<p>The textile industry’s pursuit of defect-free production has driven the demand for efficient automated inspection systems, yet low-resolution imaging and subtle defect visibility remain critical challenges. Traditional manual inspection is limited by high labor costs, low repeatability, and subjective judgments, while high-resolution camera systems impose significant hardware constraints. This study proposes a cost-efficient pipeline that integrates lightweight super-resolution (SR) enhancement and tiled defect detection to narrow the performance gap with native high-resolution imaging. The SR module, derived from ESRGAN, is optimized via two complementary modifications: reducing block depth and channel width in residual-in-residual dense blocks (RRDB) and replacing standard convolutions with depthwise-separable convolutions (SRRDB), achieving a 24–44% reduction in SR-generation latency and a 4–5<InlineEquation ID="IEq1"> <EquationSource Format="TEX">$\times $</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation> decrease in model size. Enhanced images are partitioned into 4<InlineEquation ID="IEq2"> <EquationSource Format="TEX">$\times $</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation>4 non-overlapping tiles, enabling precise defect localization and segmentation via single-stage detectors. Here we show that the pipeline achieves approximately 28.4 FPS in a high-performance computing (HPC) environment, indicating near-real-time GPU-based operation, with YOLOv8 variants achieving balanced accuracy—particularly for hole, oil stain, and object defects—with mAP50 values exceeding 0.84 on the TILDA dataset. For edge deployment on Raspberry Pi 4B/5, the lightweight SRRDB configuration reduces SR latency by 29% compared to the baseline RRDB, indicating improved feasibility on resource-constrained hardware. This work demonstrates that low-cost cameras, augmented by optimized SR and tiled detection, can provide a practical alternative to high-resolution imaging systems, advancing industrial quality control in textile production while balancing computational efficiency and detection performance. The relevant code is available at <a href="https://github.com/ahmet-metin/textile-defect-sr-pipeline.">https://github.com/ahmet-metin/textile-defect-sr-pipeline.</a></p>

Textile defect inspection: a lightweight super-resolution augmented detection pipeline

期刊论文
发表日期：

2026年7月26日

查看全文

Ahmet Metin,
Haydar Ozkan

摘要

The textile industry’s pursuit of defect-free production has driven the demand for efficient automated inspection systems, yet low-resolution imaging and subtle defect visibility remain critical challenges. Traditional manual inspection is limited by high labor costs, low repeatability, and subjective judgments, while high-resolution camera systems impose significant hardware constraints. This study proposes a cost-efficient pipeline that integrates lightweight super-resolution (SR) enhancement and tiled defect detection to narrow the performance gap with native high-resolution imaging. The SR module, derived from ESRGAN, is optimized via two complementary modifications: reducing block depth and channel width in residual-in-residual dense blocks (RRDB) and replacing standard convolutions with depthwise-separable convolutions (SRRDB), achieving a 24–44% reduction in SR-generation latency and a 4–5 $\times $ $\times$ decrease in model size. Enhanced images are partitioned into 4 $\times $ $\times$ 4 non-overlapping tiles, enabling precise defect localization and segmentation via single-stage detectors. Here we show that the pipeline achieves approximately 28.4 FPS in a high-performance computing (HPC) environment, indicating near-real-time GPU-based operation, with YOLOv8 variants achieving balanced accuracy—particularly for hole, oil stain, and object defects—with mAP50 values exceeding 0.84 on the TILDA dataset. For edge deployment on Raspberry Pi 4B/5, the lightweight SRRDB configuration reduces SR latency by 29% compared to the baseline RRDB, indicating improved feasibility on resource-constrained hardware. This work demonstrates that low-cost cameras, augmented by optimized SR and tiled detection, can provide a practical alternative to high-resolution imaging systems, advancing industrial quality control in textile production while balancing computational efficiency and detection performance. The relevant code is available at https://github.com/ahmet-metin/textile-defect-sr-pipeline.