DaN: A Comprehensive Semi-Real Dataset for Extreme Low-Light Image Enhancement

Abstract

Extreme low-light image enhancement (ELLIE) targets the restoration of visual quality under ultra-dim environments (<0.1 lux). Conventional image signal processing (ISP) pipelines often fail in such scenarios due to the limitations of heuristic, hand-crafted algorithms. While deep learning has advanced the field via end-to-end mapping, existing models suffer from constrained generalization and suboptimal perceptual fidelity, primarily stemming from the scarcity of large-scale, high-diversity datasets. To bridge this gap, we present the Day and Night (DaN) dataset, a semi-synthetic benchmark synthesized through a rigorous physics-based noise model. This approach effectively captures authentic noise characteristics while enabling the scalable generation of paired samples across multifaceted illumination conditions and scenes. Furthermore, we propose No Longer Vigil (NLV), a fully differentiable AI-ISP framework. By replacing traditional rigid blocks with adaptive non-linear networks, NLV facilitates scene-dependent transformations without requiring manual priors. Comprehensive evaluations demonstrate that our method significantly outshines state-of-the-art approaches, yielding a 4.15 dB gain in PSNR and a 0.026 improvement in SSIM.

1. Introduction

Extreme low-light image enhancement (ELLIE) has become essential for full-color imaging in environments with illumination below 0.1 lux, such as night surveillance, autonomous systems, and medical diagnostics. The main challenge lies in the suboptimal performance of ineffective hand-crafted algorithms. For example, although traditional image signal processing (ISP) methods use infrared cameras in dark scenes, the lack of color data prevents precise color representation [1,2,3,4,5], and while increasing ISO sensitivity or extending exposure time can improve photon capture, these hand-crafted algorithms ultimately result in increased noise and motion blur, making them ineffective for dynamic scenes [6,7,8,9,10,11,12,13].

Recent deep learning-based approaches have circumvented the difficulties of hand-crafted algorithm design through end-to-end learning. These approaches effectively suppress noise by learning to map noisy images to clear ones [14,15,16]. Moreover, there has been a notable shift in network training data from RGB three-channel images to single-channel raw Bayer array images, which contain more comprehensive information [17,18,19,20,21,22,23,24,25].

The existing methods mentioned above fundamentally depend on the quality of the training datasets [26,27,28,29,30,31,32]. While foundational benchmarks such as SID [18] and ELD [19] propelled early advancements in raw-domain restoration, their empirical utility is constrained by severe environmental homogeneity. Specifically, the static nature and restricted scene variability within these collections lead to poor algorithmic generalization when facing dynamic, semi-real low-light conditions. Furthermore, existing corpora predominantly neglect fine-grained structural restoration (e.g., text and complex textures) under sub-0.1 lux illuminance, leaving a significant gap in robust model evaluation. Specifically, the current LOL series datasets (illumination > 10 lux) are inadequate for extremely low-light conditions (<0.1 lux). Furthermore, these datasets are missing elements such as text or objects, leading to blurry images, as illustrated in Figure 1. Moreover, the existing metric scores do not accurately represent visual quality: The test datasets employ shooting strategies similar to the training datasets, including ISO, exposure time, and camera. This results in high metrics that do not align with visual image quality.

The limitations of the existing datasets highlight the need for a comprehensive dataset dedicated to extreme low-light image enhancement. In response, we propose Day and Night (DaN), a large-scale semi-synthetic dataset comprising 4200 raw format images spanning diverse scenarios, City, Suburban, Indoor, and Wild, captured at multiple times of day. We partitioned DaN into two components, among which 100 pairs were utilized for testing, and 4000 images combined with the noise model were employed for training. DaN integrates a physics-based noise model to simulate the multimodal Gaussian noise distribution in the real world, significantly reducing the time cost of data collection and expanding the data volume to infinity. Meanwhile, the independent distribution of Gaussian noise guarantees independence among the data. In addition, DaN contains data instances that exhibit character-like elements or analogous structures, which are formally defined as fine-grained features following the methodology established in [33,34]. A comparative analysis among the proposed DaN and previous datasets [18,19,24,35] is presented in

Table 1. Comparisons between our DaN dataset and existing ones. DaN comprises four primary categories throughout the entire day, from morning till night. The fine-grained features signify the existence of data dedicated to fine-grained learning. A “✓” indicates that the dataset contains images of this category, while an “✗” indicates that it does not.

Dataset	Scenarios	Morning	Daytime	Night	Fine-Grained	Illuminance	Ground Truth	Total
ELD [19]	One	✗	✓	✗	None	>0.5 lux	10	$+\infty$
LOL [35]	One	✗	✓	✗	None	>10 lux	500	1000
LOLV2-real [24]	Two	✗	✓	✓	Few	>10 lux	789	1578
LOLV2-sync [24]	Two	✓	✓	✗	None	>10 lux	1000	2000
SID [18]	Two	✗	✗	✓	Few	>0.1 lux	424	5094
DaN (Ours)	Four	✓	✓	✓	Plentiful	<0.1 lux	4200	$+\infty$

.

Furthermore, existing methods [18,19,25,36,37,38] stack multiple UNet-based deep neural networks (DNN) for denoising while disregarding other ISP components, such as white balance and color correction. This oversight results in unreal color distortion when deployed in the real world. Thus, we introduce No Longer Vigil (NLV), a pure AI-ISP pipeline. The proposed NLV comprises three components: Denoising Networks (DNs), White Balance (WB), and Color Correction (CC). These components are realized by neural networks instead of manually fixed parameters. For example, white balance is a fixed parameter matrix ${\mathbf{M}}_{WB}\in {\mathbb{R}}^{H\times 1\times 1}$ in traditional ISP, whereas in our NLV, it is predicted by a fully connected layer. The integration of ISP theory and deep learning eliminates the priors and compensates for the inexplicability of neural networks.

We conducted comprehensive experiments on the proposed DaN dataset and the NLV pipeline. We first trained and tested our NLV method on the existing SID dataset. The results demonstrate state-of-the-art performance, with our NLV outperforming existing LED methods by +4.15 dB in PSNR and +0.026 in SSIM. Subsequently, six existing methods [18,19,25,36,37,38] and our NLV method were trained on our DaN dataset and tested on the existing SID dataset. The results show that the same method performs better when trained on our DaN dataset. Finally, we conducted tests on our DaN dataset. The results demonstrated that methods with comparable performance on the existing SID dataset exhibit a considerable gap on our DaN dataset.

Our contributions are summarized below:

• We have developed a semi-synthetic dataset termed Day and Night (DaN). To our knowledge, DaN represents the first large-scale dataset specifically focused on extreme low-light image enhancement (ELLIE). Our DaN surpasses existing datasets in comprehensiveness. Its diverse and fine-grained features render it more challenging.
• We introduce No Longer Vigil (NLV), a pure AI image signal processing (AI-ISP) pipeline that reconstructs the traditional ISP pipeline by deep neural networks (DNNs) instead of merely replacing some components. NLV automatically processes signals, eliminating the priors.
• Extensive experiments show that our NLV pipeline outperforms existing methods in terms of the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). Our DaN can serve as a benchmark for extremely low-light image processing.

2. Related Works

2.1. ELLIE Datasets

Existing datasets like SID [18] and LOL [35] initiated raw sensor data collection, but were restricted by their small size and lack of variety. In contrast, the LOLV2 and ELD datasets [24,35] use generative models to increase data volume. However, they involve implicit relations, making it seem as if the model has already “seen” the test sets. Our DaN dataset stands out because (1) our DaN is 10 times larger than the previous one, encompassing broader types, periods, and fine-grained features, and (2) it ensures independence without implicit mappings, making them more challenging and suitable for benchmarking.

2.2. ELLIE Approaches

Existing approaches can be classified into two categories: traditional ISP pipeline and learning-based methods.

• Traditional ISP pipeline: The traditional ISP pipeline [6,7,8,9,10,11,12,13] encompasses multiple components, such as denoising, demosaic, color correction, and white balance. These components are artificially designed and equipped with fixed parameters. The well-designed ISP pipeline is fabricated into chips and is prevalently present in major camera manufacturers. This kind of pipeline is hardware-dependent and demonstrates poor performance in extremely low light due to the constraints of CMOS.
• Deep learning-based methods: Deep learning-based methods usually stack one to three DNNs for learning implicit noisy-normal image mapping. Various models have since been put forward, notably [18,25,36,37,39,40,41]. These models merely focus on denoising yet overlook other ISP components, thereby causing blurriness, color distortion, and the loss of fine-grained features. Our NLV method, in conjunction with ISP theory, has redesigned a pure AI-ISP pipeline suitable for ELLIE. The pipeline parameters are all adaptive, bypassing the inconvenience parameter tuning.

2.3. Evaluation Metrics

The current evaluation metrics include peak signal-to-noise ratio (PSNR) [42], measuring the logarithmic ratio of peak signal power to noise level, and the structural similarity index (SSIM) [43], evaluating perceptual fidelity through structural analysis. Additionally, researchers perform subjective evaluations of visual details.

3. Materials: The Proposed DaN Dataset

This section introduces our proposed DaN (Day and Night) dataset. First, we outline the methodology for data collection and annotation. Then, we present the proposed noise model designed to generate noisy images. Finally, we provide a quantitative assessment of our DaN dataset.

3.1. Dataset Collection

We collected 4731 raw Bayer images for the training set, which were categorized into indoor and outdoor subsets. The indoor images encompassed as many chromatic objects as possible. The outdoor images were collected from cities, suburbs, and wild landscapes, enhancing generalization. The brand of cameras is Sony A7S2, and all images were stored in the “.ARW” raw format. Furthermore, we deliberately captured images containing plentiful fine-grained features. The data collection was accomplished by five people.

We analyzed the quality and quantity of the images and classified them into four categories: Indoor (Object) and Outdoor (City, Suburb, Wild). Then, we conducted a thorough review to eliminate highly repetitive images. We excluded those containing sensitive content, such as pornography and political material, and any content infringing on privacy rights. After meticulous screening, we ultimately chose 4000 raw Bayer images as the training set of our DaN.

We also gathered 100 pairs of raw Bayer images for the test set. Each pair comprises a noisy image captured in a highly dim light environment (<0.1 lux) and a clean image taken during the daytime (>100 lux) that has been manually aligned. These images were taken by five different mobile phones and are unrelated to the images in the training set.

3.2. Annotation Pipeline

We employ a noise model to generate noise and mark it on the Ground Truth as a label. Theoretically, the noise model expands the training set to an infinite size. The whole process includes three steps, noise modeling, noise parameter selection, and manual annotation, as shown in Figure 2.

3.2.1. Noise Modeling

The noise of camera imaging in low illumination environments is mainly shot noise, dark noise, and various additive noise [44]. Shot noise follows the Poisson distribution, is related to the input signal, and can be expressed as [45]

$${\mathrm{N}}_{\mathrm{shot}}\sim \mathrm{Poisson}[\mathrm{T}·\alpha ·\Phi ],$$

(1)

where T represents the exposure time, measured in second (S). The $\alpha$ denotes the photon conversion rate, and $\Phi$ indicates the luminous flux. Dark noise adheres to Poisson distribution [46]:

$${\mathrm{N}}_{\mathrm{dark}}\sim \mathrm{Poisson}[\mathrm{T}·\Psi ],$$

(2)

where $\Psi$ is the temperature, measured in Kelvin (K). Additive noise encompasses read noise and noise generated during the digital-to-analog conversion process, and it conforms to a Gaussian distribution. Additive noise is

$${\mathrm{N}}_{\mathrm{add}}\sim \mathrm{N}[0,{\sigma }^2],$$

(3)

${\sigma }^2$ represents the variance of the additive noise. Therefore, The signal of the camera is represented as follows:

$$N=g(I+N_{Shot})+g·N_{dark}+N_{add},$$

(4)

where I is the original image without noise, and g is an amplification factor of the camera, pertains to the global gain, and is proportional to ISO. N follows Gaussian distribution.

3.2.2. Noise Parameters Selection

According to formulation (4), g, the variances of $N_{add}$, $N_{shot}$, and $N_{dark}$ need to be calibrated, and the mean and variance of the signal are

$$E\left(N\right)=\mathrm{T}·(a·\Phi +\Psi )·g,$$

(5)

$$\sigma {\left(N\right)}^2=\mathrm{T}·(a·\Phi +\Psi )·g^2+N_{\mathrm{add}}^2,$$

(6)

If we take a dark frame image, the signal is 0, and thus, the mean of the image can be expressed as follows:

$$E\left(I\right)=\mathrm{T}·(a·0+\Psi )·g=\mathrm{T}·\Psi ·g=g·N_{dark},$$

(7)

this expression represents the $g·N_{dark}$, which will be used to replace the dark noise in the subsequent processing. According to Formulas (5) and (6),

$$\sigma {\left(I\right)}^2=E\left(I\right)·g+N_{\mathrm{add}}^2,$$

(8)

The mean and variance of image pixels exhibit a linear relationship, with the slope being g and the intercept representing the additive noise variance ${\sigma }_{add}^2$. So we captured many gray scale charts, calculated the mean and variance of each pixel, and plotted the mean-variance coordinate graph to complete the calibration, as shown in Figure 2. $N_{Shot}$ can be inversely deduced once g, ${\sigma }_{add}^2$, and $N_{dark}$ have been calibrated. The noise parameters are distinct under different configurations. Therefore, we have calibrated a total of 24 sets of noise parameters altogether. Parameters are randomly chosen when generating noise according to (4).

3.2.3. Manual Annotation

The noise generated by the physics-based noise model is distributed independently, and theoretically, no implicit mapping exists. In order to match the real world, we conduct an iterative selection of the noise that is most consistent with the real environment and align it spatially with the real data through cross-correlation matching [47], thereby minimizing the spatial offset and overexposure.

3.3. Dataset Statistics

Our DaN dataset comprises 4200 Ground Truths with a resolution ranging from 1440 × 2560 to 4320 × 7680 pixels, extracted from cameras in raw format. Among them, 4000 images are used for training, and in combination with the noise model, there are theoretically an infinite number of training data pairs. 100 pairs of real images are used for evaluation. The comprehensive details of the dataset are shown in Figure 3.

Ethical Statement

We hold the copyright for the photographs that we collect and adhere strictly to the requirements established in the data collection and annotation process. To further enhance our ethical compliance, we will take the following measures: (1) Implement rigorous data anonymization techniques to safeguard personal information. (2) Ensure transparency regarding data sources and collection methods in our evaluations. (3) Commit to ongoing reviews and remain prepared to delete or modify any data collected from sources that may be considered ethically inappropriate or that have not provided the requisite authorization.

4. Methods

In extremely low-light conditions, the traditional image signal processing (ISP) pipeline often produces poor images. Current learning-based models [18,25,36,37,39,40,41] that utilize multiple layers of deep neural networks (DNN) for image denoising perform only one stage within the ISP process. The inadequacy of the ISP components leads to blurred images, color distortion, and lack of fine-grained features. To tackle these problems, we introduce No Longer Vigil (NLV), an entirely AI-based ISP framework. NLV integrates and reconfigures the homogeneous linear elements found in conventional ISPs. It consists of three main components: a Denoising Network (DN), White Balance (WB), and Color Correction (CC). The primary innovations of our NLV pipeline over existing methods are: (1) Our NLV has comprehensively reorganized and redesigned the structure and sequence of traditional ISP components. (2) All parameters within NLV are self-adjusting, contrary to the static, hand-crafted parameters of traditional ISPs. The structure of the proposed NLV pipeline is illustrated in Figure 4.

4.1. Denoising Network (DN)

In the conventional ISP pipeline, the denoising process is executed on RGB images containing three channels and is typically performed towards the latter stages of the pipeline. Conversely, in the proposed NLV framework, we prioritize denoising and combine it with black level correction (BLC), lens shading correction (LSC), and wavelet transformation. The initial raw image ${\mathbf{R}}_{in}\in {\mathbb{R}}^{H\times W\times 1}$ is processed to obtain the denoised raw Bayer images ${\mathbf{R}}_{out}\in {\mathbb{R}}^{H\times W\times 1}$.

$$R_{out}=Con{v}_{up}(Con{v}_{down}(R_{in},{\phi }_{dn1}),{\phi }_{dn2}),$$

(9)

where $Con{v}_{up}$ and $Con{v}_{down}$ denote the upsampling convolutional network and downsampling convolutional network, respectively. ${\phi }_{dn}$ represents the parameters. Both $Con{v}_{up}$ and $Con{v}_{down}$ have 4 convolutional layers, each with a kernel of $(3\times 3)$, a stride of 1, and a padding of 0.

Unlike standard spatial filters, the wavelet transform facilitates a multi-scale time-frequency analysis. By projecting the features into the frequency domain, it separates complex noise distributions from the underlying image content. The mathematical formulation utilized in our network can be expressed as

$$X(x,y)=\sqrt{x}\sum _{n=-\infty }^{\infty }(y-nx)\psi \left({\displaystyle \frac{n}{x}}\right),$$

(10)

where x defines the scaling factor that dictates the resolution, y acts as the translation parameter for spatial shifting, and $\psi$ represents the selected mother wavelet function.

4.2. White Balance (WB)

At varying temperatures, the pixel values for the same color differ [48]. To correct for this, white balance is utilized. Standard signal processing systems typically compensate for chromatic aberrations by applying rigid, pre-calibrated temperature configurations, such as a standard 5500 K daylight profile, which lack adaptability in complex dark scenes. In contrast, our NLV pipeline uses a fully connected layer to estimate the color temperature parameter. The expression for WB is

$${\phi }_{WB}=F{C}_{WB}(R_{out},{\phi }_{FC}),$$

(11)

$$R_{WB}=R_{out}·{\phi }_{WB},$$

(12)

where ${\phi }_{FC}$ represents the parameter of the fully connected layer, and $R_{WB}$ is the output. The WB contains a convolutional layer with a 3 × 3 convolution kernel, a stride of 1, and a padding of 0.

4.3. Color Correction (CC)

Standard signal processors typically decouple chromatic adaptations into distinct gamut mapping and gamma adjustment phases, which can compound errors under extreme noise. To circumvent this, our NLV architecture intrinsically fuses these non-linear transformations into a cohesive, data-driven mapping step. The generalized color restoration is parameterized by a dedicated convolutional mechanism, defined as

$$R_{CC}=Con{v}_{CC}(R_{WB},{\phi }_{CC}),$$

(13)

where $R_{CC}$ denotes the output of our NLV pipeline, ${\phi }_{CC}$ represents the learnable parameters of the color correction, $Con{v}_{CC}$ represents the color correction network. The $R_{CC}$ is subsequently downsampled to a three-channel RGB image $\mathbf{I}\in {\mathbb{R}}^{\frac{H}{2}\times \frac{W}{2}\times 3}$ by the demosaic algorithm, and can be saved in image formats such as “PNG” or “JPEG”. CC consists of two convolutional layers with 3×3 convolution kernels, stride 1, and padding 0.

5. Results

5.1. Training Details

All baseline methods were trained on the existing SID dataset and our DaN dataset, respectively. For fairness, we unified the training protocols. Model optimization was driven by the L1 objective function, iteratively minimized via the Adam algorithm. We applied a continuous learning rate scheduling strategy, initiating at a base of ${10}^{-4}$ and decaying to a lower bound of ${10}^{-6}$ across 100 training cycles. To facilitate robust feature extraction, spatial dimensions of the input batches were standardized to $512\times 512$ pixel patches through a random cropping strategy. Hardware acceleration for all network training and evaluation phases was provided by a single 48 GB NVIDIA A40 Tensor Core GPU.

5.2. Comparisons with State-of-the-Art Methods

We utilized the existing SID dataset as the foundational dataset. For comparison purposes, we employed six methods [18,19,25,36,37,38] sourced from conferences and journals of 2023–2024, including ICCV, CVPR, and TPAMI. Testing on the SID dataset reveals that our NLV pipeline outperforms the current SOTA methods. Moreover, evaluations on our DaN dataset demonstrate that DaN boosts the model’s performance and provides a more comprehensive representation of the model’s capabilities, establishing it as a viable benchmark. The findings are displayed in

Table 2. Cross experiments on existing datasets and the proposed DaN dataset.

Method	Train Dataset	LOLV1			LOLV2-Real			LOLV2-Sync			SID			DaN
		PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
SID [18]	SID	27.35	0.736	0.462	28.24	0.721	0.478	29.04	0.776	0.398	30.76	0.810	0.351	20.98	0.512	0.672
	DaN (ours)	28.15	0.742	0.441	27.61	0.742	0.445	29.26	0.735	0.412	32.15	0.851	0.305	21.06	0.525	0.658
ELD [19]	SID	35.42	0.825	0.158	35.12	0.862	0.142	36.11	0.819	0.176	36.30	0.872	0.145	24.16	0.587	0.523
	DaN (ours)	37.21	0.861	0.112	37.65	0.871	0.108	38.01	0.891	0.098	39.44	0.901	0.086	24.23	0.571	0.509
ExpDiff [37]	SID	35.21	0.816	0.167	34.52	0.762	0.202	37.26	0.851	0.126	35.00	0.808	0.186	26.62	0.604	0.455
	DaN (ours)	36.04	0.821	0.149	35.26	0.801	0.185	40.02	0.910	0.072	38.02	0.847	0.121	27.02	0.627	0.423
Retinexformer [25]	SID	35.12	0.825	0.162	32.85	0.771	0.238	35.26	0.841	0.152	34.44	0.826	0.196	22.25	0.526	0.602
	DaN (ours)	35.23	0.842	0.148	33.74	0.829	0.198	37.68	0.869	0.115	36.12	0.842	0.145	22.85	0.539	0.578
PWN [36]	SID	35.95	0.827	0.144	34.92	0.826	0.156	36.82	0.879	0.108	37.87	0.834	0.132	25.12	0.638	0.412
	DaN (ours)	36.88	0.821	0.138	36.92	0.882	0.102	37.12	0.862	0.125	40.06	0.922	0.078	26.80	0.631	0.382
LED [38]	SID	36.81	0.902	0.089	37.11	0.897	0.096	35.91	0.843	0.148	39.34	0.931	0.062	26.94	0.649	0.364
	DaN (ours)	36.72	0.894	0.095	38.28	0.911	0.078	37.20	0.870	0.118	40.26	0.930	0.057	27.49	0.668	0.345
NLV (ours)	SID	37.11	0.895	0.083	37.58	0.910	0.071	36.82	0.853	0.124	42.87	0.943	0.045	28.36	0.709	0.312
	DaN (ours)	38.06	0.902	0.074	39.62	0.920	0.064	37.20	0.912	0.105	43.49	0.957	0.038	29.52	0.732	0.286

.

5.3. Case Study

To thoroughly assess the functional aspects of our proposed DaN dataset and NLV pipeline, we conducted an in-depth visual study using six representative samples. We compared these samples against three leading extreme low-light enhancement techniques: ELD [19], RetinexFormer [25], and LED [38]. Although these baseline methods show similar results on standard quantitative metrics (PSNR: 36.00 ± 1.00 dB, SSIM: 0.870 ± 0.030), our qualitative analysis highlighted significant perceptual differences in key regions. Detailed visual analysis is shown in Figure 5.

Visual inspection reveals that our NLV pipeline, when trained on the proposed DaN dataset, outperforms current methods that are trained on existing datasets. It particularly excels in color restoration, as illustrated in the sample comparison in the fourth row, and in the restoration of fine-grained features, such as the characters depicted in the first row. Existing methods are plagued by issues like blurriness, excessive exposure, and an inability to restore characters effectively. Although our NLV pipeline mitigates these challenges, it does not solve them perfectly, indicating that the DaN dataset presents significant challenges. Thus, we conclude that our NLV pipeline exhibits superior performance in the restoration of fine-grained features and color, which can be attributed to the incorporation of the Denoising Network (DN), White Balance (WB), and Color Correction (CC). Furthermore, the DaN datasets proves to be a more challenging benchmark for ELLIE.

5.4. Ablation Study

• Our NLV pipeline components: The boxes of different colors are zoom areas, used for comparing the details’ restoration. In the ablation experiment, the effects of the two components of the NLV were investigated: white balance and color correction. The test was conducted on SID and DaN dataset, and the experimental results are presented in

  Table 3. The ablation experiments of the NLV components.

  Method	SID		DaN
  	PSNR	SSIM	PSNR	SSIM
  w/o WB	22.42	0.261	20.84	0.304
  w/o CC	23.87	0.449	22.49	0.430
  NLV (Ours)	43.49	0.957	29.52	0.732

  . The results demonstrate that two components significantly influence the ISP pipeline.
• Wavelet Transform: To validate the necessity of the wavelet module within our pipeline, we benchmarked it against alternative frequency-domain techniques, namely the Fourier, Hilbert, short-time Fourier, Wigner, and Radon transforms. The quantitative results indicate that the wavelet approach preserves structural fidelity most effectively, as shown in

  Table 4. The ablation experiments of wavelet transform.

  Method	SID		DaN
  	PSNR	SSIM	PSNR	SSIM
  w/o transform	40.63	0.912	26.63	0.701
  Fourier	41.03	0.903	27.03	0.703
  Hilbert	41.12	0.891	27.12	0.719
  Short-time Fourier	40.21	0.912	27.21	0.712
  Wigner	42.51	0.942	28.51	0.715
  Radon	42.15	0.936	28.15	0.726
  Wavelet (Ours)	43.49	0.957	29.52	0.732

  . We attribute this advantage to its multi-resolution characteristics, which inherently decouple high-frequency degradation from critical image structures, rendering it highly effective for restoring textures and edges in severely degraded low-light environments.

5.5. Generalization Evaluation

To further evaluate the model’s robustness, we conducted experiments in an industrial setting—specifically, robotic gas metal arc welding (GMAW) under extremely low-light conditions, where illumination is limited to the molten pool region. Post-weld quality assessment relies heavily on visual inspection of the weld seam; however, images captured in such low-light environments suffer from severe degradation in contrast, noise, and detail fidelity, thereby compromising defect detection accuracy. To assess semi-real deployability and generalization capability, we implemented the proposed method on an embedded NVIDIA Jetson Orin platform. Quantitative and qualitative evaluations demonstrate that our model significantly enhances image visual quality, restoring structural details, improving contrast, and suppressing noise. As shown in the Figure 6, the enhanced images enable more reliable downstream analysis: defect detection accuracy improves by 12.92% compared to results obtained using the original low-quality images.

6. Discussion

Are DaN datasets genuinely significant?The experimental findings reveal that the DaN dataset adequately addresses the limitations inherent in existing datasets, resulting in a marked improvement in training effectiveness, as illustrated in Table 2. We propose that the key advantage of DaN lies in its exceptional generalization capabilities. In contrast, current SID datasets are restricted by their limited scale, a factor often leading to overfitting when applied to train complex models with numerous parameters. Such an overfitting can substantially compromise training outcomes. However, the extensive data provided by DaN substantially reduce the risk of overfitting. Furthermore, the multicategory design of DaN allows it to attain outstanding performance, even with unfamiliar data. A cross-visualization analysis was performed to examine the link between the DaN dataset and the NLV pipeline, as shown in Figure 7. The analysis indicates that both the DaN dataset and the NLV pipeline enhance the performance of the model, with a greater dependence of visual outcomes on the DaN dataset.

How does the fully processing NLV provide advantages over a single processing? The most notable aspect of NLV is its redesigned AI-ISP framework, which significantly diverges from traditional methods that focus primarily on developing denoising networks. These traditional methods, with their exclusive focus on noise reduction, overlooked the incorporation of other crucial ISP components, thus greatly constraining the overall efficacy of the model. This constraint emerges from the intricate nature of the ISP pipeline, which includes various non-linear operations that are not straightforwardly translatable to deep neural networks. Furthermore, the shared parameters native to convolutional layers add complexity to implementing essential tasks, such as white balance and color correction. Consequently, the comprehensive NLV process, which amalgamates multiple processing stages, exhibits a distinct advantage over single-process networks that concentrate only on denoising deep neural networks.

How do the components of NLV interact and collaborate? In conventional ISP systems, the individual components usually function alone without mutual connections, such as denoising and color correction. However, the components within NLV show interdependence, working harmoniously to create an integrated system, as shown in Table 3. Removing any component from this system will cause the entire model to collapse. Implicit associations between our NLV components are not present in traditional ISP.

7. Conclusions

This paper introduces Day and Night (DaN), a comprehensive dataset designed for extreme low-light image enhancement (ELLIE). When combined with our noise model, DaN provides a limitless pool of training data. Furthermore, we have developed an entirely Artificial Intelligence-based image signal processing (AI-ISP) pipeline known as No Longer Vigil (NLV). This novel pipeline has redefined the AI-ISP framework, suggesting a new path for future research. The NLV pipeline improves image quality in extremely low-light environments and is expected to be beneficial for applications like night surveillance, cave exploration, and maritime navigation. Nonetheless, our investigation of the NLV components is potentially not exhaustive; for instance, the denoising network could be complemented with conventional filtering techniques. Our future plan includes implementing the proposed NLV model in real camera systems and testing these devices across various scenarios.