Edge-Aware Illumination Enhancement for Fine-Grained Defect Detection on Railway Surfaces

Abstract

Fine-grained defects on rail surfaces are often inadequately detected by conventional vision-based object detection models in low-light environments. Although this problem can be mitigated by enhancing image brightness and contrast or employing deep learning-based object detectors, these methods frequently distort critical edge and texture information essential for accurate defect recognition. Herein, we propose a preprocessing framework that integrates two complementary modules, namely adaptive illumination enhancement (AIE) and EdgeSeal enhancement (ESE). AIE leverages contrast-limited adaptive histogram equalization and gamma correction to enhance local contrast while adjusting the global brightness distribution. ESE further refines defect visibility through morphological closing and sharpening, enhancing edge continuity and structural clarity. When integrated with the You Only Look Once v11 (YOLOv11) object detection model and evaluated on a rail defect dataset, the proposed framework achieves an ~7% improvement in mean average precision over baseline YOLOv11 and outperforms recent state-of-the-art detectors under diverse low-light and degraded-visibility conditions. The improved precision and recall across three defect classes (defects, dirt, and gaps) demonstrate the robustness of our approach. The proposed framework holds promise for real-time railway infrastructure monitoring and automation systems and is broadly applicable to low-light object detection tasks across other industrial domains.

1. Introduction

Urban railways have become a fundamental transportation mode because of their high operational speed, large passenger capacity, and environmental sustainability. However, structural defects on rail surfaces pose serious threats to operational safety and passenger comfort and can cause system-wide inefficiencies [1,2]. When not detected early, fine-grained defects may accumulate over time, increasing the risk of derailment, mechanical failure, and service disruption, which can result in severe accidents and considerable economic losses. Traditional rail defect inspections heavily rely on manual visual assessments, which are labor-intensive, inconsistent, and prone to human error [3]. Consequently, automated, accurate, and robust rail defect detection systems are highly sought after.

Automated inspection techniques can be broadly classified into physical sensor-based nondestructive testing (NDT) and vision-based approaches. NDT methods, such as ultrasonic, magnetic particle, and eddy current testing, are effective for detecting subsurface defects without physically damaging the infrastructure. For instance, guided wave-based ultrasonic structural health monitoring has been used for real-time rail defect detection [4], and three-dimensional magnetic profiling and Fourier transform-based eddy current analysis have been used for efficient surface anomaly detection [5,6]. Despite their effectiveness, these methods often suffer from high equipment costs and dependency on skilled operators.

In contrast, vision-based methods leveraging deep learning and low-cost imaging devices offer a scalable and efficient alternative [7,8]. Recent efforts have focused on using and optimizing convolutional neural networks, particularly variants of You Only Look Once (YOLO), for rail defect detection under normal lighting conditions [9,10,11]. However, under low-light scenarios common in tunnels, stations, and nighttime environments, these models struggle because of low contrast, reduced texture visibility, and loss of boundary details [12].

Various image enhancement techniques have been introduced as preprocessing steps to improve visual quality prior to detection and thus mitigate the abovementioned problems. Methods such as histogram equalization, Retinex-based enhancement, and generative adversarial network (GAN)-based restoration improve global brightness and contrast [13,14,15] but often fail to preserve the fine contours and textures critical for detecting small defects. Moreover, conventional models frequently lack adaptability to variable illumination conditions, which hinders feature extraction and lowers detection accuracy [16,17,18]. While recent low-light image enhancement methods have achieved notable improvements by combining illumination correction with structural preservation, many of these approaches either require complex network modifications or lack modularity, which limits their applicability to existing detection frameworks.

Herein, we propose a modular two-stage preprocessing framework that enhances low-light images without altering the structure of the downstream detector (YOLOv11 [19]). This framework explicitly separates illumination enhancement and structural edge preservation, addressing the limitations of prior methods. The adaptive illumination enhancement (AIE) module enhances local contrast and adjusts global brightness distributions through contrast-limited adaptive histogram equalization (CLAHE) and gamma correction, while the EdgeSeal enhancement (ESE) module reinforces structural edges and defect boundaries using morphological operations and sharpening filters.

Although the individual operations employed in AIE and ESE are well-established in the literature, the originality of this study lies in how these operations are modularly organized, adaptively tuned, and seamlessly integrated into the detection pipeline. Unlike conventional approaches that apply enhancement operators in a fixed sequence, the proposed framework optimizes the cooperation between illumination correction and edge reinforcement to ensure stability across diverse low-light environments.

This modular design offers a practical and generalizable preprocessing solution that can be embedded into existing object detection systems without architectural modification or retraining. The proposed framework is integrated with the YOLOv11 detection model and evaluated on a rail surface defect dataset, achieving an ~7% increase in mean average precision (mAP) across three defect categories (defects, dirt, and gaps) compared with the baseline model. The main novelty of this work, therefore, lies not in the use of any specific enhancement algorithm but in the design of a flexible and adaptive preprocessing architecture that significantly improves the robustness of YOLO-based rail defect detection under low-light and degraded-visibility conditions. The key contributions of this paper are as follows.

• AIE using CLAHE and gamma correction effectively improves local contrast and adjusts global brightness distributions under low-light conditions.
• ESE, based on a morphological edge-focused enhancement strategy, preserves fine defect contours and sharpens structural boundaries to support high-fidelity feature extraction.
• When integrated with YOLOv11, the proposed preprocessing framework leverages its modular and adaptive design to enhance defect detection performance across diverse low-light and degraded-visibility conditions, achieving up to 7% increases in mAP over the baseline model.

2. Related Works

2.1. Vision-Based Rail Defect Detection

Deep learning has markedly advanced automated rail surface inspection, facilitating the development of lightweight and efficient models suitable for real-time deployment. For instance, GhostMicroNet-YOLOv8n was introduced for fast and energy-efficient damage detection on rails [20]. Similarly, improved YOLOv8 architectures, optimized for precision and speed, have achieved detection accuracies of up to 94.9% [21].

Attention mechanisms and feature fusion strategies have also been employed to enhance defect localization, e.g., a CenterNet-based approach with a ResNeXt backbone was used to extract multiscale features, enabling the detection of small defects [22]. Other works, such as those using DP-YOLO, incorporated depthwise separable convolutions and attention modules to reduce computational load while maintaining high precision in fastener defect detection [23,24,25]. Multimodal learning has emerged as a promising direction, with models such as FusWay combining vision and audio signals using vision transformers and YOLOv8 to improve robustness in real-world rail inspection scenarios [26]. Although these methods demonstrate competitive performance under ideal lighting conditions, their accuracy tends to degrade in low-illumination environments because of reduced contrast, weak texture visibility, and loss of fine edge structures.

2.2. Low-Light Image Enhancement for Industrial Applications

The enhancement of image quality under poor lighting has been explored across various industrial domains to improve detection accuracy. Retinex-based approaches, such as KinD [27] and Retinexformer [28], decompose images into reflectance and illumination components to achieve better restoration. GANs have been employed to synthesize clearer and brighter images. For example, CES-GAN effectively restored low-light fastener images using a U-Net architecture with adversarial training [29]. BrightsightNet [30] proposed a lightweight U-Net–based network designed for low-light railway image enhancement, which effectively learns both structural and illumination and structural cues to support reliable defect detection without explicit preprocessing. Similarly, LFGB-YOLO [31] integrates a low-light enhancement module directly within the detection architecture, achieving improved precision and recall for rail fastener detection under challenging illumination conditions. Moreover, Enhanced YOLOv5 [32] combines illumination correction and defect detection within a unified framework, ensuring stable crack detection performance in tunnel environments with insufficient lighting. Other studies have focused on anomaly detection under weak lighting conditions, highlighting the importance of preserving texture fidelity for reliable detection [33]. However, most of these methods focus solely on image enhancement without integration into detection pipelines or suffer from oversmoothing, which results in the loss of critical boundary details.

2.3. Research Gap and Our Approach

Despite the notable advancements in the field of vision-based rail defect detection, several limitations remain.

• Global enhancement techniques often fail to preserve fine textures and boundary features essential for detecting microdefects.
• Most detection models are not optimized for illumination variability, which limits their effectiveness in real-world low-light scenarios.
• Few works combine local contrast enhancement and structural edge reinforcement within a unified preprocessing framework.

To address these limitations, we developed a two-stage preprocessing approach integrating AIE and ESE modules to simultaneously enhance image brightness and preserve defect boundary integrity. This framework, when integrated with a state-of-the-art detection model, delivered substantial accuracy improvements across various defect types and challenging environmental conditions.

3. Methodology

Low-light images of rail surfaces often suffer from degraded quality, which hinders the accurate identification of rail defects. Images captured under low-illumination conditions, e.g., at night, inside tunnels, or in backlit environments, typically exhibit low brightness, poor contrast, and blurred boundaries of rail defects, which markedly reduces detection accuracy [34]. The various image enhancement techniques used to address this issue often fail to preserve the visual information pertaining to small and fine-grained defects. Specifically, simple contrast adjustments or global filtering techniques are insufficient to retain the structural characteristics and contours of rail defects [35].

Herein, we propose a modular image preprocessing framework that combines AIE and ESE to effectively preserve and highlight visual information pertaining to rail defects, even under low-light conditions, primarily aiming to maximize the detection performance of the YOLOv11 object detection model based on enhanced image quality [36]. The proposed preprocessing-based rail defect detection framework features three main stages. AIE enhances brightness and contrast to restore the visual information pertaining to dark regions within the rail images. ESE emphasizes the contours and boundaries of rail defects, increasing the prominence of their structural features. Finally, the preprocessed images are input into the YOLOv11 object detection model to precisely detect the location and type of rail defects. The architecture of the proposed framework is shown in Figure 1, and the details of each module are described in the following sections.

3.1. Adaptive Illumination Enhancement

Low-light rail images often suffer from low overall brightness and insufficient contrast between defects and their surrounding background, which hinders visual defect identification. As a result, defect shapes and boundaries become blurred, and the performance of subsequent processes, such as binarization, defect detection, and object recognition, deteriorates. Therefore, an effective preprocessing step that compensates for the brightness and contrast characteristics of input images is crucial for enhancing rail defect detection accuracy. To address these challenges, we developed an AIE module that sequentially applies CLAHE and gamma correction. Although CLAHE and gamma correction are established image enhancement techniques, their adaptive integration within the AIE module, in coordination with the ESE module, allows the framework to optimally balance local contrast improvement and global brightness adjustment, ensuring robustness across diverse low-light environments.

3.1.1. CLAHE

CLAHE works by dividing the input image into $\mathrm{S}$ nonoverlapping subregions of size $\mathrm{m}\times \mathrm{m}$. The pixel intensity distribution within each region is independently analyzed and adjusted to improve local contrast. For each region $\mathrm{i}$, the histogram value ${\mathrm{H}}_{\mathrm{i}}\left(\mathrm{k}\right)$ is normalized as ${\mathrm{H}}_{\mathrm{i}}^{\mathrm{n}}\left(\mathrm{k}\right)={\mathrm{H}}_{\mathrm{i}}\left(\mathrm{k}\right)/{\mathrm{N}}_{\mathrm{i}}$, where $N_i$ is the total number of pixels in region $\mathrm{i}$. A cumulative distribution function (CDF) is then calculated for each region, with the cumulative probability for intensity level $\mathrm{k}$ given by

$${\mathrm{C}\mathrm{D}\mathrm{F}}_{\mathrm{i}}\left(\mathrm{k}\right)={\sum }_{\mathrm{j}=0}^{\mathrm{k}}{\mathrm{H}}_{\mathrm{i}}^{\mathrm{n}}\left(\mathrm{j}\right).$$

(1)

To avoid noise amplification, a clipping limit (${\mathrm{C}}_{\mathrm{l}\mathrm{i}\mathrm{m}}$) is applied, yielding a clipped histogram ${\mathrm{H}}_{\mathrm{i}}^{\mathrm{c}}\left(\mathrm{k}\right)=\min \left({\mathrm{H}}_{\mathrm{i}}^{\mathrm{n}}\left(\mathrm{k}\right),{\mathrm{C}}_{\mathrm{l}\mathrm{i}\mathrm{m}}\right)$. The CDF of the clipped histogram is then recalculated as

$${\mathrm{C}\mathrm{D}\mathrm{F}}_{\mathrm{i}}^{\mathrm{c}}\left(\mathrm{k}\right)={\sum }_{j=0}^k{\displaystyle \frac{{\mathrm{H}}_{\mathrm{i}}^{\mathrm{c}}\left(\mathrm{j}\right)}{{\mathrm{N}}_{\mathrm{i}}}}.$$

(2)

Finally, the pixel intensity is remapped using the minimum CDF value, ${\mathrm{C}\mathrm{D}\mathrm{F}}_{i,\mathrm{m}\mathrm{i}\mathrm{n}}$, with the result given by

$${\mathrm{I}}_{\mathrm{c}\mathrm{l}\mathrm{a}}=\left({\displaystyle \frac{\mathrm{C}\mathrm{D}{\mathrm{F}}_{\mathrm{i}}^{\mathrm{c}}\left(\mathrm{k}\right)-{\mathrm{C}\mathrm{D}\mathrm{F}}_{\mathrm{i},\mathrm{m}\mathrm{i}\mathrm{n}}}{1-{\mathrm{C}\mathrm{D}\mathrm{F}}_{\mathrm{i},\mathrm{m}\mathrm{i}\mathrm{n}}}}\right)\times \left(\mathrm{L}-1\right),$$

(3)

where $\mathrm{L}$ is the total number of possible pixel intensity levels (typically 256). This process ensures that local contrast is enhanced uniformly across each region, making defects more visually distinct in low-light rail images.

3.1.2. Gamma Correction

Gamma correction adjusts global brightness, enhancing visual information in low-light areas. Although CLAHE improves local contrast and emphasizes fine defect details, it may struggle with shadowed regions or subtle cracks, where boundaries are not clearly defined. To address this problem, gamma correction was applied after CLAHE processing to further amplify the brightness of dark regions and thus enhance defect visibility.

The image ${\mathrm{I}}_{\mathrm{c}\mathrm{l}\mathrm{a}}$ processed by CLAHE was first normalized to a range of 0–255. Then, the gamma correction was applied as

$${\mathrm{I}}_{\mathrm{g}\mathrm{m}\mathrm{a}}={255\times \left({\displaystyle \frac{{\mathrm{I}}_{\mathrm{c}\mathrm{l}\mathrm{a}}}{255}}\right)}^{\mathsf{\gamma }},$$

(4)

where $\mathsf{\gamma }$ is the gamma coefficient that is typically set to $\mathsf{\gamma }<1$ to enhance the brightness of dark pixels. The combined effect of CLAHE-based local contrast enhancement and gamma correction–based global brightness adjustment improved the overall contrast of rail images, increasing defect detection accuracy. Figure 2 compares the original, CLAHE-processed, and post-gamma-correction (final) images.

3.2. EdgeSeal Enhancement

Despite the improvement in brightness and local contrast achieved using the AIE module, the structural form and boundaries of fine defects in low-light rail images remained unclear. Cracks, wear, and damage in rail rails are often characterized by small size, low intensity, and weak contrast with the background, which hinders visual identification based on simple illumination correction techniques alone. As a result, the incidence rates of false positives (FPs) and false negatives (FNs) in defect detection increase.

To address this problem, we proposed an ESE module combining morphological closing and sharpening operations. This module was designed to preserve the structural boundaries of defects while suppressing background noise, providing clearer edge information during the feature extraction stage of object detection.

3.2.1. Morphological Closing

Morphological operations are based on the geometric structure of objects within an image and used to correct their boundaries and structural continuity. Among these, the closing operation is particularly useful for enhancing defect visibility in low-light conditions, sequentially applying dilation and erosion operations to fill small holes within an object and smooth its outer contour. This operation is mathematically defined as

$${\mathrm{I}}_{\mathrm{m}\mathrm{o}\mathrm{p}}=\left({\mathrm{I}}_{\mathrm{g}\mathrm{m}\mathrm{a}}\oplus \mathrm{B}\right)\ominus \mathrm{B},$$

(5)

where $\mathrm{B}$ is a structural kernel; $\oplus$ denotes the dilation operation, where the pixel is set to unity if any part of the image overlaps with the kernel; and $\ominus$ represents the erosion operation. The dilation operation expands the boundary of the object and is effective in filling irregular defect contours or small holes within the defect region. The erosion operation contracts the object boundary, effectively restoring the original shape while removing unnecessary weak noise or isolated pixels.

Through this sequence, the closing operation flattened low-frequency image components, enhancing the structural continuity of defects while suppressing background noise. This process was especially effective in revealing the contours of fine defects, such as microcracks on rails, which were challenging to detect in low-light images.

3.2.2. Sharpening

Although morphological closing successfully corrected the overall contours of rail defects, finer details such as boundary lines, textures, and structural information remained blurred, which is critical for detecting small defects like micro-cracks. To further enhance the clarity of these high-frequency features, sharpening was applied after morphological processing.

Sharpening emphasized the high-frequency components of images, such as boundaries and textures, thereby making the defect contours sharper and more distinct from the background. The first stage involved applying Gaussian blurring is represented as

$${\mathrm{I}}_{\mathrm{b}\mathrm{l}\mathrm{u}}={\mathrm{I}}_{\mathrm{m}\mathrm{o}\mathrm{p}}\mathrm{*}{\mathrm{G}}_{\mathsf{\sigma }},$$

(6)

where ${\mathrm{G}}_{\mathsf{\sigma }}$ is a Gaussian kernel with standard deviation $\mathsf{\sigma }$ used to blur the image and “$\mathrm{*}$” denotes the convolution operation, which applies the kernel to the image ${\mathrm{I}}_{\mathrm{m}\mathrm{o}\mathrm{p}}$. The high-frequency component ${\mathrm{I}}_{\mathrm{h}\mathrm{i}}$ was extracted by ${\mathrm{I}}_{\mathrm{m}\mathrm{o}\mathrm{p}}-{\mathrm{I}}_{\mathrm{b}\mathrm{l}\mathrm{u}}$, which contained the edges, texture, and finer structural details of the defects and captured the essential information pertaining to defect contours and small structures critical for detecting microdefects in rails. Finally, ${\mathrm{I}}_{\mathrm{h}\mathrm{i}}$ was combined with ${\mathrm{I}}_{\mathrm{m}\mathrm{o}\mathrm{p}}$ to produce the sharpened image ${\mathrm{I}}_{\mathrm{s}\mathrm{h}\mathrm{p}}$ as

$${\mathrm{I}}_{\mathrm{s}\mathrm{h}\mathrm{p}}={\mathrm{I}}_{\mathrm{m}\mathrm{o}\mathrm{p}}+\mathsf{\alpha }·{\mathrm{I}}_{\mathrm{h}\mathrm{i}},$$

(7)

where $\mathsf{\alpha }$ is a scaling factor that controls the intensity of the high-frequency component in the final sharpened image and thus helps adjust the degree to which defect details are emphasized. The visual effects of these enhancements are shown in Figure 3, which illustrates the improvement in defect visibility and enhanced sharpness of defect boundaries, validating the effectiveness of the proposed approach.

3.3. Object Detection

Object detection simultaneously estimates the class and location of objects within an image by predicting the bounding box and class information for each object. Traditional object detection techniques, which feature two separate stages, have been superseded by end-to-end integrated approaches performing both tasks simultaneously. Among these, the YOLO series is a prominent object detection model, dividing the input image into a grid and predicting the presence of an object, its class, and the bounding box coordinates within each grid cell. This architecture offers the advantages of high processing speed and accuracy and is therefore suitable for real-time applications. Furthermore, numerous structural improvements aimed at reducing model size while enhancing accuracy have been proposed [37].

Herein, we adopted the latest version of the YOLO series, YOLOv11, as the rail defect detection model. YOLOv11 is designed to achieve model lightweighting and high-precision inference performance, which is particularly beneficial for the detection of small defects and unclear boundaries, such as those in rail images captured under low-light conditions. The overall network architecture of YOLOv11 features three modules, namely the backbone, neck, and head (Figure 4).

The backbone is the module responsible for extracting features from the input image at various levels and employs several lightweight blocks that balance computational efficiency with representational power. Key components include the traditional convolution block (CBS: convolution + batch normalization + SiLU), as well as more specialized modules such as C3k2, C2PSA, and spatial pyramid pooling fast, which enhance feature extraction across multiple scales. These components allow the model to capture meaningful features at different resolutions, thereby providing a foundation for detecting defects of varying sizes.

The subsequent neck module integrates high- and low-level features to effectively generate multiscale feature maps and combines the feature pyramid network and path aggregation network to optimize feature fusion. Finally, the head module outputs the detection results, predicting the object class, bounding box coordinates, and object presence probability for each grid cell. This process is trained based on a multiloss function given by

$${\mathcal{L}}_{\mathrm{t}\mathrm{o}\mathrm{t}}={{\mathsf{\lambda }}_{\mathrm{c}\mathrm{l}\mathrm{s}}·\mathcal{L}}_{\mathrm{c}\mathrm{l}\mathrm{s}}+{\mathsf{\lambda }}_{\mathrm{b}\mathrm{o}\mathrm{x}}·{\mathcal{L}}_{\mathrm{b}\mathrm{o}\mathrm{x}}+{\mathsf{\lambda }}_{\mathrm{o}\mathrm{b}\mathrm{j}}·{\mathcal{L}}_{\mathrm{o}\mathrm{b}\mathrm{j}},$$

(8)

where ${\mathcal{L}}_{\mathrm{c}\mathrm{l}\mathrm{s}}$, ${\mathcal{L}}_{\mathrm{b}\mathrm{o}\mathrm{x}}$, and ${\mathcal{L}}_{\mathrm{o}\mathrm{b}\mathrm{j}}$ represent the classification loss, bounding box regression loss, and objectness loss, respectively. ${\mathcal{L}}_{\mathrm{c}\mathrm{l}\mathrm{s}}$ evaluates how well the predicted object class matches the actual label, ${\mathcal{L}}_{\mathrm{b}\mathrm{o}\mathrm{x}}$ quantifies the alignment between the predicted bounding box and true object position and size, and ${\mathcal{L}}_{\mathrm{o}\mathrm{b}\mathrm{j}}$ determines the presence of an object at the predicted location. ${\mathsf{\lambda }}_{\mathrm{c}\mathrm{l}\mathrm{s}}$, ${\mathsf{\lambda }}_{\mathrm{b}\mathrm{o}\mathrm{x}}$, and ${\mathsf{\lambda }}_{\mathrm{o}\mathrm{b}\mathrm{j}}$ are weights controlling the importance of each loss term. To improve the precision of bounding box localization, the complete intersection over union loss method was employed. Unlike the traditional intersection over union, this method considers not only the overlap between boxes but also the distance between their center points and their aspect ratios, enabling more accurate bounding box regression.

4. Experimental Results

To conduct the experiments, we utilized the Rail Defect dataset [38], which includes low-light images of real rails. The original and preprocessed images were tested using the same YOLOv11 model, and its defect detection performance was assessed in terms of precision, recall, and mAP. By analyzing the performance differences between various comparison models, we aimed to validate the effectiveness of the proposed method.

4.1. Experimental Setup

Experiments were conducted using a set of real rail surface images acquired in low-light environments. The dataset included various defect types, primarily structural damages and surface contaminations. Initially, eight flaw classes were annotated, namely dents, crushes, scratches, slants, damages, unknown, dirt, and gaps. Owing to high visual similarity and inherent class imbalance among six of these classes (dents, crushes, scratches, slants, damages, unknown), they were consolidated into a single class labeled “defects”. Consequently, the final classification was performed over three target classes, namely defects, dirt, and gaps. Representative samples for each class are shown in Figure 5.

The dataset (440 images), collected from operational railway environments including tunnels and nighttime scenes with low-light and uneven illumination, was divided into training (318), validation (42), and testing (80) subsets. Three defect classes were annotated for evaluation: Defect (200), Dirt (162), and Gap (78). All images were resized to 640 × 640 pixels to ensure consistency in input dimensions. The CLAHE operation was applied with a clip limit of 2.0 and a tile grid size of 8 × 8 to enhance local contrast while avoiding over-amplification of noise. Gamma correction was uniformly applied with a γ value of 1.2, effectively brightening underexposed regions. Morphological operations for edge preservation employed a 3 × 3 kernel for both erosion and dilation. Gaussian blur was implemented with a 5 × 5 kernel and a standard deviation of 1.0 to suppress high-frequency noise, followed by a sharpening factor of 1.5 to restore structural sharpness and contour integrity.

Experiments were conducted on a machine running Windows 10 and featuring two Intel Xeon Silver 4310 processors, 16 GB of DDR3 memory, and an NVIDIA GeForce RTX 1080 graphics processing unit. Model training was conducted using the Adam optimizer with a learning rate of 1 × 10⁻⁴ and weight decay of 1 × 10⁻⁵. Each model was trained for 100 epochs with a batch size of 16, and early stopping was applied with a patience of 10 epochs based on the validation mAP. To improve model generalization, data augmentation techniques including horizontal and vertical flips, random rotations of ±15°, and brightness adjustments of ±10% were applied during training.

Recall is defined as the number of correctly detected defects divided by that of all defects. This metric reflects the ability of the model to minimize detection omissions (FNs) and can also be expressed as recall = true positives (TPs, i.e., number of correctly detected defects)/(TPs + FNs). Precision is defined as the number of true defects divided by that of model-detected defects. This metric quantifies the ability of the model to minimize false detections (FPs) and can also be expressed as precision = TPs/(TPs + FPs), with higher values corresponding to lower FP rates.

4.2. Ablation Study

To quantitatively evaluate the individual and combined contributions of the proposed AIE and ESE modules, we conducted a series of ablation experiments.

Table 1. Performance comparison of individual and integrated AIE and ESE modules.

Module		Precision			Recall			mAP @50	mAP @50~95
AIE	ESE	Defect	Dirt	Gap	Defect	Dirt	Gap
-	-	0.84	0.85	0.98	0.79	0.61	1.00	0.85	0.55
√	-	0.88	0.82	0.94	0.65	0.78	1.00	0.87	0.56
-	√	0.90	0.91	0.88	0.66	0.78	1.00	0.88	0.56
√	√	0.86	0.92	0.99	0.90	0.89	1.00	0.92	0.59

summarizes the detection performance under different module configurations.

Under the baseline condition without either module, the Precision for Defect, Dirt, and Gap classes was 0.84, 0.85, and 0.98, respectively, with corresponding Recall values of 0.79, 0.61, and 1.00. The overall mAP@50 and mAP@50–95 were measured as 0.85 and 0.55, respectively.

When the AIE module was applied alone, the Precision for the Defect class increased to 0.88, while its Recall decreased to 0.65, reflecting a slight compromise in structural detail recognition. The Dirt class exhibited a small decrease in Precision (0.85 → 0.82) but a significant increase in Recall (0.61 → 0.78), while the Gap class remained similar to the baseline. Overall, the mAP@50 increased to 0.87, indicating that the AIE effectively mitigates local brightness inconsistencies and enhances contrast under low-light conditions. However, excessive enhancement may lead to texture distortion, particularly affecting fine structural anomalies.

When the ESE module was applied alone, Precision values for the Defect and Dirt classes increased to 0.90 and 0.91, respectively, while Recall remained similar to the AIE-only configuration (0.66 and 0.78). The overall mAP@50 and mAP@50–95 were 0.88 and 0.56, respectively, demonstrating comparable performance gains. These results indicate that ESE effectively sharpens structural edges and preserves object contours, suppressing false positives and improving boundary recognition. Nevertheless, over-amplification of edges may slightly distort fine boundary textures, potentially reducing structural detail recognition for specific classes.

Integrating both AIE and ESE modules led to consistent performance improvements across all defect categories. Precision and Recall increased for the Defect and Dirt classes, while the Gap class achieved near-perfect scores. The overall mAP@50 and mAP@50–95 reached 0.92 and 0.59, respectively, highlighting a clear improvement over the baseline. These results indicate that AIE and ESE complement each other: AIE enhances visual cues by adjusting local contrast, while ESE reinforces structural contours and mitigates noise amplification. The combined effect effectively balances the Precision–Recall trade-off, enabling robust and accurate detection under challenging low-light conditions.

Table 2. Repeated run performance of the proposed and baseline models.

Model	mAP@0.50
	Run 1	Run 2	Run 3	Run 4	Run 5	Mean ± Std
YOLOv11	85.5	85.8	85.2	85.3	85.4	85.4 ± 0.2
Proposed model	92.9	92.5	92.6	92.4	92.8	92.6 ± 0.2

presents repeated-run evaluations of the baseline YOLOv11 and the proposed AIE + ESE-based YOLOv11. The proposed framework achieved an average mAP@50 of 92.6 ± 0.2, significantly outperforming the baseline (85.4 ± 0.2). The low standard deviation confirms that the proposed enhancement framework delivers stable and reliable performance across multiple trials.

Collectively, these quantitative findings demonstrate that the proposed preprocessing modules substantially improve defect visibility and detection performance under low-light conditions. The complementary effects of AIE and ESE preserve structural integrity, enhance local contrast, and suppress irrelevant features, thereby supporting robust and generalizable detection across diverse defect types.

In addition, to clearly determine the relative advantages of the proposed AIE and ESE modules over existing preprocessing techniques, additional comparative experiments were conducted. For this purpose, the YOLOv11 baseline model was evaluated using datasets preprocessed by traditional image enhancement methods—CLAHE, histogram equalization (HE), and dark channel prior (DCP)—as well as the recent deep learning–based enhancement method LEGAN [39].

Table 3. Detection performance comparison of low-light image enhancement methods for rail defect detection using the YOLOv11 baseline model.

Image Enhancement	Precision			Recall			mAP @50	mAP @50–95
	Defect	Dirt	Gap	Defect	Dirt	Gap
-	0.84	0.85	0.98	0.79	0.61	1.00	0.85	0.55
HE	0.89	1.00	0.79	0.62	0.68	1.00	0.86	0.57
CLAHE	0.93	0.93	0.86	0.65	0.78	1.00	0.86	0.57
DCP	0.83	0.93	0.80	0.73	0.73	1.00	0.87	0.56
LEGAN [39]	0.79	0.93	0.88	0.72	0.74	1.00	0.86	0.55
Proposed AIE + ESE	0.86	0.92	0.99	0.90	0.89	1.00	0.92	0.59

presents the comparative detection performance results under general low-light conditions.

Traditional approaches such as HE, CLAHE, and DCP provided moderate improvements in brightness and contrast but often led to structural distortion or uneven enhancement, resulting in fluctuating precision–recall balance. The deep learning–based LEGAN method restored brightness and texture more effectively but introduced artificial details, which increased false detections in the defect class. In contrast, the proposed modular preprocessing framework, integrating AIE and ESE, achieved the most stable and superior results (mAP@50 = 0.92, mAP@50–95 = 0.59). This improvement originates from the complementary mechanisms of the two modules: AIE effectively mitigates illumination imbalance and enhances local contrast, while ESE reinforces structural edges and preserves fine defect contours. Consequently, the proposed preprocessing pipeline substantially improves the visual quality of low-light images and enables more reliable and balanced defect detection, regardless of the downstream detection model.

4.3. Comparative Evaluation

To validate the performance of the proposed image preprocessing–based detection framework, we conducted comparative experiments using several state-of-the-art object detection models, namely YOLOX [40], RT-DETR [41], YOLOv11 [19], improved YOLOX [42], and SMI-YOLOv8 [43]. This selection was made to provide a comprehensive comparison: baseline detection models (YOLOX, RT-DETR, YOLOv11) illustrate the limitations of standard architectures, improved YOLOX and SMI-YOLOv8 represent prior attempts to address low-light detection challenges, and YOLOv11 serves as our baseline for evaluating the benefit of combining it with the proposed preprocessing modules. This arrangement clarifies the goal of the experiment—to reveal the inherent limitations of detection-only models and to prove that traditional object detection models, when used without preprocessing, struggle to detect subtle and low-contrast rail surface defects. All models were evaluated under identical experimental conditions using the metrics and defect categories described above (

Table 4. Detection performances of our and existing object detection frameworks.

Model	Precision			Recall			mAP @50	mAP @50–95
	Defect	Dirt	Gap	Defect	Dirt	Gap
YOLOX [40]	0.53	0.88	0.57	0.42	0.57	0.57	0.53	0.26
RT-DETR [41]	0.89	0.77	0.94	0.73	0.62	1.00	0.84	0.51
YOLOv11 [19]	0.84	0.85	0.98	0.79	0.61	1.00	0.85	0.55
Improved YOLOX [42]	0.53	0.88	0.63	0.45	0.60	0.60	0.56	0.27
SMI-YOLOv8 [43]	0.88	0.90	0.91	0.57	0.78	1.00	0.90	0.55
Proposed model	0.86	0.92	0.99	0.90	0.89	1.00	0.92	0.59

). Specifically, all models were trained on the same dataset with the input image resolution of 640 × 640, the batch size of 16, 300 training epochs using the SGD optimizer with a momentum of 0.973 and a weight decay of 0.005, ensuring a controlled and reproducible evaluation.

The YOLOX series models [40,42] were generally inferior to recent architectures. Specifically, the accuracy of defect class detection was notably low, primarily because of the difficulties in recognizing fine-grained structural anomalies. Additionally, both YOLOX variants showed limited recall for the gap class, which indicated the frequent omission of TPs. Although high precision was achieved for the dirt class, the corresponding recall was low, suggesting a substantial number of missed TPs. Consequently, the mAP@50 scores for YOLOX and improved YOLOX remained around 0.5, underscoring the limited effectiveness of these models in low-light environments.

RT-DETR exhibited excellent detection capabilities for larger or well-defined defects, achieving high recall in the gap category. However, its performance on the dirt class was suboptimal, probably because the visual similarities between dirt patterns and the background decreased detection reliability. YOLOv11 achieved stable performance on the defect and gap classes, but its recall on the dirt class was lower, resulting in an overall mAP performance similar to that of RT-DETR.

In contrast, the proposed model outperformed all baseline models across the three defect categories. Specifically, it achieved a recall of 0.90 for the defect class, indicating a high sensitivity to subtle and fine-grained structural anomalies. For the dirt class, both precision and recall were notably improved, reaching 0.92 and 0.89, respectively. Furthermore, our model attained perfect recall for the gap class. These results contributed to the highest overall mAP@50 score of 0.92 and were attributed to the synergistic effect of the proposed preprocessing pipeline. The AIE module effectively corrected local brightness inconsistencies and enhanced contrast in underexposed regions, thereby improving the visibility of defects in low-light conditions, while the ESE module preserved and sharpened defect contours, helping the detection model learn more discriminative features. Collectively, these enhancements mitigated the limitations imposed by challenging lighting conditions and substantially improved the robustness and generalization ability of the model across various defect types.

Figure 6, Figure 7 and Figure 8 present qualitative detection results for defect, dirt, and gap classes, respectively, with blue bounding boxes representing ground truth (GT) annotations, green boxes indicating correct detections (TPs), and red boxes denoting misdetections (FPs). Figure 6 shows that the YOLOX model frequently failed to detect fine-grained defects or generated multiple overlapping bounding boxes for a single defect instance, indicating poor localization capability. Other baseline models showed improved localization capabilities but lacked precision in delineating defect boundaries. In contrast, the proposed model exhibited superior boundary adherence and spatial accuracy, successfully detecting subtle defects with minimal FPs. Figure 7 shows that YOLOX and improved YOLOX detected only a limited subset of dirt instances, thus featuring high FN rates. YOLOv11 showed improved recall but failed to detect certain dirt regions. Although RT-DETR successfully identified all dirt regions, it also produced several FPs in nondefective areas, probably because of background similarities. The proposed model accurately detected all dirt instances without FPs, demonstrating high precision and robust discriminative capability. Figure 8 shows that YOLOX entirely failed to detect gap defects, while improved YOLOX generated several FPs on nondefective areas. In contrast, RT-DETR, YOLOv11, and the proposed model accurately detected gaps. However, the proposed model demonstrated the most precise alignment between predicted bounding boxes and the GT with no false detections and accurate localization.

Collectively, these qualitative results reveal that the proposed preprocessing framework substantially improved defect visibility under low-light conditions, enabling more accurate and robust detection across all defect categories. The ability of our model to preserve structural integrity and suppress irrelevant features is particularly effective in visually challenging environments.

To better reflect real-world railway operating conditions, we extended the simulation beyond low-light environments to include additional visibility degradation factors such as fog and shadows, which commonly occur in tunnels, during early morning hours, or under rainy conditions and often obscure the visual features of defects and hinder accurate detection. Consequently, evaluating model performance solely under low-light conditions is insufficient to ensure practical applicability. Hence, further experiments were conducted using the object detection models listed in Table 3 for test images modified to simulate fog and shadow conditions (

Table 5. Detection performances of our models and existing object detection models under low-light conditions with simulated fog.

Model	Precision			Recall			mAP @50	mAP @50–95
	Defect	Dirt	Gap	Defect	Dirt	Gap
YOLOX [40]	0.85	1.00	1.00	0.01	0.05	0.42	0.26	0.05
RT-DETR [41]	0.51	1.00	1.00	0.02	0.10	0.73	0.39	0.13
YOLOv11 [19]	0.76	0.31	0.76	0.28	0.10	0.75	0.39	0.20
Improved YOLOX [42]	0.70	0.17	0.91	0.42	0.10	1.00	0.50	0.32
SMI-YOLOv8 [43]	0.84	0.87	1.00	0.59	0.68	0.73	0.85	0.46
Proposed model	0.95	0.97	0.95	0.64	0.73	1.00	0.87	0.52

). The proposed model maintained its high detection performance across various visual disturbances, which verified its robustness and generalization capability in real-world operating environments.

The baseline models, including the YOLOX family [40,42], RT-DETR [41], and YOLOv11 [19], exhibited notable precision and recall degradation across defect and dirt classes under fog conditions. Notably, YOLOX [40] demonstrated a severe decline in defect recall, which decreased to 0.01 and indicated an almost complete inability to detect defects in fog-affected images. In contrast, the proposed model maintained a substantially higher performance, achieving precision and recall values of 0.95 and 0.64, respectively, for the defect class and 0.97 and 0.73, respectively, for the dirt class. For the gap class, the proposed model preserved near-optimal performance, achieving precision and recall values (0.95 and 1.00, respectively) comparable with those observed under clear conditions.

Overall, the proposed model attained an mAP@50 of 0.87 and mAP@50–95 of 0.52, exhibiting a slight performance decline relative to clear conditions but still substantially outperforming all baseline models. These findings demonstrate that the integration of the AIE and ESE modules effectively preserved and enhanced the visual characteristics of defects even in visually degraded environments such as fog and confirm the robustness and generalization capability of the proposed approach in handling diverse real-world visibility challenges frequently encountered in railway operations.

Figure 9 illustrates qualitative detection results for each model under fog-augmented low-light conditions, corresponding to the comparative analysis of the defect class detailed in Table 4. All baseline models exhibited a notable number of FPs and FNs, particularly for subtle or small-scale defects. These results highlight the sensitivity of conventional object detection models to visibility degradation. In contrast, the proposed model delivered robust and accurate detection, maintaining clear localization and minimizing FPs and FNs, which confirmed its effectiveness in adverse visual environments.

5. Discussion

The preprocessing framework designed to enhance the efficiency of rail surface defect detection under low-light conditions integrated AIE and ESE modules to address the common challenges of poor visibility, low contrast, and blurred defect boundaries in real-world railway environments. When integrated with YOLOv11, this framework achieved a notable detection performance improvement, as revealed by an ~7% increase in mAP@50 compared with the baseline model.

This performance gain was primarily attributed to the synergistic effect of the preprocessing steps: CLAHE-based local contrast enhancement improved the visibility of faint defect regions; gamma correction effectively adjusted global brightness distributions, particularly benefiting dark areas; and morphological sharpening preserved edge integrity and enhanced structural features. The proposed model maintained high precision and recall across all tested defect classes, which confirmed its robustness and practical potential for deployment in real rail inspection systems. In comparative evaluations, our model consistently outperformed other advanced models, including YOLOX variants and transformer-based detectors such as RT-DETR. Moreover, the model retained its performance advantage under degraded-visibility conditions with simulated fog and shadow effects, which further validated its generalization capability in challenging real-world scenarios.

A notable drawback of our preprocessing technique is that it enhances all visual features indiscriminately, including both defect regions and nondefect elements such as gravel, bolts, or debris surrounding the tracks. This feature can lead to increased FPs, as nondefective regions with similar edge or texture patterns can be mistakenly classified as defects. This trade-off highlights the need for more selective feature enhancement methods in future framework iterations. Additionally, the evaluation was conducted on a relatively small-scale dataset comprising 440 images. This limited dataset may not fully reflect the diversity of real-world railway conditions, which can widely vary in terms of lighting, weather, background noise, and defect types. To establish broader applicability and improve robustness, the framework should be validated on larger and more diverse datasets representing various operating environments. A further limitation lies in the image-dependent nature of the enhancement procedure, which may introduce performance variability under extreme lighting conditions. Although the proposed enhancement modules operate based on the characteristics of input illumination, their parameters were empirically optimized to ensure consistent performance under various lighting conditions. The modular structure also allows straightforward adjustment of enhancement parameters for new datasets or environments, mitigating potential variability and enhancing the framework’s generalizability.

Based on the abovementioned findings and limitations, the following future research directions are proposed. First, segmented learning strategies or class-specific enhancement techniques should be introduced to better learn the visual feature differences between background and defect objects. Second, the preprocessing technique should be extended from a static to a dynamic learning–based approach adapting according to the conditions to improve adaptability to varying illumination conditions. Third, beyond rail defect detection, one should explore the applicability of our framework to other industrial fields, such as road crack detection, industrial pipeline leak monitoring, and aircraft surface inspection.

In conclusion, although the proposed framework offers notable improvements in low-light rail defect detection, continued refinement and expansion are required to fully realize its potential for real-world deployment in safety-critical infrastructure monitoring systems.

6. Conclusions

A preprocessing-based detection framework is constructed to enhance defect detection performance in real-world railway environments affected by poor illumination. This framework consists of AIE and ESE modules, which improve the visual quality of images through local contrast enhancement and nonlinear brightness correction and markedly promote the recognition of fine defects by preserving defect contours and increasing structural sharpness. The proposed approach outperforms state-of-the-art object detection models, including YOLOv11, achieving competitive results with an mAP@50 of 92% and mAP@50–95 of 59%. Importantly, this method enhances detection performance without requiring modifications to the underlying object detection architecture and is therefore broadly applicable to various models and use cases.

The proposed framework enables reliable and accurate rail defect detection under challenging lighting conditions and holds promise as a core technology for the real-time monitoring, safety management, and automated maintenance of railway infrastructure. Future works should focus on developing more refined preprocessing strategies considering various illumination and environmental conditions, as well as expanding the framework to other industrial fields.