基于YOLOv3模型压缩的交通标志实时检测算法

YOLOv3目标检测算法检测精度高，检测速度快，能够实现对交通标志的实时检测。但由于YOLOv3模型要求设备具有较强的运算能力及较大的内存，难以直接部署在车辆等资源受限平台上。针对此问题，提出了一种Strong Tiny-YOLOv3目标检测模型，该模型通过引入FireModule层进行通道变换，在减小模型参数的同时能够加深模型深度。同时，模型在FireModule层之间加入short-cut来增强网络的特征提取能力。实验结果表明，模型在保持较高检测精度的前提下，能够极大减小YOLOv3模型对设备的依赖。与Tiny-YOLOv3模型相比，Strong Tiny-YOLOv3模型的参数量减少了87.3%，实际内存大小减少了77.9%，在GeForce 940MX上的检测速度提高了22.8%，且在GTSDB和CCTSDB交通标志数据集上的检测mAP分别提高了12%和3.8%。

Compression Algorithm of Traffic Sign Real-Time Detection Based on YOLOv3 Model

YOLOv3 target detection algorithm has high detection accuracy and fast detection speed, which can realize the real-time detection of traffic signs. However, because YOLOv3 model requires the equipment to have strong computing power and large memory, it is difficult to be directly deployed on the platform with limited resources such as vehicles. To solve this problem, a Strong Tiny-YOLOv3 target detection model is proposed. By introducing FireModule layer for channel transformation, the Strong Tiny-YOLOv3 can not only reduce the model parameters, but also deepen the model depth. At the same time, the model adds short-cut between FireModule layers to enhance the ability of network feature extraction. The experimental results show that the module can greatly reduce the dependence of YOLOv3 model on equipment while maintaining high detection accuracy. Compared with the Tiny-YOLOv3 model, the parameters of Strong Tiny-YOLOv3 model is reduced by 87.3%, the actual memory size is reduced by 77.9%, the detection speed on GeForce 940MX is increased by 22.8%, and the detection mAP on GTSDB and CCTSDB traffic sign data sets is increased by 12% and 3.8%, respectively.