Steel Surface Defect Detection Based on Dynamic Receptive Field and Multi-Scale Features Fusion

Hongkai Zhang, Xue Song, Sixian Chan, Jianan Chen, Suqiang Li, Chao Li, Fengguang LiuResearch Square 2025 · DOI

Abstract

Steel surface defect detection is vital for guaranteeing product quality in contemporary manufacturing. However, traditional steel surface defect detection algorithms often face challenges due to insufficient resilience in feature extraction under complex backgrounds. To address this, we present a framework for defect detection which boosts feature extraction through a multi-path optimization strategy, markedly enhancing both accuracy and efficiency. Firstly, we introduce a dynamic receptive field (DRF) module which employs the spatial kernel selection mechanism to enable the network for dynamic perception according to defect scales. Meanwhile, a multi-scale feature fusion (MFF) module is designed to combine shallow and deep contextual information, minimizing information loss and enhancing feature representation. Finally, comprehensive experiments on the GC10-DET, NEU-DET, and APDDD datasets show that our model achieves a mean average precision of 71.4%, 82.0%, and 68.3%, respectively, outperforming state-of-the-art methods, while keeping efficient inference and minimal computational cost for real-time industrial applications.

Introduction

In modern steel production, identifying surface defects like scratches, crazing, or inclusions is critical for quality control. Manual inspection is slow and subjective, while traditional machine vision often fails when backgrounds are complex or lighting changes. While deep learning—specifically the YOLO family—has revolutionized this field, two major hurdles remain: scale variation and information loss. Defects can range from microscopic pits to long, sprawling scratches, and fixed-size convolutional kernels often struggle to capture both extremes simultaneously. Furthermore, as data moves deeper into a neural network, fine-grained spatial details essential for small defects are often lost during high-level semantic processing.

DM-YOLO Architecture

The proposed framework, DM-YOLO, is built upon the YOLOv8 architecture and introduces two primary architectural innovations: the Dynamic Receptive Field (DRF) and the Multi-Scale Feature Fusion (MFF) modules. The model is designed to be "lightweight," requiring less computational power (GFLOPs) and memory, making it suitable for deployment on edge devices directly on the factory floor.

The overall architecture of DM-YOLO, illustrating the Backbone, Neck, and Head components, alongside the specific sub-modules like C2f_DRF and SPPF.

Core Innovations

Dynamic Receptive Field (DRF) Module

Instead of using a static "one-size-fits-all" convolutional kernel, the DRF employs a spatial kernel selection mechanism. It utilizes Partial Convolutions (PConv) to maintain efficiency and adaptively adjusts the receptive field size based on the defect's scale. This allows the network to "zoom" its focus dynamically, effectively capturing features across a wide range of defect sizes.

Detailed schematic of the Dynamic Receptive Field (DRF) module, showing the large kernel decomposition and spatial selection mechanism used to adapt to varying defect scales.

Multi-Scale Feature Fusion (MFF) Module

This module acts as a bridge between shallow layers (which hold rich detail) and deep layers (which understand context). By using channel attention and cross-attention enhancement, the MFF ensures that high-level semantic information guides the extraction of low-level details, and vice versa. This bidirectional enhancement minimizes information loss during the fusion process, which is particularly beneficial for detecting small and medium-sized defects.

Experimental Results

The researchers evaluated DM-YOLO across three diverse public datasets to prove its robustness:

  • NEU-DET: A benchmark for steel surface defects containing six categories.
  • GC10-DET: A complex industrial dataset with ten types of steel plate defects.
  • APDDD: An aluminum profile defect dataset used to test generalization ability.

DM-YOLO outperformed several state-of-the-art models, including variants of YOLOv11 and YOLOv12, in both accuracy and efficiency. It achieved 82.0% mAP on NEU-DET, 71.4% mAP on GC10-DET, and 68.3% mAP on APDDD. Crucially, the model maintained an inference speed of 75.5 FPS with only 2.56M parameters, comfortably exceeding the requirements for real-time industrial monitoring.

Conclusion

DM-YOLO demonstrates that smarter architectural choices, such as the DRF module, can actually reduce parameter counts while increasing accuracy. The study highlights that spatial attention is a powerful tool for filtering background noise in industrial settings. Future work will focus on edge optimization through quantization and hardware-aware tuning to further boost speed on resource-limited devices, as well as exploring multi-modal data fusion for even greater resilience.

Summary updated

Abstract

<title>Abstract</title> Steel surface defect detection is vital for guaranteeing product quality in contemporary manufacturing. However, traditional steel surface defect detection algorithms often face challenges due to insufficient resilience in feature extraction under complex backgrounds. To address this, we present a framework for defect detection which boosts feature extraction through a multi-path optimization strategy, markedly enhancing both accuracy and efficiency. Firstly, we introduce a dynamic receptive field (DRF) module which employs the spatial kernel selection mechanism to enable the network for dynamic perception according to defect scales. Meanwhile, a multi-scale feature fusion (MFF) module is designed to combine shallow and deep contextual information, minimizing information loss and enhancing feature representation. Finally, comprehensive experiments on the GC10-DET, NEU-DET, and APDDD datasets show that our model achieves a mean average precision of 71.4%, 82.0%, and 68.3%, respectively, outperforming state-of-the-art methods, while keeping efficient inference and minimal computational cost for real-time industrial applications. The source codes are at https://github.com/ssjddb/DM-YOLO.git.
BibTeX
@inproceedings{zhang2025steel,
  author = {Hongkai Zhang and Xue Song and Sixian Chan and Jianan Chen and Suqiang Li and Chao Li and Fengguang Liu},
  title = {{Steel Surface Defect Detection Based on Dynamic Receptive Field and Multi-Scale Features Fusion}},
  booktitle = {Research Square},
  year = {2025},
  doi = {10.21203/rs.3.rs-7718767/v1},
}
Figure