Improving the accuracy of image classification is the basis of automatic production. The work aims to proposes a more accurate image classification method to make automatic packaging and production more efficient. Based on the idea of ResNeSt feature graph group, by introducing the channel domain and spatial domain attention mechanism and introducing the idea of adaptive convolution kernel and gem pooling into the spatial domain attention module, the network could use different sensory fields for different pictures in the spatial domain attention mechanism to focus on more important parts. An image classification network model structure with channel domain and spatial domain attention mechanism and good portability was proposed. This method improved the accuracy of image classification. On ImageNet data set, the accuracy of top-1 was 81.39%. The ResNeSkt algorithm framework proposed in this paper is superior to the current mainstream image classification methods. At the same time, the overall network structure has good portability, and can be used as the backbone network in other image research fields such as image detection and semantic segmentation.