基于深度学习的轻量级目标检测算法的研究

Research on Lightweight Target Detection Algorithm Based on Deep Learning

铁路异物侵限检测技术在视频监控中起着重大作用,而现有的目标检测网络计算成本高,模型存储大,因为硬件成本和计算力存在矛盾,所以导致检测速度不高等问题。针对上述问题,选取YOLOv4 tiny作为基础网络并进行了改进。首先,在CSP Darknet53 tiny中,利用深度可分离卷积替换部分标准卷积,减少了参数数量和计算量;其次,将训练后权重进行转换,使得转换后权重可被Tensor RT优化推理器推理加速;然后,引入合并运算和半精度量化处理;最后,将TensorRT-yolov4 tiny部署至嵌入式设备Jeston nano上。实验表明,在2555张铁路数据集上进行实验,其检测速度提升了50%,达到了平均0.036一张检测图片,帧率达到了30.12。MAP达到了82.13%。证明了提出的方法的在部署至嵌入式设备后速度性能上具有的优越性。

This paper selects yolov4 tiny as the basic network and improves it.Firstly,in CSP darknet53 tin,the depth separable convolution is used to replace part of the standard convolution,which reduces the number of parameters and the amount of calculation. Secondly,the weight after training is converted,so that the converted weight can be accelerated by tensor RT optimization reasoner.Then,merge operation and semi precision quantization processing are introduced.Finally,deploy tensorrt-yolov4 tiny to the embedded device Jeston nano.The experiment shows that the detection speed is increased by 50% on 2555 railway data sets,reaching an average of 0.036 a detection picture and a frame rate of 30.12.Map reached82.13%.It is proved that the method proposed in this paper has advantages in speed and performance after being deployed to embedded devices.