基于轻量化网络的YOLOv₄检测算法研究

Research on YOLOv₄ Detection Algorithm Based on Lightweight Network

摘要: 针对YOLOv₄算法网络模型大、运算量大、运行效率低的问题，本文提出一种基于GhostNet的YOLOv₄轻量化模型，使用GhostNet代替原网络中的CSPDarknet-53，通过一些简单的线性运算代替部分卷积，减少卷积运算，从而减少参数量以及浮点运算量。然后，在GhostNet中引入CBAM模块，结合通道注意力机制和通道注意力机制，增强模型对有效特征关注。另外使用HDC代替原始网路中的SPPNet，减少浅层网络特征的丢失。最后，将改进的算法与基于其他轻量化网络的YOLOv₄算法对比。实验结果证明，与原始YOLOv₄相比，在精度损失较小的情况下，基于GhostNet以及注意力机制的YOLOv₄轻量化模型的大小得到了压缩，检测速度有了明显的提升。

Abstract: In view of the problems of large network model, large amount of computation and low operation efficiency of YOLOv₄ algorithm, this thesis proposes a lightweight model of YOLOv₄ based on Ghost-Net. The CSPDarknet-53 in the original network is replaced by GhostNet. And partial convolution is replaced by some simple linear operations to reduce convolution operation, thus reducing the number of arguments and floating point operations. Then, CBAM module is introduced into Ghost-Net, combining channel attention mechanism and channel attention mechanism, to enhance the model’s attention to effective features. In addition, HDC is used instead of SPPNet in the original network to reduce the loss of shallow network features. Finally, the improved algorithm is compared with YOLOv₄ algorithm based on other lightweight networks. Experimental results show that, compared with the original YOLOv₄, the size of the lightweight model of YOLOv₄ based on GhostNet and attention mechanism is compressed and the detection speed is significantly improved under the condition of less accuracy loss.