Dark Lighter Pro-Net: A Deep Learning Model for Enhancement of Low Light Underwater Images

The water holds a wealth of useful materials. vice maintenance, object detection, search and salvage, and image processing are all operations that rely on underwater images [1]. Due to the presence of suspended particles and the attenuation of light under water, the underwater optical images exhibit poor quality. These defects include color distortion, low contrast, and blurred details [2]. Improving and restoring the deteriorating underwater pictures is necessary to acquire more underwater information [3].

Nevertheless, capturing clear, detailed underwater images isn't without its challenges. Color shifts, distorted details, and reduced contrast are common complaints from underwater images as a result of light dispersion and the challenging underwater conditions [4]. Clearness, color accuracy, and sharpness are all negatively affected, especially in murky or deep water conditions [5]. Consequently, Underwater Image Enhancement (UIE) techniques are vital for improving underwater image quality, providing clearer pictures for uses like monitoring and inspections.

There was a time, over a decade ago, when most methods for improving underwater images relied on physical models or did not use physical models at all [6]. But when convolutional neural networks first came out, researchers started to pay more attention to deep learning-based UIE techniques. Using historical data to estimate the parameters utilized in image generation, physical-model-based approaches can increase the quality of underwater images by turning the physical process backwards [7]. However, due to the ever-changing nature of the marine environment, these stated traits may not be applicable in all situations [8]. To increase features like brightness, saturation, and contrast, most non-physical model-based techniques use image processing to change the pixel values [9].

Although these methods work, they have numerous drawbacks, such as being very conditional, losing information, experiencing color shifts, and having trouble restoring visual attributes that have been lost [10]. Improved underwater photography is a direct result of the widespread use of deep learning techniques, which have proven effective in fields such as high-resolution images, segmentation, and object recognition [11]. In this domain, Generative Adversarial Network (GAN) and Convolutional Neural Network (CNN) architectures are king when it comes to deep learning models [12]. In order to directly comprehend the connection between real and degraded images from training data, they make use of their powerful fitting skills [13]. Many issues remain unresolved in UIE that relies on deep learning, despite considerable progress in this area [14]. Uneven augmentation may result from the fact that most existing approaches focus on improving images in the spatial domain, which is where most underwater images are degraded [15]. As a result, there may be cases where the image's brightness is increased at the cost of noise, or where the image's clarity is diminished to improve color restoration [16]. The capturing of underwater images is a difficult task, and the process of underwater image capturing and its effects are shown in Figure 1.

Figure 1. Underwater image capturing

Based on the study, this work uses CNN's deep learning design to present a UWI improvement technique based on enhanced DarkLighter Pro-Net. Low-quality underwater photos are fed into the end-to-end framework of the approach, which produces processed, enhanced underwater photos as output. The following is a summary of the study's main contributions:

• An underwater image enhancement structure is designed, and a Dark Lighter Pro Net (DLPN) framework is proposed.
• The network model is processed on the standard Underwater Image Enhancement Benchmark (UIEB) dataset. The images in this dataset are captured under different environmental conditions underwater.
• Attention Guided Enhancement Network (AGE) is designed to perform enhancement on darker areas utilizing the channel and spatial network.
• After processing the images using the DLPN model, the metric is evaluated to prove the effectiveness of the designed model.

The proposed DLPN is unique since it is a CNN framework that integrates attention-guided enhancement, physics-based color restoration, and multi-stage residual refinement. Instead of using uniform enhancement like traditional models do, DLPN incorporates an AGE module that uses spatial and channel attention to specifically improve darker parts while preventing overexposure in brighter areas. Also, the Physics-Guided Color Restoration (PGCR) part is great for restoring natural colors that get messed up in low-light underwater settings since it precisely models water's light absorption and wavelength-dependent dispersion. Using skip connections, a Multi-Stage Residual Refinement (MRR) technique progressively refines tiny features over various scales, further preserving structural integrity. Unlike previous approaches, DLPN takes a holistic view, allowing it to tackle light loss, color distortion, and texture deterioration all at once.

To circumvent these restrictions, the proposed DLPN framework incorporates three supplementary techniques. To avoid overexposing lighter areas, AGE adaptively brightens darker ones. In contrast to earlier deep learning methods, PGCR takes absorption and scattering effects into account while restoring natural hues [19]. Last but not least, MRR solves the problem of texture degradation that often occurs in CNN and GAN-based approaches by gradually refining tiny details across scales. Through this integration, DLPN can tackle illumination loss, color distortion, and structural detail preservation all at once, outperforming state-of-the-art methods [20].

The images captured underwater suffer from a lower level of illumination [21], distortion in the colour of the image due to water scattering, and the noise present in the water environment [22]. Some of the traditional enhancement algorithms discussed in Section 2 failed to perform in extremely low-light conditions. The improvement in contrast and visibility is achieved successfully using a deep learning enhancement model. The darker lighter pro net model focuses on lower light and illumination-degraded images. The framework and design structure of DLPN are shown in Figure 2.

Figure 2. Framework of the proposed model

The mathematical notations are discussed here:

$\mathrm{I} \in[0,1]^{\mathrm{H} \times \mathrm{W} \times 3}$ indicates an observed low-light underwater RGB image (input).

$\mathrm{J} \in[0,1]^{\mathrm{H} \times \mathrm{W} \times 3 \mathrm{~J}}$ indicates a restored/enhanced RGB image (ground truth or target).

J represents network output (enhanced image).

X indicates pixel location, $\mathrm{x} \in \Omega$, where $\Omega$ indexes the $\mathrm{H} \times W$ grid.

$\mathrm{F}^{(1)} \in \mathrm{R}^{\mathrm{H} 1 \times \mathrm{W} 1 \times \mathrm{Cl}}$ indicates the feature tensor at layer 1.

$\mathrm{A}_5(\mathrm{x}) \in[0,1]$ represents the spatial attention scalar at pixel x.

$\mathrm{A}_{\mathrm{c}} \in[0,1]^{\mathrm{C}}$ represents the channel attention vector (perchannel gating).

$\mathrm{A}(\mathrm{x})$ indicates combined attention applied to features (explained below).

$\mathrm{d}(\mathrm{x})$ represents the depth/scene distance at pixel x (meters or relative scale).

$\beta_\lambda$ and $\alpha_\lambda$ indicate the wavelength-dependent attenuation coefficient for color channel $\lambda \in\{\mathrm{R}, \mathrm{G}, \mathrm{B}\}$.

$\mathrm{B}_\lambda$ represents the global background.

L indicates the total training loss.

3.1 Preprocessing

In this stage, the noise in the input data is reduced, and the values of the pixels are normalized for robust enhancement. The underwater images have noise because of absorption in water and the scattering problem. A non-local mean denoising filter is used to remove the noise in the image, and the structure of the image is preserved. The denoising of the image is performed using Eq. (1).

$I_D(x, y)=\sum_{p \in \mathcal{N}(x, y)} w(x, y, p) I(p)$            (1)

Here, the term $\mathcal{N}(x, y)$ is the neighbour pixel of the image, and the term $w(x, y, p)$ is the similar range of weights between the pixels.

The normalization helps the value of pixels to fall in a specific area for further processing of the image and ensures to have a consistent progress when working with a neural network. The pixel values are normalized between the range of 0 to 1, which allows the training process in networks to be stable. The normalization of the image given in Eq. (2).

$I_N=\frac{I_D-I_{\min }}{I_{\max }-I_{\min }}$           (2)

3.2 Design of DLPN

The DLPN model is designed utilizing many aspects, like attention, physics-based restoration, and residual refinement for improving the quality of the enhanced image.

A. Hybrid Multiscale Feature Extraction (HMFE)

In this process, the hierarchical features are extracted using the convolutional layers and transformer layers. This stage extracts fine features and coarse structure of the image by preserving the global structure of the image. The convolution extraction of the feature process utilizes three layers and is defined as Eqs. (3)-(5):

$F_1=\sigma\left(W_1 \times I_N+b_1\right)$           (3)

$F_2=\sigma\left(W_2 \times F_1+b_2\right)$         (4)

$F_3=\sigma\left(W_3 \times F_2+b_3\right)$      (5)

qThe transformer extraction of features helps in capturing of global illumination patterns of the image. The use of self-attention enhances the global structure of the image without losing the spatial details. The delf-attention factor is given in Eq. (6).

$A(Q, K, V)=\operatorname{Softmax}\left(\frac{Q K^T}{\sqrt{d_k}}\right) V$         (6)

The final features in this stage are a combination of CNN and the transformer features and are given in Eq. (7).

$ Final_F= Concat \left(F_3, A(Q, K, V)\right)$          (7)

B. Adaptive Dual Batch Normalization (ADBN)

The contrast of the input image is normalized adaptively using ADBN. This improves the training process of underwater images in the network. This allows us to normalize the contrast of the image by applying a specific batch normalization for the regions that are darker and for brighter regions. By this process, the over-enhancement of the bright region is prevented by enhancing the dark regions. The ADBN is evaluated using Eq. (8).

$I_{\text {Norm}}=\gamma_d \frac{\left(I-\mu_d\right)}{\sigma_d}+\beta_d+\gamma_d \frac{\left(I-\mu_b\right)}{\sigma_b}+\beta_b$              (8)

C. AGE

In this stage, the regions of the image are darker and enhanced with the help of attention. Two attention models are considered, one is spatial attention $\left(S_A\right)$, which improves the darker region of the image, and the other one is channel attention $\left(C_A\right)$  which enhances the colour contrast of the image. The combination of spatial and channel attention provides a final attention map.

$S_A=\sigma\left(W_s * I_{\text {norm}}+b_s\right)$          (9)

$C_A=\operatorname{Softmax}\left(W_c * G A P\left(I_{\text {noxm}}\right)\right)$            (10)

The final enhanced attention model is given in Eq. (11).

$I_{enhanced}=S_A \cdot C_A \cdot I_{norm}$                (11)

D. Physics Guided Colour Restoration (PGCR)

The PGCR allows a model of physics in evaluating factors like light absorption and scattering. This helps in restoring the natural colour in the image by understanding the concept of physics in the underwater conditions. The restoration of true colours of the underwater image is given in Eq. (12).

$I_{restared}=I_{enhanced} \times e^{d \cdot \lambda}$              (12)

where, the term d is said to be the depth attenuation factor and term λ is the wavelength-dependent coefficient of absorption.

The colours of the image restoration underwater will be performed depending on the physics.

E. MRR

The details of the image are refined using residual learning. The MRR uses a three-level residual learning, which provides a higher scope of detailed enhancement. At every level, a skip connection in the residual network refines the details of the image. The final output of the enhanced image is given by Eq. (13).

$I_{\text {final }}=I_{\text {restored}}+R_1+R_2+R_3$            (13)

Here, the terms $R_1, R_2, R_3$  are the residual enhancement at different scales.

3.3 Loss function

The loss function is evaluated to minimize the errors in the enhancements. The loss function is a measure of the output image obtained from the model with respect to the ground truth image results. The loss function in this model is evaluated using L1 loss, SSIM, and perceptual loss and is given in Eq. (14).

$ Loss =\lambda_1 \times L 1+\lambda_2 \times S S I M+\lambda_3 \times Perceptual \, loss $          (14)

L1 is given as L1 = Mean [abs(true output-predicted output)], and SSIM is given as SSIM = 1 - SSIM (true output, predicted output).

Here the terms $\lambda_1, \lambda_2$, and $\lambda_3$ are the constant values. In our design the term $\lambda_1=0.6, \lambda_2=0.3$ and $\lambda_3=0.1$ are considered to evaluate the loss function.

At different spatial dimensions, degradations occur underwater in different ways. On one hand, ambient lighting and color casts are large-scale processes, whereas small suspended particles induce local dispersion. Both global illumination patterns and fine texture may go unnoticed by a CNN operating on a single scale. By merging convolutional multi-scale features with a transformer/self-attention branch that records long-range lighting patterns, HMFE can extract both local (fine) and global (coarse) information. This makes the correction of colors and lighting more consistent worldwide while also improving local detail recovery (edges, textures).

Because degradations under the water might manifest on both a local and global scale, the first part, HMFE, is crucial. Local scattering is caused by suspended particles, while global color shifts and illumination loss are caused by light attenuation. By utilizing transformer-based self-attention [23], HMFE can simulate global lighting patterns and extract local texture features using convolutional layers. This synergy strikes a good balance between regional sharpness and general aesthetic cohesion by restoring fine details to the improved image without sacrificing global color uniformity [24].

In order to fix the issue of underwater photographs having inconsistent lighting, the ADBN module is used. When using a traditional batch normalization method, just one normalization is applied to the entire image. This can cause some parts to be too enhanced while others are under-normalized [25]. With the use of a learnt mask, ADBN introduces independent normalization statistics for light and dark areas, allowing for adaptive normalization to take place. This guarantees steady training and better contrast preservation by preventing over-exposure of already bright areas and allowing effective correction of darker portions.

To fix the areas that need it the most, the AGE module applies enhancements selectively. Global amplification can make previously clear parts of an underwater picture noisier, and not all parts of the image require equal improvement. AGE prevents the network from excessively amplifying noise while still prioritizing crucial structures and color adjustments. A composite loss function is used to train DLPN in an end-to-end fashion. Losses at the pixel level (L) are a part of the overall loss. Rather than relying on individual supervision, the attention modules improve their weights with the rest of the network through backpropagation. This guarantees that both the spatial and channel attention adjust dynamically to the data.

The PGCR module is a crucial component of the DLPN framework because it corrects colour distortions by explicitly modeling underwater light absorption and scattering. In this model, the observed pixel intensity is expressed as Iλ(x) = Jλ(x)tλ(x) + Bλ(1 − tλ(x)) is the captured intensity in channel λÎ{R,G,B} is the true scene radiance, B_λ is the global background light, and t_λ(x) = exp(−βλd(x)) is the transmission that depends on the scene depth d(x) and the wavelength-dependent attenuation coefficient β_λ. By inverting this model, the true colour values J_λ(x) can be recovered. A central challenge in PGCR is estimating the depth d(x) or transmission t_λ(x), since ground-truth depth maps are rarely available for real underwater datasets. To address this, several strategies are employed. One approach is to train a small CNN sub-network to estimate relative depth maps directly, enabling the network to learn depth cues in a weakly supervised way. Alternatively, the model can bypass explicit depth estimation by directly predicting transmission maps for each channel through a lightweight prediction head, ensuring differentiability and physical interpretability. Depth can also be approximated using priors such as the red-channel or dark-channel assumptions, which exploit the faster attenuation of red light as a depth cue. In some cases, external monocular depth estimators trained on large-scale natural datasets can provide approximate depth maps that serve as additional guidance. Hybrid methods that combine learned predictions with physics-inspired priors or smoothness regularization often achieve the most robust performance. By integrating these strategies, PGCR ensures physically consistent colour correction, avoids unrealistic over-saturation, and generalizes effectively across diverse underwater environments.