基于GhostNet的改进模型轻量化方法

Improved model lightweighting method based on GhostNet

为了降低深度卷积神经网络的部署成本,优化模型的检测性能,提出一种改进的轻量化主干网络算法S-GhostNet.该算法通过引入特征图生成优化的Ghost Module结构降低卷积操作的计算量,并结合改进类残差模块提升模型的精确度.S-GhostNet具有较强的即插即用性,可以应用于多数卷积神经网络模型.实验结果表明:在目标分类以及目标检测任务中,S-GhostNet相较于MobileNetV2、ShuffleNetV2以及GhostNet,模型计算量更小,模型的精确度持平,甚至更高.

In order to reduce the deployment cost of deep convolutional neural networks and optimize the detection performance of the models,an improved lightweight backbone network algorithm S-GhostNet is proposed.The algorithm reduces the computational effort of convolutional operations by introducing a Ghost Module structure optimized for feature map generation.It improves the accuracy of the models by combining with an improved class of residual modules.S-GhostNet has a strong plug-and-play property and can be applied to most convolutional neural network models.Experimental results show that S-GhostNet is less computationally intensive than MobileNetV2,ShuffleNetV2 and GhostNet in target classification and target detection tasks.Also the accuracy of the model is similar or even higher.