增强感受野的轻量化零部件检测方法

Lightweight Parts Detection Method Based on Enhanced Receptive Field

针对工业生产复杂工况下零部件检测任务中,传统检测方法不能兼顾识别精度和检测实时性的问题,提出一种增强感受野的轻量化网络。首先,使用ShuffleNetv2网络作为网络主干,在平衡精度和速度的情况下,通过控制宽度因子缩小网络模型;其次,引入了改进型的并行空洞金字塔轻量化模型,扩大感受野,提高网络特征提取能力;最后,在路径聚合网络设计了GSCSP模块促进特征信息传递,增强了网络融合能力。在工业汽车机舱检测任务中,该算法mAP达到96.1%,网络模型大小只有4.3 MB,相比于YOLOv5网络提升了24帧/s。

In the actual industrial production complex conditions,traditional detection methods can not take into account the recognition accuracy and real-time detection in the parts detection task,this paper proposed a lightweight network to enhance the receptive field.Firstly,the ShuffleNetv2 network was used as the network backbone to reduce the network model by controlling the width factor while balancing accuracy and speed;secondly,the improved atrous spatial pyramid pooling module was introduced to expand the receptive field and improve the network feature extraction capability;finally,the GSCSP module was designed in the path aggregation network.It was helpful to facilitate feature information transfer and enhanced the network fusion capability.In the industrial engine compartment identification task,the algorithm proposed in this paper achieved 96.1%mAP and the network model size was only 4.3 M,which was a 24 fps/s improvement compared to the YOLOv5 network.