YOLO-MDFR-Light: a lightweight defect detection network with multi-dimensional frequency attention for industrial edge deployment

Abstract

Modern deep learning object detectors with over 20 million parameters are difficult to deploy on edge devices with 2--8 GB memory. In this work, we compress YOLO11m to 7.89 M parameters, achieving a 60.8% parameter reduction with only a 1.01% mAP@50 drop for binary industrial defect classification. The backbone adopts the proposed LightweightC3k2 module with depth-wise separable convolutions and a width multiplier of 0.7, which reduces backbone parameters by 62.8%. A DFRamit frequency-domain attention block after the SPPF module enhances feature representation of tiny and weak defects. Dual ECA and Coordinate Attention (CA) modules are embedded to recover lost channel interaction and spatial localization precision. Ablation experiments verify that the combination of all proposed modules is essential for effective feature compensation. Evaluated on a cigarette burst bead dataset containing 1,625 images and 6 classes, our method achieves 98.21% mAP@50 for binary classification and 92.63% mAP@50 for multi-category classification.