基于AGD-YOLO的钢材表面缺陷检测算法

Algorithm for Steel Surface Defect Detection Based on AGD-YOLO

摘要:

钢材在制造业中应用广泛,其表面缺陷严重影响其质量。缺陷形态的多样性以及检测背景的干扰给现有检测模型带来挑战,为此本文提出钢材表面缺陷检测算法AGD-YOLO。首先设计了自适应多尺度下采样AMSD模块,该模块利用膨胀卷积与多池化操作捕捉并融合多尺度特征信息,并集成到主干网络中以提升缺陷识别能力；其次结合膨胀卷积和空间金字塔池化提出了增强型全局语义空间池化金字塔EGC-SPP模块以整合全局背景与边缘特征；最后设计了双流融合网络DSFN增强特征表示,提升算法上下文互补性和细节特征的识别能力。实验结果表明,改进后的算法在NEU-DET数据集上的检测精度提升了7.1% mAP,有效解决了钢表面缺陷检测的难点。

Abstract:

Steel is widely used in manufacturing, and surface defects significantly affect its quality. The diversity of defect shapes and interference from the detection background present challenges to existing detection models. To address this, this paper proposes the AGD-YOLO algorithm for detecting surface defects in steel.First, we designed the Adaptive Multi-Scale Downsampling (AMSD) module, which utilizes dilated convolutions and multi-pooling operations to capture and integrate multi-scale feature information. This module is incorporated into the backbone network to enhance defect recognition capability. Next, we introduced the Enhanced Global Context-Space Pooling Pyramid (EGC-SPP) module, which combines dilated convolutions and spatial pyramid pooling to integrate global background and edge features.Finally, we designed a Dual-Stream Fusion Network (DSFN) to enhance feature representation, improving the algorithm"s contextual complementarity and the ability to recognize detailed features. Experimental results indicate that the improved algorithm achieves a 7.1% increase in mAP on the NEU- DET dataset, effectively addressing the challenges in detecting surface defects on steel.