融合注意力和多级残差的引线框架表面缺陷检测方法

Fusion of attention and multi-level residuals for detection of surface defects in lead frame

摘要:

针对引线框架表面不规则缺陷难以进行准确检测的问题,提出一种融合注意力和多级残差的引线框架表面缺陷检测方法。首先,提出一种多尺度全局注意力模块,通过捕获缺陷边缘区域的通道和空间信息,进一步获取引线框架全局信息以提高分割精度。其次,为了实现缺陷信息的多尺度融合,设计一种多级残差融合注意力网络模块,提取表面划痕缺陷的全局语义信息。此外,编码器采用平滑最大化单元(smooth maximum unit,SMU) 激活函数,以改善检测时的细节缺失现象。对比实验结果表明,所提引线框架表面缺陷检测方法的平均交并比(mean intersection over union, MIoU)指标在自制引线框架表面缺陷数据集上,相比于4种典型方法分别提升了25.05%、26.79%、12.11%、21.02%；消融实验证明所提检测方法具有较好的缺陷检测性能,能获得更多的有效缺陷信息。

Abstract:

Aiming at the problem that it is difficult to accurately detect irregular defects on the surface of lead frames,a method for detecting defects on the surface of lead frames is proposed by integrating attention and multi-level residuals.First,a multi-scale global attention module is proposed to further acquire the global information of the lead frame and improve the segmentation accuracy by capturing the channel and spatial information of the defective edge region.Then,in order to realize the multi-scale fusion of defect information,a multi-level residual fusion attention network module is designed to extract the global semantic information of surface scratch defects.In addition,the encoder employs a smooth maximum unit (SMU) activation function to improve the detail missing phenomenon during detection.The comparative experimental results indicate that the mean intersection over union (MIoU) metrics of the proposed lead frame surface defect detection method are improved by 25.05%,26.79%,12.11% and 21.02% compared with the four typical methods on the homemade lead frame surface defect dataset,respectively.The ablation experiments prove that the proposed method has better defect detection performance and can obtain more effective defect information.