IR-YOLOv7-Tiny: A Lightweight and Robust Framework for Fabric-Defect Detection

TL;DR

This paper introduces IR-YOLOv7-Tiny, a lightweight fabric-defect detection system that improves accuracy and robustness in industrial settings.

Contribution

A novel lightweight framework combining DWT and modified YOLOv7-Tiny for interference-resilient fabric-defect detection.

Findings

IR-YOLOv7-Tiny achieves 96.8% and 98.8% mAP@0.5 on TILDA and DAGM datasets.

The model has only 3.5 M parameters and outperforms baselines by 2.2% and 3.9% in mAP@0.5.

Abstract

To tackle the challenges of missed detections, false alarms, electromagnetic noise, and constrained deployment resources in fabric-defect inspection, we propose a lightweight and interference-resilient fabric-defect detector based on the Discrete Wavelet Transform (DWT). First, a color-space channel separation filter leverages Hue–Saturation–Value (HSV) decomposition to suppress illumination and electromagnetic interference while preserving fabric structural details. Second, DWT is employed to extract directional texture features (horizontal, vertical, and diagonal) from complex woven structures. Third, the backbone of the You Only Look Once version 7 Tiny (YOLOv7-Tiny) is modified by replacing pooling with a Spatial Pyramid Dilated Convolution (SPD) block, which maintains fine-grained detail during downsampling. For upsampling, an inverted SPD block with channel concatenation is introduced to mitigate background redundancy caused by interpolation. Experimental results on the TILDA and DAGM datasets show that the proposed IR-YOLOv7-Tiny achieves mAP@0.5 of 96.8% and 98.8%, respectively, with only 3.5 M parameters. Outperforming baseline models achieved 2.2% and 3.9% in the mean Average Precision (mAP) at Intersection over Union (IoU) 0.5 (mAP@0.5). The results demonstrate excellent effectiveness and high deployability for resource-constrained industrial scenarios.

Figures