轻量化网络已成为面向工业场景部署的关键技术.为进一步提升Ghost module的特征提取能力并减少参数量,提出了一种改进的S-Ghost瓶颈模块(Small Ghost Bottleneck).此瓶颈模块采用 1×1 卷积通道与Ghost module并联的结构,缩减Ghost module的通道数以压缩参数量,并用与之并联的 1×1 卷积进行通道扩充;在模块输出端引入通道混洗(Channel Shuffle)操作以保证通道间信息交互.实验结果表明,利用该瓶颈结构设计的图像分类网络LGSNet(Light Ghost Networks,LGSNet),其计算量和参数量显著降低,同时网络精度与推理速度未受影响,甚至在一些测试中取得最优,此网络设计满足工业需求.这为面向工业场景的轻量化网络架构设计提供了新的解决方案和思路.
Lightweight networks have become a key technology for industrial scenarios.On the purpose to improve the feature representation ability and reduce the number of parameters of the Ghost module,we proposed an improved S-Ghost Bottleneck module.The 1×1 convolution channel is introduced in parallel with the original Ghost module.The number of channels in the Ghost module is reduces to compress the parameter scale.The paralleled 1×1 convolution channel is on the purpose to expand the number of channels.The channel shuffle operation is adopted at the output of the proposed module to enhance the channel interaction.Experimental results show that,the image classification network LGSNet(Light Ghost Networks,LGSNet)composed with the proposed structure has significantly reduced the computational complexity and the scale of the parameters.The accuracy of the proposed network has no significant change,and even achieves the best in some tests.The result shows that the proposed method is a promising solution for lightweight network design for industrial scenarios.