This stage begins with contrast enhancement, where image visibility and color fidelity are improved through a three-step sequential process. The output of each step serves as the input for the next, creating a cumulative enhancement effect. The sequence of PCGF, AHE, ALSM, RIDNet/DRHNet, and ATRS was selected to progressively improve image visibility before restoration and repair. Early illumination enhancement exposes hidden structural details and improves scene visibility, while the subsequent restoration modules suppress degradation artifacts such as noise, haze, and rain. Finally, the inpainting stage is applied after restoration to ensure that missing regions are reconstructed using enhanced and artifact-reduced contextual information.

2.1.1. Gamma Correction

The process begins with Gamma Correction, which non-linearly adjusts the global luminance of the image [29]. For dehazing, it is defined as

The technique involves inverting the input picture x to the value 1 − x , applying gamma correction, then reversing the result back to get the final image y . One may get inhomogeneous results by applying gamma correction to hazy images, or one can get consistent results by applying a gamma correction previously. Dong et al. [30] initially suggest the inverse technique, which is subsequently applied for low-light image enhancement [7]. Following the idea presented in Equation (1), Li et al. [13] define two new functions:

Then combine (1) and (3) to form an inherent whole, resulting in the following pair of complementary gamma functions [13]:

A simple blending technique allows us to obtain the optimal enhanced image using the Pair of Complementary Gamma Functions (PCGF).

where y i is the mapped image from the original low-light picture as defined by (4), and the weighting factor is derived as follows:

where, v i _ is the mean value of the HSV transformation of the image y i . Figure 2 shows the visual effect after applying the PCGF.

Figure 2. (a) Original low-light images; (b) After applying PCGF.

It is clear that PCGF is a powerful tool for enhancing images with limited lighting. It increases the overall visibility of dark images and improves the finer details by compressing the light areas and expanding the dark ones. Unfortunately, the local contrast and particulars still need to be further improved. Li et al. [13] provide a straightforward fusion technique to enhance quality further. Image fusion is often used for multi-focus image processing [12], low-light image enhancement [10], and underwater image enhancement [31].

The quality of an image is greatly affected by the amount of light it receives. So, using (5) and (6), we choose an exposedness weight map to generate the areas that will be used to assess the degree of exposure, which is represented as a Gaussian-modeled distance from the average normalized value of 0.5, as shown below [10], [32]:

where, I i ( x , y ) stands for the i th image in the input and the coordinates of the i th pixel ( x , y ) . Pixels that are closer to 0.5 are given higher values. Underexposed and overexposed pixels receive low values from this weight map, while properly exposed pixels receive high values. As a result, it produces an unbiased assignment. Finally, the following provides the normalized weight maps for fusion:

Using the pyramid fusing technique suggested by Burt and Adelson [33], we can eliminate the halos and artifacts that emerge from using a straightforward blending strategy for fusion. The enhanced image is finally provided by:

where the number of pyramid layers is denoted by l , U d is the up-sampling operator with d = 2 l − 1 , G ( . ) and L ( . ) stand for Gaussian and Laplace operators. With little computational load, it may provide excellent performance [13], as shown in Figure 3.

Figure 3. (a) Original low-light images; (b) After applying PCGF by Fusion.

2.1.2. Histogram Equalization

Following gamma correction, we apply an Adaptive Histogram Equalization (AHE) method [16] to improve local contrast without amplifying noise significantly. To improve visual contrast, the Adaptive Histogram Equalization (AHE) method begins by extending a color channel in the RGB color space and then converting the input RGB color space into the HSI color space. The intensity channel is then given a nonlinear adaptive gamma correction using a weighted histogram distribution method to improve the image’s details. The enhanced image is then returned to RGB for optimal display, as shown in Figure 4.

Figure 4. Framework for the histogram equalization method [16].

Consider the following input color image: I ( u , v ) = R ( u , v ) , G ( u , v ) , B ( u , v ) , where R , G , and B are the red, green, and blue color spaces, and ( u , v ) are the pixel coordinates. In order to correct color distortion from an unfavorable capturing situation, the input images are first stretched on each color channel. To provide the observer with the optimal set of color information, a magnitude or channel stretching mechanism expands the input color channels into their respective tolerated maximum ranges [16]. The color value for stretched red is stated as,

The same distortion is applied to the green and blue channels. Stretching enhances the image’s color richness and acts as a pre-processing step for contrast improvement. After magnitude stretching, all color signals are transformed into the hue-saturation-intensity (HSI) space. The system processes the I-signal, which represents the intensity or brightness. The H-channel and S-channel components are not altered since they have no direct effect on an image’s human-perceived contrast [16]. The transformed RGB input image is converted into HSI color space according to Equation (11).

where, T R G B H S I stands for the RGB-to-HSI color conversion procedure. The picture can be improved by keeping the H and S channel components but enhancing the intensity channel. The intensity channel in the HSI space is denoted as [16],

The main goal of this step of enhancement is to improve visual contrast while retaining as much information content as possible near the original image. The converted pixel intensity is determined as follows [16]:

For integer-valued output in [0, 255], Equation 13(a) is applied as:

Figure 5. Contrast Enhancement results and their histograms. (a) Original image; (b) HE-ed image; (c) Gamma-corrected image; and (d) AHE-ed image.

Figure 6. Histogram equalization results, (a) Original images and (b) AHE-processed images.

Figure 5 shows the result after contrast enhancement with its histogram. To produce the improved image, combine the processed intensity image [ T { I ( u , v ) } ] with the hue space and then convert the image back to the RGB color space [16] as shown in Figure 6. It is stated as,

2.1.3. Physical Illumination Modeling

The final step in this module is the application of the Absorption Light Scattering Model (ALSM) [17]. This model, an extension of the atmospheric scattering model, realistically illuminates shadowed regions by modeling light absorption. It recovers the scene radiance for each pixel, revealing hidden details in low-light areas that may not have been fully corrected by the previous steps.

McCartney, Narasimhan, and Nayar suggested a monochromatic ASM [34] that adequately describes the effects of particle scattering on visible light imaging based on the attenuation and ambient light model. Because sunlight may pass through the atmosphere and illuminate objects, atmospheric light is seldom the primary source of lighting [23]. Assuming a uniform spatial distribution of light, we may write ASM as follows:

where, I is the observed image’s 3D matrix, and R is the scene’s reflection. A pixel’s location ( x , y ) is represented by the 2D vector x . A 3D RGB vector at pixel x is denoted as I ( x ) . The color channel c ’s 2D matrix is denoted by the symbol I c . The color channel c at pixel x is represented by the scalar I c ( x ) . The coordinates R ( x ) denote a 3D vector at the given pixel x . The scalar representation of the color channel c at pixel x is denoted by R c ( x ) . Each point x has a medium transmittance, denoted by the scalar t ( x ) , which is in the interval [0, 1]. S L and A stand for the overall ambient light of the scene, provided by the sun. It is often believed that A and S L , both 3D matrices, are spatially constant; that is, they are identical matrices for every channel c in I .

Following Equation (14), after completing monotonic analysis and utilizing the minimal channel constraint, ALSM obtains the enhancement result for each pixel. Finally, the scene radiance at pixel x of the original low-light image is given by:

where, A c is the color channel c ’s 2D matrix, a scalar of the color channel c at a pixel x is represented by I c ( x ) and R c ( x ) , the medium transmittance of any point x is given by t ( x ) , and t 0 is the lowest limit of t ( x ) for presenting this term as indivisible. A common fixed value for t 0 is 0.1.

A physical finding of the imaging process of the absorption light scattering image is just one of the benefits of ALSM; the method also reproduces the hidden details of low-light images and provides new characteristics for analysis [17]. Figure 7 depicts the result of using this information to determine the brightest pixel value.

Tasks such as denoising, dehazing, and deraining provide important benchmarks for assessing image priors and evaluating visual quality.

Figure 7. Physical Illumination Modeling (a) Original Images and (b) ALSM results with T = 6, where Tdenotes the number of pyramid decomposition levels used in the multi-scale fusion step.

2.2.1. Denoising with RIDNet

If noise is detected in the preprocessing stage, we employ the Real Image Denoising Network (RIDNet) [20], a robust CNN-based model. Its residual-on-residual architecture with Enhancement Attention Modules (EAMs) allows it to effectively remove complex, real-world noise while preserving fine textures and edges.

The RIDNet model comprises three main modules: feature extraction, residual-on-residual feature learning, and image reconstruction, as illustrated in Figure 8.

Figure 8. Architecture of the CNN-based RIDNet model [20].

In this architecture, the convolutional layers with a kernel size of 1×1 are represented, while the varying green shades indicate different dilation rates in the convolutional layers. The second column provides a detailed view of the structure of each Enhancement Attention Module (EAM), as illustrated in Figure 9.

Figure 9. Feature attention mechanism [20].

Figure 9 illustrates the feature attention mechanism that focuses on the most relevant characteristics. In this process, g p denotes the global pooling layer, H D and H u represent the channel reduction and upsampling operators, respectively, and α is the sigmoid activation function. The approach incorporates four Enhancement Attention Modules (EAMs). Except for the last convolutional layer in the improved residual block, which uses a 1 × 1 kernel, all convolutional layers utilize a 3 × 3 kernel size, with zero padding applied to maintain feature map dimensions. Each convolutional layer, except during feature attention downsampling, contains 64 channels. These channels are reduced by a factor of 16, resulting in four feature maps. Depending on the input, the final convolutional layer outputs either one or three feature maps. The model executes in approximately 0.2 seconds for a 512 × 512 pixel image. The results are demonstrated in Figure 10.

Figure 10. Image denoising (a) Noisy images; (b) RIDNet images.

2.2.2. Dehazing & Deraining with DRHNet

If haze or rain is detected, we utilize the Deep Residual Haze Network (DRHNet) [35]. This end-to-end framework learns to predict a negative residual map (the haze or rain layer), which is then subtracted from the input image to restore a clear version. Its design ensures strong performance on both tasks and prevents quality degradation when applied to clear images.

Figure 11 illustrates the three components of the Deep Residual Haze Network (DRHNet): the context-aware encoder (left), the transformation module (center), and the haze decoder (right). By reformulating its layers to learn residual functions, DRHNet effectively captures the differences between clear and hazy images, which also helps to reduce computational complexity [36].

Figure 11. The framework of the DRHNet model [35].

The dehazed result that is produced by DRHNet can be obtained in the following manner:

where J denotes the dehazed image, DRHNet(I) is the output of the DRHNet model given the hazy input image I. The model employs a context-aware feature extraction module to encode contextual information into feature maps. These maps are then decoded by the haze decoder to estimate the residual map. Since the residual between hazy and haze-free images is mostly negative, DRHNet generates a negative residual map by inverting each pixel value in the residual map. The clear image is recovered by subtracting the hazy input from this negative residual. Similarly, DRHNet directly estimates the residual between clear and rainy images using the same formulation. Sample dehazing and deraining results are shown in Figure 12. Wang et al. [35] demonstrated that DRHNet achieves strong deraining performance on benchmark datasets. Moreover, applying DRHNet to haze-free images has minimal adverse effects and does not degrade image quality.

Figure 12. Image dehazing and deraining (a) Hazy image; (b) DRHNet result of Hazy image; (c) Rainy image; and (d) DRHNet result of rainy image.

The final stage addresses object removal or the repair of damaged regions. We use the Adaptive Two-Round Search (ATRS) strategy [29], an exemplar-based method. It improves upon traditional techniques by first finding a best-match patch using the Sum of Squared Differences (SSD), and then, if a potential mismatch is detected, initiating a second search round that also considers the color difference (DBP), ensuring better structural and textural consistency.

2.3.1. Priority Computation

Here we follow [28] to calculate the patch priority and decide on the filling order. For each patch Ψ p centered at the point P ( P ∈ δ   Ω ) that has a patch priority P ( p ) , it is denoted as:

where, C ( p ) and D ( p ) represent the confidence term and the data term, respectively. After determining the most crucial target patch, it searches the source area for the exemplar patch that is most similar to the target patch based on the matching rule.

2.3.2. Definition of DBP

The Sum of Squared Differences (SSD) is used to determine how similar a target patch is to an exemplar patch in the standard exemplar-based inpainting process. According to the definition [28],

where Ψ q is the exemplar patch, and Ψ p is the target patch. M is the binary mask, and it utilizes one to designate which pixels must be filled. Differences between Patches (DBP) are defined as follows, using Equation (18).

where the first item determines the average value of pixels in the target patch, the second determines the average value of pixels utilized to fill the exemplar patch.

2.3.3. Adaptive Two-Round Search (ATRS)

To find the mismatch error, two adaptive thresholds, β and γ , are given to compare it with SSD and DBP. It uses a straightforward and effective way to calculate the thresholds. SSD and DBP are first calculated for all matching patches, yielding ssd_data and dbp_data, respectively, as two arrays. Next, we’ll apply an ascending sort to each array, and then we’ll have ssd_data_sort and dbp_data_sort. After doing that, we can get the values for B and Y by doing the following [29]:

where, λ 1 and λ 2 are the coefficients of proportionality, the number of similar patches is denoted by n u m . The two-round search approach is executed adaptively if there is a discrepancy between the exemplar and target patches. It uses a two-round research approach to rediscover the exemplar patch that resembles the original. The new matching rule that incorporates SSD and DBP is denoted as follows [29]:

where, ω is a multiplier that makes the SSD and DBP the same order of magnitude apart, it is defined as ω = ∑ M _ .

In Algorithm 1, the complete workflow of our proposed framework is summarized in the pseudocode. The restoration modules are conditionally activated based on the degradation analysis performed during preprocessing. When multiple degradations are detected simultaneously, the denoising stage (RIDNet) is executed before the dehazing/deraining stage (DRHNet). This ordering is intended to suppress noise prior to atmospheric restoration, thereby reducing the possibility of noise amplification and improving the stability of the subsequent restoration process.

Algorithm 1. Proposed image enhancement pipeline.

Figure 13 illustrates the effectiveness of our connected pipeline on a single input image suffering from low light, noise, and containing an object for removal. The original image (a) is dark and lacks detail. After applying our contrast enhancement module (b), the brightness and local contrast are significantly improved, revealing details in the shadows. The subsequent conditional restoration step (c) effectively removes noise and atmospheric haze, resulting in a much cleaner image. Finally, the inpainting stage (d) seamlessly removes the marked object, producing a visually coherent and high-quality final image. This step-by-step visualization confirms that the components work synergistically, with each stage building upon the previous one to achieve a comprehensive restoration.

Figure 13. Demonstration of the integrated pipeline. (a) Original degraded image; (b) After Stage 1: Contrast Enhancement; (c) After Stage 2: Conditional Denoising & Dehazing; (d) Final result after Stage 3: Inpainting.

We compared our proposed method against several leading individual algorithms: LIME [7] for low-light enhancement, FFA-Net [37] for dehazing, and Restormer [38] for general restoration. We used the Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) as our quantitative metrics, where higher values indicate better performance. The results, averaged over a test set of 50 benchmark images from the LOL dataset [7] (low-light), RESIDE-β (haze), and Rain100H (rain) with synthetic degradations applied to clean reference images to enable full-reference PSNR/SSIM evaluation, are presented in Table 1. Our proposed approach demonstrates competitive performance compared with the selected baseline methods and the individual methods, demonstrating the benefit of a holistic pipeline that addresses multiple degradations simultaneously.

Table 1. Quantitative comparison with state-of-the-art methods.

For the inpainting task, we compared our ATRS-based module with a modern deep learning approach. In Figure 14, we show the original image (a) with a region marked for removal (b). The result from a Contextual Attention GAN [26] (c), based on two discrimination networks, provides a plausible restoration but introduces some textural artifacts. Our proposed Integrated Algorithm (d) demonstrates visually improved restoration quality in the presented example in this case, producing a more natural and structurally consistent result by effectively leveraging information from the surrounding pixels.

Figure 14. Inpainting comparison. (a) Original image; (b) Target removal region (marked in gray); (c) Result using contextual attention GAN [26] based on two-discrimination networks; and (d) Result using our proposed Integrated Algorithm.