Abstract: In order to solve the problem that the parameters of traditional deep convolution neural network are too large to be used in mobile devices, a lightweight convolution neural network architecture GResNets based on group residual structure is proposed. Using the bottleneck structure with three convolution layers, the output feature maps of the previous layer are divided into four equal groups. Three lightweight convolution neural network architectures are designed according to the way of adding identical mapping to the bottleneck module in the group and whether or not the adjacent module outside the group joins residual learning. In the experimental stage, the performance of the three network architectures was evaluated on three image recognition datasets, Caltech-256, Food-101 and GTSRB. The experimental results show that GResNets has the same or even better classification performance than traditional deep convolution neural networks under the condition of fewer network parameters, and is suitable for mobile devices.