Abstract: Traditional image recognition is often realized by large convolutional networks with complex structure and high computational power. With the increasing popularity of mobile embedded devices, lightweight convolutional networks have the characteristics of less parameters, less computational complexity, and the model can be deployed and applied to small devices. Based on the ShuffleNetV2 unit, the network width was changed and SENet attention mechanism was introduced to form a new network module to build an improved Shufflet-K5-SENet network. The network could extract feature information more efficiently while retaining the lightweight. The experimental results show that compared with ShuffleNetV2 network on Flowers Recognition dataset, the accuracy is improved by 1.88 percentage points and the Loss value is decreased by 3%. Compared with MobileNetV2 network on CIFAR-100 dataset, the accuracy is improved by 1.84 percentage points, the FLOPs is reduced by 23.89% and the number of parameters is reduced by 55.82%.