一种改进MobileNet_YOLOv3网络的快速目标检测方法

Fast Target Detection Method for Improving MobileNet_YOLOv3 Network

基于卷积神经网络的目标检测在智能机器人、无人机等领域有着重要的应用,但其模型普遍结构复杂、参数量大、占用资源多,难以满足嵌入式目标检测任务中的实时性需求.针对此问题,本文提出一种多尺度特征融合注意力网络(MSFAN:Multi-Scale Feature-fusion Attention Network)模型,该模型基于MobileNet_YOLOv3网络模型,并结合多尺度特征融合等改进措施,在高实时性的同时增强了模型对小目标的检测能力,能够更好地在嵌入式终端设备上实现实时高效目标检测,从而为边缘计算场景下的目标检测应用提供可能的方案.实验结果表明:MSFAN模型在识别精确度和性能消耗上均获得了较好表现,在NVIDIA Jetson TX2上检测速度可达46FPS,相比MobileNet_YOLOv3在速度相当的情况下精度有明显提升.

Object detection based on convolutional neural networks has important applications in the fields of intelligent robots and unmanned aerial vehicles,but its models are generally complex in structure,large in parameters,and occupy a lot of resources,making it difficult to meet the real-time requirements of embedded object detection tasks.In response to this problem,this paper proposes an MSFAN lightweight object detection model,which is based on the MobileNet_YOLOv3 network model,combined with attention mechanism and other improvement measures,while enhancing the real-time performance of the model at the same time to enhance the ability to detect small objects,can be better Realize efficient and real-time object detection on embedded terminal devices,thus providing possible solutions for object detection applications in edge computing scenarios.The experimental results show that the MSFAN model has achieved good performance in both recognition accuracy and performance consumption.The detection speed can reach 46FPS on NVIDIA Jetson TX2,and the accuracy is significantly improved compared to MobileNet_YOLOv3 at the same speed.