轻量化的YOLOv4目标检测算法

YOLOv4目标检测算法主干网络庞大且参数量和计算量过多，难以部署在算力和存储资源有限的移动端嵌入式设备上。提出一种改进的YOLOv4目标检测算法，使用轻量化的ShuffleNet V2网络作为主干特征提取网络，更换模型激活函数及扩大卷积核，同时将YOLOv4网络中的普通卷积替换为深度可分离卷积，降低算法参数量、计算量和模型占用空间。在ShuffleNet V2网络结构的改进过程中分析并剪裁其基本组件，利用2个3 × 3卷积核级联的方式增强网络感受野，并使用Mish激活函数进一步提升网络检测精度和模型推理速度。在GPU平台和VisDrone 2020数据集上的实验结果表明，与YOLOv4算法相比，改进的YOLOv4算法在牺牲1.8个百分点的检测精度情况下，提高了27%的检测速度，压缩了23.7%的模型容量，并且能够充分发挥ZYNQ平台并行高速数据处理及低功耗的优势。

Lightweight YOLOv4 Target Detection Algorithm

The YOLOv4 algorithm has a large backbone network and involves a large amount of parameters, due to which it is difficult to use the algorithmfor mobile embedded devices with limited computing power and storage.Aiming at solving these disadvantages, this study proposes an improved YOLOv4 target detection algorithm.It uses the lightweight ShuffleNetV2 network as the backbone feature extraction network, replaces the model activation function, and expands the convolution kernel.Moreover, the ordinary convolutions in the YOLOv4 network are replaced with depthwise separable convolutions, which reduce the amount of algorithm parameters and computations, and the size of the network model.Using the improved ShuffleNetV2 network structure, the basic components of the network structure are analyzed and clipped, the network receptive field is improved by cascading two 3×3 convolution cores, and the Mish activation function is used to further improve the network accuracy and speed of model reasoning.The experimental results on the GPU platform and the VisDrone 2020 dataset show that compared with the YOLOv4 algorithm, the proposed YOLOv4 algorithm improves the detection speed by 27% and compresses the algorithm capacity by 23.7%;however, detection accuracy is reduced by 1.8 percentage points.At the same time, the proposed algorithm can give full play to the advantages of parallel high-speed data processing and low power consumption of the ZYNQ platform.