In digital image processing, color image enhancement plays a crucial role. This process involves adjusting the brightness and color of a color image to meet human perception needs. Compared to grayscale images, color images contain richer information. RGB and HSV are two widely used color spaces for representing colors. In these two, there is a close correlation between the RGB components, and processing them separately may lead to image distortion, making it unsuitable for independent enhancement. The HSV space comprises three channel components: hue, saturation, and value. Hue dictates the color inclination toward red, green, or blue, and diverse hue ranges yield distinct colors. Saturation directly affects the color of the image, with deeper colors indicating higher saturation and vice versa. The brightness of the image is closely related to its luminance. In the HSV space, the advantage lies in the ability to separate the luminance component and color information, making it more conducive for observers [4] . HSV finds extensive applications in image enhancement, color selection, segmentation, and more. Therefore, choosing the HSV color space is beneficial for low-light PCB image enhancement [5] .

The core concept of the Retinex theory lies in the idea that an object’s color is not solely confined to the values of reflected light intensity but varies with the amount of light it receives [6] . This theory aims to ensure color consistency for objects under uneven lighting conditions, providing color constancy. Retinex can achieve a stable state in the presence of non-uniform characteristics, such as compressing the range of abnormal features, enhancing image edges, and stabilizing colors. Therefore, it can adaptively enhance various types of images [7] [8]. Its primary task is to preserve image details and textures while adjusting the brightness and contrast of the image. The Retinex algorithm is implemented through three primary methods: Single-Scale Retinex (SSR), Multi-Scale Retinex (MSR), and Multi-Scale Retinex with Color Restoration (MSRCR). The resulting image can be represented by the following formula, where the expression is analogous to Formula (1):

I ( x , y ) = L ( x , y ) ∗ R ( x , y ) (1)

In this context, I ( x , y ) represents the received signal, L ( x , y ) is the illuminance from the ambient light, and R ( x , y ) denotes the reflectance of the target object carrying the image features. Due to certain limitations of the Single-Scale Retinex (SSR) algorithm in restoring high-frequency components of images, the Multi-Scale Retinex (MSR) algorithm was proposed. It aims to perform filtering operations on low, medium, and high spatial frequency scales of the source image. The reflectance image obtained using the MSR algorithm can be expressed by Formula (2):

r ( x , y ) = ∑ i = 1 n ω i { log I ( x , y ) − log [ g i ( x , y ) ∗ I ( x , y ) ] } (2)

In the formula, g i represents the Gaussian surround function at different scales with other conditions being the same; n is the difference in the number of functions, typically set to n = 3; ω is the weight coefficient in the algorithm, usually set to 1/3. The Multi-Scale Retinex (MSR) algorithm may introduce color distortion as it doesn’t consider the intrinsic relationships between the RGB channels. There is a close correlation between the RGB components, and processing them independently can lead to image deformation, making it unsuitable for standalone image enhancement. To tackle this problem, the Multi-Scale Retinex with Color Restoration (MSRCR) algorithm, which incorporates color restoration functionality, was introduced. It is referred to as the MSRCR algorithm. To mitigate color distortion, a color restoration factor C i is introduced. By adjusting the variation of the color restoration factor C, the MSR algorithm’s proportion is changed to restore it to the original proportion values, thereby achieving color compensation and resolving color distortion phenomena. The mathematical expression for the MSRCR algorithm is given by Formula (3):

r M S R C R i ( x , y ) = C i ( x , y ) ⋅ r M S R i (3)

To convert an image to the HSV space, it is necessary to first transform the color in the RGB space. Therefore, colors defined by (h, s, v) are closely related to the colors in the RGB image. The colors correspond differently in various channels, with RGB color space and HSV color space being numerically equivalent [9] . The conversion relationship can be mathematically expressed by Equations (4)-(6). Where (r, g, b) are the components of the corresponding color channels, and the maximum (max) and minimum (min) values of the RGB image are maintained consistently [10] . The formula for RGB to HSV conversion is as follows:

H = COS − 1 { ( R − G ) + ( R − B ) 2 [ ( R − G ) 2 + ( R − B ) ( R − G ) 2 ] } 360 (4)

S = 1 − 3 R + G + B [ min ( R , G , B ) ] (5)

V = 1 3 ( R + G + B ) (6)

As a non-linear filter, the core of bilateral filtering involves processing the internal structure of the target image by compromising the similarity of its pixels. It also considers factors like grayscale similarity to achieve edge-preserving denoising. Bilateral filtering exhibits significant advantages, especially in enhancing the edge processing of images. In contrast to traditional approaches such as the Retinex algorithm, which typically uses Gaussian filtering functions to enhance local features of images, bilateral filtering excels in preserving details without introducing the noise that can lead to local blurring. Therefore, this paper opts for the bilateral filtering function, known for its strong detail-preserving capabilities, to replace the original Gaussian function in the algorithm. The formula is given by (7):

f ( x , y ) = ∑ ( i , j ) ∈ S x , y ω s ( i , j ) ⋅ ω r ( i , j ) ⋅ I ( i , j ) ∑ ( i , j ) ∈ S x , y ω s ( i , j ) ⋅ ω r ( i , j ) ­ (7)

In the formula, S x , y represents the central pixel and its neighborhood, I ( i , j ) denotes different pixels in the neighborhood, ω s is the spatial similarity factor, and ω r is the intensity similarity factor. The expressions for the similarity factors are given by (8)-(9):

ω s ( i , j ) = 1 2 π σ s exp [ − ( x − x c ) 2 + ( y − y c ) 2 2 σ s 2 ] (8)

ω r ( i , j ) = 1 2 π σ r exp { − [ I ( x , y ) − I ( x c , y c ) ] 2 2 σ r 2 } (9)

In the formula, σ s and σ r are scale factors.

The central surrounding function in the traditional Retinex algorithm is typically a Gaussian filter. However, when using Gaussian filtering to process images, halo phenomena may occur. This phenomenon arises because in a Gaussian filter, as the distance between neighboring pixels decreases, their relationship becomes closer; conversely, as the distance increases, their relationship becomes more distant. This situation results in the phenomenon of blurred edges in the image. To address this issue, in the theoretical framework of Gaussian filtering, the more characteristic bilateral filtering is incorporated to aim at preserving the edge features of the image. Bilateral filtering considers differences in both value and spatial domains. By enhancing the details of the image while simultaneously achieving edge denoising, bilateral filtering can effectively eliminate halo phenomena. In this paper, bilateral filtering is introduced as a new central surrounding function in the MSR algorithm with color restoration, replacing the Gaussian filter that was originally used in the traditional algorithm. This change aims to improve the experimental results of the image, particularly by reducing the occurrence of color over-enhancement while preserving detailed information.

In the HSV color space, when the hue (H parameter) remains constant, the parameters of saturation and brightness are crucial for the quality of an image. However, increasing brightness alone may lead to certain color distortions in the image. It is necessary to process the image’s saturation [11] . To make the edges of the image clearer, optimization of the processed saturation (S channel) is essential. Due to insufficient or uneven illumination intensity, low-light color images not only darken the image and blur details but also result in a decrease in color depth and saturation. In the process of image enhancement, excessive enhancement of image brightness may lead to color distortion. To avoid this phenomenon, linear stretching of contrast should be performed [12] [13] [14] , adjusting pixel values to the [0, 255] range. The mathematical expression for contrast stretching is given by (10):

I ( x , y ) = I ( x , y ) − Imin Imax − Imin ( MAX − MIN ) + MIN (10)

The minimum and maximum values in the grayscale space following the algorithmic stretching process are referred to as min and max.

The In our paper, we propose a novel algorithm for enhancing low-light images by jointly processing them in the HSV channel and the RGB space using the Retinex algorithm. The improved algorithm follows these steps: Step: 1 involves the application of an improved MSRCR image enhancement algorithm to the initial RGB image, highlighting its impact on details, including edges. In Step 2, the original image undergoes processing in the HSV space, achieved by segregating the RGB image into three channels and subsequently transforming it into the HSV space. Step 3: Optimize the image collected by its S channel using an improved bilateral filtering, followed by saturation stretching transformation. Finally, normalize the processed components. Step 4: Apply adaptive gamma correction to the brightness channel (V component) to correct the enhanced image. While effectively retaining image information, this step achieves the purpose of image correction. Further enhance brightness using the CLAHE algorithm. Step 5: Reconstruct the optimized HSV image by combining the optimized S channel, V channel, and the unchanged H channel. Convert the result back to an RGB image. Step 6: Fuse the two sets of enhanced RGB images to generate the final image. The workflow of the algorithm is illustrated in Figure 1.

From a subjective analytical standpoint, the image processed by the algorithm outlined in this paper demonstrates a discernible enhancement in overall brightness and heightened contrast when juxtaposed with the original image. More detailed information is noted to be richer, especially in annotated areas such as circuit lines. In contrast to the other four algorithms, the algorithm presented in this paper significantly amplifies the overall brightness of the PCB circuit board, thereby rendering the details on the distant circuit board more prominent. For the PCB circuit board, contrast is improved, and the patterns become clearer after processing with the algorithm in this paper. In line with human visual characteristics, components and other parts become brighter, revealing more detailed information. It’s noteworthy that the color representation of the circuit board almost remains distortion-free after processing with the algorithm in this paper, and the colors are closer to the original. In contrast, after enhancement with different algorithms, while component images are improved, there is noticeable detail blurring and image-wide distortion. Images processed with the MSRCR algorithm exhibit improved contrast but reveal instability in color processing. This highlights the effectiveness of the algorithm in this paper in enhancing low-light images, preserving and presenting more image features.

Subjective evaluation involves the observer’s intuitive visual perception and relies on multiple evaluation criteria or individual experience. Its main drawback

is the lack of consistency. In contrast, objective evaluation adopts a consistent assessment method, resulting in higher stability of results. In this study, the enhancement effects of low-light images were objectively evaluated using multiple evaluation indicators. Primarily, the standard deviation serves as an indicator of the spread of pixel values from the mean, illustrating the distribution of pixel values within the image. A higher standard deviation indicates better image quality, suggesting greater overall differences in brightness and color, potentially leading to a richer and more detailed perception for observers. Secondly, the average gradient is an indicator of image sharpness and texture variation. A higher average gradient value indicates richer texture features, making the image easier to identify and clearer, reflecting more pronounced changes in local regions of the image. Finally, the gray value reflects the overall brightness of the processed image. A higher gray value corresponds to higher brightness. Increasing the gray value during image enhancement helps improve the brightness contrast of the image, making it easier to observe and understand. Through these objective evaluation indicators, we can comprehensively understand the performance of low-light image enhancement methods in different aspects, providing better guidance for optimizing and improving image processing. For the three groups of image enhancement results using different algorithms, this paper conducts objective analysis using the aforementioned evaluation indicators. Specific results can be found in Tables 1-3. From Table 1, it can be concluded that the standard deviation improvement of the three groups of images enhanced by the algorithm in this paper is superior to other algorithms, highlighting a significant improvement

in the quality of the original images. Table 2 shows that the algorithm in this paper maintains a high level of image mean values, successfully enhancing the brightness of low-light images. The average gradient in Table 3 more intuitively illustrates the significant optimization effect of this algorithm on clarity.