印刷电路板(PCB)作为现代电子设备的核心组成部分,随着技术的进步,在多个领域得到了广泛的应用,包括消费电子、通信、医疗设备、汽车、工业控制等。针对PCB传统缺陷检测方法过程中的成本高、效率低、漏检率高,并且常见PCB缺陷微小,检测...印刷电路板(PCB)作为现代电子设备的核心组成部分,随着技术的进步,在多个领域得到了广泛的应用,包括消费电子、通信、医疗设备、汽车、工业控制等。针对PCB传统缺陷检测方法过程中的成本高、效率低、漏检率高,并且常见PCB缺陷微小,检测精度低等问题,提出一种基于改进YOLOv8的深度学习检测方法。1) 引进ASF-YOLO网络,并且在此基础上添加小目标检测层;2) 在主干网络中加入SimAM注意力机制;3) 改进损失函数为Wise_CIoUv3。实验表明,改进后的模型的平均精度mAP达到90.4%,相比基线模型提高3.55%。另外,模型参数量下降17.19%,模型大小减少0.8 MB,实现了模型部分轻量化。为此领域提供了参考和应用价值。As a core component of modern electronic equipment, printed circuit board (PCB) has been widely used in many fields with the progress of technology, including consumer electronics, communications, medical equipment, automobiles, industrial control and so on. Aiming at the problems of high cost, low efficiency, high missed detection rate, small PCB defects and low detection accuracy in traditional PCB defect detection methods, a deep learning detection method based on improved YOLOv8 was proposed. 1) Introduce ASF-YOLO network, and add small target detection layer on this basis;2) Add the SimAM attention mechanism to the backbone network;3) Improve the loss function to Wise_CIoUv3. Experiments show that the average precision mAP of the improved model reaches 90.4%, which is 3.55% higher than that of the baseline model. In addition, the number of model parameters decreased by 17.19%, the size of the model decreased by 0.8 MB, and the initial lightweight of the model was realized. It provides reference and application value in this field.展开更多
随着电子设备的广泛应用,印刷电路板(Printed Circuit Board,PCB)在电子制造行业中具有重要意义.然而,由于制造过程中的不完美和环境因素的干扰,PCB上可能存在微小的缺陷.因此,开发高效准确的缺陷检测算法对于确保产品质量至关重要.针对...随着电子设备的广泛应用,印刷电路板(Printed Circuit Board,PCB)在电子制造行业中具有重要意义.然而,由于制造过程中的不完美和环境因素的干扰,PCB上可能存在微小的缺陷.因此,开发高效准确的缺陷检测算法对于确保产品质量至关重要.针对PCB微小缺陷检测问题,本文提出了一种基于多维注意力机制的高精度PCB微小缺陷检测算法.为降低网络的模型参数量和计算量,引入部分卷积(Partial Convolution,PConv),将ELAN(Efficient Layer Aggregation Network)模块设计为更加高效的P-ELAN,同时,为增强网络对微小缺陷的特征提取能力,引入多维注意力机制(Multi-Dimensional Attention Mechanism,MDAM)的全维动态卷积(Omni-dimensional Dynamic Convolution,ODConv)并结合部分卷积,设计了POD-CSP(Partial ODconv-Cross Stage Partial)和POD-MP(Partial ODconv-Max Pooling)跨阶段部分网络模块,提出了OD-Neck结构.最后,本文基于(Youo Only Look Once version 7,YOLOv7)提出了对小目标更加高效的YOLO-POD模型,并在训练阶段采用一种新颖的Alpha-SIoU损失函数对网络进行优化.实验结果表明,YOLO-POD的检测精确率和召回率分别达到了98.31%和97.09%,并在多个指标上取得了领先优势,尤其是对于更严格的(mean Average Precision at IoU threshold of 0.75,mAP75)指标,比原始的YOLOv7模型提高28%.验证了YOLO-POD在PCB缺陷检测性能中具有较高的准确性和鲁棒性,满足高精度的检测要求,可为PCB制造行业提供有效的检测解决方案.展开更多
针对钢材表面缺陷检测中小目标缺陷检测效果不理想、特征提取不充分的问题,以YOLOv5算法为基础,提出一种YOLOv5s-ADW算法。将自注意力与卷积混合模块(a mixed model of self-attention and convolution,ACmix)融入主干网络层,增强模型...针对钢材表面缺陷检测中小目标缺陷检测效果不理想、特征提取不充分的问题,以YOLOv5算法为基础,提出一种YOLOv5s-ADW算法。将自注意力与卷积混合模块(a mixed model of self-attention and convolution,ACmix)融入主干网络层,增强模型的特征敏感度;在特征融合层中加入可变形大内核注意力机制(deformable large kernel attention,D-LKA),增强模型对图像中不规则缺陷的捕捉能力;将原损失函数替换为Wise-IoU损失函数,降低数据集中低质量示例对模型检测效果的影响并提升小目标缺陷检测能力,在NEU-DET上进行实验验证。实验验证结果表明:YOLOv5s-ADW算法的平均精度均值(mean average precision,mAP)达到88.3%,相较原始模型提升了14.4%;小目标缺陷和漏检率高的缺陷平均精度(average precision,AP)也有较大提升,相比其他主流算法,能够更好解决上述问题。展开更多
文摘印刷电路板(PCB)作为现代电子设备的核心组成部分,随着技术的进步,在多个领域得到了广泛的应用,包括消费电子、通信、医疗设备、汽车、工业控制等。针对PCB传统缺陷检测方法过程中的成本高、效率低、漏检率高,并且常见PCB缺陷微小,检测精度低等问题,提出一种基于改进YOLOv8的深度学习检测方法。1) 引进ASF-YOLO网络,并且在此基础上添加小目标检测层;2) 在主干网络中加入SimAM注意力机制;3) 改进损失函数为Wise_CIoUv3。实验表明,改进后的模型的平均精度mAP达到90.4%,相比基线模型提高3.55%。另外,模型参数量下降17.19%,模型大小减少0.8 MB,实现了模型部分轻量化。为此领域提供了参考和应用价值。As a core component of modern electronic equipment, printed circuit board (PCB) has been widely used in many fields with the progress of technology, including consumer electronics, communications, medical equipment, automobiles, industrial control and so on. Aiming at the problems of high cost, low efficiency, high missed detection rate, small PCB defects and low detection accuracy in traditional PCB defect detection methods, a deep learning detection method based on improved YOLOv8 was proposed. 1) Introduce ASF-YOLO network, and add small target detection layer on this basis;2) Add the SimAM attention mechanism to the backbone network;3) Improve the loss function to Wise_CIoUv3. Experiments show that the average precision mAP of the improved model reaches 90.4%, which is 3.55% higher than that of the baseline model. In addition, the number of model parameters decreased by 17.19%, the size of the model decreased by 0.8 MB, and the initial lightweight of the model was realized. It provides reference and application value in this field.
文摘随着电子设备的广泛应用,印刷电路板(Printed Circuit Board,PCB)在电子制造行业中具有重要意义.然而,由于制造过程中的不完美和环境因素的干扰,PCB上可能存在微小的缺陷.因此,开发高效准确的缺陷检测算法对于确保产品质量至关重要.针对PCB微小缺陷检测问题,本文提出了一种基于多维注意力机制的高精度PCB微小缺陷检测算法.为降低网络的模型参数量和计算量,引入部分卷积(Partial Convolution,PConv),将ELAN(Efficient Layer Aggregation Network)模块设计为更加高效的P-ELAN,同时,为增强网络对微小缺陷的特征提取能力,引入多维注意力机制(Multi-Dimensional Attention Mechanism,MDAM)的全维动态卷积(Omni-dimensional Dynamic Convolution,ODConv)并结合部分卷积,设计了POD-CSP(Partial ODconv-Cross Stage Partial)和POD-MP(Partial ODconv-Max Pooling)跨阶段部分网络模块,提出了OD-Neck结构.最后,本文基于(Youo Only Look Once version 7,YOLOv7)提出了对小目标更加高效的YOLO-POD模型,并在训练阶段采用一种新颖的Alpha-SIoU损失函数对网络进行优化.实验结果表明,YOLO-POD的检测精确率和召回率分别达到了98.31%和97.09%,并在多个指标上取得了领先优势,尤其是对于更严格的(mean Average Precision at IoU threshold of 0.75,mAP75)指标,比原始的YOLOv7模型提高28%.验证了YOLO-POD在PCB缺陷检测性能中具有较高的准确性和鲁棒性,满足高精度的检测要求,可为PCB制造行业提供有效的检测解决方案.
文摘针对钢材表面缺陷检测中小目标缺陷检测效果不理想、特征提取不充分的问题,以YOLOv5算法为基础,提出一种YOLOv5s-ADW算法。将自注意力与卷积混合模块(a mixed model of self-attention and convolution,ACmix)融入主干网络层,增强模型的特征敏感度;在特征融合层中加入可变形大内核注意力机制(deformable large kernel attention,D-LKA),增强模型对图像中不规则缺陷的捕捉能力;将原损失函数替换为Wise-IoU损失函数,降低数据集中低质量示例对模型检测效果的影响并提升小目标缺陷检测能力,在NEU-DET上进行实验验证。实验验证结果表明:YOLOv5s-ADW算法的平均精度均值(mean average precision,mAP)达到88.3%,相较原始模型提升了14.4%;小目标缺陷和漏检率高的缺陷平均精度(average precision,AP)也有较大提升,相比其他主流算法,能够更好解决上述问题。
