基于改进YOLOv4的轻量化目标检测算法

为解决YOLOv4目标检测网络结构复杂、参数多、训练所需的配置高以及实时检测图片的传输帧数低,难以实现工业上的应用普及等问题,提出一种基于YOLOv4改进的轻量化算法SL-YOLO。在原始的YOLOv4网络上进行改进和优化,使用ShuffleNetv2轻量级网络替换YOLOv4原始骨干网络,将SENet模块融入ShuffleNetv2,降低网络计算复杂度,在网络层中加入Swish激活函数,使模型收敛效果更好;同时用简化后的加权双向特征金字塔结构改进原模型的特征融合网络,优化目标检测精度;通过消融实验判定各通道的重要性,对冗余剪枝,将模型进行压缩。在PASCAL VOC和MS COCO数据集上进行对比实验,改进后的模型与原始YOLOv4相比,模型内存减少89.4%,浮点运算量下降88.4%,检测速度提升了近2倍。实验结果表明,改进后的YOLOv4模型能够在保持较高的精度下有效减少模型推理计算量,大大提升模型检测速度。

Improved lightweight YOLOv4 target detection algorithm

In order to solve the problems of structurally complex, numerous parameters, high configuration required for training, low transmission frames of real-time detection pictures and difficult to achieve industrial application popularization of YOLOv4 target detection network, a lightweight target detection network SL-YOLO based on YOLOv4 is proposed. It improve and optimizes the original YOLOv4 network, and replaces the original backbone network of YOLOv4 with ShuffleNetv2 lightweight network, integrates SENet module into ShuffleNetv2, reduce the network computing complexity, add Swish activation function to the network layer to make the model convergence effect better; at the same time, the simplified weighted bidirectional feature pyramid structure is used to replace the feature fusion network of YOLOv4, aims to optimize the target detection accuracy; the importance of each channel was determined, thus the redundant pruning was performed, and the model was compressed. The result of a comparative experiment on open data set PASCAL VOC and MS COCO shows that the memory of the model is compressed by 89.4%, the amount of floating-point operations of the model is reduced by 88.4%, and the detection speed of the model is increased by nearly two times, which indicates the SL-YOLO lightweight network can effectively reduce the amount of model reasoning calculation and improve the model detection speed simultaneously, and greatly improve the speed of model detection.