Fine-grained image classification is an important task in the field of computer vision, which aims to accurately classify objects that belong to the same category but have minor differences. In this paper, an improved fine-grained image classification method is proposed for the problems of difficulty in capturing subtle features and loss of information in fine-grained image classification, which introduces the BAM (Bottleneck Attention Module) attention mechanism in the ResNet50 model, which utilizes the channel and spatial attention mechanism to purposefully enhance the representation ability of the key features. Meanwhile, the interference of noise and unimportant features is suppressed to improve the discriminative ability model of the model; the method also introduces the PPM (Pyramid Pooling Module) pyramid module, which captures different levels of semantic information from feature maps at multiple scales and fuses them to help the network better understand and capture fine-grained features in fine-grained classification tasks. Experiments on multiple fine-grained datasets show that the algorithm in this paper classifies well.