基于FPGA的快速橡胶异物检测方法研究

Research on fast rubber foreign matter detection method based on FPGA

为解决目前目标检测神经网络算法存在运算量大、占用资源多与实时性等问题,提出了一种基于改进的YOLOv3-tiny算法检测橡胶中异物,并使用现场可编程门阵列(FPGA)进行硬件加速。首先,在网络中采用深度可分离卷积,降低网络参数量;然后在网络中引入卷积注意力机制模块(CBAM),增强检测网络的准确性;最后将DE-YOLOv3-tiny网络移植到FPGA芯片中,使用定制参数位宽和数据流水线设计等优化处理,使其在顺丁橡胶异物检测中达到高效轻量化的效果。实验结果表明,相同条件下,其检测精度为96.7%,网络模型大小仅为9.79 MB,运行功率为7.8 W,单帧检测速度为246 ms,能够满足顺丁橡胶异物检测工程检测需求。

In order to solve the problems of large amount of computation, large amount of resources and real-time in the current target detection neural network algorithm, a new method based on the improved YOLOv3 tiny algorithm(DE-YOLOv3-tiny) is proposed to detect foreign objects in rubber, and FPGA is used for hardware acceleration. Firstly, deep separable convolution is used in the network to reduce the network parameters. Then the attention mechanism(CBAM) is introduced into the network to enhance the accuracy of the detection network. Finally, the DE-YOLOv3-tiny network is transplanted to the FPGA chip, and optimized by using customized parameter bit width and data pipeline design, so that it can achieve the effect of high efficiency and lightweight in the detection of CIS polybutadiene rubber foreign bodies. The experimental results show that under the same conditions, the detection accuracy is 96.7%, the network model size is only 9.79 MB, the running power is 7.8 W, and the single frame detection speed is 246 ms, which can meet the detection requirements of CIS polybutadiene rubber foreign matter detection engineering.