Fabric defect detection is a fundamental quality control process in textile manufacturing, yet achieving accurate

and reliable automated inspection remains difficult because of complex background textures, subtle defect

patterns, and substantial variation in defect scale and shape. Although deep learning–based detectors have

improved inspection performance, many lightweight models still suffer from limited feature discrimination,

particularly in real-time industrial environments where computational efficiency is critical. To address this

limitation, this study proposes an enhanced fabric defect detection framework by integrating an Efficient

Pyramid Split Attention (EPSA) mechanism into a YOLOv5-based convolutional network. The EPSA module

is designed to adaptively recalibrate multi-scale feature responses, enabling the network to emphasize defect-

pyramid-based attention can substantially improve feature representation without imposing significant

computational overhead. The proposed approach offers a practical and efficient solution for automated textile

inspection and provides a useful foundation for future research on lightweight attention modeling for industrial

vision systems.

Keywords: fabric defect detection; attention mechanism; EPSA; convolutional neural networks

Fabric defect detection is a fundamental function in textile quality assurance because surface defects directly

affect product grade, commercial value, downstream processing, and customer acceptance. Defects such as holes,

stains, broken yarns, knots, abrasion marks, and texture irregularities reduce fabric usability and can significantly

systems must operate not only with acceptable detection accuracy, but also with consistency, speed, and practical

deployability under real production constraints [2], [3]. It further positions fabric defect inspection as part of the

broader intelligent manufacturing transition, in which automated visual inspection systems are becoming

mainstream because they enable contactless operation, modular deployment, greater integration, and improved

responsiveness in industrial quality-control workflows [2]. In this sense, AI-driven fabric inspection is not merely

a technical improvement in defect recognition, but a strategic enabler of productivity, quality standardization,

and smarter textile manufacturing systems [2], [3].

Despite its industrial importance, fabric inspection in many production settings still depends heavily on manual

visual examination. This traditional workflow is inherently subjective because inspectors must continuously

judgment [1]. These limitations make manual inspection increasingly incompatible with modern textile

manufacturing, where wide fabric surfaces, high production speeds, and subtle defect patterns demand objective,

scalable automated solutions [1], [2].

Earlier automated approaches based on texture descriptors, statistical methods, spectral analysis, and handcrafted

features partially improved inspection under controlled conditions, but their robustness remained limited when

confronted with repetitive fabric textures, illumination changes, irregular defect morphology, and low-contrast

backgrounds [3]–[6]. The literature structure reflects this progression from texture- and statistics-based methods

to machine learning and then deep learning, indicating that conventional feature-engineering approaches struggle

to generalize across complex industrial scenarios [2], [7].

Among these developments, one-stage object detectors from the YOLO family have become especially attractive

for industrial deployment because they offer a practical balance between detection accuracy and inference speed

[5]–[7]. In inspection environments, the ability to localize and classify defects in a single forward pass is highly

valuable for continuous monitoring. It therefore, adopts a YOLOv5-based framework as the core detector and

evaluates performance not only by accuracy, but also by model size and real-time suitability, reflecting the

industrial need to balance predictive quality with deployability [3], [8]. The further argues that YOLO-based

detectors are well-suited to textile inspection because industrial fabric defects exhibit large size variation, diverse

distribution density, and strong demands for both low miss-detection rates and fast inference [8].

At the same time, it shows that real-world fabric defects exhibit several difficult characteristics, including small

target size, unbalanced aspect ratios, and weak contrast between the defect and the surrounding fabric

Although deep learning has improved fabric defect detection, a major unresolved problem is that lightweight

detectors often struggle to emphasize defect-relevant features while suppressing repetitive background textures.

This is especially problematic in textile inspection because many defects are subtle, irregular, and visually

entangled with the woven pattern itself. Small and low-contrast defects can be overlooked when feature

extraction is insufficiently discriminative, even if the overall architecture is computationally efficient [1], [2].

The issue as a core limitation of existing lightweight CNN-based inspection systems and uses it as a key

motivation for introducing pyramid-aware attention enhancement [2].

A second problem is the accuracy–efficiency trade-off. Many studies achieve stronger laboratory performance

by adopting deeper networks, larger backbones, or more computationally expensive enhancement modules.

However, these choices increase latency, memory consumption, and deployment cost, thereby reducing

suitability for continuous industrial use [3], [4]. Conversely, when architectures are aggressively simplified to

preserve speed, the resulting detector may lose robustness precisely on the difficult defects that matter most in

spatial regions. Combined channel–spatial mechanisms have also shown useful gains in visual recognition and

detection tasks [1]–[3]. However, as noted in the original enhancement, many attention modules are designed

for general vision tasks and do not explicitly address the multi-scale nature of fabric defects, where useful cues

may appear differently across feature pyramid levels [4]. In textile inspection, this limitation is important because

defect cues are often weak, irregular, and scale-dependent, making generic attention insufficient for robust

lightweight detection [4], [5].

This creates a specific research gap. While pyramid-based feature representations are already recognized as

important for handling scale variation in detection networks [6], [7], pyramid-aware attention mechanisms

remain underexplored in fabric defect detection, particularly in lightweight detectors intended for real-time

industrial deployment [4], [5]. Existing studies often improve accuracy by increasing architectural complexity

rather than by efficient, scale-sensitive feature recalibration. The enhancing paper, therefore, positions Efficient

Pyramid Split Attention (EPSA) as a relevant candidate because it can adaptively recalibrate multi-scale

responses while preserving computational efficiency [4], [8].

To further supports this direction by organizing its experiments around feature fusion, attention mechanisms,

convolutional operations, state-of-the-art comparison, and ablation analysis. This structure shows that attention

design is not an isolated addition, but part of a broader effort to improve lightweight detection under the

can enhance feature discrimination in a lightweight fabric defect detector without compromising real-time

performance. The central objective is to improve fabric defect detection by integrating EPSA into a YOLOv5-

based convolutional framework, thereby making the detector more sensitive to subtle multi-scale defects while

remaining practical for industrial use [1], [2]. This objective is consistent with the broader goal of improving

accuracy, speed, and robustness in lightweight textile inspection models, particularly under conditions involving

complex textures, variable defect scales, and industrial deployment constraints [1], [3].

Accordingly, this addresses three research questions. First, how does EPSA-based attention affect feature

discrimination in fabric defect detection? Second, does EPSA outperform conventional attention mechanisms

when detecting small, low-contrast, and structurally complex defects? Third, can EPSA be integrated into a

lightweight detector while preserving inference performance suitable for real-time industrial deployment? These

questions align with the experimental logic of the original enhancing draft, in which comparative attention

analysis, robustness evaluation, and efficiency analysis are central to the results section [2], [4].

This paper contributes to the fabric defect detection literature by presenting an EPSA-enhanced lightweight

detection framework that specifically addresses the challenge of limited feature discrimination in complex textile

robustness without incurring excessive computational overhead. This emphasis on efficiency-aware

enhancement is consistent with the conclusion that industrial AI systems should be evaluated not only by raw

accuracy, but also by deployability, responsiveness, and computational sustainability [1], [3].

Fabric defect detection has evolved from manual, handcrafted-feature inspection to data-driven visual

intelligence systems. Early automated approaches relied on classical image processing and texture analysis

methods, such as Fourier-based inspection, statistical texture transformation, morphological analysis, adaptive

wavelets, and low-rank decomposition, to identify deviations from the normal fabric structure [1]–[5]. These

methods were useful under controlled conditions because they could detect relatively obvious texture disruptions,

but their performance was often unstable when fabric patterns, illumination, noise, and defect appearance varied

across production conditions [2], [3], [5]. As a result, conventional handcrafted approaches were generally

sensitive to environmental interference and lacked the robustness required for real-world industrial deployment

[2], [4].

The problem is also more complex than simple binary texture anomaly detection. Fabric surfaces often feature

or localized color inconsistencies. To address these challenges, several later studies introduced improved two-

stage detectors, background-suppression models, and reconstruction-based strategies. For example, EDSR-

enhanced Faster R-CNN was used to strengthen defect detail representation, while weighted double low-rank

The transition to deep learning significantly advanced the field. CNN-based methods began to outperform

handcrafted feature pipelines because they learn hierarchical visual representations directly from image data,

Table 1. Comparative review of related studies on fabric defect detection

dependency by learning channel-wise importance weights, whereas spatial attention highlights informative

regions within the feature map [1], [2]. Combined mechanisms such as the Convolutional Block Attention

Module (CBAM) attempt to exploit both channel and spatial selectivity, thereby improving representational