Fabric defect detection is a critical task in textile manufacturing, where manual inspection remains inconsistent,

detectors have shown strong potential, many existing models are too computationally demanding for practical

deployment in real-time industrial inspection systems. This study proposes a lightweight deformable YOLO-

based framework for accurate and efficient fabric defect detection. The model is built on YOLOv5s and

enhanced through three efficiency-oriented architectural improvements: Bidirectional Feature Pyramid

Network (BiFPN) for improved multi-scale feature fusion, Deformable Convolutional Networks (DCNv2) for

improve detection accuracy while preserving the lightweight characteristics required for industrial

implementation. The proposed framework offers a practical solution for automated fabric inspection and

provides a useful reference for efficiency-oriented defect detection in smart manufacturing environments.

Keywords: lightweight deep learning; real-time inspection; fabric defect detection; YOLOv5

Fabric defect detection is a fundamental component of quality assurance in textile manufacturing because it

directly influences product reliability, operational efficiency, and commercial value. Defects such as holes,

broken yarns, stains, and surface texture irregularities can substantially degrade fabric quality and marketability.

underscoring the economic importance of early and reliable inspection [1]. Despite this, inspection practices in

Although these approaches offered some success under controlled settings, they generally lacked robustness

when confronted with complex fabric patterns, varying illumination, and irregular defect characteristics [4], [5].

In recent years, deep learning, particularly convolutional neural networks (CNNs), has significantly advanced

the field of fabric defect detection by enabling models to learn hierarchical and discriminative feature

representations directly from data [6], [7]. Among these developments, one-stage object detectors such as the

YOLO family have gained considerable attention for their end-to-end detection capabilities and strong potential

for real-time industrial inspection.

real-time operational requirements.

Against this background, this study proposes a lightweight deformable YOLO-based framework for efficient

fabric defect detection. The proposed model is designed to improve the balance between detection accuracy

and deployability through targeted architectural enhancement rather than network expansion. Specifically, the

framework integrates Bidirectional Feature Pyramid Network (BiFPN) to strengthen multi-scale feature fusion,

Deformable Convolutional Networks (DCNv2) to improve adaptability to irregular defect geometry, and

Efficient Pyramid Split Attention (EPSA) to enhance feature discrimination with limited computational

overhead. The proposed model was evaluated on the Alibaba Tianchi fabric defect dataset, which contains

5,913 annotated images. After category consolidation, 20 defect categories were used for training and testing.

Experimental evaluation focused on mean Average Precision (mAP), model size, and real-time suitability

systems, embedded platforms, or resource-constrained manufacturing environments [6], [9]. As a result, there

is a persistent gap between algorithmic performance reported in research settings and the practical requirements

of industrial implementation.

Second, efforts to reduce computational complexity through lightweight convolutional operations, pruning

strategies, or model compression frequently lead to a decline in detection accuracy, especially when defects are

small, irregular, or embedded within repetitive and low-contrast fabric textures [7], [10]. This reveals a

longstanding trade-off between efficiency and robustness in fabric defect detection. Moreover, many existing

studies examine isolated improvements without offering a systematic assessment of how multiple architectural

components can be jointly optimized to preserve detection quality while improving real-time efficiency.

Therefore, a clear research gap exists in developing a fabric defect detection framework explicitly designed

higher accuracy through greater model depth and complexity, this research investigates how architectural

optimization can achieve a more effective balance between computational efficiency and detection robustness.

The rationale for the study is to demonstrate that lightweight design does not necessarily entail a substantial

sacrifice in detection capability when model components are carefully selected and integrated.

From an academic perspective, this work contributes to the broader field of industrial computer vision by

In particular, it highlights the significance of multi-scale bidirectional feature fusion, adaptive convolution, and

computationally economical attention mechanisms in improving lightweight detector performance. From a

practical perspective, the proposed framework offers textile manufacturers a more deployable solution for

automated real-time defect inspection, thereby supporting smart manufacturing, digital quality assurance, and

Industry 4.0 implementation. The findings also provide useful design guidance for researchers and engineers

developing AI-based visual inspection systems under hardware and latency constraints.

performance.

iii. Evaluate the proposed model against existing detectors in terms of efficiency and detection robustness.

Accordingly, this study addresses the following research questions:

i. How can a lightweight detection architecture be designed to satisfy real-time inspection requirements in

textile manufacturing?

ii. What is the impact of feature fusion, adaptive convolution, and attention mechanisms on detection

efficiency and robustness?

iii. How does the proposed lightweight model compare with existing fabric defect detection approaches in

terms of real-time suitability and detection performance?

The purpose of this literature review is to critically examine existing research on fabric defect detection with a

specific focus on lightweight architectures and real-time efficiency. While significant progress has been made

in improving detection accuracy through deep learning, the practical deployment of these models in industrial

environments remains constrained by computational cost, inference latency, and hardware limitations. As

textile production lines demand continuous, high-speed inspection, real-time feasibility has become as critical

as detection accuracy.

Accordingly, this review is structured around four key themes: (i) traditional fabric defect detection approaches

(iii) lightweight and real-time object detection architectures, and (iv) architectural optimization strategies for

models fabric surfaces as repetitive, structured patterns, with defects interpreted as structural or statistical

deviations from normal texture distributions. Within this framework, methods such as gray-level co-occurrence

matrices, Fourier transforms, and wavelet analysis were widely used to characterize local and global texture

abnormalities [1], [2]. These methods were computationally efficient and relatively interpretable, but they

depended heavily on handcrafted features and assumed stable texture regularity, which limited their robustness

under varying industrial conditions.

level features such as edges, contours, and textures in early layers, while deeper layers encode more abstract

heterogeneous defect characteristics, making CNNs highly effective for visual inspection tasks. However, most

standard CNN-based detection frameworks were originally designed for general object detection benchmarks

rather than computationally constrained industrial systems [4]. As a result, high representational power is often

achieved at the cost of increased parameter count, memory usage, and inference time.

In the context of the present study, the theoretical framework extends hierarchical feature learning toward

efficiency-oriented architectural design. Specifically, the study assumes that robust defect detection can be

achieved not merely through deeper networks, but through more effective architectural organization. Multi-

scale feature fusion supports the detection of defects of varying sizes, adaptive convolution improves sensitivity

features under computational constraints. Thus, this study is theoretically grounded in adapting hierarchical

feature learning to real-time industrial inspection requirements.

Traditional fabric defect detection approaches primarily relied on handcrafted feature extraction and rule-based

classification. These methods typically used statistical texture descriptors, structural regularity measures, or

Among deep learning approaches, one-stage object detectors, such as YOLO-based architectures, have received

considerable attention for their ability to combine end-to-end learning with fast inference. Their unified

detection pipeline makes them well-suited for applications requiring real-time performance. Nevertheless,

many deep learning models continue to prioritize accuracy over computational efficiency, often through deeper

backbones, more complex feature aggregation modules, and heavier parameter configurations. Although such

architectures achieve promising results under research conditions, their deployment on textile production lines

To address deployment constraints, recent research has increasingly focused on lightweight and real-time

detection architectures. Common strategies include depthwise separable convolutions, model pruning,

quantization, and simplified backbone designs [9]. These methods aim to reduce parameter count, memory

usage, and computational complexity, thereby improving inference speed and enabling operation on resource-

constrained platforms.