基于深度学习的实时图像目标检测系统设计

Design of Real-time Image Object Detection System Based on Deep Learning

针对图像目标检测的嵌入式实时应用需求,采用合并计算层的方法对基于MobileNet和单发多框检测器(SSD)的深度学习目标检测算法进行了优化,并采用软硬件结合的设计方法,基于ZYNQ可扩展处理平台设计了实时图像目标检测系统;在系统中,根据优化后的算法设计了一款多处理器核的深度学习算法加速器,并采用PYTHON语言设计了系统的软件;经过多个实验测试,深度学习目标检测系统处理速度可以达到45FPS,是深度学习软件框架在CPU上运行速度的4.9倍,在GPU上的1.7倍,完全满足实时图像目标检测的需求。

Aiming at the requirements of the embedded real-time application of image object detection, the deep learning object detection algorithm based on MobileNet and Single Shot Multi-Box Detector (SSD) is optimized by the method of combining computational layers, and the real-time image object detection system is designed by using software and hardware combination method based on ZYNQ scalable processing platform. In the system, a multi-processor core deep learning algorithm accelerator is designed according to the optimized algorithm, and the software of the system is designed by PYTHON language. After several experiments, the processing speed of deep learning object detection system can reach 45 FPS, which is 4.9X faster than deep learning framework running on CPU and 1.7X faster than on GPU. It fully meets the requirements of real-time image object detection.