轻量化的印刷电路板缺陷检测网络 Multi-CR YOLO

针对印刷电路板表面缺陷目标小，检测精度低问题，设计了印刷电路板表面缺陷检测网络Multi-CR YOLO,满足实时检测速度的前提下，有效提高了检测精度。首先，由3个Multi-CR块组成的主干特征提取网络Multi-CR backbone对印刷电路板小目标缺陷进行特征提取。其次，SDDT-FPN特征融合模块使层级高的特征层向层级低的特征层进行特征融合，同时为小目标预测头YOLO Head-P3所在特征融合层加强特征融合，进一步增强低层特征层的表达能力。PCR模块加强主干特征提取网络与SDDT-FPN特征融合模块不同尺度的特征层的特征融合机制，且防止模块之间进行特征融合时信息丢失。C₅ECA模块负责自适应调节特征权重和自适应注意小目标缺陷信息的要求，进一步提高了特征融合模块的自适应特征提取能力。最后，3个YOLO-Head负责针对不同尺度的小目标缺陷进行预测。实验表明，Multi-CR YOLO网络模型检测mAP达到98.55%,模型大小为8.90 MB,达到轻量化要求，检测速度达到了95.85 fps,满足小目标缺陷实时检测的应用需求。

Lightweight PCB defect detection network Multi-CR YOLO

Aiming at the problem of small target and low detection accuracy of printed circuit board surface defects, Multi-CR YOLO, aprinted circuit board surface defect detection network, is designed to meet the premise of real-time detection speed and effectivelyimprove the detection accuracy. Firstly, the backbone feature extraction network Multi-CR backbone, which consists of three Multi-CRresidual blocks, performs feature extraction for small target defects on printed circuit boards. Secondly, the SDDT-FPN feature fusionmodule enables the feature fusion from the high level feature layer to the low level feature layer, and at the same time strengthens thefeature fusion for the feature fusion layer where the small target prediction head YOLO Head-P3 is located, to further enhance theexpressive ability of the low level feature layer. The PCR module strengthens the feature fusion mechanism of the different scales of thebackbone feature extraction network and the feature layer of the SDDT-FPN feature fusion module, and prevents the fusion mechanismbetween the modules. The C5ECA module is responsible for adaptively adjusting the feature weights and adaptively paying attention tothe requirement of small target defect information, which further improves the adaptive feature extraction capability of the feature fusionmodule. Finally, the three YOLO-Head are responsible for predicting small target defects for different scales. The experiments show thatthe detection mAP of the Multi-CR YOLO network model reaches 98. 55%, the model size is 8. 90 MB, which meets the lightweightrequirement, and the detection speed reaches 95. 85 fps, which meets the application requirements of real-time detection of small-targetdefects.