Zero-Reference Depth Curve Estimation-Based Low-Light Image Enhancement Method for Coating Workshop Inspection

Abstract

To address the challenges of poor image quality and low detection accuracy in low-light environments during coating workshop inspections, this paper proposes a low-light image enhancement method based on zero-reference depth curve estimation, termed Zero-PTDCE. A low-light image dataset, PT-LLIE, tailored for coating workshop scenarios is constructed, encompassing various industrial inspection conditions under different lighting environments to enhance model adaptability. Furthermore, an enhancement network integrating a lightweight denoising module and depthwise separable dilated convolution is designed to reduce noise interference, expand the receptive field, and improve image detail restoration. The network training process employs a multi-constraint strategy by incorporating perceptual loss (L_p), color loss (L_c), spatial consistency loss (L_s), exposure loss (L_e), and total variation smoothness loss (L_tv) to ensure balanced brightness, natural color reproduction, and structural integrity in the enhanced images. Experimental results demonstrate that, compared to existing low-light image enhancement methods, the proposed approach achieves superior performance in terms of peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and mean absolute error (MAE), while maintaining high computational efficiency. Beyond general visual enhancement, Zero-PTDCE significantly improves the visibility of fine surface features and defect patterns under low-light conditions, which is crucial for the accurate assessment of coating quality, including defect identification such as uneven thickness, delamination, and surface abrasion. This work provides a reliable image enhancement solution for intelligent inspection systems and supports both the automated operation and material quality evaluation in modern coating workshops, contributing to the broader goals of intelligent manufacturing and material characterization.

1. Introduction

Coating equipment serves as a fundamental component in industrial production processes such as lithium battery electrode coating and optical film coating. Its operational status directly impacts the stability of the coating process and the performance of the final product [1]. Therefore, regular inspection of coating equipment is crucial for promptly identifying anomalies and potential failures, ensuring production efficiency and product quality [2]. However, with the increasing challenges posed by an aging workforce, the manufacturing industry faces growing labor shortages. Traditional coating workshop inspection management still heavily relies on manual inspections, leading to high costs, low efficiency, and susceptibility to human errors [3]. These limitations have become a significant bottleneck in improving production efficiency and enhancing market competitiveness in the coating industry. This trend underscores the urgent need for industrial upgrading and digital transformation in coating workshops. In particular, replacing manual inspections with intelligent robotic systems has become an inevitable trend to enhance production efficiency, reduce labor costs, and improve overall operational effectiveness [4]. Nevertheless, coating workshops present complex environmental challenges, particularly in terms of suboptimal lighting conditions. Structural components such as coating machines and support frames can obstruct light, resulting in uneven illumination and alternating bright and dark regions in inspection areas. Moreover, to minimize external light interference, coating workshops often operate in low-light environments, where reflective surfaces of equipment further exacerbate visibility issues [5]. Consequently, images captured for inspection frequently suffer from low brightness, poor contrast, and blurred details, severely compromising the accuracy and reliability of machine vision-based equipment status recognition and fault diagnosis [6]. Given these challenges, research on low-light image enhancement in coating workshops is of paramount importance. Effective enhancement techniques can significantly improve the reliability of inspection systems, laying a solid foundation for intelligent inspection and facilitating the automation and digital transformation of coating workshops. Furthermore, enhanced imaging quality also facilitates the identification of coating-related defects such as uneven thickness, surface scratches, and material delamination. These visual anomalies are closely related to the physical and chemical performance of the final coating layer. Therefore, low-light image enhancement plays a pivotal role not only in equipment inspection but also in the visual characterization of coating quality, bridging the gap between intelligent manufacturing and material performance assessment.

Current research on low-light image enhancement, both domestically and internationally, can be broadly categorized into traditional methods and deep-learning-based approaches [7]. Traditional low-light image enhancement methods primarily improve image quality through adjustments of brightness and contrast, as well as non-linear techniques such as gamma correction [8]. Gamma correction is widely used for its ability to enhance mid-tone details by modifying luminance in a non-linear fashion. However, despite their effectiveness in relatively uniform lighting conditions, these traditional methods often struggle to cope with complex illumination variations, particularly in industrial environments such as coating workshops, where lighting fluctuates dramatically and significant local brightness differences exist. In such cases, traditional approaches may result in overenhancement, noise amplification, or loss of structural details, thereby limiting their practical applicability. In recent years, deep-learning-based methods have made remarkable progress in low-light image enhancement. Many of these approaches employ end-to-end training to enable automatic feature learning and image enhancement [9]. However, existing deep learning methods still face several challenges when processing extremely low-light or extremely low-illumination images, including loss of details, noise amplification, and difficulties in preserving the natural appearance of enhanced images. Moreover, many deep learning models rely heavily on large amounts of annotated data for training, making the training process computationally intensive and reducing their adaptability to different real-world scenarios. Consequently, there remains a strong demand for efficient, robust, and adaptable low-light image enhancement techniques that can effectively improve image quality under complex industrial conditions.

To address the aforementioned challenges, this paper proposes a zero-reference depth curve estimation-based low-light image enhancement method for coating workshop inspection, termed Zero-PTDCE. This method is specifically designed to adapt to the low-light conditions in coating workshops without requiring reference images. It enhances images based solely on their intrinsic information while mitigating issues such as uneven brightness and overexposure in enhanced images. Moreover, it significantly reduces noise amplification, improving image quality while minimizing reliance on large amounts of annotated data and enhancing overall performance. The main contributions of this work can be summarized as follows:

(1) Integration of a lightweight denoising module at the input stage. To improve the quality of input images, a lightweight denoising module is introduced at the network’s input stage. This module consists of a 3 × 3 convolutional kernel followed by a ReLU activation function, effectively reducing noise interference. By providing higher-quality image data for subsequent feature extraction and enhancement, this module ensures a more robust enhancement process.

(2) Introduction of depthwise separable dilated convolution. Zero-PTDCE incorporates depthwise separable dilated convolution, which not only reduces model parameters and computational complexity but also enhances the extraction of both local and global image features. Additionally, by expanding the receptive field of the convolutional kernel, dilated convolution further improves the network’s ability to capture fine image details, leading to more effective enhancement results.

(3) Incorporation of perceptual loss for high-level feature consistency. Zero-PTDCE introduces a perceptual loss function (Lp) to measure the difference between the enhanced image and the original image in a high-level feature space. This approach enables the network to better capture structural and detail information during enhancement, ensuring that the enhanced images maintain visual consistency with real-world images in terms of perceptual quality.

2. Related Work

Due to insufficient ambient lighting during image acquisition, captured images often suffer from severe degradation in quality, including poor visibility, color distortion, and increased noise levels. To address these challenges, researchers worldwide have proposed various low-light image enhancement methods aimed at improving image visibility and visual quality. One of the earliest image processing techniques for low-light enhancement is histogram equalization (HE) [10], which enhances brightness and contrast by redistributing pixel intensity values. However, since HE operates independently on each pixel without considering the overall structure of the image, it often leads to overexposure in certain regions or noise amplification, making it difficult to achieve satisfactory enhancement results in complex scenes. Another major category of low-light image enhancement methods is based on Retinex theory [11], which stems from human visual perception and aims to decompose an image into reflectance and illumination components to simulate how the human eye perceives images under different lighting conditions. For instance, Wang et al. [12] proposed an enhancement algorithm using double logarithmic transformation, which balances detail preservation and naturalness. Fu et al. [13] introduced a weighted variational model, jointly estimating the reflectance and illumination components while using the reflectance as the final enhanced output. Ren et al. [14] optimized illumination estimation through a robust Retinex model and developed a sequential algorithm to suppress noise in the reflectance map. Li et al. [15] further improved this approach by incorporating a noise mapping mechanism, effectively reducing the impact of noise on enhancement performance. However, despite its effectiveness, the Retinex model requires a large number of parameters and involves high computational complexity, which limits its performance improvements in practical applications.

With the rapid advancements in deep learning for computer vision, researchers have increasingly explored deep learning-based approaches for low-light image enhancement. One of the most commonly used strategies is end-to-end learning, where a deep neural network is trained to map low-light images to their corresponding well-lit reference images, achieving automated enhancement. One of the earliest deep-learning-based methods in this field is LLNet, proposed by Lore et al. [16], which utilizes a deep autoencoder network for low-light image enhancement. This work marked an early exploration of deep learning techniques in low-light enhancement. Chen et al. [17] later introduced a bi-directional generative adversarial network (GAN)-based unpaired learning framework, which incorporates an adaptive weighting mechanism to improve GAN stability, thereby achieving high-quality enhancement results. More recently, deep learning models inspired by Retinex theory have gained significant attention. For instance, Ren et al. [18] proposed a deep hybrid network that employs a dual-stream structure to simultaneously capture global image content and local salient structures. Zhang et al. [19] introduced the KinD network, which decomposes an image into illumination and reflectance components, allowing independent processing for illumination adjustment and noise suppression. However, traditional deep learning methods based on paired training data suffer from high data collection costs and complex training procedures. To overcome this limitation, unsupervised learning methods such as EnlightenGAN [20] leverage unpaired low-light and normal-light images during training, reducing dependency on paired data to a certain extent. Furthermore, Yang et al. [21] proposed a semi-supervised enhancement model that integrates frequency band representations with adversarial learning, improving model generalization while reducing reliance on paired datasets. Despite their advantages, semi-supervised learning frameworks still face challenges, including the risk of overfitting to paired data and high memory consumption. Additionally, Yuan et al. [22] introduced an automatic exposure correction method, which employs a global optimization algorithm to estimate an S-curve for the given image. This curve is then used to map each segmented region of the image to its optimal brightness range, achieving effective enhancement. While this method improves image brightness and contrast, it has certain limitations, as it requires manual curve adjustments and may lead to overadjustment or detail loss when applied to complex scenes, thus limiting its effectiveness in dynamic environments.

3. Low-Light Image Enhancement for Coating Workshop Inspection

3.1. Zero-DCE

Zero-DCE is a deep-learning-based low-light image enhancement algorithm that incorporates reference-free loss functions and a luminance enhancement curve (LE-curve) to achieve automatic brightness adjustment without requiring paired training data [23]. This method formulates the low-light image enhancement process as a task of adjusting image brightness through a non-linear curve transformation. The core idea of Zero-DCE is to employ a deep curve estimation network (DCE-Net) to estimate the optimal illumination enhancement curve parameters for each pixel, which are then iteratively applied to adjust image brightness. Unlike traditional methods, Zero-DCE does not rely on paired training data but instead enhances image quality through adaptive curve estimation and loss function optimization. The framework of Zero-DCE is shown in Figure 1.

The design of the luminance enhancement curve (LE-curve) in Zero-DCE is inspired by brightness adjustment curves in image editing software, ensuring that the pixel values of the enhanced image remain within the range of [0, 1] while maintaining monotonicity to preserve image contrast. Additionally, the curve’s simple mathematical form and differentiability ensure stability during gradient backpropagation. The LE-curve dynamically adapts to various lighting conditions, adjusting brightness at the pixel level, thereby preventing overenhancement or underenhancement issues that often arise in global mapping-based methods.

The mathematical expression of the luminance enhancement (LE) curve is given as follows:

$${\mathit{LE}}_n(x)={\mathit{LE}}_{n-1}(x)+\mathcal{A}(x){\mathit{LE}}_{n-1}(x)(1-{\mathit{LE}}_{n-1}(x))$$

(1)

where: x represents the pixel coordinate, n denotes the number of iterations in the image enhancement process, $\mathcal{A}$ represents the parameter mapping, which has the same spatial dimensions as the given input image, ${\mathit{LE}}_{n-1}(x)$ is the enhanced image of ${\mathit{LE}}_n(x)$.

The original DCE-Net adopts a symmetric cascaded architecture consisting of seven convolutional layers, where each layer comprises 32 convolutional kernels with a kernel size of 3 × 3 and a stride of 1, followed by a ReLU activation function. The final convolutional layer also consists of 32 convolutional kernels of size 3 × 3 with a stride of 1, followed by a Tanh activation function. The network generates 24 curve parameter mappings over 8 iterations, with each iteration producing three parameter mappings corresponding to the R, G, and B channels. To preserve the relationships between adjacent pixels, downsampling layers and batch normalization layers were omitted. Later, in DCE++, standard convolution operations were replaced with depthwise separable convolutions to reduce computational complexity [24]. In this variant, the depthwise convolution kernels have a size of 3 × 3 with a stride of 1. The pointwise convolution kernels have a size of 1 × 1 with a stride of 1. The output layer generates only three curve parameter maps, which are reused across different iterations to mitigate the risk of network overfitting. However, when enhancing low-light images, significant noise artifacts may appear, adversely affecting image quality. Despite the improvements introduced by the Zero-DCE image enhancement algorithm, the enhanced images may still exhibit overbrightening and loss of fine details compared to normally illuminated images. This poses challenges for the subsequent analysis and recognition of weakly illuminated images in coating workshop inspections.

3.2. Zero-PTDCE

This paper optimizes and innovates the Zero-DCE image enhancement algorithm, proposing the Zero-PTDCE algorithm to better adapt to the characteristics of low-light images in coating workshop inspections. The network architecture of PTDCE-Net is shown in Figure 2.

The detailed configuration of the proposed PTDCE-Net architecture is as follows:

The input image is first processed by a lightweight denoising module, which consists of a single convolutional layer with a 3 × 3 kernel, stride 1, and padding 1, followed by a ReLU activation function. This module aims to reduce noise interference while preserving structural information.

The enhancement module is composed of seven depthwise separable dilated convolution blocks: the first four layers have 32 output channels each, using a 3 × 3 kernel, stride 1, and a dilation rate of 2. These layers leverage depthwise separable convolution to extract local features and expand the receptive field efficiently. The fifth and sixth layers take concatenated feature maps from the previous layers as input, resulting in 64 input channels, and produce 32 output channels each. The final layer outputs a 3-channel enhanced image using a depthwise separable convolution, followed by a Tanh activation function, which maps the output to the range (−1, 1).

All intermediate convolutional layers utilize ReLU activation functions, while the final enhancement output employs Tanh activation to ensure numerical stability.

The Zero-PTDCE algorithm, designed for coating workshop inspection, improves the original network architecture through the following enhancements:

• Pre-Denoising Module

Zero-DCE does not account for noise interference, which may lead to detail loss and amplification of existing noise during enhancement. To address this issue, this paper introduces a denoising module before image enhancement, using a single-layer convolution for denoising. The pre-convolutional layer effectively separates noise from image content, ensuring that the enhanced image improves brightness and contrast while preserving details as much as possible. The single-layer convolution is simple yet efficient, effectively removing noise while preserving fine details, making it well-suited for low-light image enhancement in coating workshops.

2.

Depthwise Separable Dilated Convolution

In image processing tasks, receptive field size plays a crucial role in capturing global information. To further improve enhancement performance, PTDCE-Net integrates dilated convolution with depthwise separable convolution: depthwise separable convolution effectively reduces parameter count and computational complexity while maintaining image detail integrity. Dilated convolution expands the receptive field, enhancing global feature extraction capabilities. These two operations work synergistically: depthwise separable convolution focuses on efficient local feature extraction, while dilated convolution expands the receptive field, allowing the network to analyze a broader image context. This reduces local detail loss while improving overall image quality. Consequently, PTDCE-Net replaces standard convolution layers with depthwise separable dilated convolution, with its structural diagram shown in Figure 3. The convolution kernel size is 3 × 3 with a stride of 1 and a dilation rate of 2, while the pointwise convolution kernel has a size of 1 × 1 with a stride of 1. This design preserves fine local details while improving global information capture, making it particularly effective for low-light image enhancement tasks.

3.

Perceptual Loss Function Lp

This paper introduces the perceptual loss function L_p, utilizing a pre-trained VGG16 network as a feature extractor to measure the difference between enhanced and original images in a high-level feature space. Specifically, L_p extracts feature maps from both input and enhanced images using intermediate convolutional layers of VGG16 and calculates the mean squared error (MSE) between them, assessing the perceptual difference of the enhanced image.

By optimizing this loss function, the enhancement network learns to focus on local pixel differences while preserving the overall structural information. The enhanced images exhibit higher visual consistency and naturalness. With perceptual loss, Zero-PTDCE not only enhances low-light images but also better retains original structural information, effectively reducing overexposure issues caused by excessive enhancement. This ensures that brightness adjustments remain balanced and natural, meeting practical requirements for coating workshop inspections. The mathematical formulation is detailed in Section 3.3.

4.

PT-LLIE: A Low-Light Image Dataset for Coating Workshops

To better support low-light image enhancement in coating workshops, this paper constructs the PT-LLIE dataset. This dataset comprises 2188 images, including: real-world low-light images captured by industrial cameras in coating workshops and publicly available low-light datasets, featuring images with various exposure levels. The industrial camera images cover typical working environments in coating workshops, ensuring data authenticity and representativeness. During dataset design, each image includes samples with different exposure scales to prevent overenhancement in areas with already optimal brightness. The dataset includes low-light conditions, localized overexposure, and high dynamic range (HDR) scenarios, providing a challenging and practical training set for low-light image enhancement tasks.

3.3. Loss Function

In this work, the perceptual loss function L_p is incorporated into the non-reference loss function and is trained together with the original network’s color loss L_c, spatial consistency loss L_s, exposure loss L_e, and smoothness loss L_tv. The perceptual loss L_p enhances the network’s ability to capture high-level image features, preserving the overall structural information after enhancement while preventing excessive brightness that may lead to overexposure. This effectively improves the visual quality and detail restoration of the enhanced image. The specific formulas for each loss function are as follows:

(1)

Perceptual Loss Function L_p

The perceptual loss L_p is designed to enhance the network’s ability to capture high-level semantic features, ensuring that the enhanced image maintains visual consistency with the real image. It extracts high-level feature representations of the enhanced image using a pre-trained VGG16 network. The formula is given as:

$$L_p={\displaystyle \frac{1}{CHW}}||\varphi (x)-\varphi (y)|{|}^2$$

(2)

In the formula: x and y represent the input image and target image, respectively. C, H, W denote the channel count, height, and width of the feature maps at a specific VGG16 layer.

(2)

Color Loss Lc

The color loss function is introduced to maintain color stability in the enhanced image, reducing color discrepancies between different channels and preventing color distortion. The formula is:

$$L_c=\sum _{\forall (p,q)\in \epsilon }{(J^p-J^q)}^2,\epsilon =\{(R,G),(R,B),(G,B)\}$$

(3)

where $J^p$ represents the mean intensity of channel p and (p, q) denotes a pair of channels.

(3)

Spatial Consistency Loss Ls

This loss function ensures that the spatial structure of the enhanced image aligns with the original image by comparing local gradient differences. The formula is:

$$L_s={\displaystyle \frac{1}{K}}\sum _{i=1}^K\sum _{j\in \mathsf{\Omega }(i)}{(|(Y_i-Y_j)|-|(I_i-I_j)|)}^2$$

(4)

where: K is the number of local regions, set to 44. $\mathsf{\Omega }(i)$ represents the four adjacent regions (top, bottom, left, right) of region i. Y and I denote the enhanced and input image’s local mean intensity values, respectively.

(4)

Exposure Loss Le

Exposure loss constrains the global brightness distribution of the enhanced image, preventing overexposure or underexposure by measuring the deviation of local intensity values from an optimal exposure level E. The formula is:

$$L_e={\displaystyle \frac{1}{M}}\sum _{k=1}^M|Y_k-E|$$

(5)

where: M represents the number of non-overlapping local regions of size 16 × 16. Y is the mean intensity of the local region in the enhanced image. E is the desired gray-level value in RGB color space, set to 0.6.

(5)

Smoothness Loss Ltv

The illumination smoothness loss is applied to each curve parameter map $\mathcal{A}$ to enforce monotonicity constraints between adjacent pixels, thereby reducing noise and artifacts. The formulation is:

$$L_{tv}={\displaystyle \frac{1}{N}}\sum _{n=1}^N\sum _{c\in \xi }{(|{\mathit{\nabla }}_x{\mathcal{A}}_n^c|+|{\mathit{\nabla }}_y{\mathcal{A}}_n^c|)}^2,\xi =\{R,G,B\}$$

(6)

where: N represents the number of iterations. ${\mathit{\nabla }}_x$ and ${\mathit{\nabla }}_y$ denote horizontal and vertical gradient operations, respectively.

(6)

Total Loss Function L_Total

The final total loss function combines all the above losses with corresponding weighting coefficients to balance different optimization objectives:

$$L_{Total}={\lambda }_{tv}{L}_{tv}+L_s+{\lambda }_c{L}_c+{\lambda }_e{L}_e+L_p$$

(7)

where ${\lambda }_{tv}, {\lambda }_c, {\lambda }_e$ are weighting coefficients that control the contribution of each loss term.

3.4. Application Framework

To effectively control the enhancement magnitude and prevent overexposure caused by excessive enhancement, Zero-PTDCE introduces a low-light brightness threshold T_L (T_L = 60) to distinguish between normal-light images and low-light images. When the average brightness of the input image falls below this threshold, the system automatically processes the image using the zero-reference deep curve estimation network (PTDCE-Net) to enhance brightness and reveal more details. Conversely, if the input image’s brightness exceeds the threshold, the system directly outputs the original image to avoid unnecessary enhancement, ensuring natural and realistic final results. The proposed low-light image enhancement method provides high-quality visual data support for coating workshop inspections, improving the intelligence and automation level of the inspection system. The application framework for low-light image enhancement in coating workshop inspection is illustrated in Figure 4.

4. Experiment Design and Analysis

4.1. Training Setup

The proposed network was trained using the PyTorch 1.12.1 framework and accelerated with an NVIDIA GeForce RTX 2060 GPU. The batch size for training was set to 8, and the weights of each convolutional filter were initialized using a Gaussian distribution with zero mean and a standard deviation of 0.02. The bias terms were initialized as constants. The Adam optimizer with default parameters was chosen to optimize network weights, and a fixed learning rate was used throughout training. To fully utilize wide dynamic range adjustments, the training dataset included both low-light and overexposed images, constructing a high-quality dataset that aligns with low-light conditions in the coating workshop. The 2188 images in the dataset were split into training (70%) and validation (30%) sets, with 1531 images used for training and the remainder for validation.

4.2. Benchmark Evaluation

To evaluate the performance of the proposed low-light image enhancement method, we compared it with URetinex-Net, SCI, Zero-DCE, and Zero-DCE++ on the coating workshop low-light dataset (PT-LLIE), the LOL dataset, and the SICE Part2 dataset. Among these methods, URetinex-Net is a supervised learning approach, SCI is an unsupervised learning method, and Zero-DCE and Zero-DCE++ are zero-reference learning techniques. The PT-LLIE dataset consists of real-world low-light images captured in a coating workshop environment, whereas the LOL dataset is a publicly available paired dataset.

4.2.1. Visual and Perceptual Comparison

To assess the enhancement performance of the proposed method on real-world low-light images from the coating workshop, we compared it with four other low-light enhancement methods. These test images were not included in training and were used solely to evaluate the adaptability of the proposed approach. The visual comparison of typical low-light image enhancement results in the coating workshop is shown in Figure 5.

From the visual comparison in Figure 4, we can observe the following: URetinex-Net enhances the overall brightness compared to the original image but introduces overexposure artifacts, such as the halo effect in Scene 3, leading to significant detail loss. SCI demonstrates better detail preservation during enhancement but struggles with low brightness and contrast, as seen in Scene 1, which remains relatively dark, indicating its limited effectiveness in low-light environments. Zero-DCE improves brightness but introduces a greenish tint, which distorts the natural color of metallic materials. Zero-DCE++ enhances overall contrast but produces a cool color tone, making metallic surfaces appear less natural. Zero-PTDCE, on the other hand, delivers consistent performance across multiple scenes, including workshops, control panels, mechanical equipment, and instrument dashboards. It effectively enhances brightness while preserving accurate colors and details and also suppresses noise, making it well-suited for industrial visual tasks.

Similarly, Figure 6 presents a comparison of the proposed method with four other low-light image enhancement approaches on the LOL dataset. The experimental results show that Zero-PTDCE excels in brightness enhancement, detail restoration, color fidelity, contrast optimization, and noise suppression: URetinex-Net significantly increases brightness but suffers from color distortions and local overexposure. SCI struggles with insufficient contrast in dark regions. Zero-DCE and Zero-DCE++ show improvements in detail recovery and color fidelity, but Zero-DCE introduces a blue cast, while Zero-DCE++ risks excessive sharpening in certain areas. In contrast, Zero-PTDCE ensures smoother brightness adjustments and optimized contrast tuning, producing more natural enhancement results while avoiding noise amplification due to overenhancement.

Figure 7 presents a qualitative comparison of the proposed method and four representative low-light image enhancement approaches on the SICE Part2 dataset. The results clearly demonstrate that Zero-PTDCE achieves superior enhancement performance in terms of brightness correction, structural detail preservation, and overall visual naturalness. URetinex-Net enhances brightness effectively, especially in extremely dark regions, but it tends to produce overexposed highlights and washed-out textures, especially around reflective or bright areas. SCI maintains good color balance but suffers from low contrast and incomplete detail recovery, particularly in shadow-heavy regions. Zero-DCE restores structural content better but introduces an obvious bluish color cast, which affects the visual realism of the output. Zero-DCE++ improves upon Zero-DCE with enhanced clarity and better contrast tuning, yet in some cases, it exhibits slight oversharpening and halo artifacts around edges. In contrast, Zero-PTDCE offers a balanced enhancement, showing stable brightness adjustment, fine-grained texture recovery, and natural color reproduction. It effectively avoids common pitfalls such as overenhancement or noise amplification, delivering visually pleasing and realistic results across diverse low-light scenarios present in SICE Part2.

Through comparative analysis across different datasets, the proposed Zero-PTDCE method demonstrates promising generalization and robustness within the tested domains. In terms of visual quality and perceptual enhancement, Zero-PTDCE consistently outperforms the other four methods, delivering visually natural and structurally enhanced results. It effectively improves brightness while maintaining true scene colors and fine details, making it well-suited for low-light industrial inspection tasks, particularly in coating workshop environments.

4.2.2. Quantitative Comparison

For the low-light image enhancement comparison experiments, we adopt full-reference image quality assessment metrics, including PSNR, SSIM, and MAE, to quantitatively evaluate the performance of different methods on the coating workshop low-light image test set and the LOL dataset.

PSNR is a key metric for measuring image quality, assessing the similarity between the enhanced image and the reference image. A higher PSNR value indicates that the enhanced image is closer to the reference image, with less distortion. The PSNR calculation formula is as follows:

$$PSNR=10\cdot {\log }_{10}\left({\displaystyle \frac{{\mathit{MAX}}^2}{\mathit{MSE}}}\right)$$

(8)

where MAX represents the maximum pixel value of the image. Mean squared error (MSE) measures the pixel-level error between the original image and the enhanced image.

SSIM evaluates the similarity between the enhanced and reference images in terms of structure, brightness, and contrast. Compared to PSNR, SSIM better aligns with human visual perception. The SSIM calculation formula is:

$$SSIM(x,y)={\displaystyle \frac{(2{\mu }_x{\mu }_y+C_1)(2{\sigma }_{xy}+C_2)}{({\mu }_x^2+{\mu }_y^2+C_1)({\sigma }_x^2+{\sigma }_y^2+C_2)}}$$

(9)

where: ${\mu }_x$, ${\mu }_y$ represent the mean values of the two images (brightness information). ${\sigma }_x$, ${\sigma }_y$ represent the variances of the two images (contrast information). ${\sigma }_{xy}$ is the covariance between the two images (structural information). $C_1$, $C_2$ are constants used to prevent division by zero.

MAE directly measures the pixel-level error between the enhanced and reference images, and its formula is:

$$MAE={\displaystyle \frac{1}{M\times N}}\sum _{i=1}^M\sum _{j=1}^N|I(i,j)-I^{\prime }(i,j)|$$

(10)

where: $I(i,j)$, $I^{\prime }(i,j)$ represent the grayscale values of the original and enhanced images at a specific pixel. M and N are the width and height of the image.

In general, a higher PSNR indicates better image quality with less noise, SSIM values range from 0 to 1, where values closer to 1 indicate higher structural similarity and enhancement quality, and a lower MAE suggests smaller pixel-level errors and higher image fidelity.

Table 1. Quantitative evaluation of different low-light enhancement methods.

Method	PSNR (dB) ↑		SSIM ↑		MAE ↓
	PT-LLIE	LOL	PT-LLIE	LOL	PT-LLIE	LOL
URetinex-Net	16.42	15.77	0.51	0.46	121.82	133.25
SCI	17.54	14.78	0.56	0.52	123.73	120.63
Zero-DCE	16.55	14.96	0.58	0.56	104.35	110.32
Zero-DCE++	16.58	15.84	0.58	0.56	103.67	109.36
Zero-PTDCE	19.61	15.23	0.63	0.58	107.75	108.56

presents the quantitative evaluation results of the proposed method compared to four other low-light image enhancement methods, with the best performance values highlighted in bold.

From Table 1, we observe that the proposed Zero-PTDCE outperforms other methods on the PT-LLIE dataset, achieving a PSNR of 19.61 dB and an SSIM of 0.63, demonstrating superior image quality and structural preservation. Additionally, on the LOL dataset, Zero-PTDCE achieves an SSIM of 0.58, outperforming other methods, which confirms its enhancement stability. While its MAE is slightly higher than that of Zero-DCE++, the overall error remains at a low level, ensuring high detail fidelity in the enhanced images. Overall, the proposed Zero-PTDCE exhibits superior comprehensive performance in low-light image enhancement tasks, making it well-suited for coating workshop environments while maintaining strong generalization capabilities.

4.2.3. Time Performance Evaluation

To assess the computational efficiency of different methods, we measured the average processing time per image under the same experimental environment. The tested image resolution is 1200 × 900. The results are shown in

Table 2. Average processing time per image for different algorithms.

Method	URetinex-Net	SCI	Zero-DCE	Zero-DCE++	Zero-PTDCE
Runtime (s)	2.1325	0.0516	0.0026	0.0013	0.0012
Platform	PyTorch (GPU)	PyTorch (GPU)	PyTorch (GPU)	PyTorch (GPU)	PyTorch (GPU)

. Experimental results indicate that Zero-PTDCE achieves high processing efficiency while ensuring enhanced image quality. Compared to URetinex-Net and SCI, its processing time is significantly reduced. Additionally, when compared with Zero-DCE++, Zero-PTDCE maintains a similar computational time, further demonstrating that the proposed method improves image enhancement quality without introducing extra computational overhead, making it highly practical for real-world applications.

4.2.4. Subjective Evaluation

To further evaluate the perceptual quality of images enhanced by Zero-PTDCE, we conducted a subjective scoring experiment. The objective of this experiment was to assess the visual naturalness of enhanced images produced by URetinex-Net, SCI, Zero-DCE, Zero-DCE++, and Zero-PTDCE, in terms of brightness, detail fidelity, and visual comfort. A total of 15 participants were invited to independently score the enhanced images on a scale of 1 to 5, where 1 indicates poor image quality and 5 indicates excellent quality.

We selected enhanced results from the PT-LLIE dataset. Each enhanced image was displayed on the screen alongside the original input image for reference. All participants independently scored the visual quality of the results, with specific attention to the following aspects:

(1)

Whether the result contains artifacts of over/underexposure or regions that are over/underenhanced;

(2)

Whether the result introduces any color distortion;

(3)

Whether the textures appear unnatural or noticeable noise is present.

The subjective evaluation scores for each enhancement method are summarized in

Table 3. Subjective evaluation scores on PT-LLIE, LOL, and SICE Part2 datasets.

Method	URetinex-Net	SCI	Zero-DCE	Zero-DCE++	Zero-PTDCE
PT-LLIE	3.32	4.02	3.18	3.43	4.15
LOL	3.25	3.85	3.23	3.55	3.95
SICE Part2	3.12	3.58	3.55	3.92	4.11

. As shown in the table, Zero-PTDCE achieved the highest scores across all three evaluation criteria. These results further demonstrate the superior perceptual quality of Zero-PTDCE-enhanced images, especially in industrial applications where preservation of fine details is crucial for accurate inspection and analysis.

4.3. Ablation Study

To validate the effectiveness of Zero-PTDCE in low-light image enhancement, an ablation study was conducted by systematically evaluating the contributions of different improvement modules. The experiments utilized real-world low-light images captured in the coating workshop as test data, with various model versions defined as follows: Z0 represents the baseline Zero-DCE model, while Z1 corresponds to the upgraded Zero-DCE++ version. Building upon Z1, Z2 incorporates a pre-denoising module to reduce noise interference and enhance input image quality. Further extending this, Z3 integrates depthwise separable dilated convolution to expand the receptive field while maintaining computational efficiency, thereby improving the extraction of fine details and structural information. The final version, Z4, introduces a perceptual loss function to refine the enhancement process by preserving high-level structural and perceptual details. Different improvement combinations are shown in Figure 8. The enhancement performance of each model variant was quantitatively assessed using standard image quality metrics, with the results summarized in

Table 4. Quantitative evaluation metrics of different improvement combinations (PT-LLIE dataset).

Improvement Combinations	PSNR (dB) ↑	SSIM ↑	MAE ↓
Z0	16.55	0.58	104.35
Z1	16.58	0.58	103.67
Z2	16.52	0.60	105.85
Z3	18.22	0.60	105.35
Z4	19.61	0.63	103.75

, demonstrating the progressive improvements achieved through the incorporation of each module.

Independent Contribution of the Pre-Denoising Module (Z1 → Z2):

The pre-denoising module is primarily designed to suppress noise artifacts commonly present in low-light environments prior to image enhancement. When comparing Z2 to Z1, the SSIM increases from 0.58 to 0.60, indicating improved structural fidelity of the enhanced image. However, the PSNR slightly decreases from 16.58 dB to 16.52 dB, which suggests a potential trade-off between noise suppression and the preservation of fine details. This trade-off is typical in low-light enhancement tasks, especially in scenarios involving severe noise contamination, such as reflective metallic surfaces or shadowed regions in industrial settings. Despite the marginal drop in PSNR, the pre-denoising module significantly improves the overall perceptual quality of the image by producing clearer contours and reducing noise perception in critical regions.

Independent Contribution of the Depthwise Separable Dilated Convolution (Z2 → Z3):

Building upon Z2, the integration of depthwise separable dilated convolution yields a notable performance improvement. The PSNR increases from 16.52 dB to 18.22 dB, while the SSIM remains stable at a relatively high level of 0.60. By introducing dilation rates, this module effectively expands the receptive field without significantly increasing computational complexity, thereby enhancing the network’s ability to capture both local details and global contextual features. This is particularly beneficial in coating inspection images that often contain complex textures, repetitive patterns, and fine structural elements. The inclusion of this module leads to enhanced sharpness, improved contrast, and more consistent structural representation in the enhanced outputs.

Independent Contribution of the Perceptual Loss Function (Z3 → Z4):

Further incorporating the perceptual loss function on top of Z3 leads to significant improvements in both objective metrics and subjective visual quality. The PSNR rises from 18.22 dB to 19.61 dB, and the SSIM improves from 0.60 to 0.63. Perceptual loss encourages the network to focus on high-level semantic consistency and texture fidelity during training, thereby enhancing the naturalness of the output. Qualitative observations indicate that Z4 produces images with smoother gradient transitions, richer texture hierarchies, and more realistic details, especially in surface patterns. Additionally, MAE is reduced from 105.35 to 103.75, indicating that the enhanced images are quantitatively closer to the ground truth. This confirms the effectiveness of the perceptual loss in guiding the network toward generating more perceptually compelling results.

In summary, the ablation study demonstrates that each module contributes complementary benefits to the Zero-PTDCE framework: the pre-denoising module enhances clarity by reducing noise, the dilated convolution improves the balance between local and global feature representation, and the perceptual loss further optimizes visual realism and semantic quality. When combined, these components (Z4) achieve the best performance across all evaluation metrics, yielding the most balanced and high-quality enhancement results.

5. Summary

Aiming at the issue of poor image quality and low detection accuracy in the inspection process of coating workshops under low-light conditions, this study proposes a low-light image enhancement method based on zero-reference deep curve estimation, termed Zero-PTDCE. The main conclusions are as follows:

(1) This study constructs a low-light image dataset, PT-LLIE, specifically designed for coating workshop inspection, covering industrial detection scenarios under varying illumination conditions. Experimental results demonstrate that this dataset effectively enhances the model’s generalization ability in complex low-light environments.

(2) A novel enhancement network, PTDCE-Net, is designed, integrating a lightweight denoising module and depthwise separable dilated convolution. The pre-denoising module effectively reduces noise interference and improves input image quality, while depthwise separable dilated convolution expands the receptive field while reducing computational complexity, thereby enhancing image details and structural information. Experimental results indicate that this design significantly improves the clarity and contrast of enhanced images.

(3) A multi-constraint loss strategy is proposed, incorporating perceptual loss L_p, color loss L_c, spatial consistency loss L_s, exposure loss L_e, and total variation loss L_tv. Experiments show that this strategy effectively prevents issues such as uneven brightness, color distortion, and structural information loss in enhanced images, ensuring natural brightness, balanced colors, and well-preserved structural integrity.

(4) Compared to existing low-light image enhancement methods, Zero-PTDCE achieves superior performance in terms of PSNR, SSIM, and MAE metrics. The enhanced images exhibit improved brightness, contrast, and detail restoration, making them closer to real-world scenes. Additionally, while maintaining high enhancement quality, the computational efficiency of Zero-PTDCE is comparable to that of Zero-DCE++ and outperforms URetinex-Net and SCI, demonstrating its practicality and efficiency in coating workshop inspection tasks.

In summary, the proposed Zero-PTDCE method enables image enhancement based solely on the input image itself, without requiring reference images. It significantly improves the quality of low-light images in coating workshops, providing a reliable image enhancement solution for intelligent inspection and supporting the industry’s transition toward intelligent automation. Beyond its utility in inspection tasks, Zero-PTDCE also demonstrates strong potential in the visual quality assessment of coatings, where enhanced imaging facilitates the early identification of surface anomalies such as uneven thickness, scratches, and delamination under suboptimal lighting. This improvement is critical for ensuring the performance, uniformity, and reliability of coated materials. As such, the proposed method contributes not only to operational efficiency but also to advancing intelligent material characterization and quality control, aligning well with the broader scope of modern coating technology research.