A random selection global diversified classification network model is presented to deal with the difficulty of capturing the potential distinguishable features in fine-grained image classification and the tendency to ignore the relationship between different features. Firstly, the ConvNeXt is taken as the backbone to improve classification performance, and a random elimination boosting selection(REBS) strategy is designed to promote network learning more image information and capture potential distinguishable features through the enhancement of the interaction between the feature elimination and feature boosting branches. After that, the global diversification module(GDM) is proposed focusing on modelling feature maps of different levels interactively to enhance the comparison ability of network. Meanwhile, the dataset of logo imprint image is established, and the fine-grained algorithm is applied to conduct authenticity identification, which realizes the practical application of fine-grained image classification task in natural scenes. This network achieves 91.9%, 93.8% and 93.5% accuracy on three open datasets, CUB-200-2011, Stanford Cars and FGVC-Aircraft, respectively. Compared with other advanced comparison methods, the presented method greatly upgrades the comparison performance. The accuracy of 96.8% is achieved on the self-built dataset, which indicates the capacity of the network in accurate classification of true and fake images.