融合ECA机制的轻量化YOLOv4检测模型

Lightweight YOLOv4 Detection Model Incorporating ECA Mechanism

近年来,卷积神经网络已在人脸识别、无人驾驶等领域取得重大突破。随着智能移动设备的普及,高精度的大型网络往往伴随着参数量多、计算量大等问题,无法部署在这些资源有限的移动设备平台上。GhostNet通过简单的线性操作生成更多特征映射,可大幅减少计算成本。为此,提出了一种改进轻量化YOLOv4的GhostNet-YOLOv4网络模型,该模型将YOLOv4的主干网络替换为GhostNet残差结构,借助即插即用的Ghost模块升级卷积神经网络,并使用Mosaic数据增强技术对数据集进行预处理,融合ECA机制,加入Focal Loss焦点损失函数,在保证一定精度的前提下大幅减少了模型的参数量和计算量。相对于改进前的GhostNet-YOLOv4模型,改进后的GhostNet-YOLOv4在PASCAL VOC 2007数据集上的mAP(mean Average Precision)提高1.65百分点,达到85.09%,且模型参数量只有11.429 M,相对于原YOLOv4模型减少了约80%,具有更好的综合性能。

In recent years,convolutional neural networks have made breakthroughs in applications such as face recognition and autonomous driving.With the popularity of intelligent mobile devices,high-precision large-scale networks are often accompanied by problems such as a large number of parameters and a large amount of computing,and cannot be deployed on these resource-limited mobile device platforms.GhostNet generates more feature maps with simple linear operations,drastically reducing the computational cost.Therefore,we propose an improved lightweight YOLOv4 network model.The backbone network of YOLOv4 is replaced with the GhostNet residual structure,which can upgrade the convolutional neural networks,and the Mosaic data enhancement is used to pre-process the dataset.ECA mechanism is integrated and Focal Loss function is added.On the premise of ensuring certain accuracy,the number of parameters and calculation amount of the model are greatly reduced.Compared with the GhostNet-YOLOv4 before the improvement,the mAP(mean Average Precision)of the improved GhostNet-YOLOv4 on the PASCAL VOC 2007 dataset is increased by 1.65%to 85.09%,and the parameters of the model are only 11.429 M,which reduces by about 80%compared with the original YOLOv4 network,which indicates that the improved model has better overall performance.