基于YOLOv5的轻量化端到端手机检测方法

针对监控图像中手机尺寸较小、分辨率低且特征不明显等问题，给检测算法研究带来了困难。提出了一种改进的YOLOv5网络模型方法用来识别手机的使用。改进的检测算法引入轻量级网络GhostNet作为主干提取网络，将GhostConv模块、C3Ghost模块分别代替主干网络中的Conv基本卷积模块和C3模块，减小网络参数和复杂度；同时，将注意力机制CBAM引入到主干网络中，减少融合后冗余特征的影响，提取到目标区域中更加关键的特征信息；使用四尺度特征检测，在原算法基础上对应的增加检测层，用以提高更小目标的检测精度。实验结果表明，改进后的YOLOv5算法准确率达到95.7%,平均精度达到97.1%,比改进前训练的准确率和平均精度分别提升了2.5%和1.8%,运算量和参数量较改进前分别减少了14.3%和24.5%。改进的YOLOv5算法不仅具有轻量化优势，同时保证了准确率和平均精度。该方法为智能监测技术行业违规使用手机提供了理论依据和技术参考。

Lightweight end to end mobile phone detection method based on YOLOv5

For problems such as small size, low resolution and not obvious features of mobile phones in the monitoring images, it has brought difficulties to the study of the detection algorithm. This article proposes an improved YOLOv5 network model method to identify the use of mobile phones. The improved detection algorithm introduces the lightweight network GhostNet as the main extraction network, and the GhostConv module, the C3Ghost module instead of the Conv basic convolution module and the C3 module in the main network to reduce the network parameters and complexity; at the same time, the attention mechanism CBAM is introduced into the main network, reducing the effects of redundant characteristics after fusion, and extraction of more critical feature information in the target area; using four scale feature detection, the corresponding increase of the detection layer on the basis of the algorithm, to improve the detection accuracy of smaller targets. The experimental results show that, the accuracy of the improved YOLOv5 algorithm is 95.7%, and the mAP is 97.1%, the accuracy and mAP of the training increased by 2.5% and 1.8%, the calculation and parameters were reduced by 14.3% and 24.5%. The improved YOLOv5 algorithm not only has the advantages of lightweight, but also ensures the mAP and accuracy. This method provides theoretical basis and technical reference for the use of mobile phones in the intelligent monitoring technology industry.