297318]

2. Zhang, Y.; Carballo, A.; Yang, H.; Takeda, K. Perception and sensing for autonomous vehicles under adverse weather conditions: A survey. ISPRS J. Photogramm. Remote Sens.; 2021; 196, pp. 146-177. [DOI: https://dx.doi.org/10.1016/j.isprsjprs.2022.12.021]

3. Wei, Y.; Zeng, Z.; He, T.; Yu, S.; Du, Y.; Zhao, C. An Adaptive Vehicle Detection Model for Traffic Surveillance of Highway Tunnels Considering Luminance Intensity. Sensors; 2024; 24, 5912. [DOI: https://dx.doi.org/10.3390/s24185912]

4. Li, Z.; Xiang, J.; Duan, J. A low illumination target detection method based on a dynamic gradient gain allocation strategy. Dent. Sci. Rep.; 2024; 14, 29058. [DOI: https://dx.doi.org/10.1038/s41598-024-80265-w]

5. Lyu, S. Application of YOLOv8 in Monocular Downward Multiple Car Target Detection. Appl. Comput. Eng.; 2025; 121, pp. 173-178. [DOI: https://dx.doi.org/10.54254/2755-2721/2025.20045]

6. Wang, J.; Duan, S.; Zhou, M. A Lidar-Vision Fusion Target Detection Model for Low-Light Environments. Proceedings of the ICARCV; Dubai, United Arab Emirates, 12–15 December 2024; pp. 1035-1040. [DOI: https://dx.doi.org/10.1109/icarcv63323.2024.10821581]

7. Hussain, H.; Tamizharasan, P.S.; Yadav, P.K. OptiRet-Net: An Optimized Low-Light Image Enhancement Technique for CV-based Applications in Resource-Constrained Environments. ACM Trans. Intell. Syst. Technol.; 2024; 15, pp. 1-30. [DOI: https://dx.doi.org/10.1145/3700136]

8. Zhang, C.; Li, J.; Hu, H.; Bai, X. Deep learning-based low-light image enhancement method for driving behavior recognition. Proceedings of the 2024 IEEE 6th World Conference on Robotics, Computing, Sensing and Artificial Intelligence (WRCSARA); Beijing, China, 23–23 August 2024; pp. 344-349. [DOI: https://dx.doi.org/10.1109/wrcsara64167.2024.10685674]

9. Zhao, Y.; Qiao, X.; Li, N.; Pan, Q. Polarization vision. Sci. Sin. Informationis; 2024; 54, pp. 1620-1645.

10. Pierangeli, D.; Volpe, G.; Conti, C. Deep Learning Enabled Transmission of Full-Stokes Polarization Images Through Complex Media. Laser Photonics Rev.; 2024; 18, 240062. [DOI: https://dx.doi.org/10.1002/lpor.202400626]

11. Qiao, Z.; Liu, Z. Polarization Imaging. Computational Optical Imaging; Springer: Singapore, 2024; pp. 391-415. [DOI: https://dx.doi.org/10.1007/978-981-97-1455-1_14]

12. Chen, S.; Liu, P.; Ma, W.; Luo, D.; Tan, Y.; Chen, L.; Wang, S.; Zhao, Q.; Jiao, G.; Chen, W. Polarization-Enhanced Underwater Laser Range-Gated Imaging for Subaquatic Applications. Sensors; 2024; 24, 6681. [DOI: https://dx.doi.org/10.3390/s24206681]

13. Wu, L.; Zhang, X.; Chang, J.; Li, B. Underwater polarization image de-scattering utilizing a physics-driven deep learning method. Opt. Express; 2024; 32, pp. 30670-30686. [DOI: https://dx.doi.org/10.1364/OE.529711] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/39573222]

14. Wang, C.; Wu, Z.; Han, Z.; Wang, J.; Hu, H.; Li, X. Rethinking of Underwater Image Restoration Based on Circularly Polarized Light. Photonics; 2024; 11, 773. [DOI: https://dx.doi.org/10.3390/photonics11080773]

15. Zhang, H.; Gong, J.; Ren, M.; Zhou, N.; Wang, H.; Meng, Q.; Zhang, Y. Active Polarization Imaging for Cross-Linear Image Histogram Equalization and Noise Suppression in Highly Turbid Water. Photonics; 2023; 10, 145. [DOI: https://dx.doi.org/10.3390/photonics10020145]

16. Zhang, S.; Liang, J.; Jiang, Y.; Ren, L. Polarization-Based De-Scattering Imaging in Turbid Tissue-like Scattering Media. Photonics; 2023; 10, 1374. [DOI: https://dx.doi.org/10.3390/photonics10121374]

17. Xu, Y.; Fu, Y.; Li, B.; Yuan, M.; Chen, S. Experimental study on the depolarization characteristics of scattered light from rough surfaces based on near-infrared polarized supercontinuum. Proceedings of the Sixth International Conference on Optoelectronic Science and Materials (ICOSM 2024); Hefei, China, 6–8 September 2024; SPIE: Bellingham, WA, USA, 2024; 4. [DOI: https://dx.doi.org/10.1117/12.3052595]

18. Huang, F.; Zheng, J.; Liu, X.; Shen, Y.; Chen, J. Polarization of road target detection under complex weather conditions. Dent. Sci. Rep.; 2024; 14, 30348. [DOI: https://dx.doi.org/10.1038/s41598-024-80830-3]

19. Zhang, Y.; Zhang, X.; Su, Y.; Li, X.; Ma, S.; Zhang, S.; Ren, W.; Li, K. Tunnel Lining Crack Detection Method Based on Polarization 3D Imaging. Photonics; 2023; 10, 1085. [DOI: https://dx.doi.org/10.3390/photonics10101085]

20. Rahman, Z.; Yi-Fei, P.; Aamir, M.; Wali, S.; Guan, Y. Efficient image enhancement model for correcting uneven illumination images. IEEE Access; 2020; 8, pp. 109038-109053. [DOI: https://dx.doi.org/10.1109/ACCESS.2020.3001206]

21. Yang, X.; Xu, K.; Song, Y.; Zhang, Q.; Wei, X.; Lau, R.W.H. Image Correction via Deep Reciprocating HDR Transformation. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR); Long Beach, CA, USA, 15–20 June 2019; pp. 1798-1807. [DOI: https://dx.doi.org/10.1109/CVPR.2019.00189]

22. Wang, S.; Huang, X. Analysis of Methods on Evaluating the “Black Hole Effect” in Tunnel Entrance Zones. J. Highw. Transp. Res. Dev.; 2015; 32, pp. 120-126.

23. Kim, J. Vehicle Detection Using Deep Learning Technique in Tunnel Road Environments. Symmetry; 2020; 12, 2012. [DOI: https://dx.doi.org/10.3390/sym12122012]

24. Li, S.; Tan, D.; Lin, B.; Jun, G. Tunnel Luminance Detection System Based on Video. Proceedings of the 2018 IEEE International Conference on Information and Automation (ICIA); Wuyishan, China, 11–13 August 2018; pp. 957-962. [DOI: https://dx.doi.org/10.1109/ICInfA.2018.8812550]

25. Mehta, I.; Srinivasa, G.; Radhakrishnan, S.; Chawla, A.; Wilson, R.C. Gradient-Based Auto-Exposure Control Applied to a Self-Driving Car. Proceedings of the 2020 17th Conference on Computer and Robot Vision (CRV); Ottawa, ON, Canada, 13–15 May 2020; pp. 84-91. [DOI: https://dx.doi.org/10.1109/CRV50864.2020.00019]

26. Jeon, Y.; Jung, S.; Na, I.; Jeong, G.; Kim, J. Design and Implementation of Multi Exposure Smart Vehicular Camera Applying Auto Exposure Control Algorithm Based on Region of Interest. J. Korean Inst. Commun. Inf. Sci.; 2017; 42, pp. 181-192. [DOI: https://dx.doi.org/10.7840/kics.2017.42.1.181]

27. Mukherjee, R.; Chalmers, A.; Perrett, T.; Zhang, X.; Wang, Q.; Li, B. Object Detection Under Challenging Lighting Conditions Using High Dynamic Range Imagery. IEEE Access; 2021; 9, pp. 129152-129163. [DOI: https://dx.doi.org/10.1109/ACCESS.2021.3082293]

28. Mukherjee, R.; Bessa, M.; Chalmers, A. Backward Compatible Object Detection Using HDR Image Content. IEEE Access; 2020; 8, pp. 151543-151556. [DOI: https://dx.doi.org/10.1109/ACCESS.2020.3010340]

29. Wei, C.; Wang, W.; Yang, W.; Liu, J. Deep Retinex Decomposition for Low-Light Enhancement. Proceedings of the British Machine Vision Conference (BMVC); Newcastle, UK, 3–6 September 2018; pp. 1-12.

30. Jiang, Y.; Gong, X.; Liu, D.; Cheng, Y.; Fang, C.; Shen, X.; Yang, J.; Zhou, P.; Wang, Z. EnlightenGAN: Deep Light Enhancement Without Paired Supervision. IEEE Trans. Image Process.; 2021; 30, pp. 2340-2349. [DOI: https://dx.doi.org/10.1109/TIP.2021.3051462]

31. Guo, C.; Li, C.; Guo, J.; Loy, C.C.; Hou, J.; Kwong, S.; Cong, R. Zero-Reference Deep Curve Estimation for Low-Light Image Enhancement. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); Seattle, WA, USA, 13–19 June 2020; pp. 1780-1789. [DOI: https://dx.doi.org/10.1109/CVPR42600.2020.00185]

32. Zhang, Y.; Yang, L.; Fang, J.; Zhang, Q.; Zhou, Z.; Zhao, C. Lighting Environment Optimization of Highway Tunnel Entrance Based on Simulation Research. Int. J. Environ. Res. Public Health; 2019; 16, 2195. [DOI: https://dx.doi.org/10.3390/ijerph16122195] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31234339]

33. Wang, H.; Shan, Y.; Hao, T.; Zhao, X.M.; Song, S.Z.; Huang, H.; Zhang, J. Vehicle-Road Environment Perception Under Low-Visibility Condition Based on Polarization Features via Deep Learning. IEEE Trans. Intell. Transp. Syst.; 2022; 23, pp. 15353-15363. [DOI: https://dx.doi.org/10.1109/TITS.2022.3157901]

34. Sheppard, C.J.R.; Bendandi, A.; Le Gratiet, A.; Diaspro, A. Characterization of the Mueller Matrix: Purity Space and Reflectance Imaging. Photonics; 2022; 9, 88. [DOI: https://dx.doi.org/10.3390/photonics9020088]

35. Qiao, X.; Zhao, Y.; Chen, L.; Kong, S.G.; Chan, J.C. Mosaic gradient histogram for object tracking in DoFP infrared polarization imaging. ISPRS J. Photogramm. Remote Sens.; 2022; 194, pp. 108-118. [DOI: https://dx.doi.org/10.1016/j.isprsjprs.2022.10.006]

36. Zhou, H.; Zeng, X.; Bing, L.; Li, D.; Shah, S.; Liu, B.; Guo, K.Q.; Guo, Z. Polarization motivating high-performance imaging weak targets. Opt. Express; 2024; 32, pp. 12283-12298. [DOI: https://dx.doi.org/10.1364/oe.504918]

37. Zhou, C.; Teng, M.; Lyu, Y.; Li, S.; Xu, C.; Shi, B. Polarization-Aware Low-Light Image Enhancement. Proc. AAAI Conf. Artif. Intell.; 2023; 37, pp. 3666-3674. [DOI: https://dx.doi.org/10.1609/aaai.v37i3.25486]

38. Lin, B.; Fan, X.; Li, D.; Guo, Z. High-Performance Polarization Imaging Reconstruction in Scattering System under Natural Light Conditions with an Improved U-Net. Photonics; 2023; 10, 204. [DOI: https://dx.doi.org/10.3390/photonics10020204]

39. Wang, H.; Shan, Y.; Song, S.Z.; Huang, H. Low-Visibility Vehicle-Road Environment Perception Based on the Multi-Modal Visual Features Fusion of Polarization and Infrared. IEEE Trans. Intell. Transp. Syst.; 2023; 24, pp. 11599-11609. [DOI: https://dx.doi.org/10.1109/TITS.2023.3286541]

40. Wu, R.; Zhao, Y.; Li, N.; Kong, S.G. Polarization image demosaicking using polarization channel difference prior. Opt. Express; 2021; 29, pp. 22066-22079. [DOI: https://dx.doi.org/10.1364/OE.424457] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/34265979]

41. Li, N.; Zhao, Y.; Pan, Q.; Kong, S.G. No-Reference Physics-Based Quality Assessment of Polarization Images and Its Application to Demosaicking. IEEE Trans. Image Process.; 2021; 30, pp. 8983-8998. [DOI: https://dx.doi.org/10.1109/TIP.2021.3122085] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/34705645]

42. Wang, B.; Zhao, Y.; Yang, L.; Long, T. A New Polarization Image Demosaicking Method in the Frequency Domain for DoFP Polarization Imaging. Proceedings of the 2020 39th Chinese Control Conference (CCC); Shenyang, China, 27–29 July 2020; IEEE: Piscataway, NJ, USA, 2020; pp. 6644-6649. [DOI: https://dx.doi.org/10.23919/CCC50068.2020.9188824]

© 2025 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.

Polarization-Guided Deep Fusion for Real-Time Enhancement of Day–Night Tunnel Traffic Scenes: Dataset, Algorithm, and Network

Content area