Abstract: Limited by the memory and computing power of the embedded devices and mobile devices,the deployment of deep convolutional neural networks(CNN)is hindered by the large amount of parameters and slow inference speed.Therefore,lightweight network research has attracted more and more attention.This paper constructs a dual-path architecture with shallow and deep layers based on ResNet-50.The compression of the model could be achieved by adjusting the channel dimension ratio of the two paths.In addition,a feature separation module was proposed,which divided the feature maps into two groups based on channel attention.One group entered the deep path,and the other group entered the shallow path.Precise grouping could make feature extraction more efficient.This architecture surpassed the current best pruning methods and lightweight design models on the ImageNet dataset.