基于SCC-YOLO的指针式仪表轻量化检测方法

Lightweight Detection Method for Pointer Meters Based on SCC-YOLO

针对指针式仪表检测模型结构复杂、占用内存量高、参数计算量大而导致的不易部署问题,提出一种基于YOLOv5的轻量化仪表目标检测网络SCC-YOLO。采用轻量化主干ShuffleBlock_lite结构重新设计网络主干,引入卷积核重构的深度可分离卷积,通过SimAM无参注意力机制模块进一步提升特征提取能力。融合坐标卷积CoordConv与CARAFE轻量化上采样模块提高模型特征融合性能。利用数据增强技术构建真实场景下和复杂场景下的指针式仪表图像数据集。对比实验结果表明:SCC-YOLO模型能大幅提升指针式仪表的检测效率,模型的参数量平均降低27.3%,计算量平均降低54.8%,精度上综合提升1.3%。轻量化的设计使其能够在移动端与边缘设备更容易部署,能够满足真实场景下的指针式仪表检测任务要求。

To address the issue of difficult deployment caused by the complex structure,high memory usage,and large parameter calculation of pointer instrument detection,a lightweight instrumentation target detection network SCC-YOLO based on YOLOv5 was proposed.The network backbone was redesigned by using the lightweight backbone ShuffleBlock_lite structure,and the depth separable convolution reconstructed by convolution kernel was introduced to further improve the feature extraction capability through the SimAM parameter-free attention mechanism module to further enhance the feature extraction capability.Fusing coordinate convolution CoordConv with CARAFE lightweight upsampling module improved the model feature fusion performance.Data enhancement techniques were utilized to construct pointer gauge image datasets in real scenes and in complex scenes.Comparative experimental results show that the SCC-YOLO model can significantly improve the detection efficiency of pointer gauges,with an average reduction of 27.3%in the number of parameters of the model,an average reduction of 54.8%in the computation,and an integrated improvement of 1.3%in the accuracy.The lightweight design makes it easier to be deployed on mobile and edge devices,and can meet the requirements of pointer meter detection tasks in real scenarios.