结合注意力机制的带钢表面缺陷检测模型

Detection of surface defects in strip steel in combination with attention mechanisms

带钢的表面缺陷对带钢的质量影响极大。针对由于带钢缺陷存在类间相似且容易受背景干扰,导致检测模型的精度较差的问题,提出了一种基于YOLOv8n的轻量级实时检测模型EDD-YOLO(EnhanceDefectsDe-tectionYOLO)。首先,构造了一个特殊卷积ECConv(EnhanceCoordConv),其使用额外的坐标通道更加清晰地感受待检测目标的空间位置信息;其次,将轻量级注意力机制EMA(EfficientMulti-ScaleAttention)嵌入特征融合网络中,使得计算资源高效且合理分配,增强特征融合能力;然后,采用CARAFE(Content-AwareReAssembly of Features Extraction)替代原融合网络的上采样模块;最后,在预测部分使用WIOU改进原损失函数,加速模型收敛。实验数据表明,该模型相较于YOLOv8n,检测精度提高3.6%,检测速度保持在166fps,并且模型大小、空间复杂度与原模型基本持平,更好地满足了复杂工业场景下带钢缺陷的实时检测要求。

The surface defects of the strip have a great impact on the quality of the strip.A lightweight real-time detection algorithm EDD-YOLO(Enhance Defects Detection YOLO)based on YOLOv8n was proposed for the problems of poor accuracy of detection algorithms due to the existence of inter-class similarity and susceptibility to background interference of strip defects.First,a special convolutional Enhance Coord Conv was constructed,which used additional coordinate channels to provide a clearer sense of the spatial location information of the target to be detected.Secondly,Efficient Multi-Scale Attention was embedded in the feature fusion'network,which made the computational resources efficient and rational allocation of computational resources to enhance feature fusion.Then,Content-Aware ReAssembly of Features Extraction was used to replace the upsampling module of the original fusion network.Finally,the original loss function was improved using WIOU in the prediction part to accelerate the convergence of the model.The experimental data shows that the detection accuracy of the model is 3.6%higher than that of YOLOv8n,the detection speed is maintained at 166 fps,and the model size and spatial complexity are basically the same as the original model,which better meets the requirements of real-time detection of strip defects in complex industrial scenarios.